mirror_zfs/module/zfs/arc.c

5666 lines
161 KiB
C
Raw Normal View History

2008-11-20 23:01:55 +03:00
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2011, 2014 by Delphix. All rights reserved.
* Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
* Copyright 2014 Nexenta Systems, Inc. All rights reserved.
2008-11-20 23:01:55 +03:00
*/
/*
* DVA-based Adjustable Replacement Cache
*
* While much of the theory of operation used here is
* based on the self-tuning, low overhead replacement cache
* presented by Megiddo and Modha at FAST 2003, there are some
* significant differences:
*
* 1. The Megiddo and Modha model assumes any page is evictable.
* Pages in its cache cannot be "locked" into memory. This makes
* the eviction algorithm simple: evict the last page in the list.
* This also make the performance characteristics easy to reason
* about. Our cache is not so simple. At any given moment, some
* subset of the blocks in the cache are un-evictable because we
* have handed out a reference to them. Blocks are only evictable
* when there are no external references active. This makes
* eviction far more problematic: we choose to evict the evictable
* blocks that are the "lowest" in the list.
*
* There are times when it is not possible to evict the requested
* space. In these circumstances we are unable to adjust the cache
* size. To prevent the cache growing unbounded at these times we
* implement a "cache throttle" that slows the flow of new data
* into the cache until we can make space available.
*
* 2. The Megiddo and Modha model assumes a fixed cache size.
* Pages are evicted when the cache is full and there is a cache
* miss. Our model has a variable sized cache. It grows with
* high use, but also tries to react to memory pressure from the
* operating system: decreasing its size when system memory is
* tight.
*
* 3. The Megiddo and Modha model assumes a fixed page size. All
* elements of the cache are therefore exactly the same size. So
2008-11-20 23:01:55 +03:00
* when adjusting the cache size following a cache miss, its simply
* a matter of choosing a single page to evict. In our model, we
* have variable sized cache blocks (rangeing from 512 bytes to
* 128K bytes). We therefore choose a set of blocks to evict to make
2008-11-20 23:01:55 +03:00
* space for a cache miss that approximates as closely as possible
* the space used by the new block.
*
* See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
* by N. Megiddo & D. Modha, FAST 2003
*/
/*
* The locking model:
*
* A new reference to a cache buffer can be obtained in two
* ways: 1) via a hash table lookup using the DVA as a key,
* or 2) via one of the ARC lists. The arc_read() interface
* uses method 1, while the internal arc algorithms for
* adjusting the cache use method 2. We therefore provide two
2008-11-20 23:01:55 +03:00
* types of locks: 1) the hash table lock array, and 2) the
* arc list locks.
*
* Buffers do not have their own mutexes, rather they rely on the
* hash table mutexes for the bulk of their protection (i.e. most
* fields in the arc_buf_hdr_t are protected by these mutexes).
2008-11-20 23:01:55 +03:00
*
* buf_hash_find() returns the appropriate mutex (held) when it
* locates the requested buffer in the hash table. It returns
* NULL for the mutex if the buffer was not in the table.
*
* buf_hash_remove() expects the appropriate hash mutex to be
* already held before it is invoked.
*
* Each arc state also has a mutex which is used to protect the
* buffer list associated with the state. When attempting to
* obtain a hash table lock while holding an arc list lock you
* must use: mutex_tryenter() to avoid deadlock. Also note that
* the active state mutex must be held before the ghost state mutex.
*
* Arc buffers may have an associated eviction callback function.
* This function will be invoked prior to removing the buffer (e.g.
* in arc_do_user_evicts()). Note however that the data associated
* with the buffer may be evicted prior to the callback. The callback
* must be made with *no locks held* (to prevent deadlock). Additionally,
* the users of callbacks must ensure that their private data is
* protected from simultaneous callbacks from arc_clear_callback()
2008-11-20 23:01:55 +03:00
* and arc_do_user_evicts().
*
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
* It as also possible to register a callback which is run when the
* arc_meta_limit is reached and no buffers can be safely evicted. In
* this case the arc user should drop a reference on some arc buffers so
* they can be reclaimed and the arc_meta_limit honored. For example,
* when using the ZPL each dentry holds a references on a znode. These
* dentries must be pruned before the arc buffer holding the znode can
* be safely evicted.
*
2008-11-20 23:01:55 +03:00
* Note that the majority of the performance stats are manipulated
* with atomic operations.
*
* The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
*
* - L2ARC buflist creation
* - L2ARC buflist eviction
* - L2ARC write completion, which walks L2ARC buflists
* - ARC header destruction, as it removes from L2ARC buflists
* - ARC header release, as it removes from L2ARC buflists
*/
#include <sys/spa.h>
#include <sys/zio.h>
#include <sys/zio_compress.h>
2008-11-20 23:01:55 +03:00
#include <sys/zfs_context.h>
#include <sys/arc.h>
#include <sys/vdev.h>
2009-07-03 02:44:48 +04:00
#include <sys/vdev_impl.h>
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
#include <sys/dsl_pool.h>
2008-11-20 23:01:55 +03:00
#ifdef _KERNEL
#include <sys/vmsystm.h>
#include <vm/anon.h>
#include <sys/fs/swapnode.h>
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
#include <sys/zpl.h>
#include <linux/mm_compat.h>
2008-11-20 23:01:55 +03:00
#endif
#include <sys/callb.h>
#include <sys/kstat.h>
#include <sys/dmu_tx.h>
#include <zfs_fletcher.h>
#include <sys/arc_impl.h>
Remove duplicate typedefs from trace.h Older versions of GCC (e.g. GCC 4.4.7 on RHEL6) do not allow duplicate typedef declarations with the same type. The trace.h header contains some typedefs to avoid 'unknown type' errors for C files that haven't declared the type in question. But this causes build failures for C files that have already declared the type. Newer versions of GCC (e.g. v4.6) allow duplicate typedefs with the same type unless pedantic error checking is in force. To support the older versions we need to remove the duplicate typedefs. Removal of the typedefs means we can't built tracepoints code using those types unless the required headers have been included. To facilitate this, all tracepoint event declarations have been moved out of trace.h into separate headers. Each new header is explicitly included from the C file that uses the events defined therein. The trace.h header is still indirectly included form zfs_context.h and provides the implementation of the dprintf(), dbgmsg(), and SET_ERROR() interfaces. This makes those interfaces readily available throughout the code base. The macros that redefine DTRACE_PROBE* to use Linux tracepoints are also still provided by trace.h, so it is a prerequisite for the other trace_*.h headers. These new Linux implementation-specific headers do introduce a small divergence from upstream ZFS in several core C files, but this should not present a significant maintenance burden. Signed-off-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2953
2014-12-13 05:07:39 +03:00
#include <sys/trace_arc.h>
2008-11-20 23:01:55 +03:00
#ifndef _KERNEL
/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
boolean_t arc_watch = B_FALSE;
#endif
2008-11-20 23:01:55 +03:00
static kmutex_t arc_reclaim_thr_lock;
static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
static uint8_t arc_thread_exit;
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
/* number of bytes to prune from caches when at arc_meta_limit is reached */
int zfs_arc_meta_prune = 1048576;
2008-11-20 23:01:55 +03:00
typedef enum arc_reclaim_strategy {
ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
ARC_RECLAIM_CONS /* Conservative reclaim strategy */
} arc_reclaim_strategy_t;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
/*
* The number of iterations through arc_evict_*() before we
* drop & reacquire the lock.
*/
int arc_evict_iterations = 100;
2008-11-20 23:01:55 +03:00
/* number of seconds before growing cache again */
int zfs_arc_grow_retry = 5;
2008-11-20 23:01:55 +03:00
Disable aggressive arc_p growth by default For specific workloads consisting mainly of mfu data and new anon data buffers, the aggressive growth of arc_p found in the arc_get_data_buf() function can have detrimental effects on the mfu list size and ghost list hit rate. Running a workload consisting of two processes: * Process 1 is creating many small files * Process 2 is tar'ing a directory consisting of many small files I've seen arc_p and the mru grow to their maximum size, while the mru ghost list receives 100K times fewer hits than the mfu ghost list. Ideally, as the mfu ghost list receives hits, arc_p should be driven down and the size of the mfu should increase. Given the specific workload I was testing with, the mfu list size should grow to a point where almost no mfu ghost list hits would occur. Unfortunately, this does not happen because the newly dirtied anon buffers constancy drive arc_p to its maximum value and keep it there (effectively prioritizing the mru list and starving the mfu list down to a negligible size). The logic to increment arc_p from within the arc_get_data_buf() function was introduced many years ago in this upstream commit: commit 641fbdae3a027d12b3c3dcd18927ccafae6d58bc Author: maybee <none@none> Date: Wed Dec 20 15:46:12 2006 -0800 6505658 target MRU size (arc.p) needs to be adjusted more aggressively and since I don't fully understand the motivation for the change, I am reluctant to completely remove it. As a way to test out how it's removal might affect performance, I've disabled that code by default, but left it tunable via a module option. Thus, if its removal is found to be grossly detrimental for certain workloads, it can be re-enabled on the fly, without a code change. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-11 21:40:13 +04:00
/* disable anon data aggressively growing arc_p */
int zfs_arc_p_aggressive_disable = 1;
/* disable arc_p adapt dampener in arc_adapt */
int zfs_arc_p_dampener_disable = 1;
2009-02-18 23:51:31 +03:00
/* log2(fraction of arc to reclaim) */
int zfs_arc_shrink_shift = 5;
2009-02-18 23:51:31 +03:00
2008-11-20 23:01:55 +03:00
/*
* minimum lifespan of a prefetch block in clock ticks
* (initialized in arc_init())
*/
int zfs_arc_min_prefetch_lifespan = HZ;
/* disable arc proactive arc throttle due to low memory */
int zfs_arc_memory_throttle_disable = 1;
/* disable duplicate buffer eviction */
int zfs_disable_dup_eviction = 0;
2008-11-20 23:01:55 +03:00
/* average block used to size buf_hash_table */
int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
/*
* If this percent of memory is free, don't throttle.
*/
int arc_lotsfree_percent = 10;
2008-11-20 23:01:55 +03:00
static int arc_dead;
/* expiration time for arc_no_grow */
static clock_t arc_grow_time = 0;
/*
* The arc has filled available memory and has now warmed up.
*/
static boolean_t arc_warm;
2008-11-20 23:01:55 +03:00
/*
* These tunables are for performance analysis.
*/
unsigned long zfs_arc_max = 0;
unsigned long zfs_arc_min = 0;
unsigned long zfs_arc_meta_limit = 0;
2008-11-20 23:01:55 +03:00
/* The 6 states: */
static arc_state_t ARC_anon;
static arc_state_t ARC_mru;
static arc_state_t ARC_mru_ghost;
static arc_state_t ARC_mfu;
static arc_state_t ARC_mfu_ghost;
static arc_state_t ARC_l2c_only;
typedef struct arc_stats {
kstat_named_t arcstat_hits;
kstat_named_t arcstat_misses;
kstat_named_t arcstat_demand_data_hits;
kstat_named_t arcstat_demand_data_misses;
kstat_named_t arcstat_demand_metadata_hits;
kstat_named_t arcstat_demand_metadata_misses;
kstat_named_t arcstat_prefetch_data_hits;
kstat_named_t arcstat_prefetch_data_misses;
kstat_named_t arcstat_prefetch_metadata_hits;
kstat_named_t arcstat_prefetch_metadata_misses;
kstat_named_t arcstat_mru_hits;
kstat_named_t arcstat_mru_ghost_hits;
kstat_named_t arcstat_mfu_hits;
kstat_named_t arcstat_mfu_ghost_hits;
kstat_named_t arcstat_deleted;
kstat_named_t arcstat_recycle_miss;
/*
* Number of buffers that could not be evicted because the hash lock
* was held by another thread. The lock may not necessarily be held
* by something using the same buffer, since hash locks are shared
* by multiple buffers.
*/
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_mutex_miss;
/*
* Number of buffers skipped because they have I/O in progress, are
* indrect prefetch buffers that have not lived long enough, or are
* not from the spa we're trying to evict from.
*/
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_evict_skip;
kstat_named_t arcstat_evict_l2_cached;
kstat_named_t arcstat_evict_l2_eligible;
kstat_named_t arcstat_evict_l2_ineligible;
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_hash_elements;
kstat_named_t arcstat_hash_elements_max;
kstat_named_t arcstat_hash_collisions;
kstat_named_t arcstat_hash_chains;
kstat_named_t arcstat_hash_chain_max;
kstat_named_t arcstat_p;
kstat_named_t arcstat_c;
kstat_named_t arcstat_c_min;
kstat_named_t arcstat_c_max;
kstat_named_t arcstat_size;
kstat_named_t arcstat_hdr_size;
2009-02-18 23:51:31 +03:00
kstat_named_t arcstat_data_size;
kstat_named_t arcstat_meta_size;
2009-02-18 23:51:31 +03:00
kstat_named_t arcstat_other_size;
kstat_named_t arcstat_anon_size;
kstat_named_t arcstat_anon_evict_data;
kstat_named_t arcstat_anon_evict_metadata;
kstat_named_t arcstat_mru_size;
kstat_named_t arcstat_mru_evict_data;
kstat_named_t arcstat_mru_evict_metadata;
kstat_named_t arcstat_mru_ghost_size;
kstat_named_t arcstat_mru_ghost_evict_data;
kstat_named_t arcstat_mru_ghost_evict_metadata;
kstat_named_t arcstat_mfu_size;
kstat_named_t arcstat_mfu_evict_data;
kstat_named_t arcstat_mfu_evict_metadata;
kstat_named_t arcstat_mfu_ghost_size;
kstat_named_t arcstat_mfu_ghost_evict_data;
kstat_named_t arcstat_mfu_ghost_evict_metadata;
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_l2_hits;
kstat_named_t arcstat_l2_misses;
kstat_named_t arcstat_l2_feeds;
kstat_named_t arcstat_l2_rw_clash;
2009-02-18 23:51:31 +03:00
kstat_named_t arcstat_l2_read_bytes;
kstat_named_t arcstat_l2_write_bytes;
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_l2_writes_sent;
kstat_named_t arcstat_l2_writes_done;
kstat_named_t arcstat_l2_writes_error;
kstat_named_t arcstat_l2_writes_hdr_miss;
kstat_named_t arcstat_l2_evict_lock_retry;
kstat_named_t arcstat_l2_evict_reading;
kstat_named_t arcstat_l2_free_on_write;
kstat_named_t arcstat_l2_cdata_free_on_write;
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_l2_abort_lowmem;
kstat_named_t arcstat_l2_cksum_bad;
kstat_named_t arcstat_l2_io_error;
kstat_named_t arcstat_l2_size;
kstat_named_t arcstat_l2_asize;
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_l2_hdr_size;
kstat_named_t arcstat_l2_compress_successes;
kstat_named_t arcstat_l2_compress_zeros;
kstat_named_t arcstat_l2_compress_failures;
2008-11-20 23:01:55 +03:00
kstat_named_t arcstat_memory_throttle_count;
kstat_named_t arcstat_duplicate_buffers;
kstat_named_t arcstat_duplicate_buffers_size;
kstat_named_t arcstat_duplicate_reads;
kstat_named_t arcstat_memory_direct_count;
kstat_named_t arcstat_memory_indirect_count;
kstat_named_t arcstat_no_grow;
kstat_named_t arcstat_tempreserve;
kstat_named_t arcstat_loaned_bytes;
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
kstat_named_t arcstat_prune;
kstat_named_t arcstat_meta_used;
kstat_named_t arcstat_meta_limit;
kstat_named_t arcstat_meta_max;
2008-11-20 23:01:55 +03:00
} arc_stats_t;
static arc_stats_t arc_stats = {
{ "hits", KSTAT_DATA_UINT64 },
{ "misses", KSTAT_DATA_UINT64 },
{ "demand_data_hits", KSTAT_DATA_UINT64 },
{ "demand_data_misses", KSTAT_DATA_UINT64 },
{ "demand_metadata_hits", KSTAT_DATA_UINT64 },
{ "demand_metadata_misses", KSTAT_DATA_UINT64 },
{ "prefetch_data_hits", KSTAT_DATA_UINT64 },
{ "prefetch_data_misses", KSTAT_DATA_UINT64 },
{ "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
{ "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
{ "mru_hits", KSTAT_DATA_UINT64 },
{ "mru_ghost_hits", KSTAT_DATA_UINT64 },
{ "mfu_hits", KSTAT_DATA_UINT64 },
{ "mfu_ghost_hits", KSTAT_DATA_UINT64 },
{ "deleted", KSTAT_DATA_UINT64 },
{ "recycle_miss", KSTAT_DATA_UINT64 },
{ "mutex_miss", KSTAT_DATA_UINT64 },
{ "evict_skip", KSTAT_DATA_UINT64 },
{ "evict_l2_cached", KSTAT_DATA_UINT64 },
{ "evict_l2_eligible", KSTAT_DATA_UINT64 },
{ "evict_l2_ineligible", KSTAT_DATA_UINT64 },
2008-11-20 23:01:55 +03:00
{ "hash_elements", KSTAT_DATA_UINT64 },
{ "hash_elements_max", KSTAT_DATA_UINT64 },
{ "hash_collisions", KSTAT_DATA_UINT64 },
{ "hash_chains", KSTAT_DATA_UINT64 },
{ "hash_chain_max", KSTAT_DATA_UINT64 },
{ "p", KSTAT_DATA_UINT64 },
{ "c", KSTAT_DATA_UINT64 },
{ "c_min", KSTAT_DATA_UINT64 },
{ "c_max", KSTAT_DATA_UINT64 },
{ "size", KSTAT_DATA_UINT64 },
{ "hdr_size", KSTAT_DATA_UINT64 },
2009-02-18 23:51:31 +03:00
{ "data_size", KSTAT_DATA_UINT64 },
{ "meta_size", KSTAT_DATA_UINT64 },
2009-02-18 23:51:31 +03:00
{ "other_size", KSTAT_DATA_UINT64 },
{ "anon_size", KSTAT_DATA_UINT64 },
{ "anon_evict_data", KSTAT_DATA_UINT64 },
{ "anon_evict_metadata", KSTAT_DATA_UINT64 },
{ "mru_size", KSTAT_DATA_UINT64 },
{ "mru_evict_data", KSTAT_DATA_UINT64 },
{ "mru_evict_metadata", KSTAT_DATA_UINT64 },
{ "mru_ghost_size", KSTAT_DATA_UINT64 },
{ "mru_ghost_evict_data", KSTAT_DATA_UINT64 },
{ "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 },
{ "mfu_size", KSTAT_DATA_UINT64 },
{ "mfu_evict_data", KSTAT_DATA_UINT64 },
{ "mfu_evict_metadata", KSTAT_DATA_UINT64 },
{ "mfu_ghost_size", KSTAT_DATA_UINT64 },
{ "mfu_ghost_evict_data", KSTAT_DATA_UINT64 },
{ "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 },
2008-11-20 23:01:55 +03:00
{ "l2_hits", KSTAT_DATA_UINT64 },
{ "l2_misses", KSTAT_DATA_UINT64 },
{ "l2_feeds", KSTAT_DATA_UINT64 },
{ "l2_rw_clash", KSTAT_DATA_UINT64 },
2009-02-18 23:51:31 +03:00
{ "l2_read_bytes", KSTAT_DATA_UINT64 },
{ "l2_write_bytes", KSTAT_DATA_UINT64 },
2008-11-20 23:01:55 +03:00
{ "l2_writes_sent", KSTAT_DATA_UINT64 },
{ "l2_writes_done", KSTAT_DATA_UINT64 },
{ "l2_writes_error", KSTAT_DATA_UINT64 },
{ "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
{ "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
{ "l2_evict_reading", KSTAT_DATA_UINT64 },
{ "l2_free_on_write", KSTAT_DATA_UINT64 },
{ "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
2008-11-20 23:01:55 +03:00
{ "l2_abort_lowmem", KSTAT_DATA_UINT64 },
{ "l2_cksum_bad", KSTAT_DATA_UINT64 },
{ "l2_io_error", KSTAT_DATA_UINT64 },
{ "l2_size", KSTAT_DATA_UINT64 },
{ "l2_asize", KSTAT_DATA_UINT64 },
2008-11-20 23:01:55 +03:00
{ "l2_hdr_size", KSTAT_DATA_UINT64 },
{ "l2_compress_successes", KSTAT_DATA_UINT64 },
{ "l2_compress_zeros", KSTAT_DATA_UINT64 },
{ "l2_compress_failures", KSTAT_DATA_UINT64 },
{ "memory_throttle_count", KSTAT_DATA_UINT64 },
{ "duplicate_buffers", KSTAT_DATA_UINT64 },
{ "duplicate_buffers_size", KSTAT_DATA_UINT64 },
{ "duplicate_reads", KSTAT_DATA_UINT64 },
{ "memory_direct_count", KSTAT_DATA_UINT64 },
{ "memory_indirect_count", KSTAT_DATA_UINT64 },
{ "arc_no_grow", KSTAT_DATA_UINT64 },
{ "arc_tempreserve", KSTAT_DATA_UINT64 },
{ "arc_loaned_bytes", KSTAT_DATA_UINT64 },
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
{ "arc_prune", KSTAT_DATA_UINT64 },
{ "arc_meta_used", KSTAT_DATA_UINT64 },
{ "arc_meta_limit", KSTAT_DATA_UINT64 },
{ "arc_meta_max", KSTAT_DATA_UINT64 },
2008-11-20 23:01:55 +03:00
};
#define ARCSTAT(stat) (arc_stats.stat.value.ui64)
#define ARCSTAT_INCR(stat, val) \
atomic_add_64(&arc_stats.stat.value.ui64, (val))
2008-11-20 23:01:55 +03:00
#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
2008-11-20 23:01:55 +03:00
#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
#define ARCSTAT_MAX(stat, val) { \
uint64_t m; \
while ((val) > (m = arc_stats.stat.value.ui64) && \
(m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
continue; \
}
#define ARCSTAT_MAXSTAT(stat) \
ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
/*
* We define a macro to allow ARC hits/misses to be easily broken down by
* two separate conditions, giving a total of four different subtypes for
* each of hits and misses (so eight statistics total).
*/
#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
if (cond1) { \
if (cond2) { \
ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
} else { \
ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
} \
} else { \
if (cond2) { \
ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
} else { \
ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
} \
}
kstat_t *arc_ksp;
static arc_state_t *arc_anon;
2008-11-20 23:01:55 +03:00
static arc_state_t *arc_mru;
static arc_state_t *arc_mru_ghost;
static arc_state_t *arc_mfu;
static arc_state_t *arc_mfu_ghost;
static arc_state_t *arc_l2c_only;
/*
* There are several ARC variables that are critical to export as kstats --
* but we don't want to have to grovel around in the kstat whenever we wish to
* manipulate them. For these variables, we therefore define them to be in
* terms of the statistic variable. This assures that we are not introducing
* the possibility of inconsistency by having shadow copies of the variables,
* while still allowing the code to be readable.
*/
#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
#define arc_c ARCSTAT(arcstat_c) /* target size of cache */
#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
#define arc_no_grow ARCSTAT(arcstat_no_grow)
#define arc_tempreserve ARCSTAT(arcstat_tempreserve)
#define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
2008-11-20 23:01:55 +03:00
#define L2ARC_IS_VALID_COMPRESS(_c_) \
((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
static list_t arc_prune_list;
static kmutex_t arc_prune_mtx;
2008-11-20 23:01:55 +03:00
static arc_buf_t *arc_eviction_list;
static kmutex_t arc_eviction_mtx;
static arc_buf_hdr_t arc_eviction_hdr;
static void arc_get_data_buf(arc_buf_t *buf);
static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
static int arc_evict_needed(arc_buf_contents_t type);
static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes,
arc_buf_contents_t type);
static void arc_buf_watch(arc_buf_t *buf);
2008-11-20 23:01:55 +03:00
static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
2008-11-20 23:01:55 +03:00
#define GHOST_STATE(state) \
((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
(state) == arc_l2c_only)
/*
* Private ARC flags. These flags are private ARC only flags that will show up
* in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
* be passed in as arc_flags in things like arc_read. However, these flags
* should never be passed and should only be set by ARC code. When adding new
* public flags, make sure not to smash the private ones.
*/
#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
#define ARC_INDIRECT (1 << 14) /* this is an indirect block */
#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
2008-11-20 23:01:55 +03:00
#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
2009-02-18 23:51:31 +03:00
#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
2008-11-20 23:01:55 +03:00
#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
(hdr)->b_l2hdr != NULL)
2008-11-20 23:01:55 +03:00
#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
/*
* Other sizes
*/
#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
/*
* Hash table routines
*/
#define HT_LOCK_ALIGN 64
#define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
2008-11-20 23:01:55 +03:00
struct ht_lock {
kmutex_t ht_lock;
#ifdef _KERNEL
unsigned char pad[HT_LOCK_PAD];
2008-11-20 23:01:55 +03:00
#endif
};
#define BUF_LOCKS 8192
2008-11-20 23:01:55 +03:00
typedef struct buf_hash_table {
uint64_t ht_mask;
arc_buf_hdr_t **ht_table;
struct ht_lock ht_locks[BUF_LOCKS];
} buf_hash_table_t;
static buf_hash_table_t buf_hash_table;
#define BUF_HASH_INDEX(spa, dva, birth) \
(buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
#define HDR_LOCK(hdr) \
(BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
2008-11-20 23:01:55 +03:00
uint64_t zfs_crc64_table[256];
/*
* Level 2 ARC
*/
#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
#define L2ARC_HEADROOM 2 /* num of writes */
/*
* If we discover during ARC scan any buffers to be compressed, we boost
* our headroom for the next scanning cycle by this percentage multiple.
*/
#define L2ARC_HEADROOM_BOOST 200
2009-02-18 23:51:31 +03:00
#define L2ARC_FEED_SECS 1 /* caching interval secs */
#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
2008-11-20 23:01:55 +03:00
#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
/* L2ARC Performance Tunables */
unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
int l2arc_nocompress = B_FALSE; /* don't compress bufs */
int l2arc_feed_again = B_TRUE; /* turbo warmup */
int l2arc_norw = B_FALSE; /* no reads during writes */
2008-11-20 23:01:55 +03:00
/*
* L2ARC Internals
*/
static list_t L2ARC_dev_list; /* device list */
static list_t *l2arc_dev_list; /* device list pointer */
static kmutex_t l2arc_dev_mtx; /* device list mutex */
static l2arc_dev_t *l2arc_dev_last; /* last device used */
static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
static list_t L2ARC_free_on_write; /* free after write buf list */
static list_t *l2arc_free_on_write; /* free after write list ptr */
static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
static uint64_t l2arc_ndev; /* number of devices */
typedef struct l2arc_read_callback {
arc_buf_t *l2rcb_buf; /* read buffer */
spa_t *l2rcb_spa; /* spa */
blkptr_t l2rcb_bp; /* original blkptr */
zbookmark_phys_t l2rcb_zb; /* original bookmark */
int l2rcb_flags; /* original flags */
enum zio_compress l2rcb_compress; /* applied compress */
2008-11-20 23:01:55 +03:00
} l2arc_read_callback_t;
struct l2arc_buf_hdr {
/* protected by arc_buf_hdr mutex */
l2arc_dev_t *b_dev; /* L2ARC device */
uint64_t b_daddr; /* disk address, offset byte */
/* compression applied to buffer data */
enum zio_compress b_compress;
/* real alloc'd buffer size depending on b_compress applied */
uint32_t b_hits;
uint64_t b_asize;
/* temporary buffer holder for in-flight compressed data */
void *b_tmp_cdata;
2008-11-20 23:01:55 +03:00
};
typedef struct l2arc_data_free {
/* protected by l2arc_free_on_write_mtx */
void *l2df_data;
size_t l2df_size;
void (*l2df_func)(void *, size_t);
list_node_t l2df_list_node;
} l2arc_data_free_t;
static kmutex_t l2arc_feed_thr_lock;
static kcondvar_t l2arc_feed_thr_cv;
static uint8_t l2arc_thread_exit;
static void l2arc_read_done(zio_t *zio);
static void l2arc_hdr_stat_add(void);
static void l2arc_hdr_stat_remove(void);
static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
enum zio_compress c);
static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
2008-11-20 23:01:55 +03:00
static uint64_t
2009-02-18 23:51:31 +03:00
buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
2008-11-20 23:01:55 +03:00
{
uint8_t *vdva = (uint8_t *)dva;
uint64_t crc = -1ULL;
int i;
ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
for (i = 0; i < sizeof (dva_t); i++)
crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
2009-02-18 23:51:31 +03:00
crc ^= (spa>>8) ^ birth;
2008-11-20 23:01:55 +03:00
return (crc);
}
#define BUF_EMPTY(buf) \
((buf)->b_dva.dva_word[0] == 0 && \
(buf)->b_dva.dva_word[1] == 0 && \
(buf)->b_cksum0 == 0)
2008-11-20 23:01:55 +03:00
#define BUF_EQUAL(spa, dva, birth, buf) \
((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
((buf)->b_birth == birth) && ((buf)->b_spa == spa)
static void
buf_discard_identity(arc_buf_hdr_t *hdr)
{
hdr->b_dva.dva_word[0] = 0;
hdr->b_dva.dva_word[1] = 0;
hdr->b_birth = 0;
hdr->b_cksum0 = 0;
}
2008-11-20 23:01:55 +03:00
static arc_buf_hdr_t *
buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
2008-11-20 23:01:55 +03:00
{
const dva_t *dva = BP_IDENTITY(bp);
uint64_t birth = BP_PHYSICAL_BIRTH(bp);
2008-11-20 23:01:55 +03:00
uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
arc_buf_hdr_t *buf;
mutex_enter(hash_lock);
for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
buf = buf->b_hash_next) {
if (BUF_EQUAL(spa, dva, birth, buf)) {
*lockp = hash_lock;
return (buf);
}
}
mutex_exit(hash_lock);
*lockp = NULL;
return (NULL);
}
/*
* Insert an entry into the hash table. If there is already an element
* equal to elem in the hash table, then the already existing element
* will be returned and the new element will not be inserted.
* Otherwise returns NULL.
*/
static arc_buf_hdr_t *
buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
{
uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
arc_buf_hdr_t *fbuf;
uint32_t i;
ASSERT(!DVA_IS_EMPTY(&buf->b_dva));
ASSERT(buf->b_birth != 0);
2008-11-20 23:01:55 +03:00
ASSERT(!HDR_IN_HASH_TABLE(buf));
*lockp = hash_lock;
mutex_enter(hash_lock);
for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
fbuf = fbuf->b_hash_next, i++) {
if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
return (fbuf);
}
buf->b_hash_next = buf_hash_table.ht_table[idx];
buf_hash_table.ht_table[idx] = buf;
buf->b_flags |= ARC_IN_HASH_TABLE;
/* collect some hash table performance data */
if (i > 0) {
ARCSTAT_BUMP(arcstat_hash_collisions);
if (i == 1)
ARCSTAT_BUMP(arcstat_hash_chains);
ARCSTAT_MAX(arcstat_hash_chain_max, i);
}
ARCSTAT_BUMP(arcstat_hash_elements);
ARCSTAT_MAXSTAT(arcstat_hash_elements);
return (NULL);
}
static void
buf_hash_remove(arc_buf_hdr_t *buf)
{
arc_buf_hdr_t *fbuf, **bufp;
uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
ASSERT(HDR_IN_HASH_TABLE(buf));
bufp = &buf_hash_table.ht_table[idx];
while ((fbuf = *bufp) != buf) {
ASSERT(fbuf != NULL);
bufp = &fbuf->b_hash_next;
}
*bufp = buf->b_hash_next;
buf->b_hash_next = NULL;
buf->b_flags &= ~ARC_IN_HASH_TABLE;
/* collect some hash table performance data */
ARCSTAT_BUMPDOWN(arcstat_hash_elements);
if (buf_hash_table.ht_table[idx] &&
buf_hash_table.ht_table[idx]->b_hash_next == NULL)
ARCSTAT_BUMPDOWN(arcstat_hash_chains);
}
/*
* Global data structures and functions for the buf kmem cache.
*/
static kmem_cache_t *hdr_cache;
static kmem_cache_t *buf_cache;
static kmem_cache_t *l2arc_hdr_cache;
2008-11-20 23:01:55 +03:00
static void
buf_fini(void)
{
int i;
#if defined(_KERNEL) && defined(HAVE_SPL)
/*
* Large allocations which do not require contiguous pages
* should be using vmem_free() in the linux kernel\
*/
vmem_free(buf_hash_table.ht_table,
(buf_hash_table.ht_mask + 1) * sizeof (void *));
#else
2008-11-20 23:01:55 +03:00
kmem_free(buf_hash_table.ht_table,
(buf_hash_table.ht_mask + 1) * sizeof (void *));
#endif
2008-11-20 23:01:55 +03:00
for (i = 0; i < BUF_LOCKS; i++)
mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
kmem_cache_destroy(hdr_cache);
kmem_cache_destroy(buf_cache);
kmem_cache_destroy(l2arc_hdr_cache);
2008-11-20 23:01:55 +03:00
}
/*
* Constructor callback - called when the cache is empty
* and a new buf is requested.
*/
/* ARGSUSED */
static int
hdr_cons(void *vbuf, void *unused, int kmflag)
{
arc_buf_hdr_t *buf = vbuf;
bzero(buf, sizeof (arc_buf_hdr_t));
refcount_create(&buf->b_refcnt);
cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
list_link_init(&buf->b_arc_node);
list_link_init(&buf->b_l2node);
2009-02-18 23:51:31 +03:00
arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
2008-11-20 23:01:55 +03:00
return (0);
}
/* ARGSUSED */
static int
buf_cons(void *vbuf, void *unused, int kmflag)
{
arc_buf_t *buf = vbuf;
bzero(buf, sizeof (arc_buf_t));
mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
2009-02-18 23:51:31 +03:00
arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
return (0);
}
2008-11-20 23:01:55 +03:00
/*
* Destructor callback - called when a cached buf is
* no longer required.
*/
/* ARGSUSED */
static void
hdr_dest(void *vbuf, void *unused)
{
arc_buf_hdr_t *buf = vbuf;
ASSERT(BUF_EMPTY(buf));
2008-11-20 23:01:55 +03:00
refcount_destroy(&buf->b_refcnt);
cv_destroy(&buf->b_cv);
mutex_destroy(&buf->b_freeze_lock);
2009-02-18 23:51:31 +03:00
arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
2008-11-20 23:01:55 +03:00
}
/* ARGSUSED */
static void
buf_dest(void *vbuf, void *unused)
{
arc_buf_t *buf = vbuf;
mutex_destroy(&buf->b_evict_lock);
2009-02-18 23:51:31 +03:00
arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
}
2008-11-20 23:01:55 +03:00
static void
buf_init(void)
{
uint64_t *ct;
uint64_t hsize = 1ULL << 12;
int i, j;
/*
* The hash table is big enough to fill all of physical memory
* with an average block size of zfs_arc_average_blocksize (default 8K).
* By default, the table will take up
* totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
2008-11-20 23:01:55 +03:00
*/
while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
2008-11-20 23:01:55 +03:00
hsize <<= 1;
retry:
buf_hash_table.ht_mask = hsize - 1;
#if defined(_KERNEL) && defined(HAVE_SPL)
/*
* Large allocations which do not require contiguous pages
* should be using vmem_alloc() in the linux kernel
*/
buf_hash_table.ht_table =
vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
#else
2008-11-20 23:01:55 +03:00
buf_hash_table.ht_table =
kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
#endif
2008-11-20 23:01:55 +03:00
if (buf_hash_table.ht_table == NULL) {
ASSERT(hsize > (1ULL << 8));
hsize >>= 1;
goto retry;
}
hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
2008-11-20 23:01:55 +03:00
buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
l2arc_hdr_cache = kmem_cache_create("l2arc_buf_hdr_t", L2HDR_SIZE,
0, NULL, NULL, NULL, NULL, NULL, 0);
2008-11-20 23:01:55 +03:00
for (i = 0; i < 256; i++)
for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
*ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
for (i = 0; i < BUF_LOCKS; i++) {
mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
NULL, MUTEX_FSTRANS, NULL);
2008-11-20 23:01:55 +03:00
}
}
#define ARC_MINTIME (hz>>4) /* 62 ms */
static void
arc_cksum_verify(arc_buf_t *buf)
{
zio_cksum_t zc;
if (!(zfs_flags & ZFS_DEBUG_MODIFY))
return;
mutex_enter(&buf->b_hdr->b_freeze_lock);
if (buf->b_hdr->b_freeze_cksum == NULL ||
(buf->b_hdr->b_flags & ARC_IO_ERROR)) {
mutex_exit(&buf->b_hdr->b_freeze_lock);
return;
}
fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
panic("buffer modified while frozen!");
mutex_exit(&buf->b_hdr->b_freeze_lock);
}
static int
arc_cksum_equal(arc_buf_t *buf)
{
zio_cksum_t zc;
int equal;
mutex_enter(&buf->b_hdr->b_freeze_lock);
fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
mutex_exit(&buf->b_hdr->b_freeze_lock);
return (equal);
}
static void
arc_cksum_compute(arc_buf_t *buf, boolean_t force)
{
if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
return;
mutex_enter(&buf->b_hdr->b_freeze_lock);
if (buf->b_hdr->b_freeze_cksum != NULL) {
mutex_exit(&buf->b_hdr->b_freeze_lock);
return;
}
Use KM_PUSHPAGE in l2arc_write_buffers There is potential for deadlock in the l2arc_feed thread if KM_PUSHPAGE is not used for the allocations made in l2arc_write_buffers. Specifically, if KM_PUSHPAGE is not used for these allocations, it is possible for reclaim to be triggered which can cause the l2arc_feed thread to deadlock itself on the ARC_mru mutex. An example of this is demonstrated in the following backtrace of the l2arc_feed thread: crash> bt 4123 PID: 4123 TASK: ffff88062f8c1500 CPU: 6 COMMAND: "l2arc_feed" 0 [ffff88062511d610] schedule at ffffffff814eeee0 1 [ffff88062511d6d8] __mutex_lock_slowpath at ffffffff814f057e 2 [ffff88062511d748] mutex_lock at ffffffff814f041b 3 [ffff88062511d768] arc_evict at ffffffffa05130ca [zfs] 4 [ffff88062511d858] arc_adjust at ffffffffa05139a9 [zfs] 5 [ffff88062511d878] arc_shrink at ffffffffa0513a95 [zfs] 6 [ffff88062511d898] arc_kmem_reap_now at ffffffffa0513be8 [zfs] 7 [ffff88062511d8c8] arc_shrinker_func at ffffffffa0513ccc [zfs] 8 [ffff88062511d8f8] shrink_slab at ffffffff8112a17a 9 [ffff88062511d958] do_try_to_free_pages at ffffffff8112bfdf 10 [ffff88062511d9e8] try_to_free_pages at ffffffff8112c3ed 11 [ffff88062511da98] __alloc_pages_nodemask at ffffffff8112431d 12 [ffff88062511dbb8] kmem_getpages at ffffffff8115e632 13 [ffff88062511dbe8] fallback_alloc at ffffffff8115f24a 14 [ffff88062511dc68] ____cache_alloc_node at ffffffff8115efc9 15 [ffff88062511dcc8] __kmalloc at ffffffff8115fbf9 16 [ffff88062511dd18] kmem_alloc_debug at ffffffffa047b8cb [spl] 17 [ffff88062511dda8] l2arc_feed_thread at ffffffffa0511e71 [zfs] 18 [ffff88062511dea8] thread_generic_wrapper at ffffffffa047d1a1 [spl] 19 [ffff88062511dee8] kthread at ffffffff81090a86 20 [ffff88062511df48] kernel_thread at ffffffff8100c14a Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-04-10 21:55:17 +04:00
buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
KM_SLEEP);
2008-11-20 23:01:55 +03:00
fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
buf->b_hdr->b_freeze_cksum);
mutex_exit(&buf->b_hdr->b_freeze_lock);
arc_buf_watch(buf);
}
#ifndef _KERNEL
void
arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
{
panic("Got SIGSEGV at address: 0x%lx\n", (long) si->si_addr);
}
#endif
/* ARGSUSED */
static void
arc_buf_unwatch(arc_buf_t *buf)
{
#ifndef _KERNEL
if (arc_watch) {
ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size,
PROT_READ | PROT_WRITE));
}
#endif
}
/* ARGSUSED */
static void
arc_buf_watch(arc_buf_t *buf)
{
#ifndef _KERNEL
if (arc_watch)
ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size, PROT_READ));
#endif
2008-11-20 23:01:55 +03:00
}
void
arc_buf_thaw(arc_buf_t *buf)
{
if (zfs_flags & ZFS_DEBUG_MODIFY) {
if (buf->b_hdr->b_state != arc_anon)
panic("modifying non-anon buffer!");
if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
panic("modifying buffer while i/o in progress!");
arc_cksum_verify(buf);
}
mutex_enter(&buf->b_hdr->b_freeze_lock);
if (buf->b_hdr->b_freeze_cksum != NULL) {
kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
buf->b_hdr->b_freeze_cksum = NULL;
}
2008-11-20 23:01:55 +03:00
mutex_exit(&buf->b_hdr->b_freeze_lock);
arc_buf_unwatch(buf);
2008-11-20 23:01:55 +03:00
}
void
arc_buf_freeze(arc_buf_t *buf)
{
kmutex_t *hash_lock;
2008-11-20 23:01:55 +03:00
if (!(zfs_flags & ZFS_DEBUG_MODIFY))
return;
hash_lock = HDR_LOCK(buf->b_hdr);
mutex_enter(hash_lock);
2008-11-20 23:01:55 +03:00
ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
buf->b_hdr->b_state == arc_anon);
arc_cksum_compute(buf, B_FALSE);
mutex_exit(hash_lock);
2008-11-20 23:01:55 +03:00
}
static void
add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
{
ASSERT(MUTEX_HELD(hash_lock));
if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
(ab->b_state != arc_anon)) {
uint64_t delta = ab->b_size * ab->b_datacnt;
list_t *list = &ab->b_state->arcs_list[ab->b_type];
uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
mutex_enter(&ab->b_state->arcs_mtx);
ASSERT(list_link_active(&ab->b_arc_node));
list_remove(list, ab);
if (GHOST_STATE(ab->b_state)) {
ASSERT0(ab->b_datacnt);
2008-11-20 23:01:55 +03:00
ASSERT3P(ab->b_buf, ==, NULL);
delta = ab->b_size;
}
ASSERT(delta > 0);
ASSERT3U(*size, >=, delta);
atomic_add_64(size, -delta);
mutex_exit(&ab->b_state->arcs_mtx);
/* remove the prefetch flag if we get a reference */
2008-11-20 23:01:55 +03:00
if (ab->b_flags & ARC_PREFETCH)
ab->b_flags &= ~ARC_PREFETCH;
}
}
static int
remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
{
int cnt;
arc_state_t *state = ab->b_state;
ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
ASSERT(!GHOST_STATE(state));
if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
(state != arc_anon)) {
uint64_t *size = &state->arcs_lsize[ab->b_type];
ASSERT(!MUTEX_HELD(&state->arcs_mtx));
mutex_enter(&state->arcs_mtx);
ASSERT(!list_link_active(&ab->b_arc_node));
list_insert_head(&state->arcs_list[ab->b_type], ab);
ASSERT(ab->b_datacnt > 0);
atomic_add_64(size, ab->b_size * ab->b_datacnt);
mutex_exit(&state->arcs_mtx);
}
return (cnt);
}
/*
* Returns detailed information about a specific arc buffer. When the
* state_index argument is set the function will calculate the arc header
* list position for its arc state. Since this requires a linear traversal
* callers are strongly encourage not to do this. However, it can be helpful
* for targeted analysis so the functionality is provided.
*/
void
arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
{
arc_buf_hdr_t *hdr = ab->b_hdr;
arc_state_t *state = hdr->b_state;
memset(abi, 0, sizeof (arc_buf_info_t));
abi->abi_flags = hdr->b_flags;
abi->abi_datacnt = hdr->b_datacnt;
abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
abi->abi_state_contents = hdr->b_type;
abi->abi_state_index = -1;
abi->abi_size = hdr->b_size;
abi->abi_access = hdr->b_arc_access;
abi->abi_mru_hits = hdr->b_mru_hits;
abi->abi_mru_ghost_hits = hdr->b_mru_ghost_hits;
abi->abi_mfu_hits = hdr->b_mfu_hits;
abi->abi_mfu_ghost_hits = hdr->b_mfu_ghost_hits;
abi->abi_holds = refcount_count(&hdr->b_refcnt);
if (hdr->b_l2hdr) {
abi->abi_l2arc_dattr = hdr->b_l2hdr->b_daddr;
abi->abi_l2arc_asize = hdr->b_l2hdr->b_asize;
abi->abi_l2arc_compress = hdr->b_l2hdr->b_compress;
abi->abi_l2arc_hits = hdr->b_l2hdr->b_hits;
}
if (state && state_index && list_link_active(&hdr->b_arc_node)) {
list_t *list = &state->arcs_list[hdr->b_type];
arc_buf_hdr_t *h;
mutex_enter(&state->arcs_mtx);
for (h = list_head(list); h != NULL; h = list_next(list, h)) {
abi->abi_state_index++;
if (h == hdr)
break;
}
mutex_exit(&state->arcs_mtx);
}
}
2008-11-20 23:01:55 +03:00
/*
* Move the supplied buffer to the indicated state. The mutex
* for the buffer must be held by the caller.
*/
static void
arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
{
arc_state_t *old_state = ab->b_state;
int64_t refcnt = refcount_count(&ab->b_refcnt);
uint64_t from_delta, to_delta;
ASSERT(MUTEX_HELD(hash_lock));
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
ASSERT3P(new_state, !=, old_state);
2008-11-20 23:01:55 +03:00
ASSERT(refcnt == 0 || ab->b_datacnt > 0);
ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
2008-11-20 23:01:55 +03:00
from_delta = to_delta = ab->b_datacnt * ab->b_size;
/*
* If this buffer is evictable, transfer it from the
* old state list to the new state list.
*/
if (refcnt == 0) {
if (old_state != arc_anon) {
int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
uint64_t *size = &old_state->arcs_lsize[ab->b_type];
if (use_mutex)
mutex_enter(&old_state->arcs_mtx);
ASSERT(list_link_active(&ab->b_arc_node));
list_remove(&old_state->arcs_list[ab->b_type], ab);
/*
* If prefetching out of the ghost cache,
* we will have a non-zero datacnt.
2008-11-20 23:01:55 +03:00
*/
if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
/* ghost elements have a ghost size */
ASSERT(ab->b_buf == NULL);
from_delta = ab->b_size;
}
ASSERT3U(*size, >=, from_delta);
atomic_add_64(size, -from_delta);
if (use_mutex)
mutex_exit(&old_state->arcs_mtx);
}
if (new_state != arc_anon) {
int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
uint64_t *size = &new_state->arcs_lsize[ab->b_type];
if (use_mutex)
mutex_enter(&new_state->arcs_mtx);
list_insert_head(&new_state->arcs_list[ab->b_type], ab);
/* ghost elements have a ghost size */
if (GHOST_STATE(new_state)) {
ASSERT(ab->b_datacnt == 0);
ASSERT(ab->b_buf == NULL);
to_delta = ab->b_size;
}
atomic_add_64(size, to_delta);
if (use_mutex)
mutex_exit(&new_state->arcs_mtx);
}
}
ASSERT(!BUF_EMPTY(ab));
if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
2008-11-20 23:01:55 +03:00
buf_hash_remove(ab);
/* adjust state sizes */
if (to_delta)
atomic_add_64(&new_state->arcs_size, to_delta);
if (from_delta) {
ASSERT3U(old_state->arcs_size, >=, from_delta);
atomic_add_64(&old_state->arcs_size, -from_delta);
}
ab->b_state = new_state;
/* adjust l2arc hdr stats */
if (new_state == arc_l2c_only)
l2arc_hdr_stat_add();
else if (old_state == arc_l2c_only)
l2arc_hdr_stat_remove();
}
void
2009-02-18 23:51:31 +03:00
arc_space_consume(uint64_t space, arc_space_type_t type)
2008-11-20 23:01:55 +03:00
{
2009-02-18 23:51:31 +03:00
ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
switch (type) {
default:
break;
2009-02-18 23:51:31 +03:00
case ARC_SPACE_DATA:
ARCSTAT_INCR(arcstat_data_size, space);
break;
case ARC_SPACE_META:
ARCSTAT_INCR(arcstat_meta_size, space);
break;
2009-02-18 23:51:31 +03:00
case ARC_SPACE_OTHER:
ARCSTAT_INCR(arcstat_other_size, space);
break;
case ARC_SPACE_HDRS:
ARCSTAT_INCR(arcstat_hdr_size, space);
break;
case ARC_SPACE_L2HDRS:
ARCSTAT_INCR(arcstat_l2_hdr_size, space);
break;
}
if (type != ARC_SPACE_DATA)
ARCSTAT_INCR(arcstat_meta_used, space);
2008-11-20 23:01:55 +03:00
atomic_add_64(&arc_size, space);
}
void
2009-02-18 23:51:31 +03:00
arc_space_return(uint64_t space, arc_space_type_t type)
2008-11-20 23:01:55 +03:00
{
2009-02-18 23:51:31 +03:00
ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
switch (type) {
default:
break;
2009-02-18 23:51:31 +03:00
case ARC_SPACE_DATA:
ARCSTAT_INCR(arcstat_data_size, -space);
break;
case ARC_SPACE_META:
ARCSTAT_INCR(arcstat_meta_size, -space);
break;
2009-02-18 23:51:31 +03:00
case ARC_SPACE_OTHER:
ARCSTAT_INCR(arcstat_other_size, -space);
break;
case ARC_SPACE_HDRS:
ARCSTAT_INCR(arcstat_hdr_size, -space);
break;
case ARC_SPACE_L2HDRS:
ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
break;
}
if (type != ARC_SPACE_DATA) {
ASSERT(arc_meta_used >= space);
if (arc_meta_max < arc_meta_used)
arc_meta_max = arc_meta_used;
ARCSTAT_INCR(arcstat_meta_used, -space);
}
2008-11-20 23:01:55 +03:00
ASSERT(arc_size >= space);
atomic_add_64(&arc_size, -space);
}
arc_buf_t *
arc_buf_alloc(spa_t *spa, uint64_t size, void *tag, arc_buf_contents_t type)
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *hdr;
arc_buf_t *buf;
VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
2008-11-20 23:01:55 +03:00
hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
ASSERT(BUF_EMPTY(hdr));
hdr->b_size = size;
hdr->b_type = type;
hdr->b_spa = spa_load_guid(spa);
2008-11-20 23:01:55 +03:00
hdr->b_state = arc_anon;
hdr->b_arc_access = 0;
hdr->b_mru_hits = 0;
hdr->b_mru_ghost_hits = 0;
hdr->b_mfu_hits = 0;
hdr->b_mfu_ghost_hits = 0;
hdr->b_l2_hits = 0;
2008-11-20 23:01:55 +03:00
buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
buf->b_hdr = hdr;
buf->b_data = NULL;
buf->b_efunc = NULL;
buf->b_private = NULL;
buf->b_next = NULL;
hdr->b_buf = buf;
arc_get_data_buf(buf);
hdr->b_datacnt = 1;
hdr->b_flags = 0;
ASSERT(refcount_is_zero(&hdr->b_refcnt));
(void) refcount_add(&hdr->b_refcnt, tag);
return (buf);
}
2009-07-03 02:44:48 +04:00
static char *arc_onloan_tag = "onloan";
/*
* Loan out an anonymous arc buffer. Loaned buffers are not counted as in
* flight data by arc_tempreserve_space() until they are "returned". Loaned
* buffers must be returned to the arc before they can be used by the DMU or
* freed.
*/
arc_buf_t *
arc_loan_buf(spa_t *spa, uint64_t size)
2009-07-03 02:44:48 +04:00
{
arc_buf_t *buf;
buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
atomic_add_64(&arc_loaned_bytes, size);
return (buf);
}
/*
* Return a loaned arc buffer to the arc.
*/
void
arc_return_buf(arc_buf_t *buf, void *tag)
{
arc_buf_hdr_t *hdr = buf->b_hdr;
ASSERT(buf->b_data != NULL);
(void) refcount_add(&hdr->b_refcnt, tag);
(void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
2009-07-03 02:44:48 +04:00
atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
}
/* Detach an arc_buf from a dbuf (tag) */
void
arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
{
arc_buf_hdr_t *hdr;
ASSERT(buf->b_data != NULL);
hdr = buf->b_hdr;
(void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
(void) refcount_remove(&hdr->b_refcnt, tag);
buf->b_efunc = NULL;
buf->b_private = NULL;
atomic_add_64(&arc_loaned_bytes, hdr->b_size);
}
2008-11-20 23:01:55 +03:00
static arc_buf_t *
arc_buf_clone(arc_buf_t *from)
{
arc_buf_t *buf;
arc_buf_hdr_t *hdr = from->b_hdr;
uint64_t size = hdr->b_size;
ASSERT(hdr->b_state != arc_anon);
2008-11-20 23:01:55 +03:00
buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
buf->b_hdr = hdr;
buf->b_data = NULL;
buf->b_efunc = NULL;
buf->b_private = NULL;
buf->b_next = hdr->b_buf;
hdr->b_buf = buf;
arc_get_data_buf(buf);
bcopy(from->b_data, buf->b_data, size);
/*
* This buffer already exists in the arc so create a duplicate
* copy for the caller. If the buffer is associated with user data
* then track the size and number of duplicates. These stats will be
* updated as duplicate buffers are created and destroyed.
*/
if (hdr->b_type == ARC_BUFC_DATA) {
ARCSTAT_BUMP(arcstat_duplicate_buffers);
ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
}
2008-11-20 23:01:55 +03:00
hdr->b_datacnt += 1;
return (buf);
}
void
arc_buf_add_ref(arc_buf_t *buf, void* tag)
{
arc_buf_hdr_t *hdr;
kmutex_t *hash_lock;
/*
* Check to see if this buffer is evicted. Callers
* must verify b_data != NULL to know if the add_ref
* was successful.
2008-11-20 23:01:55 +03:00
*/
mutex_enter(&buf->b_evict_lock);
if (buf->b_data == NULL) {
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
return;
}
hash_lock = HDR_LOCK(buf->b_hdr);
2008-11-20 23:01:55 +03:00
mutex_enter(hash_lock);
hdr = buf->b_hdr;
ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
add_reference(hdr, hash_lock, tag);
2009-02-18 23:51:31 +03:00
DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2008-11-20 23:01:55 +03:00
arc_access(hdr, hash_lock);
mutex_exit(hash_lock);
ARCSTAT_BUMP(arcstat_hits);
ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
data, metadata, hits);
}
static void
arc_buf_free_on_write(void *data, size_t size,
void (*free_func)(void *, size_t))
{
l2arc_data_free_t *df;
df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
df->l2df_data = data;
df->l2df_size = size;
df->l2df_func = free_func;
mutex_enter(&l2arc_free_on_write_mtx);
list_insert_head(l2arc_free_on_write, df);
mutex_exit(&l2arc_free_on_write_mtx);
}
2008-11-20 23:01:55 +03:00
/*
* Free the arc data buffer. If it is an l2arc write in progress,
* the buffer is placed on l2arc_free_on_write to be freed later.
*/
static void
arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *hdr = buf->b_hdr;
2008-11-20 23:01:55 +03:00
if (HDR_L2_WRITING(hdr)) {
arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_l2_free_on_write);
} else {
free_func(buf->b_data, hdr->b_size);
2008-11-20 23:01:55 +03:00
}
}
/*
* Free up buf->b_data and if 'remove' is set, then pull the
* arc_buf_t off of the the arc_buf_hdr_t's list and free it.
*/
static void
arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
{
l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
ASSERT(MUTEX_HELD(&l2arc_buflist_mtx));
if (l2hdr->b_tmp_cdata == NULL)
return;
ASSERT(HDR_L2_WRITING(hdr));
arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size,
zio_data_buf_free);
ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
l2hdr->b_tmp_cdata = NULL;
}
2008-11-20 23:01:55 +03:00
static void
arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove)
2008-11-20 23:01:55 +03:00
{
arc_buf_t **bufp;
/* free up data associated with the buf */
if (buf->b_data) {
arc_state_t *state = buf->b_hdr->b_state;
uint64_t size = buf->b_hdr->b_size;
arc_buf_contents_t type = buf->b_hdr->b_type;
arc_cksum_verify(buf);
arc_buf_unwatch(buf);
2008-11-20 23:01:55 +03:00
if (!recycle) {
if (type == ARC_BUFC_METADATA) {
arc_buf_data_free(buf, zio_buf_free);
arc_space_return(size, ARC_SPACE_META);
2008-11-20 23:01:55 +03:00
} else {
ASSERT(type == ARC_BUFC_DATA);
arc_buf_data_free(buf, zio_data_buf_free);
arc_space_return(size, ARC_SPACE_DATA);
2008-11-20 23:01:55 +03:00
}
}
if (list_link_active(&buf->b_hdr->b_arc_node)) {
uint64_t *cnt = &state->arcs_lsize[type];
ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
ASSERT(state != arc_anon);
ASSERT3U(*cnt, >=, size);
atomic_add_64(cnt, -size);
}
ASSERT3U(state->arcs_size, >=, size);
atomic_add_64(&state->arcs_size, -size);
buf->b_data = NULL;
/*
* If we're destroying a duplicate buffer make sure
* that the appropriate statistics are updated.
*/
if (buf->b_hdr->b_datacnt > 1 &&
buf->b_hdr->b_type == ARC_BUFC_DATA) {
ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
}
2008-11-20 23:01:55 +03:00
ASSERT(buf->b_hdr->b_datacnt > 0);
buf->b_hdr->b_datacnt -= 1;
}
/* only remove the buf if requested */
if (!remove)
2008-11-20 23:01:55 +03:00
return;
/* remove the buf from the hdr list */
for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
continue;
*bufp = buf->b_next;
buf->b_next = NULL;
2008-11-20 23:01:55 +03:00
ASSERT(buf->b_efunc == NULL);
/* clean up the buf */
buf->b_hdr = NULL;
kmem_cache_free(buf_cache, buf);
}
static void
arc_hdr_destroy(arc_buf_hdr_t *hdr)
{
l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
2008-11-20 23:01:55 +03:00
ASSERT(refcount_is_zero(&hdr->b_refcnt));
ASSERT3P(hdr->b_state, ==, arc_anon);
ASSERT(!HDR_IO_IN_PROGRESS(hdr));
if (l2hdr != NULL) {
boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
/*
* To prevent arc_free() and l2arc_evict() from
* attempting to free the same buffer at the same time,
* a FREE_IN_PROGRESS flag is given to arc_free() to
* give it priority. l2arc_evict() can't destroy this
* header while we are waiting on l2arc_buflist_mtx.
*
* The hdr may be removed from l2ad_buflist before we
* grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
*/
if (!buflist_held) {
2008-11-20 23:01:55 +03:00
mutex_enter(&l2arc_buflist_mtx);
l2hdr = hdr->b_l2hdr;
2008-11-20 23:01:55 +03:00
}
if (l2hdr != NULL) {
list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
arc_buf_l2_cdata_free(hdr);
ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
vdev_space_update(l2hdr->b_dev->l2ad_vdev,
-l2hdr->b_asize, 0, 0);
kmem_cache_free(l2arc_hdr_cache, l2hdr);
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
if (hdr->b_state == arc_l2c_only)
l2arc_hdr_stat_remove();
hdr->b_l2hdr = NULL;
}
if (!buflist_held)
mutex_exit(&l2arc_buflist_mtx);
2008-11-20 23:01:55 +03:00
}
if (!BUF_EMPTY(hdr)) {
ASSERT(!HDR_IN_HASH_TABLE(hdr));
buf_discard_identity(hdr);
2008-11-20 23:01:55 +03:00
}
while (hdr->b_buf) {
arc_buf_t *buf = hdr->b_buf;
if (buf->b_efunc) {
mutex_enter(&arc_eviction_mtx);
mutex_enter(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
ASSERT(buf->b_hdr != NULL);
arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
hdr->b_buf = buf->b_next;
buf->b_hdr = &arc_eviction_hdr;
buf->b_next = arc_eviction_list;
arc_eviction_list = buf;
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
mutex_exit(&arc_eviction_mtx);
} else {
arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
}
}
if (hdr->b_freeze_cksum != NULL) {
kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
hdr->b_freeze_cksum = NULL;
}
ASSERT(!list_link_active(&hdr->b_arc_node));
ASSERT3P(hdr->b_hash_next, ==, NULL);
ASSERT3P(hdr->b_acb, ==, NULL);
kmem_cache_free(hdr_cache, hdr);
}
void
arc_buf_free(arc_buf_t *buf, void *tag)
{
arc_buf_hdr_t *hdr = buf->b_hdr;
int hashed = hdr->b_state != arc_anon;
ASSERT(buf->b_efunc == NULL);
ASSERT(buf->b_data != NULL);
if (hashed) {
kmutex_t *hash_lock = HDR_LOCK(hdr);
mutex_enter(hash_lock);
hdr = buf->b_hdr;
ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2008-11-20 23:01:55 +03:00
(void) remove_reference(hdr, hash_lock, tag);
if (hdr->b_datacnt > 1) {
2008-11-20 23:01:55 +03:00
arc_buf_destroy(buf, FALSE, TRUE);
} else {
ASSERT(buf == hdr->b_buf);
ASSERT(buf->b_efunc == NULL);
2008-11-20 23:01:55 +03:00
hdr->b_flags |= ARC_BUF_AVAILABLE;
}
2008-11-20 23:01:55 +03:00
mutex_exit(hash_lock);
} else if (HDR_IO_IN_PROGRESS(hdr)) {
int destroy_hdr;
/*
* We are in the middle of an async write. Don't destroy
* this buffer unless the write completes before we finish
* decrementing the reference count.
*/
mutex_enter(&arc_eviction_mtx);
(void) remove_reference(hdr, NULL, tag);
ASSERT(refcount_is_zero(&hdr->b_refcnt));
destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
mutex_exit(&arc_eviction_mtx);
if (destroy_hdr)
arc_hdr_destroy(hdr);
} else {
if (remove_reference(hdr, NULL, tag) > 0)
2008-11-20 23:01:55 +03:00
arc_buf_destroy(buf, FALSE, TRUE);
else
2008-11-20 23:01:55 +03:00
arc_hdr_destroy(hdr);
}
}
boolean_t
2008-11-20 23:01:55 +03:00
arc_buf_remove_ref(arc_buf_t *buf, void* tag)
{
arc_buf_hdr_t *hdr = buf->b_hdr;
kmutex_t *hash_lock = NULL;
boolean_t no_callback = (buf->b_efunc == NULL);
2008-11-20 23:01:55 +03:00
if (hdr->b_state == arc_anon) {
ASSERT(hdr->b_datacnt == 1);
2008-11-20 23:01:55 +03:00
arc_buf_free(buf, tag);
return (no_callback);
}
hash_lock = HDR_LOCK(hdr);
2008-11-20 23:01:55 +03:00
mutex_enter(hash_lock);
hdr = buf->b_hdr;
ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2008-11-20 23:01:55 +03:00
ASSERT(hdr->b_state != arc_anon);
ASSERT(buf->b_data != NULL);
(void) remove_reference(hdr, hash_lock, tag);
if (hdr->b_datacnt > 1) {
if (no_callback)
arc_buf_destroy(buf, FALSE, TRUE);
} else if (no_callback) {
ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
ASSERT(buf->b_efunc == NULL);
2008-11-20 23:01:55 +03:00
hdr->b_flags |= ARC_BUF_AVAILABLE;
}
ASSERT(no_callback || hdr->b_datacnt > 1 ||
refcount_is_zero(&hdr->b_refcnt));
mutex_exit(hash_lock);
return (no_callback);
}
uint64_t
2008-11-20 23:01:55 +03:00
arc_buf_size(arc_buf_t *buf)
{
return (buf->b_hdr->b_size);
}
/*
* Called from the DMU to determine if the current buffer should be
* evicted. In order to ensure proper locking, the eviction must be initiated
* from the DMU. Return true if the buffer is associated with user data and
* duplicate buffers still exist.
*/
boolean_t
arc_buf_eviction_needed(arc_buf_t *buf)
{
arc_buf_hdr_t *hdr;
boolean_t evict_needed = B_FALSE;
if (zfs_disable_dup_eviction)
return (B_FALSE);
mutex_enter(&buf->b_evict_lock);
hdr = buf->b_hdr;
if (hdr == NULL) {
/*
* We are in arc_do_user_evicts(); let that function
* perform the eviction.
*/
ASSERT(buf->b_data == NULL);
mutex_exit(&buf->b_evict_lock);
return (B_FALSE);
} else if (buf->b_data == NULL) {
/*
* We have already been added to the arc eviction list;
* recommend eviction.
*/
ASSERT3P(hdr, ==, &arc_eviction_hdr);
mutex_exit(&buf->b_evict_lock);
return (B_TRUE);
}
if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
evict_needed = B_TRUE;
mutex_exit(&buf->b_evict_lock);
return (evict_needed);
}
2008-11-20 23:01:55 +03:00
/*
* Evict buffers from list until we've removed the specified number of
* bytes. Move the removed buffers to the appropriate evict state.
* If the recycle flag is set, then attempt to "recycle" a buffer:
* - look for a buffer to evict that is `bytes' long.
* - return the data block from this buffer rather than freeing it.
* This flag is used by callers that are trying to make space for a
* new buffer in a full arc cache.
*
* This function makes a "best effort". It skips over any buffers
* it can't get a hash_lock on, and so may not catch all candidates.
* It may also return without evicting as much space as requested.
*/
static void *
2009-02-18 23:51:31 +03:00
arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
2008-11-20 23:01:55 +03:00
arc_buf_contents_t type)
{
arc_state_t *evicted_state;
uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
arc_buf_hdr_t *ab, *ab_prev = NULL;
list_t *list = &state->arcs_list[type];
kmutex_t *hash_lock;
boolean_t have_lock;
void *stolen = NULL;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
arc_buf_hdr_t marker = {{{ 0 }}};
int count = 0;
2008-11-20 23:01:55 +03:00
ASSERT(state == arc_mru || state == arc_mfu);
evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
top:
2008-11-20 23:01:55 +03:00
mutex_enter(&state->arcs_mtx);
mutex_enter(&evicted_state->arcs_mtx);
for (ab = list_tail(list); ab; ab = ab_prev) {
ab_prev = list_prev(list, ab);
/* prefetch buffers have a minimum lifespan */
if (HDR_IO_IN_PROGRESS(ab) ||
(spa && ab->b_spa != spa) ||
(ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
ddi_get_lbolt() - ab->b_arc_access <
zfs_arc_min_prefetch_lifespan)) {
2008-11-20 23:01:55 +03:00
skipped++;
continue;
}
/* "lookahead" for better eviction candidate */
if (recycle && ab->b_size != bytes &&
ab_prev && ab_prev->b_size == bytes)
continue;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
/* ignore markers */
if (ab->b_spa == 0)
continue;
/*
* It may take a long time to evict all the bufs requested.
* To avoid blocking all arc activity, periodically drop
* the arcs_mtx and give other threads a chance to run
* before reacquiring the lock.
*
* If we are looking for a buffer to recycle, we are in
* the hot code path, so don't sleep.
*/
if (!recycle && count++ > arc_evict_iterations) {
list_insert_after(list, ab, &marker);
mutex_exit(&evicted_state->arcs_mtx);
mutex_exit(&state->arcs_mtx);
kpreempt(KPREEMPT_SYNC);
mutex_enter(&state->arcs_mtx);
mutex_enter(&evicted_state->arcs_mtx);
ab_prev = list_prev(list, &marker);
list_remove(list, &marker);
count = 0;
continue;
}
2008-11-20 23:01:55 +03:00
hash_lock = HDR_LOCK(ab);
have_lock = MUTEX_HELD(hash_lock);
if (have_lock || mutex_tryenter(hash_lock)) {
ASSERT0(refcount_count(&ab->b_refcnt));
2008-11-20 23:01:55 +03:00
ASSERT(ab->b_datacnt > 0);
while (ab->b_buf) {
arc_buf_t *buf = ab->b_buf;
if (!mutex_tryenter(&buf->b_evict_lock)) {
missed += 1;
break;
}
2008-11-20 23:01:55 +03:00
if (buf->b_data) {
bytes_evicted += ab->b_size;
if (recycle && ab->b_type == type &&
ab->b_size == bytes &&
!HDR_L2_WRITING(ab)) {
stolen = buf->b_data;
recycle = FALSE;
}
}
if (buf->b_efunc) {
mutex_enter(&arc_eviction_mtx);
arc_buf_destroy(buf,
buf->b_data == stolen, FALSE);
ab->b_buf = buf->b_next;
buf->b_hdr = &arc_eviction_hdr;
buf->b_next = arc_eviction_list;
arc_eviction_list = buf;
mutex_exit(&arc_eviction_mtx);
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
} else {
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
arc_buf_destroy(buf,
buf->b_data == stolen, TRUE);
}
}
if (ab->b_l2hdr) {
ARCSTAT_INCR(arcstat_evict_l2_cached,
ab->b_size);
} else {
if (l2arc_write_eligible(ab->b_spa, ab)) {
ARCSTAT_INCR(arcstat_evict_l2_eligible,
ab->b_size);
} else {
ARCSTAT_INCR(
arcstat_evict_l2_ineligible,
ab->b_size);
}
}
if (ab->b_datacnt == 0) {
arc_change_state(evicted_state, ab, hash_lock);
ASSERT(HDR_IN_HASH_TABLE(ab));
ab->b_flags |= ARC_IN_HASH_TABLE;
ab->b_flags &= ~ARC_BUF_AVAILABLE;
DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
}
2008-11-20 23:01:55 +03:00
if (!have_lock)
mutex_exit(hash_lock);
if (bytes >= 0 && bytes_evicted >= bytes)
break;
} else {
missed += 1;
}
}
mutex_exit(&evicted_state->arcs_mtx);
mutex_exit(&state->arcs_mtx);
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
if (list == &state->arcs_list[ARC_BUFC_DATA] &&
(bytes < 0 || bytes_evicted < bytes)) {
/* Prevent second pass from recycling metadata into data */
recycle = FALSE;
type = ARC_BUFC_METADATA;
list = &state->arcs_list[type];
goto top;
}
2008-11-20 23:01:55 +03:00
if (bytes_evicted < bytes)
dprintf("only evicted %lld bytes from %x\n",
(longlong_t)bytes_evicted, state->arcs_state);
2008-11-20 23:01:55 +03:00
if (skipped)
ARCSTAT_INCR(arcstat_evict_skip, skipped);
if (missed)
ARCSTAT_INCR(arcstat_mutex_miss, missed);
/*
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
* Note: we have just evicted some data into the ghost state,
* potentially putting the ghost size over the desired size. Rather
* that evicting from the ghost list in this hot code path, leave
* this chore to the arc_reclaim_thread().
2008-11-20 23:01:55 +03:00
*/
return (stolen);
}
/*
* Remove buffers from list until we've removed the specified number of
* bytes. Destroy the buffers that are removed.
*/
static void
arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes,
arc_buf_contents_t type)
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *ab, *ab_prev;
arc_buf_hdr_t marker;
list_t *list = &state->arcs_list[type];
2008-11-20 23:01:55 +03:00
kmutex_t *hash_lock;
uint64_t bytes_deleted = 0;
uint64_t bufs_skipped = 0;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
int count = 0;
2008-11-20 23:01:55 +03:00
ASSERT(GHOST_STATE(state));
bzero(&marker, sizeof (marker));
2008-11-20 23:01:55 +03:00
top:
mutex_enter(&state->arcs_mtx);
for (ab = list_tail(list); ab; ab = ab_prev) {
ab_prev = list_prev(list, ab);
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
if (ab->b_type > ARC_BUFC_NUMTYPES)
panic("invalid ab=%p", (void *)ab);
2008-11-20 23:01:55 +03:00
if (spa && ab->b_spa != spa)
continue;
/* ignore markers */
if (ab->b_spa == 0)
continue;
2008-11-20 23:01:55 +03:00
hash_lock = HDR_LOCK(ab);
/* caller may be trying to modify this buffer, skip it */
if (MUTEX_HELD(hash_lock))
continue;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
/*
* It may take a long time to evict all the bufs requested.
* To avoid blocking all arc activity, periodically drop
* the arcs_mtx and give other threads a chance to run
* before reacquiring the lock.
*/
if (count++ > arc_evict_iterations) {
list_insert_after(list, ab, &marker);
mutex_exit(&state->arcs_mtx);
kpreempt(KPREEMPT_SYNC);
mutex_enter(&state->arcs_mtx);
ab_prev = list_prev(list, &marker);
list_remove(list, &marker);
count = 0;
continue;
}
2008-11-20 23:01:55 +03:00
if (mutex_tryenter(hash_lock)) {
ASSERT(!HDR_IO_IN_PROGRESS(ab));
ASSERT(ab->b_buf == NULL);
ARCSTAT_BUMP(arcstat_deleted);
bytes_deleted += ab->b_size;
if (ab->b_l2hdr != NULL) {
/*
* This buffer is cached on the 2nd Level ARC;
* don't destroy the header.
*/
arc_change_state(arc_l2c_only, ab, hash_lock);
mutex_exit(hash_lock);
} else {
arc_change_state(arc_anon, ab, hash_lock);
mutex_exit(hash_lock);
arc_hdr_destroy(ab);
}
DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
if (bytes >= 0 && bytes_deleted >= bytes)
break;
} else if (bytes < 0) {
/*
* Insert a list marker and then wait for the
* hash lock to become available. Once its
* available, restart from where we left off.
*/
list_insert_after(list, ab, &marker);
mutex_exit(&state->arcs_mtx);
mutex_enter(hash_lock);
mutex_exit(hash_lock);
mutex_enter(&state->arcs_mtx);
ab_prev = list_prev(list, &marker);
list_remove(list, &marker);
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
} else {
2008-11-20 23:01:55 +03:00
bufs_skipped += 1;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
}
2008-11-20 23:01:55 +03:00
}
mutex_exit(&state->arcs_mtx);
if (list == &state->arcs_list[ARC_BUFC_DATA] &&
(bytes < 0 || bytes_deleted < bytes)) {
list = &state->arcs_list[ARC_BUFC_METADATA];
goto top;
}
if (bufs_skipped) {
ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
ASSERT(bytes >= 0);
}
if (bytes_deleted < bytes)
dprintf("only deleted %lld bytes from %p\n",
2008-11-20 23:01:55 +03:00
(longlong_t)bytes_deleted, state);
}
static void
arc_adjust(void)
{
2009-02-18 23:51:31 +03:00
int64_t adjustment, delta;
/*
* Adjust MRU size
*/
2008-11-20 23:01:55 +03:00
adjustment = MIN((int64_t)(arc_size - arc_c),
(int64_t)(arc_anon->arcs_size + arc_mru->arcs_size - arc_p));
2008-11-20 23:01:55 +03:00
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
if (adjustment > 0 && arc_mru->arcs_size > 0) {
delta = MIN(arc_mru->arcs_size, adjustment);
(void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2008-11-20 23:01:55 +03:00
}
2009-02-18 23:51:31 +03:00
/*
* Adjust MFU size
*/
2008-11-20 23:01:55 +03:00
2009-02-18 23:51:31 +03:00
adjustment = arc_size - arc_c;
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
if (adjustment > 0 && arc_mfu->arcs_size > 0) {
delta = MIN(arc_mfu->arcs_size, adjustment);
(void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2009-02-18 23:51:31 +03:00
}
2008-11-20 23:01:55 +03:00
2009-02-18 23:51:31 +03:00
/*
* Adjust ghost lists
*/
2008-11-20 23:01:55 +03:00
2009-02-18 23:51:31 +03:00
adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
delta = MIN(arc_mru_ghost->arcs_size, adjustment);
arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_DATA);
2009-02-18 23:51:31 +03:00
}
2008-11-20 23:01:55 +03:00
2009-02-18 23:51:31 +03:00
adjustment =
arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2008-11-20 23:01:55 +03:00
2009-02-18 23:51:31 +03:00
if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_DATA);
2008-11-20 23:01:55 +03:00
}
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
/*
* Request that arc user drop references so that N bytes can be released
* from the cache. This provides a mechanism to ensure the arc can honor
* the arc_meta_limit and reclaim buffers which are pinned in the cache
* by higher layers. (i.e. the zpl)
*/
static void
arc_do_user_prune(int64_t adjustment)
{
arc_prune_func_t *func;
void *private;
arc_prune_t *cp, *np;
mutex_enter(&arc_prune_mtx);
cp = list_head(&arc_prune_list);
while (cp != NULL) {
func = cp->p_pfunc;
private = cp->p_private;
np = list_next(&arc_prune_list, cp);
refcount_add(&cp->p_refcnt, func);
mutex_exit(&arc_prune_mtx);
if (func != NULL)
func(adjustment, private);
mutex_enter(&arc_prune_mtx);
/* User removed prune callback concurrently with execution */
if (refcount_remove(&cp->p_refcnt, func) == 0) {
ASSERT(!list_link_active(&cp->p_node));
refcount_destroy(&cp->p_refcnt);
kmem_free(cp, sizeof (*cp));
}
cp = np;
}
ARCSTAT_BUMP(arcstat_prune);
mutex_exit(&arc_prune_mtx);
}
2008-11-20 23:01:55 +03:00
static void
arc_do_user_evicts(void)
{
mutex_enter(&arc_eviction_mtx);
while (arc_eviction_list != NULL) {
arc_buf_t *buf = arc_eviction_list;
arc_eviction_list = buf->b_next;
mutex_enter(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
buf->b_hdr = NULL;
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
mutex_exit(&arc_eviction_mtx);
if (buf->b_efunc != NULL)
VERIFY0(buf->b_efunc(buf->b_private));
2008-11-20 23:01:55 +03:00
buf->b_efunc = NULL;
buf->b_private = NULL;
kmem_cache_free(buf_cache, buf);
mutex_enter(&arc_eviction_mtx);
}
mutex_exit(&arc_eviction_mtx);
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
/*
* Evict only meta data objects from the cache leaving the data objects.
* This is only used to enforce the tunable arc_meta_limit, if we are
* unable to evict enough buffers notify the user via the prune callback.
*/
static void
arc_adjust_meta(void)
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
{
int64_t adjustmnt, delta;
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
/*
* This slightly differs than the way we evict from the mru in
* arc_adjust because we don't have a "target" value (i.e. no
* "meta" arc_p). As a result, I think we can completely
* cannibalize the metadata in the MRU before we evict the
* metadata from the MFU. I think we probably need to implement a
* "metadata arc_p" value to do this properly.
*/
adjustmnt = arc_meta_used - arc_meta_limit;
if (adjustmnt > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustmnt);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
adjustmnt -= delta;
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
}
/*
* We can't afford to recalculate adjustmnt here. If we do,
* new metadata buffers can sneak into the MRU or ANON lists,
* thus penalize the MFU metadata. Although the fudge factor is
* small, it has been empirically shown to be significant for
* certain workloads (e.g. creating many empty directories). As
* such, we use the original calculation for adjustmnt, and
* simply decrement the amount of data evicted from the MRU.
*/
if (adjustmnt > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustmnt);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
}
adjustmnt = arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] - arc_meta_limit;
if (adjustmnt > 0 && arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] > 0) {
delta = MIN(adjustmnt,
arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA]);
arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_METADATA);
}
adjustmnt = arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA] - arc_meta_limit;
if (adjustmnt > 0 && arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA] > 0) {
delta = MIN(adjustmnt,
arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA]);
arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_METADATA);
}
if (arc_meta_used > arc_meta_limit)
arc_do_user_prune(zfs_arc_meta_prune);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
}
2008-11-20 23:01:55 +03:00
/*
* Flush all *evictable* data from the cache for the given spa.
* NOTE: this will not touch "active" (i.e. referenced) data.
*/
void
arc_flush(spa_t *spa)
{
2009-02-18 23:51:31 +03:00
uint64_t guid = 0;
if (spa)
guid = spa_load_guid(spa);
2009-02-18 23:51:31 +03:00
2008-11-20 23:01:55 +03:00
while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2009-02-18 23:51:31 +03:00
(void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2008-11-20 23:01:55 +03:00
if (spa)
break;
}
while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2009-02-18 23:51:31 +03:00
(void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2008-11-20 23:01:55 +03:00
if (spa)
break;
}
while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2009-02-18 23:51:31 +03:00
(void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2008-11-20 23:01:55 +03:00
if (spa)
break;
}
while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2009-02-18 23:51:31 +03:00
(void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2008-11-20 23:01:55 +03:00
if (spa)
break;
}
arc_evict_ghost(arc_mru_ghost, guid, -1, ARC_BUFC_DATA);
arc_evict_ghost(arc_mfu_ghost, guid, -1, ARC_BUFC_DATA);
2008-11-20 23:01:55 +03:00
mutex_enter(&arc_reclaim_thr_lock);
arc_do_user_evicts();
mutex_exit(&arc_reclaim_thr_lock);
ASSERT(spa || arc_eviction_list == NULL);
}
void
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
arc_shrink(uint64_t bytes)
2008-11-20 23:01:55 +03:00
{
if (arc_c > arc_c_min) {
uint64_t to_free;
to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift;
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
2008-11-20 23:01:55 +03:00
if (arc_c > arc_c_min + to_free)
atomic_add_64(&arc_c, -to_free);
else
arc_c = arc_c_min;
to_free = bytes ? bytes : arc_p >> zfs_arc_shrink_shift;
if (arc_p > to_free)
atomic_add_64(&arc_p, -to_free);
else
arc_p = 0;
2008-11-20 23:01:55 +03:00
if (arc_c > arc_size)
arc_c = MAX(arc_size, arc_c_min);
if (arc_p > arc_c)
arc_p = (arc_c >> 1);
ASSERT(arc_c >= arc_c_min);
ASSERT((int64_t)arc_p >= 0);
}
if (arc_size > arc_c)
arc_adjust();
}
static void
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2008-11-20 23:01:55 +03:00
{
size_t i;
kmem_cache_t *prev_cache = NULL;
kmem_cache_t *prev_data_cache = NULL;
extern kmem_cache_t *zio_buf_cache[];
extern kmem_cache_t *zio_data_buf_cache[];
/*
* An aggressive reclamation will shrink the cache size as well as
* reap free buffers from the arc kmem caches.
*/
if (strat == ARC_RECLAIM_AGGR)
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
arc_shrink(bytes);
2008-11-20 23:01:55 +03:00
for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
if (zio_buf_cache[i] != prev_cache) {
prev_cache = zio_buf_cache[i];
kmem_cache_reap_now(zio_buf_cache[i]);
}
if (zio_data_buf_cache[i] != prev_data_cache) {
prev_data_cache = zio_data_buf_cache[i];
kmem_cache_reap_now(zio_data_buf_cache[i]);
}
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
2008-11-20 23:01:55 +03:00
kmem_cache_reap_now(buf_cache);
kmem_cache_reap_now(hdr_cache);
}
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
/*
* Unlike other ZFS implementations this thread is only responsible for
* adapting the target ARC size on Linux. The responsibility for memory
* reclamation has been entirely delegated to the arc_shrinker_func()
* which is registered with the VM. To reflect this change in behavior
* the arc_reclaim thread has been renamed to arc_adapt.
*/
2008-11-20 23:01:55 +03:00
static void
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
arc_adapt_thread(void)
2008-11-20 23:01:55 +03:00
{
callb_cpr_t cpr;
CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
mutex_enter(&arc_reclaim_thr_lock);
while (arc_thread_exit == 0) {
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
#ifndef _KERNEL
arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
if (spa_get_random(100) == 0) {
2008-11-20 23:01:55 +03:00
if (arc_no_grow) {
if (last_reclaim == ARC_RECLAIM_CONS) {
last_reclaim = ARC_RECLAIM_AGGR;
} else {
last_reclaim = ARC_RECLAIM_CONS;
}
} else {
arc_no_grow = TRUE;
last_reclaim = ARC_RECLAIM_AGGR;
membar_producer();
}
/* reset the growth delay for every reclaim */
arc_grow_time = ddi_get_lbolt() +
(zfs_arc_grow_retry * hz);
2008-11-20 23:01:55 +03:00
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
arc_kmem_reap_now(last_reclaim, 0);
arc_warm = B_TRUE;
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
}
#endif /* !_KERNEL */
2008-11-20 23:01:55 +03:00
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
/* No recent memory pressure allow the ARC to grow. */
if (arc_no_grow &&
ddi_time_after_eq(ddi_get_lbolt(), arc_grow_time))
2008-11-20 23:01:55 +03:00
arc_no_grow = FALSE;
arc_adjust_meta();
arc_adjust();
2008-11-20 23:01:55 +03:00
if (arc_eviction_list != NULL)
arc_do_user_evicts();
/* block until needed, or one second, whichever is shorter */
CALLB_CPR_SAFE_BEGIN(&cpr);
(void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
&arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2008-11-20 23:01:55 +03:00
CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
/* Allow the module options to be changed */
if (zfs_arc_max > 64 << 20 &&
zfs_arc_max < physmem * PAGESIZE &&
zfs_arc_max != arc_c_max)
arc_c_max = zfs_arc_max;
if (zfs_arc_min > 0 &&
zfs_arc_min < arc_c_max &&
zfs_arc_min != arc_c_min)
arc_c_min = zfs_arc_min;
if (zfs_arc_meta_limit > 0 &&
zfs_arc_meta_limit <= arc_c_max &&
zfs_arc_meta_limit != arc_meta_limit)
arc_meta_limit = zfs_arc_meta_limit;
2008-11-20 23:01:55 +03:00
}
arc_thread_exit = 0;
cv_broadcast(&arc_reclaim_thr_cv);
CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
thread_exit();
}
#ifdef _KERNEL
/*
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
* Determine the amount of memory eligible for eviction contained in the
* ARC. All clean data reported by the ghost lists can always be safely
* evicted. Due to arc_c_min, the same does not hold for all clean data
* contained by the regular mru and mfu lists.
*
* In the case of the regular mru and mfu lists, we need to report as
* much clean data as possible, such that evicting that same reported
* data will not bring arc_size below arc_c_min. Thus, in certain
* circumstances, the total amount of clean data in the mru and mfu
* lists might not actually be evictable.
*
* The following two distinct cases are accounted for:
*
* 1. The sum of the amount of dirty data contained by both the mru and
* mfu lists, plus the ARC's other accounting (e.g. the anon list),
* is greater than or equal to arc_c_min.
* (i.e. amount of dirty data >= arc_c_min)
*
* This is the easy case; all clean data contained by the mru and mfu
* lists is evictable. Evicting all clean data can only drop arc_size
* to the amount of dirty data, which is greater than arc_c_min.
*
* 2. The sum of the amount of dirty data contained by both the mru and
* mfu lists, plus the ARC's other accounting (e.g. the anon list),
* is less than arc_c_min.
* (i.e. arc_c_min > amount of dirty data)
*
* 2.1. arc_size is greater than or equal arc_c_min.
* (i.e. arc_size >= arc_c_min > amount of dirty data)
*
* In this case, not all clean data from the regular mru and mfu
* lists is actually evictable; we must leave enough clean data
* to keep arc_size above arc_c_min. Thus, the maximum amount of
* evictable data from the two lists combined, is exactly the
* difference between arc_size and arc_c_min.
*
* 2.2. arc_size is less than arc_c_min
* (i.e. arc_c_min > arc_size > amount of dirty data)
*
* In this case, none of the data contained in the mru and mfu
* lists is evictable, even if it's clean. Since arc_size is
* already below arc_c_min, evicting any more would only
* increase this negative difference.
*/
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
static uint64_t
arc_evictable_memory(void) {
uint64_t arc_clean =
arc_mru->arcs_lsize[ARC_BUFC_DATA] +
arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
uint64_t ghost_clean =
arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
if (arc_dirty >= arc_c_min)
return (ghost_clean + arc_clean);
return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
}
/*
* If sc->nr_to_scan is zero, the caller is requesting a query of the
* number of objects which can potentially be freed. If it is nonzero,
* the request is to free that many objects.
*
* Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
* in struct shrinker and also require the shrinker to return the number
* of objects freed.
*
* Older kernels require the shrinker to return the number of freeable
* objects following the freeing of nr_to_free.
*/
static spl_shrinker_t
__arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
{
int64_t pages;
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
/* The arc is considered warm once reclaim has occurred */
if (unlikely(arc_warm == B_FALSE))
arc_warm = B_TRUE;
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
/* Return the potential number of reclaimable pages */
pages = btop((int64_t)arc_evictable_memory());
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
if (sc->nr_to_scan == 0)
return (pages);
/* Not allowed to perform filesystem reclaim */
if (!(sc->gfp_mask & __GFP_FS))
return (SHRINK_STOP);
/* Reclaim in progress */
if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
return (SHRINK_STOP);
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
/*
* Evict the requested number of pages by shrinking arc_c the
* requested amount. If there is nothing left to evict just
* reap whatever we can from the various arc slabs.
*/
if (pages > 0) {
arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
#ifdef HAVE_SPLIT_SHRINKER_CALLBACK
pages = MAX(pages - btop(arc_evictable_memory()), 0);
#else
pages = btop(arc_evictable_memory());
#endif
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
} else {
arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
pages = SHRINK_STOP;
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
}
/*
* When direct reclaim is observed it usually indicates a rapid
* increase in memory pressure. This occurs because the kswapd
* threads were unable to asynchronously keep enough free memory
* available. In this case set arc_no_grow to briefly pause arc
* growth to avoid compounding the memory pressure.
*/
if (current_is_kswapd()) {
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
ARCSTAT_BUMP(arcstat_memory_indirect_count);
} else {
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
arc_no_grow = B_TRUE;
arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz);
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
ARCSTAT_BUMP(arcstat_memory_direct_count);
}
mutex_exit(&arc_reclaim_thr_lock);
return (pages);
}
SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
#endif /* _KERNEL */
2008-11-20 23:01:55 +03:00
/*
* Adapt arc info given the number of bytes we are trying to add and
* the state that we are comming from. This function is only called
* when we are adding new content to the cache.
*/
static void
arc_adapt(int bytes, arc_state_t *state)
{
int mult;
if (state == arc_l2c_only)
return;
ASSERT(bytes > 0);
/*
* Adapt the target size of the MRU list:
* - if we just hit in the MRU ghost list, then increase
* the target size of the MRU list.
* - if we just hit in the MFU ghost list, then increase
* the target size of the MFU list by decreasing the
* target size of the MRU list.
*/
if (state == arc_mru_ghost) {
mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
if (!zfs_arc_p_dampener_disable)
mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2008-11-20 23:01:55 +03:00
arc_p = MIN(arc_c, arc_p + bytes * mult);
2008-11-20 23:01:55 +03:00
} else if (state == arc_mfu_ghost) {
2009-02-18 23:51:31 +03:00
uint64_t delta;
2008-11-20 23:01:55 +03:00
mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
if (!zfs_arc_p_dampener_disable)
mult = MIN(mult, 10);
2008-11-20 23:01:55 +03:00
2009-02-18 23:51:31 +03:00
delta = MIN(bytes * mult, arc_p);
arc_p = MAX(0, arc_p - delta);
2008-11-20 23:01:55 +03:00
}
ASSERT((int64_t)arc_p >= 0);
if (arc_no_grow)
return;
if (arc_c >= arc_c_max)
return;
/*
* If we're within (2 * maxblocksize) bytes of the target
* cache size, increment the target cache size
*/
if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
atomic_add_64(&arc_c, (int64_t)bytes);
if (arc_c > arc_c_max)
arc_c = arc_c_max;
else if (state == arc_anon)
atomic_add_64(&arc_p, (int64_t)bytes);
if (arc_p > arc_c)
arc_p = arc_c;
}
ASSERT((int64_t)arc_p >= 0);
}
/*
* Check if the cache has reached its limits and eviction is required
* prior to insert.
*/
static int
arc_evict_needed(arc_buf_contents_t type)
{
if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
return (1);
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
if (arc_no_grow)
2008-11-20 23:01:55 +03:00
return (1);
return (arc_size > arc_c);
}
/*
* The buffer, supplied as the first argument, needs a data block.
* So, if we are at cache max, determine which cache should be victimized.
* We have the following cases:
*
* 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
* In this situation if we're out of space, but the resident size of the MFU is
* under the limit, victimize the MFU cache to satisfy this insertion request.
*
* 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
* Here, we've used up all of the available space for the MRU, so we need to
* evict from our own cache instead. Evict from the set of resident MRU
* entries.
*
* 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
* c minus p represents the MFU space in the cache, since p is the size of the
* cache that is dedicated to the MRU. In this situation there's still space on
* the MFU side, so the MRU side needs to be victimized.
*
* 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
* MFU's resident set is consuming more space than it has been allotted. In
* this situation, we must victimize our own cache, the MFU, for this insertion.
*/
static void
arc_get_data_buf(arc_buf_t *buf)
{
arc_state_t *state = buf->b_hdr->b_state;
uint64_t size = buf->b_hdr->b_size;
arc_buf_contents_t type = buf->b_hdr->b_type;
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
arc_buf_contents_t evict = ARC_BUFC_DATA;
boolean_t recycle = TRUE;
2008-11-20 23:01:55 +03:00
arc_adapt(size, state);
/*
* We have not yet reached cache maximum size,
* just allocate a new buffer.
*/
if (!arc_evict_needed(type)) {
if (type == ARC_BUFC_METADATA) {
buf->b_data = zio_buf_alloc(size);
arc_space_consume(size, ARC_SPACE_META);
2008-11-20 23:01:55 +03:00
} else {
ASSERT(type == ARC_BUFC_DATA);
buf->b_data = zio_data_buf_alloc(size);
arc_space_consume(size, ARC_SPACE_DATA);
2008-11-20 23:01:55 +03:00
}
goto out;
}
/*
* If we are prefetching from the mfu ghost list, this buffer
* will end up on the mru list; so steal space from there.
*/
if (state == arc_mfu_ghost)
state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
else if (state == arc_mru_ghost)
state = arc_mru;
if (state == arc_mru || state == arc_anon) {
uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2009-02-18 23:51:31 +03:00
state = (arc_mfu->arcs_lsize[type] >= size &&
2008-11-20 23:01:55 +03:00
arc_p > mru_used) ? arc_mfu : arc_mru;
} else {
/* MFU cases */
uint64_t mfu_space = arc_c - arc_p;
2009-02-18 23:51:31 +03:00
state = (arc_mru->arcs_lsize[type] >= size &&
2008-11-20 23:01:55 +03:00
mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
/*
* Evict data buffers prior to metadata buffers, unless we're
* over the metadata limit and adding a metadata buffer.
*/
if (type == ARC_BUFC_METADATA) {
if (arc_meta_used >= arc_meta_limit)
evict = ARC_BUFC_METADATA;
else
/*
* In this case, we're evicting data while
* adding metadata. Thus, to prevent recycling a
* data buffer into a metadata buffer, recycling
* is disabled in the following arc_evict call.
*/
recycle = FALSE;
}
if ((buf->b_data = arc_evict(state, 0, size, recycle, evict)) == NULL) {
2008-11-20 23:01:55 +03:00
if (type == ARC_BUFC_METADATA) {
buf->b_data = zio_buf_alloc(size);
arc_space_consume(size, ARC_SPACE_META);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
/*
* If we are unable to recycle an existing meta buffer
* signal the reclaim thread. It will notify users
* via the prune callback to drop references. The
* prune callback in run in the context of the reclaim
* thread to avoid deadlocking on the hash_lock.
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
* Of course, only do this when recycle is true.
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
*/
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
if (recycle)
cv_signal(&arc_reclaim_thr_cv);
2008-11-20 23:01:55 +03:00
} else {
ASSERT(type == ARC_BUFC_DATA);
buf->b_data = zio_data_buf_alloc(size);
arc_space_consume(size, ARC_SPACE_DATA);
2008-11-20 23:01:55 +03:00
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
Prioritize "metadata" in arc_get_data_buf When the arc is at it's size limit and a new buffer is added, data will be evicted (or recycled) from the arc to make room for this new buffer. As far as I can tell, this is to try and keep the arc from over stepping it's bounds (i.e. keep it below the size limitation placed on it). This makes sense conceptually, but there appears to be a subtle flaw in its current implementation, resulting in metadata buffers being throttled. When it evicts from the arc's lists, it also passes in a "type" so as to remove a buffer of the same type that it is adding. The problem with this is that once the size limit is hit, the ratio of "metadata" to "data" contained in the arc essentially becomes fixed. For example, consider the following scenario: * the size of the arc is capped at 10G * the meta_limit is capped at 4G * 9G of the arc contains "data" * 1G of the arc contains "metadata" Now, every time a new "metadata" buffer is created and added to the arc, an older "metadata" buffer(s) will be removed from the arc; preserving the 9G "data" to 1G "metadata" ratio that was in-place when the size limit was reached. This occurs even though the amount of "metadata" is far below the "metadata" limit. This can result in the arc behaving pathologically for certain workloads. To fix this, the arc_get_data_buf function was modified to evict "data" from the arc even when adding a "metadata" buffer; unless it's at the "metadata" limit. In addition, arc_evict now more closely resembles arc_evict_ghost; such that when evicting "data" from the arc, it may make a second pass over the arc lists and evict "metadata" if it cannot meet the eviction size the first time around. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-30 21:30:00 +04:00
/* Only bump this if we tried to recycle and failed */
if (recycle)
ARCSTAT_BUMP(arcstat_recycle_miss);
2008-11-20 23:01:55 +03:00
}
ASSERT(buf->b_data != NULL);
out:
/*
* Update the state size. Note that ghost states have a
* "ghost size" and so don't need to be updated.
*/
if (!GHOST_STATE(buf->b_hdr->b_state)) {
arc_buf_hdr_t *hdr = buf->b_hdr;
atomic_add_64(&hdr->b_state->arcs_size, size);
if (list_link_active(&hdr->b_arc_node)) {
ASSERT(refcount_is_zero(&hdr->b_refcnt));
atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
}
/*
* If we are growing the cache, and we are adding anonymous
* data, and we have outgrown arc_p, update arc_p
*/
Disable aggressive arc_p growth by default For specific workloads consisting mainly of mfu data and new anon data buffers, the aggressive growth of arc_p found in the arc_get_data_buf() function can have detrimental effects on the mfu list size and ghost list hit rate. Running a workload consisting of two processes: * Process 1 is creating many small files * Process 2 is tar'ing a directory consisting of many small files I've seen arc_p and the mru grow to their maximum size, while the mru ghost list receives 100K times fewer hits than the mfu ghost list. Ideally, as the mfu ghost list receives hits, arc_p should be driven down and the size of the mfu should increase. Given the specific workload I was testing with, the mfu list size should grow to a point where almost no mfu ghost list hits would occur. Unfortunately, this does not happen because the newly dirtied anon buffers constancy drive arc_p to its maximum value and keep it there (effectively prioritizing the mru list and starving the mfu list down to a negligible size). The logic to increment arc_p from within the arc_get_data_buf() function was introduced many years ago in this upstream commit: commit 641fbdae3a027d12b3c3dcd18927ccafae6d58bc Author: maybee <none@none> Date: Wed Dec 20 15:46:12 2006 -0800 6505658 target MRU size (arc.p) needs to be adjusted more aggressively and since I don't fully understand the motivation for the change, I am reluctant to completely remove it. As a way to test out how it's removal might affect performance, I've disabled that code by default, but left it tunable via a module option. Thus, if its removal is found to be grossly detrimental for certain workloads, it can be re-enabled on the fly, without a code change. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-11 21:40:13 +04:00
if (!zfs_arc_p_aggressive_disable &&
arc_size < arc_c && hdr->b_state == arc_anon &&
2008-11-20 23:01:55 +03:00
arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
arc_p = MIN(arc_c, arc_p + size);
}
}
/*
* This routine is called whenever a buffer is accessed.
* NOTE: the hash lock is dropped in this function.
*/
static void
arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
{
clock_t now;
2008-11-20 23:01:55 +03:00
ASSERT(MUTEX_HELD(hash_lock));
if (buf->b_state == arc_anon) {
/*
* This buffer is not in the cache, and does not
* appear in our "ghost" list. Add the new buffer
* to the MRU state.
*/
ASSERT(buf->b_arc_access == 0);
buf->b_arc_access = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
arc_change_state(arc_mru, buf, hash_lock);
} else if (buf->b_state == arc_mru) {
now = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
/*
* If this buffer is here because of a prefetch, then either:
* - clear the flag if this is a "referencing" read
* (any subsequent access will bump this into the MFU state).
* or
* - move the buffer to the head of the list if this is
* another prefetch (to make it less likely to be evicted).
*/
if ((buf->b_flags & ARC_PREFETCH) != 0) {
if (refcount_count(&buf->b_refcnt) == 0) {
ASSERT(list_link_active(&buf->b_arc_node));
} else {
buf->b_flags &= ~ARC_PREFETCH;
atomic_inc_32(&buf->b_mru_hits);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_mru_hits);
}
buf->b_arc_access = now;
2008-11-20 23:01:55 +03:00
return;
}
/*
* This buffer has been "accessed" only once so far,
* but it is still in the cache. Move it to the MFU
* state.
*/
if (ddi_time_after(now, buf->b_arc_access + ARC_MINTIME)) {
2008-11-20 23:01:55 +03:00
/*
* More than 125ms have passed since we
* instantiated this buffer. Move it to the
* most frequently used state.
*/
buf->b_arc_access = now;
2008-11-20 23:01:55 +03:00
DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
arc_change_state(arc_mfu, buf, hash_lock);
}
atomic_inc_32(&buf->b_mru_hits);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_mru_hits);
} else if (buf->b_state == arc_mru_ghost) {
arc_state_t *new_state;
/*
* This buffer has been "accessed" recently, but
* was evicted from the cache. Move it to the
* MFU state.
*/
if (buf->b_flags & ARC_PREFETCH) {
new_state = arc_mru;
if (refcount_count(&buf->b_refcnt) > 0)
buf->b_flags &= ~ARC_PREFETCH;
DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
} else {
new_state = arc_mfu;
DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
}
buf->b_arc_access = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
arc_change_state(new_state, buf, hash_lock);
atomic_inc_32(&buf->b_mru_ghost_hits);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_mru_ghost_hits);
} else if (buf->b_state == arc_mfu) {
/*
* This buffer has been accessed more than once and is
* still in the cache. Keep it in the MFU state.
*
* NOTE: an add_reference() that occurred when we did
* the arc_read() will have kicked this off the list.
* If it was a prefetch, we will explicitly move it to
* the head of the list now.
*/
if ((buf->b_flags & ARC_PREFETCH) != 0) {
ASSERT(refcount_count(&buf->b_refcnt) == 0);
ASSERT(list_link_active(&buf->b_arc_node));
}
atomic_inc_32(&buf->b_mfu_hits);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_mfu_hits);
buf->b_arc_access = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
} else if (buf->b_state == arc_mfu_ghost) {
arc_state_t *new_state = arc_mfu;
/*
* This buffer has been accessed more than once but has
* been evicted from the cache. Move it back to the
* MFU state.
*/
if (buf->b_flags & ARC_PREFETCH) {
/*
* This is a prefetch access...
* move this block back to the MRU state.
*/
ASSERT0(refcount_count(&buf->b_refcnt));
2008-11-20 23:01:55 +03:00
new_state = arc_mru;
}
buf->b_arc_access = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
arc_change_state(new_state, buf, hash_lock);
atomic_inc_32(&buf->b_mfu_ghost_hits);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
} else if (buf->b_state == arc_l2c_only) {
/*
* This buffer is on the 2nd Level ARC.
*/
buf->b_arc_access = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
arc_change_state(arc_mfu, buf, hash_lock);
} else {
ASSERT(!"invalid arc state");
}
}
/* a generic arc_done_func_t which you can use */
/* ARGSUSED */
void
arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
{
if (zio == NULL || zio->io_error == 0)
bcopy(buf->b_data, arg, buf->b_hdr->b_size);
VERIFY(arc_buf_remove_ref(buf, arg));
2008-11-20 23:01:55 +03:00
}
/* a generic arc_done_func_t */
void
arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
{
arc_buf_t **bufp = arg;
if (zio && zio->io_error) {
VERIFY(arc_buf_remove_ref(buf, arg));
2008-11-20 23:01:55 +03:00
*bufp = NULL;
} else {
*bufp = buf;
ASSERT(buf->b_data);
2008-11-20 23:01:55 +03:00
}
}
static void
arc_read_done(zio_t *zio)
{
arc_buf_hdr_t *hdr;
2008-11-20 23:01:55 +03:00
arc_buf_t *buf;
arc_buf_t *abuf; /* buffer we're assigning to callback */
kmutex_t *hash_lock = NULL;
2008-11-20 23:01:55 +03:00
arc_callback_t *callback_list, *acb;
int freeable = FALSE;
buf = zio->io_private;
hdr = buf->b_hdr;
/*
* The hdr was inserted into hash-table and removed from lists
* prior to starting I/O. We should find this header, since
* it's in the hash table, and it should be legit since it's
* not possible to evict it during the I/O. The only possible
* reason for it not to be found is if we were freed during the
* read.
*/
if (HDR_IN_HASH_TABLE(hdr)) {
arc_buf_hdr_t *found;
ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
ASSERT3U(hdr->b_dva.dva_word[0], ==,
BP_IDENTITY(zio->io_bp)->dva_word[0]);
ASSERT3U(hdr->b_dva.dva_word[1], ==,
BP_IDENTITY(zio->io_bp)->dva_word[1]);
found = buf_hash_find(hdr->b_spa, zio->io_bp,
&hash_lock);
ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
hash_lock == NULL) ||
(found == hdr &&
DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
(found == hdr && HDR_L2_READING(hdr)));
}
2008-11-20 23:01:55 +03:00
hdr->b_flags &= ~ARC_L2_EVICTED;
2008-11-20 23:01:55 +03:00
if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
hdr->b_flags &= ~ARC_L2CACHE;
2008-11-20 23:01:55 +03:00
/* byteswap if necessary */
callback_list = hdr->b_acb;
ASSERT(callback_list != NULL);
if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
dmu_object_byteswap_t bswap =
DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
if (BP_GET_LEVEL(zio->io_bp) > 0)
byteswap_uint64_array(buf->b_data, hdr->b_size);
else
dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size);
}
2008-11-20 23:01:55 +03:00
arc_cksum_compute(buf, B_FALSE);
arc_buf_watch(buf);
2008-11-20 23:01:55 +03:00
if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
/*
* Only call arc_access on anonymous buffers. This is because
* if we've issued an I/O for an evicted buffer, we've already
* called arc_access (to prevent any simultaneous readers from
* getting confused).
*/
arc_access(hdr, hash_lock);
}
2008-11-20 23:01:55 +03:00
/* create copies of the data buffer for the callers */
abuf = buf;
for (acb = callback_list; acb; acb = acb->acb_next) {
if (acb->acb_done) {
if (abuf == NULL) {
ARCSTAT_BUMP(arcstat_duplicate_reads);
2008-11-20 23:01:55 +03:00
abuf = arc_buf_clone(buf);
}
2008-11-20 23:01:55 +03:00
acb->acb_buf = abuf;
abuf = NULL;
}
}
hdr->b_acb = NULL;
hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
ASSERT(!HDR_BUF_AVAILABLE(hdr));
if (abuf == buf) {
ASSERT(buf->b_efunc == NULL);
ASSERT(hdr->b_datacnt == 1);
2008-11-20 23:01:55 +03:00
hdr->b_flags |= ARC_BUF_AVAILABLE;
}
2008-11-20 23:01:55 +03:00
ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
if (zio->io_error != 0) {
hdr->b_flags |= ARC_IO_ERROR;
if (hdr->b_state != arc_anon)
arc_change_state(arc_anon, hdr, hash_lock);
if (HDR_IN_HASH_TABLE(hdr))
buf_hash_remove(hdr);
freeable = refcount_is_zero(&hdr->b_refcnt);
}
/*
* Broadcast before we drop the hash_lock to avoid the possibility
* that the hdr (and hence the cv) might be freed before we get to
* the cv_broadcast().
*/
cv_broadcast(&hdr->b_cv);
if (hash_lock) {
mutex_exit(hash_lock);
} else {
/*
* This block was freed while we waited for the read to
* complete. It has been removed from the hash table and
* moved to the anonymous state (so that it won't show up
* in the cache).
*/
ASSERT3P(hdr->b_state, ==, arc_anon);
freeable = refcount_is_zero(&hdr->b_refcnt);
}
/* execute each callback and free its structure */
while ((acb = callback_list) != NULL) {
if (acb->acb_done)
acb->acb_done(zio, acb->acb_buf, acb->acb_private);
if (acb->acb_zio_dummy != NULL) {
acb->acb_zio_dummy->io_error = zio->io_error;
zio_nowait(acb->acb_zio_dummy);
}
callback_list = acb->acb_next;
kmem_free(acb, sizeof (arc_callback_t));
}
if (freeable)
arc_hdr_destroy(hdr);
}
/*
* "Read" the block at the specified DVA (in bp) via the
2008-11-20 23:01:55 +03:00
* cache. If the block is found in the cache, invoke the provided
* callback immediately and return. Note that the `zio' parameter
* in the callback will be NULL in this case, since no IO was
* required. If the block is not in the cache pass the read request
* on to the spa with a substitute callback function, so that the
* requested block will be added to the cache.
*
* If a read request arrives for a block that has a read in-progress,
* either wait for the in-progress read to complete (and return the
* results); or, if this is a read with a "done" func, add a record
* to the read to invoke the "done" func when the read completes,
* and return; or just return.
*
* arc_read_done() will invoke all the requested "done" functions
* for readers of this block.
*/
int
arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
const zbookmark_phys_t *zb)
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *hdr = NULL;
arc_buf_t *buf = NULL;
kmutex_t *hash_lock = NULL;
2008-11-20 23:01:55 +03:00
zio_t *rzio;
uint64_t guid = spa_load_guid(spa);
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
int rc = 0;
2008-11-20 23:01:55 +03:00
ASSERT(!BP_IS_EMBEDDED(bp) ||
BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
2008-11-20 23:01:55 +03:00
top:
if (!BP_IS_EMBEDDED(bp)) {
/*
* Embedded BP's have no DVA and require no I/O to "read".
* Create an anonymous arc buf to back it.
*/
hdr = buf_hash_find(guid, bp, &hash_lock);
}
if (hdr != NULL && hdr->b_datacnt > 0) {
2008-11-20 23:01:55 +03:00
*arc_flags |= ARC_CACHED;
if (HDR_IO_IN_PROGRESS(hdr)) {
if (*arc_flags & ARC_WAIT) {
cv_wait(&hdr->b_cv, hash_lock);
mutex_exit(hash_lock);
goto top;
}
ASSERT(*arc_flags & ARC_NOWAIT);
if (done) {
arc_callback_t *acb = NULL;
acb = kmem_zalloc(sizeof (arc_callback_t),
KM_SLEEP);
2008-11-20 23:01:55 +03:00
acb->acb_done = done;
acb->acb_private = private;
if (pio != NULL)
acb->acb_zio_dummy = zio_null(pio,
2009-02-18 23:51:31 +03:00
spa, NULL, NULL, NULL, zio_flags);
2008-11-20 23:01:55 +03:00
ASSERT(acb->acb_done != NULL);
acb->acb_next = hdr->b_acb;
hdr->b_acb = acb;
add_reference(hdr, hash_lock, private);
mutex_exit(hash_lock);
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
goto out;
2008-11-20 23:01:55 +03:00
}
mutex_exit(hash_lock);
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
goto out;
2008-11-20 23:01:55 +03:00
}
ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
if (done) {
add_reference(hdr, hash_lock, private);
/*
* If this block is already in use, create a new
* copy of the data so that we will be guaranteed
* that arc_release() will always succeed.
*/
buf = hdr->b_buf;
ASSERT(buf);
ASSERT(buf->b_data);
if (HDR_BUF_AVAILABLE(hdr)) {
ASSERT(buf->b_efunc == NULL);
hdr->b_flags &= ~ARC_BUF_AVAILABLE;
} else {
buf = arc_buf_clone(buf);
}
2008-11-20 23:01:55 +03:00
} else if (*arc_flags & ARC_PREFETCH &&
refcount_count(&hdr->b_refcnt) == 0) {
hdr->b_flags |= ARC_PREFETCH;
}
DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
arc_access(hdr, hash_lock);
if (*arc_flags & ARC_L2CACHE)
hdr->b_flags |= ARC_L2CACHE;
if (*arc_flags & ARC_L2COMPRESS)
hdr->b_flags |= ARC_L2COMPRESS;
2008-11-20 23:01:55 +03:00
mutex_exit(hash_lock);
ARCSTAT_BUMP(arcstat_hits);
ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
data, metadata, hits);
if (done)
done(NULL, buf, private);
} else {
uint64_t size = BP_GET_LSIZE(bp);
arc_callback_t *acb;
vdev_t *vd = NULL;
uint64_t addr = 0;
2009-02-18 23:51:31 +03:00
boolean_t devw = B_FALSE;
enum zio_compress b_compress = ZIO_COMPRESS_OFF;
uint64_t b_asize = 0;
2008-11-20 23:01:55 +03:00
/*
* Gracefully handle a damaged logical block size as a
* checksum error by passing a dummy zio to the done callback.
*/
if (size > SPA_MAXBLOCKSIZE) {
if (done) {
rzio = zio_null(pio, spa, NULL,
NULL, NULL, zio_flags);
rzio->io_error = ECKSUM;
done(rzio, buf, private);
zio_nowait(rzio);
}
rc = ECKSUM;
goto out;
}
2008-11-20 23:01:55 +03:00
if (hdr == NULL) {
/* this block is not in the cache */
arc_buf_hdr_t *exists = NULL;
2008-11-20 23:01:55 +03:00
arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
buf = arc_buf_alloc(spa, size, private, type);
hdr = buf->b_hdr;
if (!BP_IS_EMBEDDED(bp)) {
hdr->b_dva = *BP_IDENTITY(bp);
hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
exists = buf_hash_insert(hdr, &hash_lock);
}
if (exists != NULL) {
2008-11-20 23:01:55 +03:00
/* somebody beat us to the hash insert */
mutex_exit(hash_lock);
buf_discard_identity(hdr);
2008-11-20 23:01:55 +03:00
(void) arc_buf_remove_ref(buf, private);
goto top; /* restart the IO request */
}
/* if this is a prefetch, we don't have a reference */
if (*arc_flags & ARC_PREFETCH) {
(void) remove_reference(hdr, hash_lock,
private);
hdr->b_flags |= ARC_PREFETCH;
}
if (*arc_flags & ARC_L2CACHE)
hdr->b_flags |= ARC_L2CACHE;
if (*arc_flags & ARC_L2COMPRESS)
hdr->b_flags |= ARC_L2COMPRESS;
2008-11-20 23:01:55 +03:00
if (BP_GET_LEVEL(bp) > 0)
hdr->b_flags |= ARC_INDIRECT;
} else {
/* this block is in the ghost cache */
ASSERT(GHOST_STATE(hdr->b_state));
ASSERT(!HDR_IO_IN_PROGRESS(hdr));
ASSERT0(refcount_count(&hdr->b_refcnt));
2008-11-20 23:01:55 +03:00
ASSERT(hdr->b_buf == NULL);
/* if this is a prefetch, we don't have a reference */
if (*arc_flags & ARC_PREFETCH)
hdr->b_flags |= ARC_PREFETCH;
else
add_reference(hdr, hash_lock, private);
if (*arc_flags & ARC_L2CACHE)
hdr->b_flags |= ARC_L2CACHE;
if (*arc_flags & ARC_L2COMPRESS)
hdr->b_flags |= ARC_L2COMPRESS;
2008-11-20 23:01:55 +03:00
buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
buf->b_hdr = hdr;
buf->b_data = NULL;
buf->b_efunc = NULL;
buf->b_private = NULL;
buf->b_next = NULL;
hdr->b_buf = buf;
ASSERT(hdr->b_datacnt == 0);
hdr->b_datacnt = 1;
arc_get_data_buf(buf);
arc_access(hdr, hash_lock);
2008-11-20 23:01:55 +03:00
}
ASSERT(!GHOST_STATE(hdr->b_state));
acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2008-11-20 23:01:55 +03:00
acb->acb_done = done;
acb->acb_private = private;
ASSERT(hdr->b_acb == NULL);
hdr->b_acb = acb;
hdr->b_flags |= ARC_IO_IN_PROGRESS;
if (hdr->b_l2hdr != NULL &&
(vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2009-02-18 23:51:31 +03:00
devw = hdr->b_l2hdr->b_dev->l2ad_writing;
addr = hdr->b_l2hdr->b_daddr;
b_compress = hdr->b_l2hdr->b_compress;
b_asize = hdr->b_l2hdr->b_asize;
/*
* Lock out device removal.
*/
if (vdev_is_dead(vd) ||
!spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
vd = NULL;
}
if (hash_lock != NULL)
mutex_exit(hash_lock);
/*
* At this point, we have a level 1 cache miss. Try again in
* L2ARC if possible.
*/
2008-11-20 23:01:55 +03:00
ASSERT3U(hdr->b_size, ==, size);
DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
uint64_t, size, zbookmark_phys_t *, zb);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_misses);
ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
data, metadata, misses);
2009-02-18 23:51:31 +03:00
if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
2008-11-20 23:01:55 +03:00
/*
* Read from the L2ARC if the following are true:
* 1. The L2ARC vdev was previously cached.
* 2. This buffer still has L2ARC metadata.
* 3. This buffer isn't currently writing to the L2ARC.
* 4. The L2ARC entry wasn't evicted, which may
* also have invalidated the vdev.
2009-02-18 23:51:31 +03:00
* 5. This isn't prefetch and l2arc_noprefetch is set.
2008-11-20 23:01:55 +03:00
*/
if (hdr->b_l2hdr != NULL &&
2009-02-18 23:51:31 +03:00
!HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
!(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
2008-11-20 23:01:55 +03:00
l2arc_read_callback_t *cb;
DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
ARCSTAT_BUMP(arcstat_l2_hits);
atomic_inc_32(&hdr->b_l2hdr->b_hits);
2008-11-20 23:01:55 +03:00
cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
KM_SLEEP);
2008-11-20 23:01:55 +03:00
cb->l2rcb_buf = buf;
cb->l2rcb_spa = spa;
cb->l2rcb_bp = *bp;
cb->l2rcb_zb = *zb;
cb->l2rcb_flags = zio_flags;
cb->l2rcb_compress = b_compress;
2008-11-20 23:01:55 +03:00
ASSERT(addr >= VDEV_LABEL_START_SIZE &&
addr + size < vd->vdev_psize -
VDEV_LABEL_END_SIZE);
2008-11-20 23:01:55 +03:00
/*
* l2arc read. The SCL_L2ARC lock will be
* released by l2arc_read_done().
* Issue a null zio if the underlying buffer
* was squashed to zero size by compression.
2008-11-20 23:01:55 +03:00
*/
if (b_compress == ZIO_COMPRESS_EMPTY) {
rzio = zio_null(pio, spa, vd,
l2arc_read_done, cb,
zio_flags | ZIO_FLAG_DONT_CACHE |
ZIO_FLAG_CANFAIL |
ZIO_FLAG_DONT_PROPAGATE |
ZIO_FLAG_DONT_RETRY);
} else {
rzio = zio_read_phys(pio, vd, addr,
b_asize, buf->b_data,
ZIO_CHECKSUM_OFF,
l2arc_read_done, cb, priority,
zio_flags | ZIO_FLAG_DONT_CACHE |
ZIO_FLAG_CANFAIL |
ZIO_FLAG_DONT_PROPAGATE |
ZIO_FLAG_DONT_RETRY, B_FALSE);
}
2008-11-20 23:01:55 +03:00
DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
zio_t *, rzio);
ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
2008-11-20 23:01:55 +03:00
if (*arc_flags & ARC_NOWAIT) {
zio_nowait(rzio);
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
goto out;
}
2008-11-20 23:01:55 +03:00
ASSERT(*arc_flags & ARC_WAIT);
if (zio_wait(rzio) == 0)
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
goto out;
/* l2arc read error; goto zio_read() */
2008-11-20 23:01:55 +03:00
} else {
DTRACE_PROBE1(l2arc__miss,
arc_buf_hdr_t *, hdr);
ARCSTAT_BUMP(arcstat_l2_misses);
if (HDR_L2_WRITING(hdr))
ARCSTAT_BUMP(arcstat_l2_rw_clash);
spa_config_exit(spa, SCL_L2ARC, vd);
2008-11-20 23:01:55 +03:00
}
2009-02-18 23:51:31 +03:00
} else {
if (vd != NULL)
spa_config_exit(spa, SCL_L2ARC, vd);
if (l2arc_ndev != 0) {
DTRACE_PROBE1(l2arc__miss,
arc_buf_hdr_t *, hdr);
ARCSTAT_BUMP(arcstat_l2_misses);
}
2008-11-20 23:01:55 +03:00
}
rzio = zio_read(pio, spa, bp, buf->b_data, size,
arc_read_done, buf, priority, zio_flags, zb);
2008-11-20 23:01:55 +03:00
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
if (*arc_flags & ARC_WAIT) {
rc = zio_wait(rzio);
goto out;
}
2008-11-20 23:01:55 +03:00
ASSERT(*arc_flags & ARC_NOWAIT);
zio_nowait(rzio);
}
Add visibility in to arc_read This change is an attempt to add visibility into the arc_read calls occurring on a system, in real time. To do this, a list was added to the in memory SPA data structure for a pool, with each element on the list corresponding to a call to arc_read. These entries are then exported through the kstat interface, which can then be interpreted in userspace. For each arc_read call, the following information is exported: * A unique identifier (uint64_t) * The time the entry was added to the list (hrtime_t) (*not* wall clock time; relative to the other entries on the list) * The objset ID (uint64_t) * The object number (uint64_t) * The indirection level (uint64_t) * The block ID (uint64_t) * The name of the function originating the arc_read call (char[24]) * The arc_flags from the arc_read call (uint32_t) * The PID of the reading thread (pid_t) * The command or name of thread originating read (char[16]) From this exported information one can see, in real time, exactly what is being read, what function is generating the read, and whether or not the read was found to be already cached. There is still some work to be done, but this should serve as a good starting point. Specifically, dbuf_read's are not accounted for in the currently exported information. Thus, a follow up patch should probably be added to export these calls that never call into arc_read (they only hit the dbuf hash table). In addition, it might be nice to create a utility similar to "arcstat.py" to digest the exported information and display it in a more readable format. Or perhaps, log the information and allow for it to be "replayed" at a later time. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-09-07 03:09:05 +04:00
out:
spa_read_history_add(spa, zb, *arc_flags);
return (rc);
2008-11-20 23:01:55 +03:00
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
arc_prune_t *
arc_add_prune_callback(arc_prune_func_t *func, void *private)
{
arc_prune_t *p;
p = kmem_alloc(sizeof (*p), KM_SLEEP);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
p->p_pfunc = func;
p->p_private = private;
list_link_init(&p->p_node);
refcount_create(&p->p_refcnt);
mutex_enter(&arc_prune_mtx);
refcount_add(&p->p_refcnt, &arc_prune_list);
list_insert_head(&arc_prune_list, p);
mutex_exit(&arc_prune_mtx);
return (p);
}
void
arc_remove_prune_callback(arc_prune_t *p)
{
mutex_enter(&arc_prune_mtx);
list_remove(&arc_prune_list, p);
if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
refcount_destroy(&p->p_refcnt);
kmem_free(p, sizeof (*p));
}
mutex_exit(&arc_prune_mtx);
}
2008-11-20 23:01:55 +03:00
void
arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
{
ASSERT(buf->b_hdr != NULL);
ASSERT(buf->b_hdr->b_state != arc_anon);
ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
ASSERT(buf->b_efunc == NULL);
ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
2008-11-20 23:01:55 +03:00
buf->b_efunc = func;
buf->b_private = private;
}
Illumos #3805 arc shouldn't cache freed blocks 3805 arc shouldn't cache freed blocks Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Richard Elling <richard.elling@dey-sys.com> Reviewed by: Will Andrews <will@firepipe.net> Approved by: Dan McDonald <danmcd@nexenta.com> References: illumos/illumos-gate@6e6d5868f52089b9026785bd90257a3d3f6e5ee2 https://www.illumos.org/issues/3805 ZFS should proactively evict freed blocks from the cache. On dcenter, we saw that we were caching ~256GB of metadata, while the pool only had <4GB of metadata on disk. We were wasting about half the system's RAM (252GB) on blocks that have been freed. Even though these freed blocks will never be used again, and thus will eventually be evicted, this causes us to use memory inefficiently for 2 reasons: 1. A block that is freed has no chance of being accessed again, but will be kept in memory preferentially to a block that was accessed before it (and is thus older) but has not been freed and thus has at least some chance of being accessed again. 2. We partition the ARC into several buckets: user data that has been accessed only once (MRU) metadata that has been accessed only once (MRU) user data that has been accessed more than once (MFU) metadata that has been accessed more than once (MFU) The user data vs metadata split is somewhat arbitrary, and the primary control on how much memory is used to cache data vs metadata is to simply try to keep the proportion the same as it has been in the past (each bucket "evicts against" itself). The secondary control is to evict data before evicting metadata. Because of this bucketing, we may end up with one bucket mostly containing freed blocks that are very old, while another bucket has more recently accessed, still-allocated blocks. Data in the useful bucket (with still-allocated blocks) may be evicted in preference to data in the useless bucket (with old, freed blocks). On dcenter, we saw that the MFU metadata bucket was 230MB, while the MFU data bucket was 27GB and the MRU metadata bucket was 256GB. However, the vast majority of data in the MRU metadata bucket (256GB) was freed blocks, and thus useless. Meanwhile, the MFU metadata bucket (230MB) was constantly evicting useful blocks that will be soon needed. The problem of cache segmentation is a larger problem that needs more investigation. However, if we stop caching freed blocks, it should reduce the impact of this more fundamental issue. Ported-by: Richard Yao <ryao@cs.stonybrook.edu> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1503
2013-06-07 02:46:55 +04:00
/*
* Notify the arc that a block was freed, and thus will never be used again.
*/
void
arc_freed(spa_t *spa, const blkptr_t *bp)
{
arc_buf_hdr_t *hdr;
kmutex_t *hash_lock;
uint64_t guid = spa_load_guid(spa);
ASSERT(!BP_IS_EMBEDDED(bp));
hdr = buf_hash_find(guid, bp, &hash_lock);
Illumos #3805 arc shouldn't cache freed blocks 3805 arc shouldn't cache freed blocks Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Richard Elling <richard.elling@dey-sys.com> Reviewed by: Will Andrews <will@firepipe.net> Approved by: Dan McDonald <danmcd@nexenta.com> References: illumos/illumos-gate@6e6d5868f52089b9026785bd90257a3d3f6e5ee2 https://www.illumos.org/issues/3805 ZFS should proactively evict freed blocks from the cache. On dcenter, we saw that we were caching ~256GB of metadata, while the pool only had <4GB of metadata on disk. We were wasting about half the system's RAM (252GB) on blocks that have been freed. Even though these freed blocks will never be used again, and thus will eventually be evicted, this causes us to use memory inefficiently for 2 reasons: 1. A block that is freed has no chance of being accessed again, but will be kept in memory preferentially to a block that was accessed before it (and is thus older) but has not been freed and thus has at least some chance of being accessed again. 2. We partition the ARC into several buckets: user data that has been accessed only once (MRU) metadata that has been accessed only once (MRU) user data that has been accessed more than once (MFU) metadata that has been accessed more than once (MFU) The user data vs metadata split is somewhat arbitrary, and the primary control on how much memory is used to cache data vs metadata is to simply try to keep the proportion the same as it has been in the past (each bucket "evicts against" itself). The secondary control is to evict data before evicting metadata. Because of this bucketing, we may end up with one bucket mostly containing freed blocks that are very old, while another bucket has more recently accessed, still-allocated blocks. Data in the useful bucket (with still-allocated blocks) may be evicted in preference to data in the useless bucket (with old, freed blocks). On dcenter, we saw that the MFU metadata bucket was 230MB, while the MFU data bucket was 27GB and the MRU metadata bucket was 256GB. However, the vast majority of data in the MRU metadata bucket (256GB) was freed blocks, and thus useless. Meanwhile, the MFU metadata bucket (230MB) was constantly evicting useful blocks that will be soon needed. The problem of cache segmentation is a larger problem that needs more investigation. However, if we stop caching freed blocks, it should reduce the impact of this more fundamental issue. Ported-by: Richard Yao <ryao@cs.stonybrook.edu> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1503
2013-06-07 02:46:55 +04:00
if (hdr == NULL)
return;
if (HDR_BUF_AVAILABLE(hdr)) {
arc_buf_t *buf = hdr->b_buf;
add_reference(hdr, hash_lock, FTAG);
hdr->b_flags &= ~ARC_BUF_AVAILABLE;
mutex_exit(hash_lock);
arc_release(buf, FTAG);
(void) arc_buf_remove_ref(buf, FTAG);
} else {
mutex_exit(hash_lock);
}
}
2008-11-20 23:01:55 +03:00
/*
* Clear the user eviction callback set by arc_set_callback(), first calling
* it if it exists. Because the presence of a callback keeps an arc_buf cached
* clearing the callback may result in the arc_buf being destroyed. However,
* it will not result in the *last* arc_buf being destroyed, hence the data
* will remain cached in the ARC. We make a copy of the arc buffer here so
* that we can process the callback without holding any locks.
*
* It's possible that the callback is already in the process of being cleared
* by another thread. In this case we can not clear the callback.
*
* Returns B_TRUE if the callback was successfully called and cleared.
2008-11-20 23:01:55 +03:00
*/
boolean_t
arc_clear_callback(arc_buf_t *buf)
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *hdr;
kmutex_t *hash_lock;
arc_evict_func_t *efunc = buf->b_efunc;
void *private = buf->b_private;
2008-11-20 23:01:55 +03:00
mutex_enter(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
hdr = buf->b_hdr;
if (hdr == NULL) {
/*
* We are in arc_do_user_evicts().
*/
ASSERT(buf->b_data == NULL);
mutex_exit(&buf->b_evict_lock);
return (B_FALSE);
} else if (buf->b_data == NULL) {
2008-11-20 23:01:55 +03:00
/*
* We are on the eviction list; process this buffer now
* but let arc_do_user_evicts() do the reaping.
2008-11-20 23:01:55 +03:00
*/
buf->b_efunc = NULL;
mutex_exit(&buf->b_evict_lock);
VERIFY0(efunc(private));
return (B_TRUE);
2008-11-20 23:01:55 +03:00
}
hash_lock = HDR_LOCK(hdr);
mutex_enter(hash_lock);
hdr = buf->b_hdr;
ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2008-11-20 23:01:55 +03:00
ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
buf->b_efunc = NULL;
buf->b_private = NULL;
2008-11-20 23:01:55 +03:00
if (hdr->b_datacnt > 1) {
mutex_exit(&buf->b_evict_lock);
arc_buf_destroy(buf, FALSE, TRUE);
} else {
ASSERT(buf == hdr->b_buf);
hdr->b_flags |= ARC_BUF_AVAILABLE;
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
}
mutex_exit(hash_lock);
VERIFY0(efunc(private));
return (B_TRUE);
2008-11-20 23:01:55 +03:00
}
/*
* Release this buffer from the cache, making it an anonymous buffer. This
* must be done after a read and prior to modifying the buffer contents.
2008-11-20 23:01:55 +03:00
* If the buffer has more than one reference, we must make
* a new hdr for the buffer.
2008-11-20 23:01:55 +03:00
*/
void
arc_release(arc_buf_t *buf, void *tag)
{
arc_buf_hdr_t *hdr;
kmutex_t *hash_lock = NULL;
l2arc_buf_hdr_t *l2hdr;
uint64_t buf_size = 0;
2008-11-20 23:01:55 +03:00
/*
* It would be nice to assert that if it's DMU metadata (level >
* 0 || it's the dnode file), then it must be syncing context.
* But we don't know that information at this level.
*/
mutex_enter(&buf->b_evict_lock);
hdr = buf->b_hdr;
2008-11-20 23:01:55 +03:00
/* this buffer is not on any list */
ASSERT(refcount_count(&hdr->b_refcnt) > 0);
if (hdr->b_state == arc_anon) {
/* this buffer is already released */
ASSERT(buf->b_efunc == NULL);
2009-07-03 02:44:48 +04:00
} else {
hash_lock = HDR_LOCK(hdr);
mutex_enter(hash_lock);
hdr = buf->b_hdr;
ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2008-11-20 23:01:55 +03:00
}
l2hdr = hdr->b_l2hdr;
if (l2hdr) {
mutex_enter(&l2arc_buflist_mtx);
arc_buf_l2_cdata_free(hdr);
hdr->b_l2hdr = NULL;
list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
}
buf_size = hdr->b_size;
2008-11-20 23:01:55 +03:00
/*
* Do we have more than one buf?
*/
if (hdr->b_datacnt > 1) {
2008-11-20 23:01:55 +03:00
arc_buf_hdr_t *nhdr;
arc_buf_t **bufp;
uint64_t blksz = hdr->b_size;
2009-02-18 23:51:31 +03:00
uint64_t spa = hdr->b_spa;
2008-11-20 23:01:55 +03:00
arc_buf_contents_t type = hdr->b_type;
uint32_t flags = hdr->b_flags;
ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
2008-11-20 23:01:55 +03:00
/*
* Pull the data off of this hdr and attach it to
* a new anonymous hdr.
2008-11-20 23:01:55 +03:00
*/
(void) remove_reference(hdr, hash_lock, tag);
bufp = &hdr->b_buf;
while (*bufp != buf)
bufp = &(*bufp)->b_next;
*bufp = buf->b_next;
2008-11-20 23:01:55 +03:00
buf->b_next = NULL;
ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
if (refcount_is_zero(&hdr->b_refcnt)) {
uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
ASSERT3U(*size, >=, hdr->b_size);
atomic_add_64(size, -hdr->b_size);
}
/*
* We're releasing a duplicate user data buffer, update
* our statistics accordingly.
*/
if (hdr->b_type == ARC_BUFC_DATA) {
ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
ARCSTAT_INCR(arcstat_duplicate_buffers_size,
-hdr->b_size);
}
2008-11-20 23:01:55 +03:00
hdr->b_datacnt -= 1;
arc_cksum_verify(buf);
arc_buf_unwatch(buf);
2008-11-20 23:01:55 +03:00
mutex_exit(hash_lock);
nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
nhdr->b_size = blksz;
nhdr->b_spa = spa;
nhdr->b_type = type;
nhdr->b_buf = buf;
nhdr->b_state = arc_anon;
nhdr->b_arc_access = 0;
nhdr->b_mru_hits = 0;
nhdr->b_mru_ghost_hits = 0;
nhdr->b_mfu_hits = 0;
nhdr->b_mfu_ghost_hits = 0;
nhdr->b_l2_hits = 0;
2008-11-20 23:01:55 +03:00
nhdr->b_flags = flags & ARC_L2_WRITING;
nhdr->b_l2hdr = NULL;
nhdr->b_datacnt = 1;
nhdr->b_freeze_cksum = NULL;
(void) refcount_add(&nhdr->b_refcnt, tag);
buf->b_hdr = nhdr;
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
atomic_add_64(&arc_anon->arcs_size, blksz);
} else {
mutex_exit(&buf->b_evict_lock);
2008-11-20 23:01:55 +03:00
ASSERT(refcount_count(&hdr->b_refcnt) == 1);
ASSERT(!list_link_active(&hdr->b_arc_node));
ASSERT(!HDR_IO_IN_PROGRESS(hdr));
if (hdr->b_state != arc_anon)
arc_change_state(arc_anon, hdr, hash_lock);
2008-11-20 23:01:55 +03:00
hdr->b_arc_access = 0;
hdr->b_mru_hits = 0;
hdr->b_mru_ghost_hits = 0;
hdr->b_mfu_hits = 0;
hdr->b_mfu_ghost_hits = 0;
hdr->b_l2_hits = 0;
if (hash_lock)
mutex_exit(hash_lock);
2008-11-20 23:01:55 +03:00
buf_discard_identity(hdr);
2008-11-20 23:01:55 +03:00
arc_buf_thaw(buf);
}
buf->b_efunc = NULL;
buf->b_private = NULL;
if (l2hdr) {
ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
vdev_space_update(l2hdr->b_dev->l2ad_vdev,
-l2hdr->b_asize, 0, 0);
kmem_cache_free(l2arc_hdr_cache, l2hdr);
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
2008-11-20 23:01:55 +03:00
ARCSTAT_INCR(arcstat_l2_size, -buf_size);
mutex_exit(&l2arc_buflist_mtx);
}
2008-11-20 23:01:55 +03:00
}
int
arc_released(arc_buf_t *buf)
{
int released;
mutex_enter(&buf->b_evict_lock);
released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
mutex_exit(&buf->b_evict_lock);
return (released);
2008-11-20 23:01:55 +03:00
}
#ifdef ZFS_DEBUG
int
arc_referenced(arc_buf_t *buf)
{
int referenced;
mutex_enter(&buf->b_evict_lock);
referenced = (refcount_count(&buf->b_hdr->b_refcnt));
mutex_exit(&buf->b_evict_lock);
return (referenced);
2008-11-20 23:01:55 +03:00
}
#endif
static void
arc_write_ready(zio_t *zio)
{
arc_write_callback_t *callback = zio->io_private;
arc_buf_t *buf = callback->awcb_buf;
arc_buf_hdr_t *hdr = buf->b_hdr;
ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
callback->awcb_ready(zio, buf, callback->awcb_private);
2008-11-20 23:01:55 +03:00
/*
* If the IO is already in progress, then this is a re-write
* attempt, so we need to thaw and re-compute the cksum.
* It is the responsibility of the callback to handle the
* accounting for any re-write attempt.
2008-11-20 23:01:55 +03:00
*/
if (HDR_IO_IN_PROGRESS(hdr)) {
mutex_enter(&hdr->b_freeze_lock);
if (hdr->b_freeze_cksum != NULL) {
kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
hdr->b_freeze_cksum = NULL;
}
mutex_exit(&hdr->b_freeze_lock);
}
arc_cksum_compute(buf, B_FALSE);
hdr->b_flags |= ARC_IO_IN_PROGRESS;
}
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
/*
* The SPA calls this callback for each physical write that happens on behalf
* of a logical write. See the comment in dbuf_write_physdone() for details.
*/
static void
arc_write_physdone(zio_t *zio)
{
arc_write_callback_t *cb = zio->io_private;
if (cb->awcb_physdone != NULL)
cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
}
2008-11-20 23:01:55 +03:00
static void
arc_write_done(zio_t *zio)
{
arc_write_callback_t *callback = zio->io_private;
arc_buf_t *buf = callback->awcb_buf;
arc_buf_hdr_t *hdr = buf->b_hdr;
ASSERT(hdr->b_acb == NULL);
if (zio->io_error == 0) {
if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
buf_discard_identity(hdr);
} else {
hdr->b_dva = *BP_IDENTITY(zio->io_bp);
hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
}
} else {
ASSERT(BUF_EMPTY(hdr));
}
2008-11-20 23:01:55 +03:00
/*
* If the block to be written was all-zero or compressed enough to be
* embedded in the BP, no write was performed so there will be no
* dva/birth/checksum. The buffer must therefore remain anonymous
* (and uncached).
2008-11-20 23:01:55 +03:00
*/
if (!BUF_EMPTY(hdr)) {
arc_buf_hdr_t *exists;
kmutex_t *hash_lock;
ASSERT(zio->io_error == 0);
2008-11-20 23:01:55 +03:00
arc_cksum_verify(buf);
exists = buf_hash_insert(hdr, &hash_lock);
if (exists) {
/*
* This can only happen if we overwrite for
* sync-to-convergence, because we remove
* buffers from the hash table when we arc_free().
*/
if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
panic("bad overwrite, hdr=%p exists=%p",
(void *)hdr, (void *)exists);
ASSERT(refcount_is_zero(&exists->b_refcnt));
arc_change_state(arc_anon, exists, hash_lock);
mutex_exit(hash_lock);
arc_hdr_destroy(exists);
exists = buf_hash_insert(hdr, &hash_lock);
ASSERT3P(exists, ==, NULL);
} else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
/* nopwrite */
ASSERT(zio->io_prop.zp_nopwrite);
if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
panic("bad nopwrite, hdr=%p exists=%p",
(void *)hdr, (void *)exists);
} else {
/* Dedup */
ASSERT(hdr->b_datacnt == 1);
ASSERT(hdr->b_state == arc_anon);
ASSERT(BP_GET_DEDUP(zio->io_bp));
ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
}
2008-11-20 23:01:55 +03:00
}
hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
/* if it's not anon, we are doing a scrub */
if (!exists && hdr->b_state == arc_anon)
arc_access(hdr, hash_lock);
2008-11-20 23:01:55 +03:00
mutex_exit(hash_lock);
} else {
hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
}
ASSERT(!refcount_is_zero(&hdr->b_refcnt));
callback->awcb_done(zio, buf, callback->awcb_private);
2008-11-20 23:01:55 +03:00
kmem_free(callback, sizeof (arc_write_callback_t));
}
zio_t *
arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
arc_done_func_t *done, void *private, zio_priority_t priority,
int zio_flags, const zbookmark_phys_t *zb)
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *hdr = buf->b_hdr;
arc_write_callback_t *callback;
zio_t *zio;
2008-11-20 23:01:55 +03:00
ASSERT(ready != NULL);
ASSERT(done != NULL);
2008-11-20 23:01:55 +03:00
ASSERT(!HDR_IO_ERROR(hdr));
ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
ASSERT(hdr->b_acb == NULL);
if (l2arc)
hdr->b_flags |= ARC_L2CACHE;
if (l2arc_compress)
hdr->b_flags |= ARC_L2COMPRESS;
callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
2008-11-20 23:01:55 +03:00
callback->awcb_ready = ready;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
callback->awcb_physdone = physdone;
2008-11-20 23:01:55 +03:00
callback->awcb_done = done;
callback->awcb_private = private;
callback->awcb_buf = buf;
zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
arc_write_ready, arc_write_physdone, arc_write_done, callback,
priority, zio_flags, zb);
2008-11-20 23:01:55 +03:00
return (zio);
}
static int
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
arc_memory_throttle(uint64_t reserve, uint64_t txg)
2008-11-20 23:01:55 +03:00
{
#ifdef _KERNEL
if (zfs_arc_memory_throttle_disable)
return (0);
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
if (freemem <= physmem * arc_lotsfree_percent / 100) {
2008-11-20 23:01:55 +03:00
ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
return (SET_ERROR(EAGAIN));
2008-11-20 23:01:55 +03:00
}
#endif
return (0);
}
void
arc_tempreserve_clear(uint64_t reserve)
{
atomic_add_64(&arc_tempreserve, -reserve);
ASSERT((int64_t)arc_tempreserve >= 0);
}
int
arc_tempreserve_space(uint64_t reserve, uint64_t txg)
{
int error;
2009-07-03 02:44:48 +04:00
uint64_t anon_size;
2008-11-20 23:01:55 +03:00
if (reserve > arc_c/4 && !arc_no_grow)
arc_c = MIN(arc_c_max, reserve * 4);
/*
* Throttle when the calculated memory footprint for the TXG
* exceeds the target ARC size.
*/
if (reserve > arc_c) {
DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
return (SET_ERROR(ERESTART));
}
2008-11-20 23:01:55 +03:00
2009-07-03 02:44:48 +04:00
/*
* Don't count loaned bufs as in flight dirty data to prevent long
* network delays from blocking transactions that are ready to be
* assigned to a txg.
*/
anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
2008-11-20 23:01:55 +03:00
/*
* Writes will, almost always, require additional memory allocations
* in order to compress/encrypt/etc the data. We therefore need to
2008-11-20 23:01:55 +03:00
* make sure that there is sufficient available memory for this.
*/
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
error = arc_memory_throttle(reserve, txg);
if (error != 0)
2008-11-20 23:01:55 +03:00
return (error);
/*
* Throttle writes when the amount of dirty data in the cache
* gets too large. We try to keep the cache less than half full
* of dirty blocks so that our sync times don't grow too large.
* Note: if two requests come in concurrently, we might let them
* both succeed, when one of them should fail. Not a huge deal.
*/
2009-07-03 02:44:48 +04:00
if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
anon_size > arc_c / 4) {
2008-11-20 23:01:55 +03:00
dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
"anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
arc_tempreserve>>10,
arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
reserve>>10, arc_c>>10);
DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
return (SET_ERROR(ERESTART));
2008-11-20 23:01:55 +03:00
}
atomic_add_64(&arc_tempreserve, reserve);
return (0);
}
static void
arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
kstat_named_t *evict_data, kstat_named_t *evict_metadata)
{
size->value.ui64 = state->arcs_size;
evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
}
static int
arc_kstat_update(kstat_t *ksp, int rw)
{
arc_stats_t *as = ksp->ks_data;
if (rw == KSTAT_WRITE) {
return (SET_ERROR(EACCES));
} else {
arc_kstat_update_state(arc_anon,
&as->arcstat_anon_size,
&as->arcstat_anon_evict_data,
&as->arcstat_anon_evict_metadata);
arc_kstat_update_state(arc_mru,
&as->arcstat_mru_size,
&as->arcstat_mru_evict_data,
&as->arcstat_mru_evict_metadata);
arc_kstat_update_state(arc_mru_ghost,
&as->arcstat_mru_ghost_size,
&as->arcstat_mru_ghost_evict_data,
&as->arcstat_mru_ghost_evict_metadata);
arc_kstat_update_state(arc_mfu,
&as->arcstat_mfu_size,
&as->arcstat_mfu_evict_data,
&as->arcstat_mfu_evict_metadata);
arc_kstat_update_state(arc_mfu_ghost,
&as->arcstat_mfu_ghost_size,
&as->arcstat_mfu_ghost_evict_data,
&as->arcstat_mfu_ghost_evict_metadata);
}
return (0);
}
2008-11-20 23:01:55 +03:00
void
arc_init(void)
{
mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
/* Convert seconds to clock ticks */
zfs_arc_min_prefetch_lifespan = 1 * hz;
2008-11-20 23:01:55 +03:00
/* Start out with 1/8 of all memory */
arc_c = physmem * PAGESIZE / 8;
#ifdef _KERNEL
/*
* On architectures where the physical memory can be larger
* than the addressable space (intel in 32-bit mode), we may
* need to limit the cache to 1/8 of VM size.
*/
arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
/*
* Register a shrinker to support synchronous (direct) memory
* reclaim from the arc. This is done to prevent kswapd from
* swapping out pages when it is preferable to shrink the arc.
*/
spl_register_shrinker(&arc_shrinker);
2008-11-20 23:01:55 +03:00
#endif
/* set min cache to zero */
arc_c_min = 4<<20;
/* set max to 1/2 of all memory */
arc_c_max = arc_c * 4;
2008-11-20 23:01:55 +03:00
/*
* Allow the tunables to override our calculations if they are
* reasonable (ie. over 64MB)
*/
if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
arc_c_max = zfs_arc_max;
if (zfs_arc_min > 0 && zfs_arc_min <= arc_c_max)
2008-11-20 23:01:55 +03:00
arc_c_min = zfs_arc_min;
arc_c = arc_c_max;
arc_p = (arc_c >> 1);
Set "arc_meta_limit" to 3/4 arc_c_max by default Unfortunately, this change is an cheap attempt to work around a pathological workload for the ARC. A "real" solution still needs to be fleshed out, so this patch is intended to alleviate the situation in the meantime. Let me try and describe the problem.. Data buffers residing in the dbuf hash table (dbuf cache) will keep a hold on their respective dnode, this dnode will in turn keep a hold on its backing dbuf (the physical block of the dnode object backing it). Since the dnode has a hold on its backing dbuf, the arc buffer for this dbuf is unevictable. What this essentially boils down to, "data" buffers have the potential to pin "metadata" in the arc (as a result of these dnode object buffers being unevictable). This scenario becomes a real problem when the workload consists of many small files (e.g. creating millions of 4K files). With this workload, the arc's "arc_meta_used" space get filled up with buffers for any resident directories as well as buffers for the objset's dnode object. Once the "arc_meta_limit" is reached, the directory buffers will be evicted and only the unevictable dnode object buffers will reside. If the workload is simply creating new small files, these dnode object buffers will never even be needed again, whereas the directory buffers will be used constantly until the creates move to a new directory. If "arc_c" and "arc_meta_limit" are sized appropriately, this situation wont occur. This is because as the data buffers accumulate, "arc_size" will eventually approach "arc_c" (before "arc_meta_used" reaches "arc_meta_limit"); at that point the data buffers will be evicted, which releases the hold on the dnode, which releases the hold on the dnode object's dbuf, which allows that buffer to be evicted from the arc in preference to more "useful" metadata. So, to side step the issue, we simply need to ensure "arc_size" reaches "arc_c" before "arc_meta_used" reaches "arc_meta_limit". In order to pick a proper limit, we have to do some math. To make things a little easier to follow, it is assumed that there will only be a single data buffer per file (which is probably always the case for "small" files anyways). Based on the current internals of the arc, if N files residing in the dbuf cache all pin a single dnode buffer (i.e. their dnodes all share the same physical dnode object block), then the following amount of "arc_meta_used" space will be consumed: - 16K for the dnode object's block - [ 16384 bytes] - N * sizeof(dnode_t) -------------- [ N * 928 bytes] - (N + 1) * sizeof(arc_buf_t) ------ [(N + 1) * 72 bytes] - (N + 1) * sizeof(arc_buf_hdr_t) -- [(N + 1) * 264 bytes] - (N + 1) * sizeof(dmu_buf_impl_t) - [(N + 1) * 280 bytes] To simplify, these N files will pin the following amount of "arc_meta_used" space as unevictable: Pinned "arc_meta_used" bytes = 16384 + N * 928 + (N + 1) * (72 + 264 + 280) Pinned "arc_meta_used" bytes = 17000 + N * 1544 This pinned space is regardless of the size of the files, and is only dependent on the number of pinned dnodes sharing a physical block (i.e. N). For example, 32 512b files sharing a single dnode object block would consume the same "arc_meta_used" space as 32 4K files sharing a single dnode object block. Now, given a files size of S, we can determine the total amount of space that will be consumed in the arc: Total = 17000 + N * 1544 + S * N ^^^^^^^^^^^^^^^^ ^^^^^ metadata data So, given these formulas, we can generate a table which states the ratio of pinned metadata to total arc (meta + data) using different values of N (number of pinned dnodes per pinned physical dnode block) and S (size of the file). File Sizes (S) | 512 | 1024 | 2048 | 4096 | 8192 | 16384 | ---+----------+----------+----------+----------+----------+----------+ 1 | 0.973132 | 0.947670 | 0.900544 | 0.819081 | 0.693597 | 0.530921 | 2 | 0.951497 | 0.907481 | 0.830632 | 0.710325 | 0.550779 | 0.380051 | N 4 | 0.918807 | 0.849809 | 0.738842 | 0.585844 | 0.414271 | 0.261250 | 8 | 0.877541 | 0.781803 | 0.641770 | 0.472505 | 0.309333 | 0.182965 | 16 | 0.835819 | 0.717945 | 0.559996 | 0.388885 | 0.241376 | 0.137253 | 32 | 0.802106 | 0.669597 | 0.503304 | 0.336277 | 0.202123 | 0.112423 | As you can see, if we wanted to support the absolute worst case of 1 dnode per physical dnode block and 512b files, we would have to set the "arc_meta_limit" to something greater than 97.3132% of "arc_c_max". At that point, it essentially defeats the purpose of having an "arc_meta_limit" at all. This patch changes the default value of "arc_meta_limit" to be 75% of "arc_c_max", which should be good enough for "most" workloads (I think). Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2014-02-04 00:21:51 +04:00
/* limit meta-data to 3/4 of the arc capacity */
arc_meta_limit = (3 * arc_c_max) / 4;
arc_meta_max = 0;
2008-11-20 23:01:55 +03:00
/* Allow the tunable to override if it is reasonable */
if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
arc_meta_limit = zfs_arc_meta_limit;
/* if kmem_flags are set, lets try to use less memory */
if (kmem_debugging())
arc_c = arc_c / 2;
if (arc_c < arc_c_min)
arc_c = arc_c_min;
arc_anon = &ARC_anon;
arc_mru = &ARC_mru;
arc_mru_ghost = &ARC_mru_ghost;
arc_mfu = &ARC_mfu;
arc_mfu_ghost = &ARC_mfu_ghost;
arc_l2c_only = &ARC_l2c_only;
arc_size = 0;
mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
arc_anon->arcs_state = ARC_STATE_ANON;
arc_mru->arcs_state = ARC_STATE_MRU;
arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
arc_mfu->arcs_state = ARC_STATE_MFU;
arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
2008-11-20 23:01:55 +03:00
buf_init();
arc_thread_exit = 0;
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
list_create(&arc_prune_list, sizeof (arc_prune_t),
offsetof(arc_prune_t, p_node));
2008-11-20 23:01:55 +03:00
arc_eviction_list = NULL;
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
2008-11-20 23:01:55 +03:00
mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
if (arc_ksp != NULL) {
arc_ksp->ks_data = &arc_stats;
arc_ksp->ks_update = arc_kstat_update;
2008-11-20 23:01:55 +03:00
kstat_install(arc_ksp);
}
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
(void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
2008-11-20 23:01:55 +03:00
TS_RUN, minclsyspri);
arc_dead = FALSE;
arc_warm = B_FALSE;
2008-11-20 23:01:55 +03:00
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 07:01:20 +04:00
/*
* Calculate maximum amount of dirty data per pool.
*
* If it has been set by a module parameter, take that.
* Otherwise, use a percentage of physical memory defined by
* zfs_dirty_data_max_percent (default 10%) with a cap at
* zfs_dirty_data_max_max (default 25% of physical memory).
*/
if (zfs_dirty_data_max_max == 0)
zfs_dirty_data_max_max = physmem * PAGESIZE *
zfs_dirty_data_max_max_percent / 100;
if (zfs_dirty_data_max == 0) {
zfs_dirty_data_max = physmem * PAGESIZE *
zfs_dirty_data_max_percent / 100;
zfs_dirty_data_max = MIN(zfs_dirty_data_max,
zfs_dirty_data_max_max);
}
2008-11-20 23:01:55 +03:00
}
void
arc_fini(void)
{
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
arc_prune_t *p;
2008-11-20 23:01:55 +03:00
mutex_enter(&arc_reclaim_thr_lock);
#ifdef _KERNEL
spl_unregister_shrinker(&arc_shrinker);
#endif /* _KERNEL */
2008-11-20 23:01:55 +03:00
arc_thread_exit = 1;
while (arc_thread_exit != 0)
cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
mutex_exit(&arc_reclaim_thr_lock);
arc_flush(NULL);
arc_dead = TRUE;
if (arc_ksp != NULL) {
kstat_delete(arc_ksp);
arc_ksp = NULL;
}
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
mutex_enter(&arc_prune_mtx);
while ((p = list_head(&arc_prune_list)) != NULL) {
list_remove(&arc_prune_list, p);
refcount_remove(&p->p_refcnt, &arc_prune_list);
refcount_destroy(&p->p_refcnt);
kmem_free(p, sizeof (*p));
}
mutex_exit(&arc_prune_mtx);
list_destroy(&arc_prune_list);
mutex_destroy(&arc_prune_mtx);
2008-11-20 23:01:55 +03:00
mutex_destroy(&arc_eviction_mtx);
mutex_destroy(&arc_reclaim_thr_lock);
cv_destroy(&arc_reclaim_thr_cv);
list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
mutex_destroy(&arc_anon->arcs_mtx);
mutex_destroy(&arc_mru->arcs_mtx);
mutex_destroy(&arc_mru_ghost->arcs_mtx);
mutex_destroy(&arc_mfu->arcs_mtx);
mutex_destroy(&arc_mfu_ghost->arcs_mtx);
2009-01-16 00:59:39 +03:00
mutex_destroy(&arc_l2c_only->arcs_mtx);
2008-11-20 23:01:55 +03:00
buf_fini();
2009-07-03 02:44:48 +04:00
ASSERT(arc_loaned_bytes == 0);
2008-11-20 23:01:55 +03:00
}
/*
* Level 2 ARC
*
* The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
* It uses dedicated storage devices to hold cached data, which are populated
* using large infrequent writes. The main role of this cache is to boost
* the performance of random read workloads. The intended L2ARC devices
* include short-stroked disks, solid state disks, and other media with
* substantially faster read latency than disk.
*
* +-----------------------+
* | ARC |
* +-----------------------+
* | ^ ^
* | | |
* l2arc_feed_thread() arc_read()
* | | |
* | l2arc read |
* V | |
* +---------------+ |
* | L2ARC | |
* +---------------+ |
* | ^ |
* l2arc_write() | |
* | | |
* V | |
* +-------+ +-------+
* | vdev | | vdev |
* | cache | | cache |
* +-------+ +-------+
* +=========+ .-----.
* : L2ARC : |-_____-|
* : devices : | Disks |
* +=========+ `-_____-'
*
* Read requests are satisfied from the following sources, in order:
*
* 1) ARC
* 2) vdev cache of L2ARC devices
* 3) L2ARC devices
* 4) vdev cache of disks
* 5) disks
*
* Some L2ARC device types exhibit extremely slow write performance.
* To accommodate for this there are some significant differences between
* the L2ARC and traditional cache design:
*
* 1. There is no eviction path from the ARC to the L2ARC. Evictions from
* the ARC behave as usual, freeing buffers and placing headers on ghost
* lists. The ARC does not send buffers to the L2ARC during eviction as
* this would add inflated write latencies for all ARC memory pressure.
*
* 2. The L2ARC attempts to cache data from the ARC before it is evicted.
* It does this by periodically scanning buffers from the eviction-end of
* the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
* not already there. It scans until a headroom of buffers is satisfied,
* which itself is a buffer for ARC eviction. If a compressible buffer is
* found during scanning and selected for writing to an L2ARC device, we
* temporarily boost scanning headroom during the next scan cycle to make
* sure we adapt to compression effects (which might significantly reduce
* the data volume we write to L2ARC). The thread that does this is
2008-11-20 23:01:55 +03:00
* l2arc_feed_thread(), illustrated below; example sizes are included to
* provide a better sense of ratio than this diagram:
*
* head --> tail
* +---------------------+----------+
* ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
* +---------------------+----------+ | o L2ARC eligible
* ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
* +---------------------+----------+ |
* 15.9 Gbytes ^ 32 Mbytes |
* headroom |
* l2arc_feed_thread()
* |
* l2arc write hand <--[oooo]--'
* | 8 Mbyte
* | write max
* V
* +==============================+
* L2ARC dev |####|#|###|###| |####| ... |
* +==============================+
* 32 Gbytes
*
* 3. If an ARC buffer is copied to the L2ARC but then hit instead of
* evicted, then the L2ARC has cached a buffer much sooner than it probably
* needed to, potentially wasting L2ARC device bandwidth and storage. It is
* safe to say that this is an uncommon case, since buffers at the end of
* the ARC lists have moved there due to inactivity.
*
* 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
* then the L2ARC simply misses copying some buffers. This serves as a
* pressure valve to prevent heavy read workloads from both stalling the ARC
* with waits and clogging the L2ARC with writes. This also helps prevent
* the potential for the L2ARC to churn if it attempts to cache content too
* quickly, such as during backups of the entire pool.
*
* 5. After system boot and before the ARC has filled main memory, there are
* no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
* lists can remain mostly static. Instead of searching from tail of these
* lists as pictured, the l2arc_feed_thread() will search from the list heads
* for eligible buffers, greatly increasing its chance of finding them.
*
* The L2ARC device write speed is also boosted during this time so that
* the L2ARC warms up faster. Since there have been no ARC evictions yet,
* there are no L2ARC reads, and no fear of degrading read performance
* through increased writes.
*
* 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
2008-11-20 23:01:55 +03:00
* the vdev queue can aggregate them into larger and fewer writes. Each
* device is written to in a rotor fashion, sweeping writes through
* available space then repeating.
*
* 7. The L2ARC does not store dirty content. It never needs to flush
2008-11-20 23:01:55 +03:00
* write buffers back to disk based storage.
*
* 8. If an ARC buffer is written (and dirtied) which also exists in the
2008-11-20 23:01:55 +03:00
* L2ARC, the now stale L2ARC buffer is immediately dropped.
*
* The performance of the L2ARC can be tweaked by a number of tunables, which
* may be necessary for different workloads:
*
* l2arc_write_max max write bytes per interval
* l2arc_write_boost extra write bytes during device warmup
2008-11-20 23:01:55 +03:00
* l2arc_noprefetch skip caching prefetched buffers
* l2arc_nocompress skip compressing buffers
2008-11-20 23:01:55 +03:00
* l2arc_headroom number of max device writes to precache
* l2arc_headroom_boost when we find compressed buffers during ARC
* scanning, we multiply headroom by this
* percentage factor for the next scan cycle,
* since more compressed buffers are likely to
* be present
2008-11-20 23:01:55 +03:00
* l2arc_feed_secs seconds between L2ARC writing
*
* Tunables may be removed or added as future performance improvements are
* integrated, and also may become zpool properties.
2009-02-18 23:51:31 +03:00
*
* There are three key functions that control how the L2ARC warms up:
*
* l2arc_write_eligible() check if a buffer is eligible to cache
* l2arc_write_size() calculate how much to write
* l2arc_write_interval() calculate sleep delay between writes
*
* These three functions determine what to write, how much, and how quickly
* to send writes.
2008-11-20 23:01:55 +03:00
*/
2009-02-18 23:51:31 +03:00
static boolean_t
l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
{
/*
* A buffer is *not* eligible for the L2ARC if it:
* 1. belongs to a different spa.
* 2. is already cached on the L2ARC.
* 3. has an I/O in progress (it may be an incomplete read).
* 4. is flagged not eligible (zfs property).
2009-02-18 23:51:31 +03:00
*/
if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
2009-02-18 23:51:31 +03:00
HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
return (B_FALSE);
return (B_TRUE);
}
static uint64_t
l2arc_write_size(void)
2009-02-18 23:51:31 +03:00
{
uint64_t size;
/*
* Make sure our globals have meaningful values in case the user
* altered them.
*/
size = l2arc_write_max;
if (size == 0) {
cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
"be greater than zero, resetting it to the default (%d)",
L2ARC_WRITE_SIZE);
size = l2arc_write_max = L2ARC_WRITE_SIZE;
}
2009-02-18 23:51:31 +03:00
if (arc_warm == B_FALSE)
size += l2arc_write_boost;
2009-02-18 23:51:31 +03:00
return (size);
}
static clock_t
l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
{
clock_t interval, next, now;
2009-02-18 23:51:31 +03:00
/*
* If the ARC lists are busy, increase our write rate; if the
* lists are stale, idle back. This is achieved by checking
* how much we previously wrote - if it was more than half of
* what we wanted, schedule the next write much sooner.
*/
if (l2arc_feed_again && wrote > (wanted / 2))
interval = (hz * l2arc_feed_min_ms) / 1000;
else
interval = hz * l2arc_feed_secs;
now = ddi_get_lbolt();
next = MAX(now, MIN(now + interval, began + interval));
2009-02-18 23:51:31 +03:00
return (next);
}
2008-11-20 23:01:55 +03:00
static void
l2arc_hdr_stat_add(void)
{
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE);
2008-11-20 23:01:55 +03:00
ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
}
static void
l2arc_hdr_stat_remove(void)
{
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE);
2008-11-20 23:01:55 +03:00
ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
}
/*
* Cycle through L2ARC devices. This is how L2ARC load balances.
* If a device is returned, this also returns holding the spa config lock.
2008-11-20 23:01:55 +03:00
*/
static l2arc_dev_t *
l2arc_dev_get_next(void)
{
l2arc_dev_t *first, *next = NULL;
2008-11-20 23:01:55 +03:00
/*
* Lock out the removal of spas (spa_namespace_lock), then removal
* of cache devices (l2arc_dev_mtx). Once a device has been selected,
* both locks will be dropped and a spa config lock held instead.
*/
mutex_enter(&spa_namespace_lock);
mutex_enter(&l2arc_dev_mtx);
/* if there are no vdevs, there is nothing to do */
if (l2arc_ndev == 0)
goto out;
first = NULL;
next = l2arc_dev_last;
do {
/* loop around the list looking for a non-faulted vdev */
if (next == NULL) {
2008-11-20 23:01:55 +03:00
next = list_head(l2arc_dev_list);
} else {
next = list_next(l2arc_dev_list, next);
if (next == NULL)
next = list_head(l2arc_dev_list);
}
/* if we have come back to the start, bail out */
if (first == NULL)
first = next;
else if (next == first)
break;
} while (vdev_is_dead(next->l2ad_vdev));
/* if we were unable to find any usable vdevs, return NULL */
if (vdev_is_dead(next->l2ad_vdev))
next = NULL;
2008-11-20 23:01:55 +03:00
l2arc_dev_last = next;
out:
mutex_exit(&l2arc_dev_mtx);
/*
* Grab the config lock to prevent the 'next' device from being
* removed while we are writing to it.
*/
if (next != NULL)
spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
mutex_exit(&spa_namespace_lock);
2008-11-20 23:01:55 +03:00
return (next);
}
/*
* Free buffers that were tagged for destruction.
*/
static void
l2arc_do_free_on_write(void)
{
list_t *buflist;
l2arc_data_free_t *df, *df_prev;
mutex_enter(&l2arc_free_on_write_mtx);
buflist = l2arc_free_on_write;
for (df = list_tail(buflist); df; df = df_prev) {
df_prev = list_prev(buflist, df);
ASSERT(df->l2df_data != NULL);
ASSERT(df->l2df_func != NULL);
df->l2df_func(df->l2df_data, df->l2df_size);
list_remove(buflist, df);
kmem_free(df, sizeof (l2arc_data_free_t));
}
mutex_exit(&l2arc_free_on_write_mtx);
}
2008-11-20 23:01:55 +03:00
/*
* A write to a cache device has completed. Update all headers to allow
* reads from these buffers to begin.
*/
static void
l2arc_write_done(zio_t *zio)
{
l2arc_write_callback_t *cb;
l2arc_dev_t *dev;
list_t *buflist;
arc_buf_hdr_t *head, *ab, *ab_prev;
l2arc_buf_hdr_t *abl2;
2008-11-20 23:01:55 +03:00
kmutex_t *hash_lock;
int64_t bytes_dropped = 0;
2008-11-20 23:01:55 +03:00
cb = zio->io_private;
ASSERT(cb != NULL);
dev = cb->l2wcb_dev;
ASSERT(dev != NULL);
head = cb->l2wcb_head;
ASSERT(head != NULL);
buflist = dev->l2ad_buflist;
ASSERT(buflist != NULL);
DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
l2arc_write_callback_t *, cb);
if (zio->io_error != 0)
ARCSTAT_BUMP(arcstat_l2_writes_error);
mutex_enter(&l2arc_buflist_mtx);
/*
* All writes completed, or an error was hit.
*/
for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
ab_prev = list_prev(buflist, ab);
abl2 = ab->b_l2hdr;
/*
* Release the temporary compressed buffer as soon as possible.
*/
if (abl2->b_compress != ZIO_COMPRESS_OFF)
l2arc_release_cdata_buf(ab);
2008-11-20 23:01:55 +03:00
hash_lock = HDR_LOCK(ab);
if (!mutex_tryenter(hash_lock)) {
/*
* This buffer misses out. It may be in a stage
* of eviction. Its ARC_L2_WRITING flag will be
* left set, denying reads to this buffer.
*/
ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
continue;
}
if (zio->io_error != 0) {
/*
* Error - drop L2ARC entry.
2008-11-20 23:01:55 +03:00
*/
list_remove(buflist, ab);
ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
bytes_dropped += abl2->b_asize;
2008-11-20 23:01:55 +03:00
ab->b_l2hdr = NULL;
kmem_cache_free(l2arc_hdr_cache, abl2);
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
2008-11-20 23:01:55 +03:00
}
/*
* Allow ARC to begin reads to this L2ARC entry.
*/
ab->b_flags &= ~ARC_L2_WRITING;
mutex_exit(hash_lock);
}
atomic_inc_64(&l2arc_writes_done);
list_remove(buflist, head);
kmem_cache_free(hdr_cache, head);
mutex_exit(&l2arc_buflist_mtx);
vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
l2arc_do_free_on_write();
2008-11-20 23:01:55 +03:00
kmem_free(cb, sizeof (l2arc_write_callback_t));
}
/*
* A read to a cache device completed. Validate buffer contents before
* handing over to the regular ARC routines.
*/
static void
l2arc_read_done(zio_t *zio)
{
l2arc_read_callback_t *cb;
arc_buf_hdr_t *hdr;
arc_buf_t *buf;
kmutex_t *hash_lock;
int equal;
ASSERT(zio->io_vd != NULL);
ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
2008-11-20 23:01:55 +03:00
cb = zio->io_private;
ASSERT(cb != NULL);
buf = cb->l2rcb_buf;
ASSERT(buf != NULL);
hash_lock = HDR_LOCK(buf->b_hdr);
2008-11-20 23:01:55 +03:00
mutex_enter(hash_lock);
hdr = buf->b_hdr;
ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2008-11-20 23:01:55 +03:00
/*
* If the buffer was compressed, decompress it first.
*/
if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
ASSERT(zio->io_data != NULL);
2008-11-20 23:01:55 +03:00
/*
* Check this survived the L2ARC journey.
*/
equal = arc_cksum_equal(buf);
if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
mutex_exit(hash_lock);
zio->io_private = buf;
zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
2008-11-20 23:01:55 +03:00
arc_read_done(zio);
} else {
mutex_exit(hash_lock);
/*
* Buffer didn't survive caching. Increment stats and
* reissue to the original storage device.
*/
if (zio->io_error != 0) {
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_l2_io_error);
} else {
zio->io_error = SET_ERROR(EIO);
}
2008-11-20 23:01:55 +03:00
if (!equal)
ARCSTAT_BUMP(arcstat_l2_cksum_bad);
/*
* If there's no waiter, issue an async i/o to the primary
* storage now. If there *is* a waiter, the caller must
* issue the i/o in a context where it's OK to block.
2008-11-20 23:01:55 +03:00
*/
2009-02-18 23:51:31 +03:00
if (zio->io_waiter == NULL) {
zio_t *pio = zio_unique_parent(zio);
ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
buf->b_data, zio->io_size, arc_read_done, buf,
zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
2009-02-18 23:51:31 +03:00
}
2008-11-20 23:01:55 +03:00
}
kmem_free(cb, sizeof (l2arc_read_callback_t));
}
/*
* This is the list priority from which the L2ARC will search for pages to
* cache. This is used within loops (0..3) to cycle through lists in the
* desired order. This order can have a significant effect on cache
* performance.
*
* Currently the metadata lists are hit first, MFU then MRU, followed by
* the data lists. This function returns a locked list, and also returns
* the lock pointer.
*/
static list_t *
l2arc_list_locked(int list_num, kmutex_t **lock)
{
list_t *list = NULL;
2008-11-20 23:01:55 +03:00
ASSERT(list_num >= 0 && list_num <= 3);
switch (list_num) {
case 0:
list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
*lock = &arc_mfu->arcs_mtx;
break;
case 1:
list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
*lock = &arc_mru->arcs_mtx;
break;
case 2:
list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
*lock = &arc_mfu->arcs_mtx;
break;
case 3:
list = &arc_mru->arcs_list[ARC_BUFC_DATA];
*lock = &arc_mru->arcs_mtx;
break;
}
ASSERT(!(MUTEX_HELD(*lock)));
mutex_enter(*lock);
return (list);
}
/*
* Evict buffers from the device write hand to the distance specified in
* bytes. This distance may span populated buffers, it may span nothing.
* This is clearing a region on the L2ARC device ready for writing.
* If the 'all' boolean is set, every buffer is evicted.
*/
static void
l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
{
list_t *buflist;
l2arc_buf_hdr_t *abl2;
arc_buf_hdr_t *ab, *ab_prev;
kmutex_t *hash_lock;
uint64_t taddr;
int64_t bytes_evicted = 0;
2008-11-20 23:01:55 +03:00
buflist = dev->l2ad_buflist;
if (buflist == NULL)
return;
if (!all && dev->l2ad_first) {
/*
* This is the first sweep through the device. There is
* nothing to evict.
*/
return;
}
if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
2008-11-20 23:01:55 +03:00
/*
* When nearing the end of the device, evict to the end
* before the device write hand jumps to the start.
*/
taddr = dev->l2ad_end;
} else {
taddr = dev->l2ad_hand + distance;
}
DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
uint64_t, taddr, boolean_t, all);
top:
mutex_enter(&l2arc_buflist_mtx);
for (ab = list_tail(buflist); ab; ab = ab_prev) {
ab_prev = list_prev(buflist, ab);
hash_lock = HDR_LOCK(ab);
if (!mutex_tryenter(hash_lock)) {
/*
* Missed the hash lock. Retry.
*/
ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
mutex_exit(&l2arc_buflist_mtx);
mutex_enter(hash_lock);
mutex_exit(hash_lock);
goto top;
}
if (HDR_L2_WRITE_HEAD(ab)) {
/*
* We hit a write head node. Leave it for
* l2arc_write_done().
*/
list_remove(buflist, ab);
mutex_exit(hash_lock);
continue;
}
if (!all && ab->b_l2hdr != NULL &&
(ab->b_l2hdr->b_daddr > taddr ||
ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
/*
* We've evicted to the target address,
* or the end of the device.
*/
mutex_exit(hash_lock);
break;
}
if (HDR_FREE_IN_PROGRESS(ab)) {
/*
* Already on the path to destruction.
*/
mutex_exit(hash_lock);
continue;
}
if (ab->b_state == arc_l2c_only) {
ASSERT(!HDR_L2_READING(ab));
/*
* This doesn't exist in the ARC. Destroy.
* arc_hdr_destroy() will call list_remove()
* and decrement arcstat_l2_size.
*/
arc_change_state(arc_anon, ab, hash_lock);
arc_hdr_destroy(ab);
} else {
/*
* Invalidate issued or about to be issued
* reads, since we may be about to write
* over this location.
*/
if (HDR_L2_READING(ab)) {
ARCSTAT_BUMP(arcstat_l2_evict_reading);
ab->b_flags |= ARC_L2_EVICTED;
}
2008-11-20 23:01:55 +03:00
/*
* Tell ARC this no longer exists in L2ARC.
*/
if (ab->b_l2hdr != NULL) {
abl2 = ab->b_l2hdr;
ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
bytes_evicted += abl2->b_asize;
2008-11-20 23:01:55 +03:00
ab->b_l2hdr = NULL;
/*
* We are destroying l2hdr, so ensure that
* its compressed buffer, if any, is not leaked.
*/
ASSERT(abl2->b_tmp_cdata == NULL);
kmem_cache_free(l2arc_hdr_cache, abl2);
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
2008-11-20 23:01:55 +03:00
ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
}
list_remove(buflist, ab);
/*
* This may have been leftover after a
* failed write.
*/
ab->b_flags &= ~ARC_L2_WRITING;
}
mutex_exit(hash_lock);
}
mutex_exit(&l2arc_buflist_mtx);
vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0);
2008-11-20 23:01:55 +03:00
dev->l2ad_evict = taddr;
}
/*
* Find and write ARC buffers to the L2ARC device.
*
* An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
* for reading until they have completed writing.
* The headroom_boost is an in-out parameter used to maintain headroom boost
* state between calls to this function.
*
* Returns the number of bytes actually written (which may be smaller than
* the delta by which the device hand has changed due to alignment).
2008-11-20 23:01:55 +03:00
*/
2009-02-18 23:51:31 +03:00
static uint64_t
l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
boolean_t *headroom_boost)
2008-11-20 23:01:55 +03:00
{
arc_buf_hdr_t *ab, *ab_prev, *head;
list_t *list;
uint64_t write_asize, write_psize, write_sz, headroom,
buf_compress_minsz;
2008-11-20 23:01:55 +03:00
void *buf_data;
kmutex_t *list_lock = NULL;
boolean_t full;
2008-11-20 23:01:55 +03:00
l2arc_write_callback_t *cb;
zio_t *pio, *wzio;
uint64_t guid = spa_load_guid(spa);
int try;
const boolean_t do_headroom_boost = *headroom_boost;
2008-11-20 23:01:55 +03:00
ASSERT(dev->l2ad_vdev != NULL);
/* Lower the flag now, we might want to raise it again later. */
*headroom_boost = B_FALSE;
2008-11-20 23:01:55 +03:00
pio = NULL;
write_sz = write_asize = write_psize = 0;
2008-11-20 23:01:55 +03:00
full = B_FALSE;
head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
head->b_flags |= ARC_L2_WRITE_HEAD;
/*
* We will want to try to compress buffers that are at least 2x the
* device sector size.
*/
buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
2008-11-20 23:01:55 +03:00
/*
* Copy buffers for L2ARC writing.
*/
mutex_enter(&l2arc_buflist_mtx);
for (try = 0; try <= 3; try++) {
uint64_t passed_sz = 0;
2008-11-20 23:01:55 +03:00
list = l2arc_list_locked(try, &list_lock);
/*
* L2ARC fast warmup.
*
* Until the ARC is warm and starts to evict, read from the
* head of the ARC lists rather than the tail.
*/
if (arc_warm == B_FALSE)
ab = list_head(list);
else
ab = list_tail(list);
headroom = target_sz * l2arc_headroom;
if (do_headroom_boost)
headroom = (headroom * l2arc_headroom_boost) / 100;
for (; ab; ab = ab_prev) {
l2arc_buf_hdr_t *l2hdr;
kmutex_t *hash_lock;
uint64_t buf_sz;
if (arc_warm == B_FALSE)
ab_prev = list_next(list, ab);
else
ab_prev = list_prev(list, ab);
2008-11-20 23:01:55 +03:00
hash_lock = HDR_LOCK(ab);
if (!mutex_tryenter(hash_lock)) {
2008-11-20 23:01:55 +03:00
/*
* Skip this buffer rather than waiting.
*/
continue;
}
passed_sz += ab->b_size;
if (passed_sz > headroom) {
/*
* Searched too far.
*/
mutex_exit(hash_lock);
break;
}
2009-02-18 23:51:31 +03:00
if (!l2arc_write_eligible(guid, ab)) {
2008-11-20 23:01:55 +03:00
mutex_exit(hash_lock);
continue;
}
if ((write_sz + ab->b_size) > target_sz) {
full = B_TRUE;
mutex_exit(hash_lock);
break;
}
if (pio == NULL) {
/*
* Insert a dummy header on the buflist so
* l2arc_write_done() can find where the
* write buffers begin without searching.
*/
list_insert_head(dev->l2ad_buflist, head);
Use KM_PUSHPAGE in l2arc_write_buffers There is potential for deadlock in the l2arc_feed thread if KM_PUSHPAGE is not used for the allocations made in l2arc_write_buffers. Specifically, if KM_PUSHPAGE is not used for these allocations, it is possible for reclaim to be triggered which can cause the l2arc_feed thread to deadlock itself on the ARC_mru mutex. An example of this is demonstrated in the following backtrace of the l2arc_feed thread: crash> bt 4123 PID: 4123 TASK: ffff88062f8c1500 CPU: 6 COMMAND: "l2arc_feed" 0 [ffff88062511d610] schedule at ffffffff814eeee0 1 [ffff88062511d6d8] __mutex_lock_slowpath at ffffffff814f057e 2 [ffff88062511d748] mutex_lock at ffffffff814f041b 3 [ffff88062511d768] arc_evict at ffffffffa05130ca [zfs] 4 [ffff88062511d858] arc_adjust at ffffffffa05139a9 [zfs] 5 [ffff88062511d878] arc_shrink at ffffffffa0513a95 [zfs] 6 [ffff88062511d898] arc_kmem_reap_now at ffffffffa0513be8 [zfs] 7 [ffff88062511d8c8] arc_shrinker_func at ffffffffa0513ccc [zfs] 8 [ffff88062511d8f8] shrink_slab at ffffffff8112a17a 9 [ffff88062511d958] do_try_to_free_pages at ffffffff8112bfdf 10 [ffff88062511d9e8] try_to_free_pages at ffffffff8112c3ed 11 [ffff88062511da98] __alloc_pages_nodemask at ffffffff8112431d 12 [ffff88062511dbb8] kmem_getpages at ffffffff8115e632 13 [ffff88062511dbe8] fallback_alloc at ffffffff8115f24a 14 [ffff88062511dc68] ____cache_alloc_node at ffffffff8115efc9 15 [ffff88062511dcc8] __kmalloc at ffffffff8115fbf9 16 [ffff88062511dd18] kmem_alloc_debug at ffffffffa047b8cb [spl] 17 [ffff88062511dda8] l2arc_feed_thread at ffffffffa0511e71 [zfs] 18 [ffff88062511dea8] thread_generic_wrapper at ffffffffa047d1a1 [spl] 19 [ffff88062511dee8] kthread at ffffffff81090a86 20 [ffff88062511df48] kernel_thread at ffffffff8100c14a Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-04-10 21:55:17 +04:00
cb = kmem_alloc(sizeof (l2arc_write_callback_t),
KM_SLEEP);
2008-11-20 23:01:55 +03:00
cb->l2wcb_dev = dev;
cb->l2wcb_head = head;
pio = zio_root(spa, l2arc_write_done, cb,
ZIO_FLAG_CANFAIL);
}
/*
* Create and add a new L2ARC header.
*/
l2hdr = kmem_cache_alloc(l2arc_hdr_cache, KM_SLEEP);
l2hdr->b_dev = dev;
l2hdr->b_daddr = 0;
Fix inaccurate arcstat_l2_hdr_size calculations Based on the comments in arc.c we know that buffers can exist both in arc and l2arc, under this circumstance both arc_buf_hdr_t and l2arc_buf_hdr_t will be allocated. However the current logic only cares for memory that l2arc_buf_hdr takes up when the buffer's state transfers from or to arc_l2c_only. This will cause obvious deviations for illumos's zfs version since the sizeof(l2arc_buf_hdr) is larger than ZOL's. We can implement the calcuation in the following simple way: 1. When allocate a l2arc_buf_hdr_t we add its memory consumption instantly and subtract it when we free or evict the l2arc buf. 2. According to l2arc_hdr_stat_add and l2arc_hdr_stat_remove, if the buffer only stays in l2arc we should also add the memory its arc_buf_hdr_t consumes, so we only need to add HDR_SIZE to arcstat_l2_hdr_size since we already concerned with L2HDR_SIZE in step 1 and the same for transfering arc bufs from l2arc only state. The testbox has 2 4-core Intel Xeon CPUs(2.13GHz), with 16GB memory and tests were set upped in the following way: 1. Fdisked a SATA disk into two partitions, one partition for zpool storage and the other one was used as the cache device. 2. Generated some files occupying 14GB altogether in the zpool prepared in step 1 using iozone. 3. Read them all using md5sum and watched the l2arc related statistics in /proc/spl/kstat/zfs/arcstats. After the reading ended the l2_hdr_size and l2_size were shown like this: l2_size 4 4403780608 l2_hdr_size 4 0 which was weird. 4. After applying this patch and reran step 1-3, the results were as following: l2_size 4 4306443264 l2_hdr_size 4 535600 these numbers made sense, on 64-bit systems the sizeof(l2arc_buf_hdr_t) is 16 bytes. Assue all blocks cached by l2arc are 128KB, so 535600/16*128*1024=4387635200, since not all blocks are equal-sized, the theoretical result will be a little bigger, as we can see. Since I'm familiar with systemtap instrumentation tool I used it to examine what had happened. The script looked like this: probe module("zfs").function("arc_chage_state") { if ($new_state == $arc_l2_only) printf("change arc buf to arc_l2_only\n") } It will print out some information each time we call funciton arc_chage_state if the argument new_state is arc_l2_only. I gathered the trace logs and found that none of the arc bufs ran into arc state arc_l2_only when the tests was running, this was the reason why l2_hdr_size in step 3 was 0. The arc bufs fell into arc_l2_only when the pool or the filesystem was offlined. Signed-off-by: Ying Zhu <casualfisher@gmail.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2013-06-22 16:35:18 +04:00
arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS);
2008-11-20 23:01:55 +03:00
ab->b_flags |= ARC_L2_WRITING;
/*
* Temporarily stash the data buffer in b_tmp_cdata.
* The subsequent write step will pick it up from
* there. This is because can't access ab->b_buf
* without holding the hash_lock, which we in turn
* can't access without holding the ARC list locks
* (which we want to avoid during compression/writing)
*/
l2hdr->b_compress = ZIO_COMPRESS_OFF;
l2hdr->b_asize = ab->b_size;
l2hdr->b_tmp_cdata = ab->b_buf->b_data;
l2hdr->b_hits = 0;
2008-11-20 23:01:55 +03:00
buf_sz = ab->b_size;
ab->b_l2hdr = l2hdr;
list_insert_head(dev->l2ad_buflist, ab);
2008-11-20 23:01:55 +03:00
/*
* Compute and store the buffer cksum before
* writing. On debug the cksum is verified first.
*/
arc_cksum_verify(ab->b_buf);
arc_cksum_compute(ab->b_buf, B_TRUE);
mutex_exit(hash_lock);
write_sz += buf_sz;
}
mutex_exit(list_lock);
if (full == B_TRUE)
break;
}
/* No buffers selected for writing? */
if (pio == NULL) {
ASSERT0(write_sz);
mutex_exit(&l2arc_buflist_mtx);
kmem_cache_free(hdr_cache, head);
return (0);
}
/*
* Now start writing the buffers. We're starting at the write head
* and work backwards, retracing the course of the buffer selector
* loop above.
*/
for (ab = list_prev(dev->l2ad_buflist, head); ab;
ab = list_prev(dev->l2ad_buflist, ab)) {
l2arc_buf_hdr_t *l2hdr;
uint64_t buf_sz;
/*
* We shouldn't need to lock the buffer here, since we flagged
* it as ARC_L2_WRITING in the previous step, but we must take
* care to only access its L2 cache parameters. In particular,
* ab->b_buf may be invalid by now due to ARC eviction.
*/
l2hdr = ab->b_l2hdr;
l2hdr->b_daddr = dev->l2ad_hand;
if (!l2arc_nocompress && (ab->b_flags & ARC_L2COMPRESS) &&
l2hdr->b_asize >= buf_compress_minsz) {
if (l2arc_compress_buf(l2hdr)) {
/*
* If compression succeeded, enable headroom
* boost on the next scan cycle.
*/
*headroom_boost = B_TRUE;
}
}
/*
* Pick up the buffer data we had previously stashed away
* (and now potentially also compressed).
*/
buf_data = l2hdr->b_tmp_cdata;
buf_sz = l2hdr->b_asize;
/*
* If the data has not been compressed, then clear b_tmp_cdata
* to make sure that it points only to a temporary compression
* buffer.
*/
if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress))
l2hdr->b_tmp_cdata = NULL;
/* Compression may have squashed the buffer to zero length. */
if (buf_sz != 0) {
uint64_t buf_p_sz;
2008-11-20 23:01:55 +03:00
wzio = zio_write_phys(pio, dev->l2ad_vdev,
dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
ZIO_FLAG_CANFAIL, B_FALSE);
DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
zio_t *, wzio);
(void) zio_nowait(wzio);
write_asize += buf_sz;
/*
* Keep the clock hand suitably device-aligned.
*/
buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
write_psize += buf_p_sz;
dev->l2ad_hand += buf_p_sz;
2008-11-20 23:01:55 +03:00
}
}
mutex_exit(&l2arc_buflist_mtx);
2008-11-20 23:01:55 +03:00
ASSERT3U(write_asize, <=, target_sz);
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_l2_writes_sent);
ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
2008-11-20 23:01:55 +03:00
ARCSTAT_INCR(arcstat_l2_size, write_sz);
ARCSTAT_INCR(arcstat_l2_asize, write_asize);
vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
2008-11-20 23:01:55 +03:00
/*
* Bump device hand to the device start if it is approaching the end.
* l2arc_evict() will already have evicted ahead for this case.
*/
if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
2008-11-20 23:01:55 +03:00
dev->l2ad_hand = dev->l2ad_start;
dev->l2ad_evict = dev->l2ad_start;
dev->l2ad_first = B_FALSE;
}
2009-02-18 23:51:31 +03:00
dev->l2ad_writing = B_TRUE;
2008-11-20 23:01:55 +03:00
(void) zio_wait(pio);
2009-02-18 23:51:31 +03:00
dev->l2ad_writing = B_FALSE;
return (write_asize);
}
/*
* Compresses an L2ARC buffer.
* The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
* size in l2hdr->b_asize. This routine tries to compress the data and
* depending on the compression result there are three possible outcomes:
* *) The buffer was incompressible. The original l2hdr contents were left
* untouched and are ready for writing to an L2 device.
* *) The buffer was all-zeros, so there is no need to write it to an L2
* device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
* set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
* *) Compression succeeded and b_tmp_cdata was replaced with a temporary
* data buffer which holds the compressed data to be written, and b_asize
* tells us how much data there is. b_compress is set to the appropriate
* compression algorithm. Once writing is done, invoke
* l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
*
* Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
* buffer was incompressible).
*/
static boolean_t
l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
{
void *cdata;
size_t csize, len, rounded;
ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
ASSERT(l2hdr->b_tmp_cdata != NULL);
len = l2hdr->b_asize;
cdata = zio_data_buf_alloc(len);
csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
cdata, l2hdr->b_asize);
rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
if (rounded > csize) {
bzero((char *)cdata + csize, rounded - csize);
csize = rounded;
}
if (csize == 0) {
/* zero block, indicate that there's nothing to write */
zio_data_buf_free(cdata, len);
l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
l2hdr->b_asize = 0;
l2hdr->b_tmp_cdata = NULL;
ARCSTAT_BUMP(arcstat_l2_compress_zeros);
return (B_TRUE);
} else if (csize > 0 && csize < len) {
/*
* Compression succeeded, we'll keep the cdata around for
* writing and release it afterwards.
*/
l2hdr->b_compress = ZIO_COMPRESS_LZ4;
l2hdr->b_asize = csize;
l2hdr->b_tmp_cdata = cdata;
ARCSTAT_BUMP(arcstat_l2_compress_successes);
return (B_TRUE);
} else {
/*
* Compression failed, release the compressed buffer.
* l2hdr will be left unmodified.
*/
zio_data_buf_free(cdata, len);
ARCSTAT_BUMP(arcstat_l2_compress_failures);
return (B_FALSE);
}
}
/*
* Decompresses a zio read back from an l2arc device. On success, the
* underlying zio's io_data buffer is overwritten by the uncompressed
* version. On decompression error (corrupt compressed stream), the
* zio->io_error value is set to signal an I/O error.
*
* Please note that the compressed data stream is not checksummed, so
* if the underlying device is experiencing data corruption, we may feed
* corrupt data to the decompressor, so the decompressor needs to be
* able to handle this situation (LZ4 does).
*/
static void
l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
{
uint64_t csize;
void *cdata;
ASSERT(L2ARC_IS_VALID_COMPRESS(c));
if (zio->io_error != 0) {
/*
* An io error has occured, just restore the original io
* size in preparation for a main pool read.
*/
zio->io_orig_size = zio->io_size = hdr->b_size;
return;
}
if (c == ZIO_COMPRESS_EMPTY) {
/*
* An empty buffer results in a null zio, which means we
* need to fill its io_data after we're done restoring the
* buffer's contents.
*/
ASSERT(hdr->b_buf != NULL);
bzero(hdr->b_buf->b_data, hdr->b_size);
zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
} else {
ASSERT(zio->io_data != NULL);
/*
* We copy the compressed data from the start of the arc buffer
* (the zio_read will have pulled in only what we need, the
* rest is garbage which we will overwrite at decompression)
* and then decompress back to the ARC data buffer. This way we
* can minimize copying by simply decompressing back over the
* original compressed data (rather than decompressing to an
* aux buffer and then copying back the uncompressed buffer,
* which is likely to be much larger).
*/
csize = zio->io_size;
cdata = zio_data_buf_alloc(csize);
bcopy(zio->io_data, cdata, csize);
if (zio_decompress_data(c, cdata, zio->io_data, csize,
hdr->b_size) != 0)
zio->io_error = SET_ERROR(EIO);
zio_data_buf_free(cdata, csize);
}
/* Restore the expected uncompressed IO size. */
zio->io_orig_size = zio->io_size = hdr->b_size;
}
/*
* Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
* This buffer serves as a temporary holder of compressed data while
* the buffer entry is being written to an l2arc device. Once that is
* done, we can dispose of it.
*/
static void
l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
{
l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress));
if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) {
/*
* If the data was compressed, then we've allocated a
* temporary buffer for it, so now we need to release it.
*/
ASSERT(l2hdr->b_tmp_cdata != NULL);
zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
l2hdr->b_tmp_cdata = NULL;
} else {
ASSERT(l2hdr->b_tmp_cdata == NULL);
}
2008-11-20 23:01:55 +03:00
}
/*
* This thread feeds the L2ARC at regular intervals. This is the beating
* heart of the L2ARC.
*/
static void
l2arc_feed_thread(void)
{
callb_cpr_t cpr;
l2arc_dev_t *dev;
spa_t *spa;
2009-02-18 23:51:31 +03:00
uint64_t size, wrote;
clock_t begin, next = ddi_get_lbolt();
boolean_t headroom_boost = B_FALSE;
2008-11-20 23:01:55 +03:00
CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
mutex_enter(&l2arc_feed_thr_lock);
while (l2arc_thread_exit == 0) {
CALLB_CPR_SAFE_BEGIN(&cpr);
(void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
&l2arc_feed_thr_lock, next);
2008-11-20 23:01:55 +03:00
CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
next = ddi_get_lbolt() + hz;
2008-11-20 23:01:55 +03:00
/*
* Quick check for L2ARC devices.
2008-11-20 23:01:55 +03:00
*/
mutex_enter(&l2arc_dev_mtx);
if (l2arc_ndev == 0) {
mutex_exit(&l2arc_dev_mtx);
continue;
}
mutex_exit(&l2arc_dev_mtx);
begin = ddi_get_lbolt();
2008-11-20 23:01:55 +03:00
/*
* This selects the next l2arc device to write to, and in
* doing so the next spa to feed from: dev->l2ad_spa. This
* will return NULL if there are now no l2arc devices or if
* they are all faulted.
*
* If a device is returned, its spa's config lock is also
* held to prevent device removal. l2arc_dev_get_next()
* will grab and release l2arc_dev_mtx.
2008-11-20 23:01:55 +03:00
*/
if ((dev = l2arc_dev_get_next()) == NULL)
2008-11-20 23:01:55 +03:00
continue;
spa = dev->l2ad_spa;
ASSERT(spa != NULL);
2008-11-20 23:01:55 +03:00
/*
* If the pool is read-only then force the feed thread to
* sleep a little longer.
*/
if (!spa_writeable(spa)) {
next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
spa_config_exit(spa, SCL_L2ARC, dev);
continue;
}
2008-11-20 23:01:55 +03:00
/*
* Avoid contributing to memory pressure.
2008-11-20 23:01:55 +03:00
*/
Integrate ARC more tightly with Linux Under Solaris the ARC was designed to stay one step ahead of the VM subsystem. It would attempt to recognize low memory situtions before they occured and evict data from the cache. It would also make assessments about if there was enough free memory to perform a specific operation. This was all possible because Solaris exposes a fairly decent view of the memory state of the system to other kernel threads. Linux on the other hand does not make this information easily available. To avoid extensive modifications to the ARC the SPL attempts to provide these same interfaces. While this works it is not ideal and problems can arise when the ARC and Linux have different ideas about when your out of memory. This has manifested itself in the past as a spinning arc_reclaim_thread. This patch abandons the emulated Solaris interfaces in favor of the prefered Linux interface. That means moving the bulk of the memory reclaim logic out of the arc_reclaim_thread and in to the evict driven shrinker callback. The Linux VM will call this function when it needs memory. The ARC is then responsible for attempting to free the requested amount of memory if possible. Several interfaces have been modified to accomidate this approach, however the basic user space implementation remains the same. The following changes almost exclusively just apply to the kernel implementation. * Removed the hdr_recl() reclaim callback which is redundant with the broader arc_shrinker_func(). * Reduced arc_grow_retry to 5 seconds from 60. This is now used internally in the ARC with arc_no_grow to indicate that direct reclaim was recently performed. This typically indicates a rapid change in memory demands which the kswapd threads were unable to keep ahead of. As long as direct reclaim is happening once every 5 seconds arc growth will be paused to avoid further contributing to the existing memory pressure. The more common indirect reclaim paths will not set arc_no_grow. * arc_shrink() has been extended to take the number of bytes by which arc_c should be reduced. This allows for a more granual reduction of the arc target. Since the kernel provides a reclaim value to the arc_shrinker_func() this value is used instead of 1<<arc_shrink_shift. * arc_reclaim_needed() has been removed. It was used to determine if the system was under memory pressure and relied extensively on Solaris specific VM interfaces. In most case the new code just checks arc_no_grow which indicates that within the last arc_grow_retry seconds direct memory reclaim occurred. * arc_memory_throttle() has been updated to always include the amount of evictable memory (arc and page cache) in its free space calculations. This space is largely available in most call paths due to direct memory reclaim. * The Solaris pageout code was also removed to avoid confusion. It has always been disabled due to proc_pageout being defined as NULL in the Linux port. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2012-03-14 01:29:16 +04:00
if (arc_no_grow) {
ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
spa_config_exit(spa, SCL_L2ARC, dev);
2008-11-20 23:01:55 +03:00
continue;
}
2008-11-20 23:01:55 +03:00
ARCSTAT_BUMP(arcstat_l2_feeds);
size = l2arc_write_size();
2008-11-20 23:01:55 +03:00
/*
* Evict L2ARC buffers that will be overwritten.
*/
l2arc_evict(dev, size, B_FALSE);
2008-11-20 23:01:55 +03:00
/*
* Write ARC buffers.
*/
wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
2009-02-18 23:51:31 +03:00
/*
* Calculate interval between writes.
*/
next = l2arc_write_interval(begin, size, wrote);
spa_config_exit(spa, SCL_L2ARC, dev);
2008-11-20 23:01:55 +03:00
}
l2arc_thread_exit = 0;
cv_broadcast(&l2arc_feed_thr_cv);
CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
thread_exit();
}
boolean_t
l2arc_vdev_present(vdev_t *vd)
{
l2arc_dev_t *dev;
mutex_enter(&l2arc_dev_mtx);
for (dev = list_head(l2arc_dev_list); dev != NULL;
dev = list_next(l2arc_dev_list, dev)) {
if (dev->l2ad_vdev == vd)
break;
}
mutex_exit(&l2arc_dev_mtx);
return (dev != NULL);
}
2008-11-20 23:01:55 +03:00
/*
* Add a vdev for use by the L2ARC. By this point the spa has already
* validated the vdev and opened it.
*/
void
2009-07-03 02:44:48 +04:00
l2arc_add_vdev(spa_t *spa, vdev_t *vd)
2008-11-20 23:01:55 +03:00
{
l2arc_dev_t *adddev;
ASSERT(!l2arc_vdev_present(vd));
2008-11-20 23:01:55 +03:00
/*
* Create a new l2arc device entry.
*/
adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
adddev->l2ad_spa = spa;
adddev->l2ad_vdev = vd;
2009-07-03 02:44:48 +04:00
adddev->l2ad_start = VDEV_LABEL_START_SIZE;
adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
2008-11-20 23:01:55 +03:00
adddev->l2ad_hand = adddev->l2ad_start;
adddev->l2ad_evict = adddev->l2ad_start;
adddev->l2ad_first = B_TRUE;
2009-02-18 23:51:31 +03:00
adddev->l2ad_writing = B_FALSE;
list_link_init(&adddev->l2ad_node);
2008-11-20 23:01:55 +03:00
/*
* This is a list of all ARC buffers that are still valid on the
* device.
*/
adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
offsetof(arc_buf_hdr_t, b_l2node));
vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
2008-11-20 23:01:55 +03:00
/*
* Add device to global list
*/
mutex_enter(&l2arc_dev_mtx);
list_insert_head(l2arc_dev_list, adddev);
atomic_inc_64(&l2arc_ndev);
mutex_exit(&l2arc_dev_mtx);
}
/*
* Remove a vdev from the L2ARC.
*/
void
l2arc_remove_vdev(vdev_t *vd)
{
l2arc_dev_t *dev, *nextdev, *remdev = NULL;
/*
* Find the device by vdev
*/
mutex_enter(&l2arc_dev_mtx);
for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
nextdev = list_next(l2arc_dev_list, dev);
if (vd == dev->l2ad_vdev) {
remdev = dev;
break;
}
}
ASSERT(remdev != NULL);
/*
* Remove device from global list
*/
list_remove(l2arc_dev_list, remdev);
l2arc_dev_last = NULL; /* may have been invalidated */
atomic_dec_64(&l2arc_ndev);
mutex_exit(&l2arc_dev_mtx);
2008-11-20 23:01:55 +03:00
/*
* Clear all buflists and ARC references. L2ARC device flush.
*/
l2arc_evict(remdev, 0, B_TRUE);
list_destroy(remdev->l2ad_buflist);
kmem_free(remdev->l2ad_buflist, sizeof (list_t));
kmem_free(remdev, sizeof (l2arc_dev_t));
}
void
l2arc_init(void)
2008-11-20 23:01:55 +03:00
{
l2arc_thread_exit = 0;
l2arc_ndev = 0;
l2arc_writes_sent = 0;
l2arc_writes_done = 0;
mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
l2arc_dev_list = &L2ARC_dev_list;
l2arc_free_on_write = &L2ARC_free_on_write;
list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
offsetof(l2arc_dev_t, l2ad_node));
list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
offsetof(l2arc_data_free_t, l2df_list_node));
}
void
l2arc_fini(void)
2008-11-20 23:01:55 +03:00
{
/*
* This is called from dmu_fini(), which is called from spa_fini();
* Because of this, we can assume that all l2arc devices have
* already been removed when the pools themselves were removed.
*/
l2arc_do_free_on_write();
2008-11-20 23:01:55 +03:00
mutex_destroy(&l2arc_feed_thr_lock);
cv_destroy(&l2arc_feed_thr_cv);
mutex_destroy(&l2arc_dev_mtx);
mutex_destroy(&l2arc_buflist_mtx);
mutex_destroy(&l2arc_free_on_write_mtx);
list_destroy(l2arc_dev_list);
list_destroy(l2arc_free_on_write);
}
void
l2arc_start(void)
{
2009-01-16 00:59:39 +03:00
if (!(spa_mode_global & FWRITE))
return;
(void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
TS_RUN, minclsyspri);
}
void
l2arc_stop(void)
{
2009-01-16 00:59:39 +03:00
if (!(spa_mode_global & FWRITE))
return;
mutex_enter(&l2arc_feed_thr_lock);
cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
l2arc_thread_exit = 1;
while (l2arc_thread_exit != 0)
cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
mutex_exit(&l2arc_feed_thr_lock);
}
#if defined(_KERNEL) && defined(HAVE_SPL)
EXPORT_SYMBOL(arc_buf_size);
EXPORT_SYMBOL(arc_write);
EXPORT_SYMBOL(arc_read);
EXPORT_SYMBOL(arc_buf_remove_ref);
EXPORT_SYMBOL(arc_buf_info);
EXPORT_SYMBOL(arc_getbuf_func);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
EXPORT_SYMBOL(arc_add_prune_callback);
EXPORT_SYMBOL(arc_remove_prune_callback);
module_param(zfs_arc_min, ulong, 0644);
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-04 02:09:28 +04:00
MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
module_param(zfs_arc_max, ulong, 0644);
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-04 02:09:28 +04:00
MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
module_param(zfs_arc_meta_limit, ulong, 0644);
MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
module_param(zfs_arc_meta_prune, int, 0644);
Linux 3.1 compat, super_block->s_shrink The Linux 3.1 kernel has introduced the concept of per-filesystem shrinkers which are directly assoicated with a super block. Prior to this change there was one shared global shrinker. The zfs code relied on being able to call the global shrinker when the arc_meta_limit was exceeded. This would cause the VFS to drop references on a fraction of the dentries in the dcache. The ARC could then safely reclaim the memory used by these entries and honor the arc_meta_limit. Unfortunately, when per-filesystem shrinkers were added the old interfaces were made unavailable. This change adds support to use the new per-filesystem shrinker interface so we can continue to honor the arc_meta_limit. The major benefit of the new interface is that we can now target only the zfs filesystem for dentry and inode pruning. Thus we can minimize any impact on the caching of other filesystems. In the context of making this change several other important issues related to managing the ARC were addressed, they include: * The dnlc_reduce_cache() function which was called by the ARC to drop dentries for the Posix layer was replaced with a generic zfs_prune_t callback. The ZPL layer now registers a callback to drop these dentries removing a layering violation which dates back to the Solaris code. This callback can also be used by other ARC consumers such as Lustre. arc_add_prune_callback() arc_remove_prune_callback() * The arc_reduce_dnlc_percent module option has been changed to arc_meta_prune for clarity. The dnlc functions are specific to Solaris's VFS and have already been largely eliminated already. The replacement tunable now represents the number of bytes the prune callback will request when invoked. * Less aggressively invoke the prune callback. We used to call this whenever we exceeded the arc_meta_limit however that's not strictly correct since it results in over zeleous reclaim of dentries and inodes. It is now only called once the arc_meta_limit is exceeded and every effort has been made to evict other data from the ARC cache. * More promptly manage exceeding the arc_meta_limit. When reading meta data in to the cache if a buffer was unable to be recycled notify the arc_reclaim thread to invoke the required prune. * Added arcstat_prune kstat which is incremented when the ARC is forced to request that a consumer prune its cache. Remember this will only occur when the ARC has no other choice. If it can evict buffers safely without invoking the prune callback it will. * This change is also expected to resolve the unexpect collapses of the ARC cache. This would occur because when exceeded just the arc_meta_limit reclaim presure would be excerted on the arc_c value via arc_shrink(). This effectively shrunk the entire cache when really we just needed to reclaim meta data. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #466 Closes #292
2011-12-23 00:20:43 +04:00
MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-04 02:09:28 +04:00
module_param(zfs_arc_grow_retry, int, 0644);
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-04 02:09:28 +04:00
MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
Disable aggressive arc_p growth by default For specific workloads consisting mainly of mfu data and new anon data buffers, the aggressive growth of arc_p found in the arc_get_data_buf() function can have detrimental effects on the mfu list size and ghost list hit rate. Running a workload consisting of two processes: * Process 1 is creating many small files * Process 2 is tar'ing a directory consisting of many small files I've seen arc_p and the mru grow to their maximum size, while the mru ghost list receives 100K times fewer hits than the mfu ghost list. Ideally, as the mfu ghost list receives hits, arc_p should be driven down and the size of the mfu should increase. Given the specific workload I was testing with, the mfu list size should grow to a point where almost no mfu ghost list hits would occur. Unfortunately, this does not happen because the newly dirtied anon buffers constancy drive arc_p to its maximum value and keep it there (effectively prioritizing the mru list and starving the mfu list down to a negligible size). The logic to increment arc_p from within the arc_get_data_buf() function was introduced many years ago in this upstream commit: commit 641fbdae3a027d12b3c3dcd18927ccafae6d58bc Author: maybee <none@none> Date: Wed Dec 20 15:46:12 2006 -0800 6505658 target MRU size (arc.p) needs to be adjusted more aggressively and since I don't fully understand the motivation for the change, I am reluctant to completely remove it. As a way to test out how it's removal might affect performance, I've disabled that code by default, but left it tunable via a module option. Thus, if its removal is found to be grossly detrimental for certain workloads, it can be re-enabled on the fly, without a code change. Signed-off-by: Prakash Surya <surya1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #2110
2013-12-11 21:40:13 +04:00
module_param(zfs_arc_p_aggressive_disable, int, 0644);
MODULE_PARM_DESC(zfs_arc_p_aggressive_disable, "disable aggressive arc_p grow");
module_param(zfs_arc_p_dampener_disable, int, 0644);
MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener");
module_param(zfs_arc_shrink_shift, int, 0644);
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-04 02:09:28 +04:00
MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
module_param(zfs_disable_dup_eviction, int, 0644);
MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction");
module_param(zfs_arc_average_blocksize, int, 0444);
MODULE_PARM_DESC(zfs_arc_average_blocksize, "Target average block size");
module_param(zfs_arc_memory_throttle_disable, int, 0644);
MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle");
module_param(zfs_arc_min_prefetch_lifespan, int, 0644);
MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block");
module_param(l2arc_write_max, ulong, 0644);
MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
module_param(l2arc_write_boost, ulong, 0644);
MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
module_param(l2arc_headroom, ulong, 0644);
MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
module_param(l2arc_headroom_boost, ulong, 0644);
MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier");
module_param(l2arc_feed_secs, ulong, 0644);
MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
module_param(l2arc_feed_min_ms, ulong, 0644);
MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
module_param(l2arc_noprefetch, int, 0644);
MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
module_param(l2arc_nocompress, int, 0644);
MODULE_PARM_DESC(l2arc_nocompress, "Skip compressing L2ARC buffers");
module_param(l2arc_feed_again, int, 0644);
MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
module_param(l2arc_norw, int, 0644);
MODULE_PARM_DESC(l2arc_norw, "No reads during writes");
#endif