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
|
|
|
|
|
*/
|
|
|
|
|
/*
|
2010-05-29 00:45:14 +04:00
|
|
|
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
|
2012-04-08 21:10:49 +04:00
|
|
|
* Copyright 2011 Nexenta Systems, Inc. All rights reserved.
|
2013-04-11 03:58:22 +04:00
|
|
|
* Copyright (c) 2013 by Delphix. All rights reserved.
|
2012-04-08 21:10:49 +04:00
|
|
|
*/
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
#include <sys/dmu.h>
|
|
|
|
|
#include <sys/dmu_impl.h>
|
|
|
|
|
#include <sys/dbuf.h>
|
|
|
|
|
#include <sys/dmu_tx.h>
|
|
|
|
|
#include <sys/dmu_objset.h>
|
|
|
|
|
#include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
|
|
|
|
|
#include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
|
|
|
|
|
#include <sys/dsl_pool.h>
|
|
|
|
|
#include <sys/zap_impl.h> /* for fzap_default_block_shift */
|
|
|
|
|
#include <sys/spa.h>
|
2010-05-29 00:45:14 +04:00
|
|
|
#include <sys/sa.h>
|
|
|
|
|
#include <sys/sa_impl.h>
|
2008-11-20 23:01:55 +03:00
|
|
|
#include <sys/zfs_context.h>
|
2010-05-29 00:45:14 +04:00
|
|
|
#include <sys/varargs.h>
|
2014-12-13 05:07:39 +03:00
|
|
|
#include <sys/trace_dmu.h>
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
|
|
|
|
|
uint64_t arg1, uint64_t arg2);
|
|
|
|
|
|
2012-01-20 22:58:57 +04:00
|
|
|
dmu_tx_stats_t dmu_tx_stats = {
|
|
|
|
|
{ "dmu_tx_assigned", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_delay", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_error", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_suspended", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_group", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
|
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
|
|
|
{ "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
|
|
|
|
|
{ "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
|
2012-01-20 22:58:57 +04:00
|
|
|
{ "dmu_tx_quota", KSTAT_DATA_UINT64 },
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
static kstat_t *dmu_tx_ksp;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
dmu_tx_t *
|
|
|
|
|
dmu_tx_create_dd(dsl_dir_t *dd)
|
|
|
|
|
{
|
2012-05-07 21:49:51 +04:00
|
|
|
dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_PUSHPAGE);
|
2008-11-20 23:01:55 +03:00
|
|
|
tx->tx_dir = dd;
|
2013-08-28 15:45:09 +04:00
|
|
|
if (dd != NULL)
|
2008-11-20 23:01:55 +03:00
|
|
|
tx->tx_pool = dd->dd_pool;
|
|
|
|
|
list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
|
|
|
|
|
offsetof(dmu_tx_hold_t, txh_node));
|
2010-05-29 00:45:14 +04:00
|
|
|
list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
|
|
|
|
|
offsetof(dmu_tx_callback_t, dcb_node));
|
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
|
|
|
tx->tx_start = gethrtime();
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-11-20 23:01:55 +03:00
|
|
|
refcount_create(&tx->tx_space_written);
|
|
|
|
|
refcount_create(&tx->tx_space_freed);
|
|
|
|
|
#endif
|
|
|
|
|
return (tx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dmu_tx_t *
|
|
|
|
|
dmu_tx_create(objset_t *os)
|
|
|
|
|
{
|
2010-05-29 00:45:14 +04:00
|
|
|
dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
|
2008-11-20 23:01:55 +03:00
|
|
|
tx->tx_objset = os;
|
2010-05-29 00:45:14 +04:00
|
|
|
tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
|
2008-11-20 23:01:55 +03:00
|
|
|
return (tx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dmu_tx_t *
|
|
|
|
|
dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_t *tx = dmu_tx_create_dd(NULL);
|
|
|
|
|
|
|
|
|
|
ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
|
|
|
|
|
tx->tx_pool = dp;
|
|
|
|
|
tx->tx_txg = txg;
|
|
|
|
|
tx->tx_anyobj = TRUE;
|
|
|
|
|
|
|
|
|
|
return (tx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
int
|
|
|
|
|
dmu_tx_is_syncing(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
return (tx->tx_anyobj);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
int
|
|
|
|
|
dmu_tx_private_ok(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
return (tx->tx_anyobj);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static dmu_tx_hold_t *
|
|
|
|
|
dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
|
|
|
|
|
enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
dnode_t *dn = NULL;
|
|
|
|
|
int err;
|
|
|
|
|
|
|
|
|
|
if (object != DMU_NEW_OBJECT) {
|
2010-05-29 00:45:14 +04:00
|
|
|
err = dnode_hold(os, object, tx, &dn);
|
2008-11-20 23:01:55 +03:00
|
|
|
if (err) {
|
|
|
|
|
tx->tx_err = err;
|
|
|
|
|
return (NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (err == 0 && tx->tx_txg != 0) {
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
|
/*
|
|
|
|
|
* dn->dn_assigned_txg == tx->tx_txg doesn't pose a
|
|
|
|
|
* problem, but there's no way for it to happen (for
|
|
|
|
|
* now, at least).
|
|
|
|
|
*/
|
|
|
|
|
ASSERT(dn->dn_assigned_txg == 0);
|
|
|
|
|
dn->dn_assigned_txg = tx->tx_txg;
|
|
|
|
|
(void) refcount_add(&dn->dn_tx_holds, tx);
|
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2012-05-07 21:49:51 +04:00
|
|
|
txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_PUSHPAGE);
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_tx = tx;
|
|
|
|
|
txh->txh_dnode = dn;
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_type = type;
|
|
|
|
|
txh->txh_arg1 = arg1;
|
|
|
|
|
txh->txh_arg2 = arg2;
|
|
|
|
|
#endif
|
|
|
|
|
list_insert_tail(&tx->tx_holds, txh);
|
|
|
|
|
|
|
|
|
|
return (txh);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
|
|
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* If we're syncing, they can manipulate any object anyhow, and
|
|
|
|
|
* the hold on the dnode_t can cause problems.
|
|
|
|
|
*/
|
|
|
|
|
if (!dmu_tx_is_syncing(tx)) {
|
|
|
|
|
(void) dmu_tx_hold_object_impl(tx, os,
|
|
|
|
|
object, THT_NEWOBJECT, 0, 0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int
|
|
|
|
|
dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
|
|
|
|
|
{
|
|
|
|
|
int err;
|
|
|
|
|
dmu_buf_impl_t *db;
|
|
|
|
|
|
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_READER);
|
|
|
|
|
db = dbuf_hold_level(dn, level, blkid, FTAG);
|
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
|
if (db == NULL)
|
2013-03-08 22:41:28 +04:00
|
|
|
return (SET_ERROR(EIO));
|
2008-11-20 23:01:55 +03:00
|
|
|
err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
|
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
|
return (err);
|
|
|
|
|
}
|
|
|
|
|
|
2009-07-03 02:44:48 +04:00
|
|
|
static void
|
2010-05-29 00:45:14 +04:00
|
|
|
dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
|
|
|
|
|
int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
|
2009-07-03 02:44:48 +04:00
|
|
|
{
|
2010-05-29 00:45:14 +04:00
|
|
|
objset_t *os = dn->dn_objset;
|
|
|
|
|
dsl_dataset_t *ds = os->os_dsl_dataset;
|
|
|
|
|
int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
|
dmu_buf_impl_t *parent = NULL;
|
|
|
|
|
blkptr_t *bp = NULL;
|
|
|
|
|
uint64_t space;
|
|
|
|
|
|
|
|
|
|
if (level >= dn->dn_nlevels || history[level] == blkid)
|
2009-07-03 02:44:48 +04:00
|
|
|
return;
|
|
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
history[level] = blkid;
|
2009-07-03 02:44:48 +04:00
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
|
|
|
|
|
|
|
|
|
|
if (db == NULL || db == dn->dn_dbuf) {
|
|
|
|
|
ASSERT(level != 0);
|
|
|
|
|
db = NULL;
|
|
|
|
|
} else {
|
2010-08-27 01:24:34 +04:00
|
|
|
ASSERT(DB_DNODE(db) == dn);
|
2010-05-29 00:45:14 +04:00
|
|
|
ASSERT(db->db_level == level);
|
|
|
|
|
ASSERT(db->db.db_size == space);
|
|
|
|
|
ASSERT(db->db_blkid == blkid);
|
|
|
|
|
bp = db->db_blkptr;
|
|
|
|
|
parent = db->db_parent;
|
2009-07-03 02:44:48 +04:00
|
|
|
}
|
|
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
freeable = (bp && (freeable ||
|
|
|
|
|
dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
|
2009-07-03 02:44:48 +04:00
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
if (freeable)
|
|
|
|
|
txh->txh_space_tooverwrite += space;
|
|
|
|
|
else
|
|
|
|
|
txh->txh_space_towrite += space;
|
|
|
|
|
if (bp)
|
|
|
|
|
txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
|
|
|
|
|
|
|
|
|
|
dmu_tx_count_twig(txh, dn, parent, level + 1,
|
|
|
|
|
blkid >> epbs, freeable, history);
|
2009-07-03 02:44:48 +04:00
|
|
|
}
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
/* ARGSUSED */
|
|
|
|
|
static void
|
|
|
|
|
dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
|
|
|
|
|
{
|
|
|
|
|
dnode_t *dn = txh->txh_dnode;
|
|
|
|
|
uint64_t start, end, i;
|
|
|
|
|
int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
|
|
|
|
|
int err = 0;
|
2010-08-26 20:52:39 +04:00
|
|
|
int l;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
if (len == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
min_bs = SPA_MINBLOCKSHIFT;
|
|
|
|
|
max_bs = SPA_MAXBLOCKSHIFT;
|
|
|
|
|
min_ibs = DN_MIN_INDBLKSHIFT;
|
|
|
|
|
max_ibs = DN_MAX_INDBLKSHIFT;
|
|
|
|
|
|
|
|
|
|
if (dn) {
|
2010-05-29 00:45:14 +04:00
|
|
|
uint64_t history[DN_MAX_LEVELS];
|
2009-07-03 02:44:48 +04:00
|
|
|
int nlvls = dn->dn_nlevels;
|
|
|
|
|
int delta;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* For i/o error checking, read the first and last level-0
|
|
|
|
|
* blocks (if they are not aligned), and all the level-1 blocks.
|
|
|
|
|
*/
|
2008-11-20 23:01:55 +03:00
|
|
|
if (dn->dn_maxblkid == 0) {
|
2009-07-03 02:44:48 +04:00
|
|
|
delta = dn->dn_datablksz;
|
|
|
|
|
start = (off < dn->dn_datablksz) ? 0 : 1;
|
|
|
|
|
end = (off+len <= dn->dn_datablksz) ? 0 : 1;
|
|
|
|
|
if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
|
2008-12-03 23:09:06 +03:00
|
|
|
err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
|
|
|
|
|
if (err)
|
|
|
|
|
goto out;
|
2009-07-03 02:44:48 +04:00
|
|
|
delta -= off;
|
2008-12-03 23:09:06 +03:00
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
} else {
|
|
|
|
|
zio_t *zio = zio_root(dn->dn_objset->os_spa,
|
|
|
|
|
NULL, NULL, ZIO_FLAG_CANFAIL);
|
|
|
|
|
|
|
|
|
|
/* first level-0 block */
|
|
|
|
|
start = off >> dn->dn_datablkshift;
|
|
|
|
|
if (P2PHASE(off, dn->dn_datablksz) ||
|
|
|
|
|
len < dn->dn_datablksz) {
|
|
|
|
|
err = dmu_tx_check_ioerr(zio, dn, 0, start);
|
|
|
|
|
if (err)
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* last level-0 block */
|
|
|
|
|
end = (off+len-1) >> dn->dn_datablkshift;
|
2008-12-03 23:09:06 +03:00
|
|
|
if (end != start && end <= dn->dn_maxblkid &&
|
2008-11-20 23:01:55 +03:00
|
|
|
P2PHASE(off+len, dn->dn_datablksz)) {
|
|
|
|
|
err = dmu_tx_check_ioerr(zio, dn, 0, end);
|
|
|
|
|
if (err)
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* level-1 blocks */
|
2009-07-03 02:44:48 +04:00
|
|
|
if (nlvls > 1) {
|
|
|
|
|
int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
|
for (i = (start>>shft)+1; i < end>>shft; i++) {
|
2008-11-20 23:01:55 +03:00
|
|
|
err = dmu_tx_check_ioerr(zio, dn, 1, i);
|
|
|
|
|
if (err)
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
err = zio_wait(zio);
|
|
|
|
|
if (err)
|
|
|
|
|
goto out;
|
2009-07-03 02:44:48 +04:00
|
|
|
delta = P2NPHASE(off, dn->dn_datablksz);
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
2013-04-11 03:58:22 +04:00
|
|
|
min_ibs = max_ibs = dn->dn_indblkshift;
|
2009-07-03 02:44:48 +04:00
|
|
|
if (dn->dn_maxblkid > 0) {
|
|
|
|
|
/*
|
|
|
|
|
* The blocksize can't change,
|
|
|
|
|
* so we can make a more precise estimate.
|
|
|
|
|
*/
|
|
|
|
|
ASSERT(dn->dn_datablkshift != 0);
|
2008-11-20 23:01:55 +03:00
|
|
|
min_bs = max_bs = dn->dn_datablkshift;
|
2009-07-03 02:44:48 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If this write is not off the end of the file
|
|
|
|
|
* we need to account for overwrites/unref.
|
|
|
|
|
*/
|
2010-05-29 00:45:14 +04:00
|
|
|
if (start <= dn->dn_maxblkid) {
|
2010-08-26 20:52:39 +04:00
|
|
|
for (l = 0; l < DN_MAX_LEVELS; l++)
|
2010-05-29 00:45:14 +04:00
|
|
|
history[l] = -1ULL;
|
|
|
|
|
}
|
2009-07-03 02:44:48 +04:00
|
|
|
while (start <= dn->dn_maxblkid) {
|
|
|
|
|
dmu_buf_impl_t *db;
|
|
|
|
|
|
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_READER);
|
2010-05-29 00:45:14 +04:00
|
|
|
err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
|
2009-07-03 02:44:48 +04:00
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
if (err) {
|
|
|
|
|
txh->txh_tx->tx_err = err;
|
|
|
|
|
return;
|
2009-07-03 02:44:48 +04:00
|
|
|
}
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
|
|
|
|
|
history);
|
2009-07-03 02:44:48 +04:00
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
|
if (++start > end) {
|
|
|
|
|
/*
|
|
|
|
|
* Account for new indirects appearing
|
|
|
|
|
* before this IO gets assigned into a txg.
|
|
|
|
|
*/
|
|
|
|
|
bits = 64 - min_bs;
|
|
|
|
|
epbs = min_ibs - SPA_BLKPTRSHIFT;
|
|
|
|
|
for (bits -= epbs * (nlvls - 1);
|
|
|
|
|
bits >= 0; bits -= epbs)
|
|
|
|
|
txh->txh_fudge += 1ULL << max_ibs;
|
|
|
|
|
goto out;
|
|
|
|
|
}
|
|
|
|
|
off += delta;
|
|
|
|
|
if (len >= delta)
|
|
|
|
|
len -= delta;
|
|
|
|
|
delta = dn->dn_datablksz;
|
|
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* 'end' is the last thing we will access, not one past.
|
|
|
|
|
* This way we won't overflow when accessing the last byte.
|
|
|
|
|
*/
|
|
|
|
|
start = P2ALIGN(off, 1ULL << max_bs);
|
|
|
|
|
end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
|
|
|
|
|
txh->txh_space_towrite += end - start + 1;
|
|
|
|
|
|
|
|
|
|
start >>= min_bs;
|
|
|
|
|
end >>= min_bs;
|
|
|
|
|
|
|
|
|
|
epbs = min_ibs - SPA_BLKPTRSHIFT;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The object contains at most 2^(64 - min_bs) blocks,
|
|
|
|
|
* and each indirect level maps 2^epbs.
|
|
|
|
|
*/
|
|
|
|
|
for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
|
|
|
|
|
start >>= epbs;
|
|
|
|
|
end >>= epbs;
|
2009-07-03 02:44:48 +04:00
|
|
|
ASSERT3U(end, >=, start);
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_space_towrite += (end - start + 1) << max_ibs;
|
2009-07-03 02:44:48 +04:00
|
|
|
if (start != 0) {
|
|
|
|
|
/*
|
|
|
|
|
* We also need a new blkid=0 indirect block
|
|
|
|
|
* to reference any existing file data.
|
|
|
|
|
*/
|
|
|
|
|
txh->txh_space_towrite += 1ULL << max_ibs;
|
|
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
out:
|
2009-07-03 02:44:48 +04:00
|
|
|
if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
|
|
|
|
|
2 * DMU_MAX_ACCESS)
|
2013-03-08 22:41:28 +04:00
|
|
|
err = SET_ERROR(EFBIG);
|
2009-07-03 02:44:48 +04:00
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
if (err)
|
|
|
|
|
txh->txh_tx->tx_err = err;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
dmu_tx_count_dnode(dmu_tx_hold_t *txh)
|
|
|
|
|
{
|
|
|
|
|
dnode_t *dn = txh->txh_dnode;
|
2010-08-27 01:24:34 +04:00
|
|
|
dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
|
2008-11-20 23:01:55 +03:00
|
|
|
uint64_t space = mdn->dn_datablksz +
|
|
|
|
|
((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
|
|
|
|
|
|
|
|
|
|
if (dn && dn->dn_dbuf->db_blkptr &&
|
|
|
|
|
dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
|
2010-05-29 00:45:14 +04:00
|
|
|
dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_space_tooverwrite += space;
|
2009-07-03 02:44:48 +04:00
|
|
|
txh->txh_space_tounref += space;
|
2008-11-20 23:01:55 +03:00
|
|
|
} else {
|
|
|
|
|
txh->txh_space_towrite += space;
|
|
|
|
|
if (dn && dn->dn_dbuf->db_blkptr)
|
|
|
|
|
txh->txh_space_tounref += space;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
ASSERT(len < DMU_MAX_ACCESS);
|
|
|
|
|
ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
|
|
|
|
|
|
|
|
|
|
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
|
|
|
|
|
object, THT_WRITE, off, len);
|
|
|
|
|
if (txh == NULL)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
dmu_tx_count_write(txh, off, len);
|
|
|
|
|
dmu_tx_count_dnode(txh);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
|
|
|
|
|
{
|
2008-12-03 23:09:06 +03:00
|
|
|
uint64_t blkid, nblks, lastblk;
|
|
|
|
|
uint64_t space = 0, unref = 0, skipped = 0;
|
2008-11-20 23:01:55 +03:00
|
|
|
dnode_t *dn = txh->txh_dnode;
|
|
|
|
|
dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
|
|
|
|
|
spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
|
2008-12-03 23:09:06 +03:00
|
|
|
int epbs;
|
2013-01-14 22:26:31 +04:00
|
|
|
uint64_t l0span = 0, nl1blks = 0;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
if (dn->dn_nlevels == 0)
|
2008-11-20 23:01:55 +03:00
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/*
|
2008-12-03 23:09:06 +03:00
|
|
|
* The struct_rwlock protects us against dn_nlevels
|
2008-11-20 23:01:55 +03:00
|
|
|
* changing, in case (against all odds) we manage to dirty &
|
|
|
|
|
* sync out the changes after we check for being dirty.
|
2010-05-29 00:45:14 +04:00
|
|
|
* Also, dbuf_hold_impl() wants us to have the struct_rwlock.
|
2008-11-20 23:01:55 +03:00
|
|
|
*/
|
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_READER);
|
2008-12-03 23:09:06 +03:00
|
|
|
epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
|
if (dn->dn_maxblkid == 0) {
|
2008-11-20 23:01:55 +03:00
|
|
|
if (off == 0 && len >= dn->dn_datablksz) {
|
|
|
|
|
blkid = 0;
|
|
|
|
|
nblks = 1;
|
|
|
|
|
} else {
|
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
blkid = off >> dn->dn_datablkshift;
|
2008-12-03 23:09:06 +03:00
|
|
|
nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2013-08-07 22:32:46 +04:00
|
|
|
if (blkid > dn->dn_maxblkid) {
|
2008-11-20 23:01:55 +03:00
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
|
return;
|
|
|
|
|
}
|
2008-12-03 23:09:06 +03:00
|
|
|
if (blkid + nblks > dn->dn_maxblkid)
|
2013-08-07 22:32:46 +04:00
|
|
|
nblks = dn->dn_maxblkid - blkid + 1;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
}
|
2013-01-14 22:26:31 +04:00
|
|
|
l0span = nblks; /* save for later use to calc level > 1 overhead */
|
2008-12-03 23:09:06 +03:00
|
|
|
if (dn->dn_nlevels == 1) {
|
2008-11-20 23:01:55 +03:00
|
|
|
int i;
|
|
|
|
|
for (i = 0; i < nblks; i++) {
|
|
|
|
|
blkptr_t *bp = dn->dn_phys->dn_blkptr;
|
2008-12-03 23:09:06 +03:00
|
|
|
ASSERT3U(blkid + i, <, dn->dn_nblkptr);
|
2008-11-20 23:01:55 +03:00
|
|
|
bp += blkid + i;
|
2010-05-29 00:45:14 +04:00
|
|
|
if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
|
2008-11-20 23:01:55 +03:00
|
|
|
dprintf_bp(bp, "can free old%s", "");
|
2010-05-29 00:45:14 +04:00
|
|
|
space += bp_get_dsize(spa, bp);
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
unref += BP_GET_ASIZE(bp);
|
|
|
|
|
}
|
2013-01-14 22:26:31 +04:00
|
|
|
nl1blks = 1;
|
2008-11-20 23:01:55 +03:00
|
|
|
nblks = 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
lastblk = blkid + nblks - 1;
|
2008-11-20 23:01:55 +03:00
|
|
|
while (nblks) {
|
|
|
|
|
dmu_buf_impl_t *dbuf;
|
2008-12-03 23:09:06 +03:00
|
|
|
uint64_t ibyte, new_blkid;
|
|
|
|
|
int epb = 1 << epbs;
|
|
|
|
|
int err, i, blkoff, tochk;
|
|
|
|
|
blkptr_t *bp;
|
|
|
|
|
|
|
|
|
|
ibyte = blkid << dn->dn_datablkshift;
|
|
|
|
|
err = dnode_next_offset(dn,
|
|
|
|
|
DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
|
|
|
|
|
new_blkid = ibyte >> dn->dn_datablkshift;
|
|
|
|
|
if (err == ESRCH) {
|
|
|
|
|
skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (err) {
|
|
|
|
|
txh->txh_tx->tx_err = err;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (new_blkid > lastblk) {
|
|
|
|
|
skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
|
|
|
|
|
break;
|
|
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
if (new_blkid > blkid) {
|
|
|
|
|
ASSERT((new_blkid >> epbs) > (blkid >> epbs));
|
|
|
|
|
skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
|
|
|
|
|
nblks -= new_blkid - blkid;
|
|
|
|
|
blkid = new_blkid;
|
|
|
|
|
}
|
|
|
|
|
blkoff = P2PHASE(blkid, epb);
|
|
|
|
|
tochk = MIN(epb - blkoff, nblks);
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
|
|
|
|
|
if (err) {
|
|
|
|
|
txh->txh_tx->tx_err = err;
|
2008-12-03 23:09:06 +03:00
|
|
|
break;
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
txh->txh_memory_tohold += dbuf->db.db_size;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We don't check memory_tohold against DMU_MAX_ACCESS because
|
|
|
|
|
* memory_tohold is an over-estimation (especially the >L1
|
|
|
|
|
* indirect blocks), so it could fail. Callers should have
|
|
|
|
|
* already verified that they will not be holding too much
|
|
|
|
|
* memory.
|
|
|
|
|
*/
|
|
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
|
|
|
|
|
if (err != 0) {
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_tx->tx_err = err;
|
2008-12-03 23:09:06 +03:00
|
|
|
dbuf_rele(dbuf, FTAG);
|
2008-11-20 23:01:55 +03:00
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
bp = dbuf->db.db_data;
|
|
|
|
|
bp += blkoff;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < tochk; i++) {
|
2010-05-29 00:45:14 +04:00
|
|
|
if (dsl_dataset_block_freeable(ds, &bp[i],
|
|
|
|
|
bp[i].blk_birth)) {
|
2008-12-03 23:09:06 +03:00
|
|
|
dprintf_bp(&bp[i], "can free old%s", "");
|
2010-05-29 00:45:14 +04:00
|
|
|
space += bp_get_dsize(spa, &bp[i]);
|
2008-12-03 23:09:06 +03:00
|
|
|
}
|
|
|
|
|
unref += BP_GET_ASIZE(bp);
|
|
|
|
|
}
|
|
|
|
|
dbuf_rele(dbuf, FTAG);
|
|
|
|
|
|
2013-01-14 22:26:31 +04:00
|
|
|
++nl1blks;
|
2008-11-20 23:01:55 +03:00
|
|
|
blkid += tochk;
|
|
|
|
|
nblks -= tochk;
|
|
|
|
|
}
|
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
|
|
2013-01-14 22:26:31 +04:00
|
|
|
/*
|
|
|
|
|
* Add in memory requirements of higher-level indirects.
|
|
|
|
|
* This assumes a worst-possible scenario for dn_nlevels and a
|
|
|
|
|
* worst-possible distribution of l1-blocks over the region to free.
|
|
|
|
|
*/
|
|
|
|
|
{
|
|
|
|
|
uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
|
|
|
|
|
int level = 2;
|
|
|
|
|
/*
|
|
|
|
|
* Here we don't use DN_MAX_LEVEL, but calculate it with the
|
|
|
|
|
* given datablkshift and indblkshift. This makes the
|
|
|
|
|
* difference between 19 and 8 on large files.
|
|
|
|
|
*/
|
|
|
|
|
int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
|
|
|
|
|
(dn->dn_indblkshift - SPA_BLKPTRSHIFT);
|
|
|
|
|
|
|
|
|
|
while (level++ < maxlevel) {
|
2013-01-14 22:29:55 +04:00
|
|
|
txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
|
2013-01-14 22:26:31 +04:00
|
|
|
<< dn->dn_indblkshift;
|
|
|
|
|
blkcnt = 1 + (blkcnt >> epbs);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
/* account for new level 1 indirect blocks that might show up */
|
|
|
|
|
if (skipped > 0) {
|
|
|
|
|
txh->txh_fudge += skipped << dn->dn_indblkshift;
|
|
|
|
|
skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
|
|
|
|
|
txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
|
|
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_space_tofree += space;
|
|
|
|
|
txh->txh_space_tounref += unref;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
dnode_t *dn;
|
2013-07-29 22:58:53 +04:00
|
|
|
int err;
|
2008-11-20 23:01:55 +03:00
|
|
|
zio_t *zio;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
|
|
|
|
|
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
|
|
|
|
|
object, THT_FREE, off, len);
|
|
|
|
|
if (txh == NULL)
|
|
|
|
|
return;
|
|
|
|
|
dn = txh->txh_dnode;
|
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
|
|
|
dmu_tx_count_dnode(txh);
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
|
|
|
|
|
return;
|
|
|
|
|
if (len == DMU_OBJECT_END)
|
|
|
|
|
len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
|
|
|
|
|
|
2013-07-29 22:58:53 +04:00
|
|
|
dmu_tx_count_dnode(txh);
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
/*
|
2013-07-29 22:58:53 +04:00
|
|
|
* For i/o error checking, we read the first and last level-0
|
|
|
|
|
* blocks if they are not aligned, and all the level-1 blocks.
|
|
|
|
|
*
|
|
|
|
|
* Note: dbuf_free_range() assumes that we have not instantiated
|
|
|
|
|
* any level-0 dbufs that will be completely freed. Therefore we must
|
|
|
|
|
* exercise care to not read or count the first and last blocks
|
|
|
|
|
* if they are blocksize-aligned.
|
|
|
|
|
*/
|
|
|
|
|
if (dn->dn_datablkshift == 0) {
|
2013-08-21 08:11:52 +04:00
|
|
|
if (off != 0 || len < dn->dn_datablksz)
|
2013-08-30 13:19:35 +04:00
|
|
|
dmu_tx_count_write(txh, 0, dn->dn_datablksz);
|
2013-07-29 22:58:53 +04:00
|
|
|
} else {
|
|
|
|
|
/* first block will be modified if it is not aligned */
|
|
|
|
|
if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
|
|
|
|
|
dmu_tx_count_write(txh, off, 1);
|
|
|
|
|
/* last block will be modified if it is not aligned */
|
|
|
|
|
if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
|
|
|
|
|
dmu_tx_count_write(txh, off+len, 1);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Check level-1 blocks.
|
2008-11-20 23:01:55 +03:00
|
|
|
*/
|
|
|
|
|
if (dn->dn_nlevels > 1) {
|
2013-07-29 22:58:53 +04:00
|
|
|
int shift = dn->dn_datablkshift + dn->dn_indblkshift -
|
2008-11-20 23:01:55 +03:00
|
|
|
SPA_BLKPTRSHIFT;
|
2013-07-29 22:58:53 +04:00
|
|
|
uint64_t start = off >> shift;
|
|
|
|
|
uint64_t end = (off + len) >> shift;
|
|
|
|
|
uint64_t i;
|
|
|
|
|
|
|
|
|
|
ASSERT(dn->dn_indblkshift != 0);
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2014-01-31 19:52:11 +04:00
|
|
|
/*
|
|
|
|
|
* dnode_reallocate() can result in an object with indirect
|
|
|
|
|
* blocks having an odd data block size. In this case,
|
|
|
|
|
* just check the single block.
|
|
|
|
|
*/
|
|
|
|
|
if (dn->dn_datablkshift == 0)
|
|
|
|
|
start = end = 0;
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
zio = zio_root(tx->tx_pool->dp_spa,
|
|
|
|
|
NULL, NULL, ZIO_FLAG_CANFAIL);
|
|
|
|
|
for (i = start; i <= end; i++) {
|
|
|
|
|
uint64_t ibyte = i << shift;
|
2008-12-03 23:09:06 +03:00
|
|
|
err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
|
2008-11-20 23:01:55 +03:00
|
|
|
i = ibyte >> shift;
|
|
|
|
|
if (err == ESRCH)
|
|
|
|
|
break;
|
|
|
|
|
if (err) {
|
|
|
|
|
tx->tx_err = err;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
err = dmu_tx_check_ioerr(zio, dn, 1, i);
|
|
|
|
|
if (err) {
|
|
|
|
|
tx->tx_err = err;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
err = zio_wait(zio);
|
|
|
|
|
if (err) {
|
|
|
|
|
tx->tx_err = err;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dmu_tx_count_free(txh, off, len);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
2009-07-03 02:44:48 +04:00
|
|
|
dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
|
2008-11-20 23:01:55 +03:00
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
dnode_t *dn;
|
|
|
|
|
uint64_t nblocks;
|
|
|
|
|
int epbs, err;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
|
|
|
|
|
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
|
|
|
|
|
object, THT_ZAP, add, (uintptr_t)name);
|
|
|
|
|
if (txh == NULL)
|
|
|
|
|
return;
|
|
|
|
|
dn = txh->txh_dnode;
|
|
|
|
|
|
|
|
|
|
dmu_tx_count_dnode(txh);
|
|
|
|
|
|
|
|
|
|
if (dn == NULL) {
|
|
|
|
|
/*
|
|
|
|
|
* We will be able to fit a new object's entries into one leaf
|
|
|
|
|
* block. So there will be at most 2 blocks total,
|
|
|
|
|
* including the header block.
|
|
|
|
|
*/
|
|
|
|
|
dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
2012-12-14 03:24:15 +04:00
|
|
|
ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
if (dn->dn_maxblkid == 0 && !add) {
|
2012-04-08 21:10:49 +04:00
|
|
|
blkptr_t *bp;
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
/*
|
|
|
|
|
* If there is only one block (i.e. this is a micro-zap)
|
|
|
|
|
* and we are not adding anything, the accounting is simple.
|
|
|
|
|
*/
|
|
|
|
|
err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
|
|
|
|
|
if (err) {
|
|
|
|
|
tx->tx_err = err;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Use max block size here, since we don't know how much
|
|
|
|
|
* the size will change between now and the dbuf dirty call.
|
|
|
|
|
*/
|
2012-04-08 21:10:49 +04:00
|
|
|
bp = &dn->dn_phys->dn_blkptr[0];
|
2008-11-20 23:01:55 +03:00
|
|
|
if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
|
2012-04-08 21:10:49 +04:00
|
|
|
bp, bp->blk_birth))
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
|
2012-04-08 21:10:49 +04:00
|
|
|
else
|
2008-11-20 23:01:55 +03:00
|
|
|
txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
|
2012-04-08 21:10:49 +04:00
|
|
|
if (!BP_IS_HOLE(bp))
|
2009-07-03 02:44:48 +04:00
|
|
|
txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
|
2008-11-20 23:01:55 +03:00
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (dn->dn_maxblkid > 0 && name) {
|
|
|
|
|
/*
|
|
|
|
|
* access the name in this fat-zap so that we'll check
|
|
|
|
|
* for i/o errors to the leaf blocks, etc.
|
|
|
|
|
*/
|
2010-05-29 00:45:14 +04:00
|
|
|
err = zap_lookup(dn->dn_objset, dn->dn_object, name,
|
2008-11-20 23:01:55 +03:00
|
|
|
8, 0, NULL);
|
|
|
|
|
if (err == EIO) {
|
|
|
|
|
tx->tx_err = err;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
|
2009-08-18 22:43:27 +04:00
|
|
|
&txh->txh_space_towrite, &txh->txh_space_tooverwrite);
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If the modified blocks are scattered to the four winds,
|
|
|
|
|
* we'll have to modify an indirect twig for each.
|
|
|
|
|
*/
|
|
|
|
|
epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
|
for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
|
2009-07-03 02:44:48 +04:00
|
|
|
if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
|
|
|
|
|
txh->txh_space_towrite += 3 << dn->dn_indblkshift;
|
|
|
|
|
else
|
|
|
|
|
txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
|
|
|
|
|
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
|
|
|
|
|
object, THT_BONUS, 0, 0);
|
|
|
|
|
if (txh)
|
|
|
|
|
dmu_tx_count_dnode(txh);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
2013-07-23 21:32:57 +04:00
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
|
|
|
|
|
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
|
|
|
|
|
DMU_NEW_OBJECT, THT_SPACE, space, 0);
|
2013-07-23 21:32:57 +04:00
|
|
|
if (txh)
|
|
|
|
|
txh->txh_space_towrite += space;
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
int
|
|
|
|
|
dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
int holds = 0;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* By asserting that the tx is assigned, we're counting the
|
|
|
|
|
* number of dn_tx_holds, which is the same as the number of
|
|
|
|
|
* dn_holds. Otherwise, we'd be counting dn_holds, but
|
|
|
|
|
* dn_tx_holds could be 0.
|
|
|
|
|
*/
|
|
|
|
|
ASSERT(tx->tx_txg != 0);
|
|
|
|
|
|
|
|
|
|
/* if (tx->tx_anyobj == TRUE) */
|
|
|
|
|
/* return (0); */
|
|
|
|
|
|
|
|
|
|
for (txh = list_head(&tx->tx_holds); txh;
|
|
|
|
|
txh = list_next(&tx->tx_holds, txh)) {
|
|
|
|
|
if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
|
|
|
|
|
holds++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return (holds);
|
|
|
|
|
}
|
|
|
|
|
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-11-20 23:01:55 +03:00
|
|
|
void
|
|
|
|
|
dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
int match_object = FALSE, match_offset = FALSE;
|
2010-08-27 01:24:34 +04:00
|
|
|
dnode_t *dn;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2010-08-27 01:24:34 +04:00
|
|
|
DB_DNODE_ENTER(db);
|
|
|
|
|
dn = DB_DNODE(db);
|
2012-03-12 23:38:00 +04:00
|
|
|
ASSERT(dn != NULL);
|
2008-11-20 23:01:55 +03:00
|
|
|
ASSERT(tx->tx_txg != 0);
|
2010-05-29 00:45:14 +04:00
|
|
|
ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
|
2008-11-20 23:01:55 +03:00
|
|
|
ASSERT3U(dn->dn_object, ==, db->db.db_object);
|
|
|
|
|
|
2010-08-27 01:24:34 +04:00
|
|
|
if (tx->tx_anyobj) {
|
|
|
|
|
DB_DNODE_EXIT(db);
|
2008-11-20 23:01:55 +03:00
|
|
|
return;
|
2010-08-27 01:24:34 +04:00
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
/* XXX No checking on the meta dnode for now */
|
2010-08-27 01:24:34 +04:00
|
|
|
if (db->db.db_object == DMU_META_DNODE_OBJECT) {
|
|
|
|
|
DB_DNODE_EXIT(db);
|
2008-11-20 23:01:55 +03:00
|
|
|
return;
|
2010-08-27 01:24:34 +04:00
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
for (txh = list_head(&tx->tx_holds); txh;
|
|
|
|
|
txh = list_next(&tx->tx_holds, txh)) {
|
2012-03-12 23:38:00 +04:00
|
|
|
ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
|
2008-11-20 23:01:55 +03:00
|
|
|
if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
|
|
|
|
|
match_object = TRUE;
|
|
|
|
|
if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
|
|
|
|
|
int datablkshift = dn->dn_datablkshift ?
|
|
|
|
|
dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
|
|
|
|
|
int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
|
int shift = datablkshift + epbs * db->db_level;
|
|
|
|
|
uint64_t beginblk = shift >= 64 ? 0 :
|
|
|
|
|
(txh->txh_arg1 >> shift);
|
|
|
|
|
uint64_t endblk = shift >= 64 ? 0 :
|
|
|
|
|
((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
|
|
|
|
|
uint64_t blkid = db->db_blkid;
|
|
|
|
|
|
|
|
|
|
/* XXX txh_arg2 better not be zero... */
|
|
|
|
|
|
|
|
|
|
dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
|
|
|
|
|
txh->txh_type, beginblk, endblk);
|
|
|
|
|
|
|
|
|
|
switch (txh->txh_type) {
|
|
|
|
|
case THT_WRITE:
|
|
|
|
|
if (blkid >= beginblk && blkid <= endblk)
|
|
|
|
|
match_offset = TRUE;
|
|
|
|
|
/*
|
|
|
|
|
* We will let this hold work for the bonus
|
2010-05-29 00:45:14 +04:00
|
|
|
* or spill buffer so that we don't need to
|
|
|
|
|
* hold it when creating a new object.
|
2008-11-20 23:01:55 +03:00
|
|
|
*/
|
2010-05-29 00:45:14 +04:00
|
|
|
if (blkid == DMU_BONUS_BLKID ||
|
|
|
|
|
blkid == DMU_SPILL_BLKID)
|
2008-11-20 23:01:55 +03:00
|
|
|
match_offset = TRUE;
|
|
|
|
|
/*
|
|
|
|
|
* They might have to increase nlevels,
|
|
|
|
|
* thus dirtying the new TLIBs. Or the
|
|
|
|
|
* might have to change the block size,
|
|
|
|
|
* thus dirying the new lvl=0 blk=0.
|
|
|
|
|
*/
|
|
|
|
|
if (blkid == 0)
|
|
|
|
|
match_offset = TRUE;
|
|
|
|
|
break;
|
|
|
|
|
case THT_FREE:
|
2008-12-03 23:09:06 +03:00
|
|
|
/*
|
|
|
|
|
* We will dirty all the level 1 blocks in
|
|
|
|
|
* the free range and perhaps the first and
|
|
|
|
|
* last level 0 block.
|
|
|
|
|
*/
|
|
|
|
|
if (blkid >= beginblk && (blkid <= endblk ||
|
|
|
|
|
txh->txh_arg2 == DMU_OBJECT_END))
|
2008-11-20 23:01:55 +03:00
|
|
|
match_offset = TRUE;
|
|
|
|
|
break;
|
2010-05-29 00:45:14 +04:00
|
|
|
case THT_SPILL:
|
|
|
|
|
if (blkid == DMU_SPILL_BLKID)
|
|
|
|
|
match_offset = TRUE;
|
|
|
|
|
break;
|
2008-11-20 23:01:55 +03:00
|
|
|
case THT_BONUS:
|
2010-05-29 00:45:14 +04:00
|
|
|
if (blkid == DMU_BONUS_BLKID)
|
2008-11-20 23:01:55 +03:00
|
|
|
match_offset = TRUE;
|
|
|
|
|
break;
|
|
|
|
|
case THT_ZAP:
|
|
|
|
|
match_offset = TRUE;
|
|
|
|
|
break;
|
|
|
|
|
case THT_NEWOBJECT:
|
|
|
|
|
match_object = TRUE;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
ASSERT(!"bad txh_type");
|
|
|
|
|
}
|
|
|
|
|
}
|
2010-08-27 01:24:34 +04:00
|
|
|
if (match_object && match_offset) {
|
|
|
|
|
DB_DNODE_EXIT(db);
|
2008-11-20 23:01:55 +03:00
|
|
|
return;
|
2010-08-27 01:24:34 +04:00
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
2010-08-27 01:24:34 +04:00
|
|
|
DB_DNODE_EXIT(db);
|
2008-11-20 23:01:55 +03:00
|
|
|
panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
|
|
|
|
|
(u_longlong_t)db->db.db_object, db->db_level,
|
|
|
|
|
(u_longlong_t)db->db_blkid);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
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 we can't do 10 iops, something is wrong. Let us go ahead
|
|
|
|
|
* and hit zfs_dirty_data_max.
|
|
|
|
|
*/
|
|
|
|
|
hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
|
|
|
|
|
int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* We delay transactions when we've determined that the backend storage
|
|
|
|
|
* isn't able to accommodate the rate of incoming writes.
|
|
|
|
|
*
|
|
|
|
|
* If there is already a transaction waiting, we delay relative to when
|
|
|
|
|
* that transaction finishes waiting. This way the calculated min_time
|
|
|
|
|
* is independent of the number of threads concurrently executing
|
|
|
|
|
* transactions.
|
|
|
|
|
*
|
|
|
|
|
* If we are the only waiter, wait relative to when the transaction
|
|
|
|
|
* started, rather than the current time. This credits the transaction for
|
|
|
|
|
* "time already served", e.g. reading indirect blocks.
|
|
|
|
|
*
|
|
|
|
|
* The minimum time for a transaction to take is calculated as:
|
|
|
|
|
* min_time = scale * (dirty - min) / (max - dirty)
|
|
|
|
|
* min_time is then capped at zfs_delay_max_ns.
|
|
|
|
|
*
|
|
|
|
|
* The delay has two degrees of freedom that can be adjusted via tunables.
|
|
|
|
|
* The percentage of dirty data at which we start to delay is defined by
|
|
|
|
|
* zfs_delay_min_dirty_percent. This should typically be at or above
|
|
|
|
|
* zfs_vdev_async_write_active_max_dirty_percent so that we only start to
|
|
|
|
|
* delay after writing at full speed has failed to keep up with the incoming
|
|
|
|
|
* write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
|
|
|
|
|
* speaking, this variable determines the amount of delay at the midpoint of
|
|
|
|
|
* the curve.
|
|
|
|
|
*
|
|
|
|
|
* delay
|
|
|
|
|
* 10ms +-------------------------------------------------------------*+
|
|
|
|
|
* | *|
|
|
|
|
|
* 9ms + *+
|
|
|
|
|
* | *|
|
|
|
|
|
* 8ms + *+
|
|
|
|
|
* | * |
|
|
|
|
|
* 7ms + * +
|
|
|
|
|
* | * |
|
|
|
|
|
* 6ms + * +
|
|
|
|
|
* | * |
|
|
|
|
|
* 5ms + * +
|
|
|
|
|
* | * |
|
|
|
|
|
* 4ms + * +
|
|
|
|
|
* | * |
|
|
|
|
|
* 3ms + * +
|
|
|
|
|
* | * |
|
|
|
|
|
* 2ms + (midpoint) * +
|
|
|
|
|
* | | ** |
|
|
|
|
|
* 1ms + v *** +
|
|
|
|
|
* | zfs_delay_scale ----------> ******** |
|
|
|
|
|
* 0 +-------------------------------------*********----------------+
|
|
|
|
|
* 0% <- zfs_dirty_data_max -> 100%
|
|
|
|
|
*
|
|
|
|
|
* Note that since the delay is added to the outstanding time remaining on the
|
|
|
|
|
* most recent transaction, the delay is effectively the inverse of IOPS.
|
|
|
|
|
* Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
|
|
|
|
|
* was chosen such that small changes in the amount of accumulated dirty data
|
|
|
|
|
* in the first 3/4 of the curve yield relatively small differences in the
|
|
|
|
|
* amount of delay.
|
|
|
|
|
*
|
|
|
|
|
* The effects can be easier to understand when the amount of delay is
|
|
|
|
|
* represented on a log scale:
|
|
|
|
|
*
|
|
|
|
|
* delay
|
|
|
|
|
* 100ms +-------------------------------------------------------------++
|
|
|
|
|
* + +
|
|
|
|
|
* | |
|
|
|
|
|
* + *+
|
|
|
|
|
* 10ms + *+
|
|
|
|
|
* + ** +
|
|
|
|
|
* | (midpoint) ** |
|
|
|
|
|
* + | ** +
|
|
|
|
|
* 1ms + v **** +
|
|
|
|
|
* + zfs_delay_scale ----------> ***** +
|
|
|
|
|
* | **** |
|
|
|
|
|
* + **** +
|
|
|
|
|
* 100us + ** +
|
|
|
|
|
* + * +
|
|
|
|
|
* | * |
|
|
|
|
|
* + * +
|
|
|
|
|
* 10us + * +
|
|
|
|
|
* + +
|
|
|
|
|
* | |
|
|
|
|
|
* + +
|
|
|
|
|
* +--------------------------------------------------------------+
|
|
|
|
|
* 0% <- zfs_dirty_data_max -> 100%
|
|
|
|
|
*
|
|
|
|
|
* Note here that only as the amount of dirty data approaches its limit does
|
|
|
|
|
* the delay start to increase rapidly. The goal of a properly tuned system
|
|
|
|
|
* should be to keep the amount of dirty data out of that range by first
|
|
|
|
|
* ensuring that the appropriate limits are set for the I/O scheduler to reach
|
|
|
|
|
* optimal throughput on the backend storage, and then by changing the value
|
|
|
|
|
* of zfs_delay_scale to increase the steepness of the curve.
|
|
|
|
|
*/
|
|
|
|
|
static void
|
|
|
|
|
dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
|
|
|
|
|
{
|
|
|
|
|
dsl_pool_t *dp = tx->tx_pool;
|
|
|
|
|
uint64_t delay_min_bytes =
|
|
|
|
|
zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
|
|
|
|
|
hrtime_t wakeup, min_tx_time, now;
|
|
|
|
|
|
|
|
|
|
if (dirty <= delay_min_bytes)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* The caller has already waited until we are under the max.
|
|
|
|
|
* We make them pass us the amount of dirty data so we don't
|
|
|
|
|
* have to handle the case of it being >= the max, which could
|
|
|
|
|
* cause a divide-by-zero if it's == the max.
|
|
|
|
|
*/
|
|
|
|
|
ASSERT3U(dirty, <, zfs_dirty_data_max);
|
|
|
|
|
|
|
|
|
|
now = gethrtime();
|
|
|
|
|
min_tx_time = zfs_delay_scale *
|
|
|
|
|
(dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
|
|
|
|
|
min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
|
|
|
|
|
if (now > tx->tx_start + min_tx_time)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
|
|
|
|
|
uint64_t, min_tx_time);
|
|
|
|
|
|
|
|
|
|
mutex_enter(&dp->dp_lock);
|
|
|
|
|
wakeup = MAX(tx->tx_start + min_tx_time,
|
|
|
|
|
dp->dp_last_wakeup + min_tx_time);
|
|
|
|
|
dp->dp_last_wakeup = wakeup;
|
|
|
|
|
mutex_exit(&dp->dp_lock);
|
|
|
|
|
|
|
|
|
|
zfs_sleep_until(wakeup);
|
|
|
|
|
}
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
static int
|
2013-09-04 16:00:57 +04:00
|
|
|
dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
|
2008-11-20 23:01:55 +03:00
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
spa_t *spa = tx->tx_pool->dp_spa;
|
2008-12-03 23:09:06 +03:00
|
|
|
uint64_t memory, asize, fsize, usize;
|
|
|
|
|
uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2013-05-11 01:17:03 +04:00
|
|
|
ASSERT0(tx->tx_txg);
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2012-01-20 22:58:57 +04:00
|
|
|
if (tx->tx_err) {
|
|
|
|
|
DMU_TX_STAT_BUMP(dmu_tx_error);
|
2008-11-20 23:01:55 +03:00
|
|
|
return (tx->tx_err);
|
2012-01-20 22:58:57 +04:00
|
|
|
}
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
if (spa_suspended(spa)) {
|
2012-01-20 22:58:57 +04:00
|
|
|
DMU_TX_STAT_BUMP(dmu_tx_suspended);
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
/*
|
|
|
|
|
* If the user has indicated a blocking failure mode
|
|
|
|
|
* then return ERESTART which will block in dmu_tx_wait().
|
|
|
|
|
* Otherwise, return EIO so that an error can get
|
|
|
|
|
* propagated back to the VOP calls.
|
|
|
|
|
*
|
|
|
|
|
* Note that we always honor the txg_how flag regardless
|
|
|
|
|
* of the failuremode setting.
|
|
|
|
|
*/
|
|
|
|
|
if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
|
|
|
|
|
txg_how != TXG_WAIT)
|
2013-03-08 22:41:28 +04:00
|
|
|
return (SET_ERROR(EIO));
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2013-03-08 22:41:28 +04:00
|
|
|
return (SET_ERROR(ERESTART));
|
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
|
|
|
if (!tx->tx_waited &&
|
|
|
|
|
dsl_pool_need_dirty_delay(tx->tx_pool)) {
|
|
|
|
|
tx->tx_wait_dirty = B_TRUE;
|
|
|
|
|
DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
|
|
|
|
|
return (ERESTART);
|
|
|
|
|
}
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
|
|
|
|
|
tx->tx_needassign_txh = NULL;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* NB: No error returns are allowed after txg_hold_open, but
|
|
|
|
|
* before processing the dnode holds, due to the
|
|
|
|
|
* dmu_tx_unassign() logic.
|
|
|
|
|
*/
|
|
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
|
2008-11-20 23:01:55 +03:00
|
|
|
for (txh = list_head(&tx->tx_holds); txh;
|
|
|
|
|
txh = list_next(&tx->tx_holds, txh)) {
|
|
|
|
|
dnode_t *dn = txh->txh_dnode;
|
|
|
|
|
if (dn != NULL) {
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
|
if (dn->dn_assigned_txg == tx->tx_txg - 1) {
|
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
tx->tx_needassign_txh = txh;
|
2012-01-20 22:58:57 +04:00
|
|
|
DMU_TX_STAT_BUMP(dmu_tx_group);
|
2013-03-08 22:41:28 +04:00
|
|
|
return (SET_ERROR(ERESTART));
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
if (dn->dn_assigned_txg == 0)
|
|
|
|
|
dn->dn_assigned_txg = tx->tx_txg;
|
|
|
|
|
ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
|
|
|
|
|
(void) refcount_add(&dn->dn_tx_holds, tx);
|
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
}
|
|
|
|
|
towrite += txh->txh_space_towrite;
|
|
|
|
|
tofree += txh->txh_space_tofree;
|
|
|
|
|
tooverwrite += txh->txh_space_tooverwrite;
|
|
|
|
|
tounref += txh->txh_space_tounref;
|
2008-12-03 23:09:06 +03:00
|
|
|
tohold += txh->txh_memory_tohold;
|
|
|
|
|
fudge += txh->txh_fudge;
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* If a snapshot has been taken since we made our estimates,
|
|
|
|
|
* assume that we won't be able to free or overwrite anything.
|
|
|
|
|
*/
|
|
|
|
|
if (tx->tx_objset &&
|
2010-05-29 00:45:14 +04:00
|
|
|
dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
|
2008-11-20 23:01:55 +03:00
|
|
|
tx->tx_lastsnap_txg) {
|
|
|
|
|
towrite += tooverwrite;
|
|
|
|
|
tooverwrite = tofree = 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-03 23:09:06 +03:00
|
|
|
/* needed allocation: worst-case estimate of write space */
|
|
|
|
|
asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
|
|
|
|
|
/* freed space estimate: worst-case overwrite + free estimate */
|
2008-11-20 23:01:55 +03:00
|
|
|
fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
|
2008-12-03 23:09:06 +03:00
|
|
|
/* convert unrefd space to worst-case estimate */
|
2008-11-20 23:01:55 +03:00
|
|
|
usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
|
2008-12-03 23:09:06 +03:00
|
|
|
/* calculate memory footprint estimate */
|
|
|
|
|
memory = towrite + tooverwrite + tohold;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-12-03 23:09:06 +03:00
|
|
|
/*
|
|
|
|
|
* Add in 'tohold' to account for our dirty holds on this memory
|
|
|
|
|
* XXX - the "fudge" factor is to account for skipped blocks that
|
|
|
|
|
* we missed because dnode_next_offset() misses in-core-only blocks.
|
|
|
|
|
*/
|
|
|
|
|
tx->tx_space_towrite = asize +
|
|
|
|
|
spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
|
2008-11-20 23:01:55 +03:00
|
|
|
tx->tx_space_tofree = tofree;
|
|
|
|
|
tx->tx_space_tooverwrite = tooverwrite;
|
|
|
|
|
tx->tx_space_tounref = tounref;
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
if (tx->tx_dir && asize != 0) {
|
2008-12-03 23:09:06 +03:00
|
|
|
int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
|
|
|
|
|
asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
|
2008-11-20 23:01:55 +03:00
|
|
|
if (err)
|
|
|
|
|
return (err);
|
|
|
|
|
}
|
|
|
|
|
|
2012-01-20 22:58:57 +04:00
|
|
|
DMU_TX_STAT_BUMP(dmu_tx_assigned);
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
return (0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void
|
|
|
|
|
dmu_tx_unassign(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
|
|
|
|
|
if (tx->tx_txg == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
txg_rele_to_quiesce(&tx->tx_txgh);
|
|
|
|
|
|
2013-06-11 21:12:34 +04:00
|
|
|
/*
|
|
|
|
|
* Walk the transaction's hold list, removing the hold on the
|
|
|
|
|
* associated dnode, and notifying waiters if the refcount drops to 0.
|
|
|
|
|
*/
|
2008-11-20 23:01:55 +03:00
|
|
|
for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
|
|
|
|
|
txh = list_next(&tx->tx_holds, txh)) {
|
|
|
|
|
dnode_t *dn = txh->txh_dnode;
|
|
|
|
|
|
|
|
|
|
if (dn == NULL)
|
|
|
|
|
continue;
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
|
ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
|
|
|
|
|
|
|
|
|
|
if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
|
|
|
|
|
dn->dn_assigned_txg = 0;
|
|
|
|
|
cv_broadcast(&dn->dn_notxholds);
|
|
|
|
|
}
|
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
txg_rele_to_sync(&tx->tx_txgh);
|
|
|
|
|
|
|
|
|
|
tx->tx_lasttried_txg = tx->tx_txg;
|
|
|
|
|
tx->tx_txg = 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Assign tx to a transaction group. txg_how can be one of:
|
|
|
|
|
*
|
|
|
|
|
* (1) TXG_WAIT. If the current open txg is full, waits until there's
|
|
|
|
|
* a new one. This should be used when you're not holding locks.
|
2013-09-04 16:00:57 +04:00
|
|
|
* It will only fail if we're truly out of space (or over quota).
|
2008-11-20 23:01:55 +03:00
|
|
|
*
|
|
|
|
|
* (2) TXG_NOWAIT. If we can't assign into the current open txg without
|
|
|
|
|
* blocking, returns immediately with ERESTART. This should be used
|
|
|
|
|
* whenever you're holding locks. On an ERESTART error, the caller
|
|
|
|
|
* should drop locks, do a dmu_tx_wait(tx), and try again.
|
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
|
|
|
*
|
|
|
|
|
* (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
|
|
|
|
|
* has already been called on behalf of this operation (though
|
|
|
|
|
* most likely on a different tx).
|
2008-11-20 23:01:55 +03:00
|
|
|
*/
|
|
|
|
|
int
|
2013-09-04 16:00:57 +04:00
|
|
|
dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
|
2008-11-20 23:01:55 +03:00
|
|
|
{
|
|
|
|
|
int err;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 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
|
|
|
ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
|
|
|
|
|
txg_how == TXG_WAITED);
|
2008-11-20 23:01:55 +03:00
|
|
|
ASSERT(!dsl_pool_sync_context(tx->tx_pool));
|
|
|
|
|
|
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 (txg_how == TXG_WAITED)
|
|
|
|
|
tx->tx_waited = B_TRUE;
|
|
|
|
|
|
2013-09-04 16:00:57 +04:00
|
|
|
/* If we might wait, we must not hold the config lock. */
|
|
|
|
|
ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
|
|
|
|
|
dmu_tx_unassign(tx);
|
|
|
|
|
|
|
|
|
|
if (err != ERESTART || txg_how != TXG_WAIT)
|
|
|
|
|
return (err);
|
|
|
|
|
|
|
|
|
|
dmu_tx_wait(tx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
txg_rele_to_quiesce(&tx->tx_txgh);
|
|
|
|
|
|
|
|
|
|
return (0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_wait(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
spa_t *spa = tx->tx_pool->dp_spa;
|
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
|
|
|
dsl_pool_t *dp = tx->tx_pool;
|
Improve reporting of tx assignment wait times
Some callers of dmu_tx_assign() use the TXG_NOWAIT flag and call
dmu_tx_wait() themselves before retrying if the assignment fails.
The wait times for such callers are not accounted for in the
dmu_tx_assign kstat histogram, because the histogram only records
time spent in dmu_tx_assign(). This change moves the histogram
update to dmu_tx_wait() to properly account for all time spent there.
One downside of this approach is that it is possible to call
dmu_tx_wait() multiple times before successfully assigning a
transaction, in which case the cumulative wait time would not be
recorded. However, this case should not often arise in practice,
because most callers currently use one of these forms:
dmu_tx_assign(tx, TXG_WAIT);
dmu_tx_assign(tx, waited ? TXG_WAITED : TXG_NOWAIT);
The first form should make just one call to dmu_tx_delay() inside of
dmu_tx_assign(). The second form retries with TXG_WAITED if the first
assignment fails and incurs a delay, in which case no further waiting
is performed. Therefore transaction delays normally occur in one
call to dmu_tx_wait() so the histogram should be fairly accurate.
Another possible downside of this approach is that the histogram will
no longer record overhead outside of dmu_tx_wait() such as in
dmu_tx_try_assign(). While I'm not aware of any reason for concern on
this point, it is conceivable that lock contention, long list
traversal, etc. could cause assignment delays that would not be
reflected in the histogram. Therefore the histogram should strictly
be used for visibility in to the normal delay mechanisms and not as a
profiling tool for code performance.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1915
2014-03-01 03:07:00 +04:00
|
|
|
hrtime_t before;
|
2008-11-20 23:01:55 +03:00
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
2013-09-04 16:00:57 +04:00
|
|
|
ASSERT(!dsl_pool_config_held(tx->tx_pool));
|
2008-11-20 23:01:55 +03:00
|
|
|
|
Improve reporting of tx assignment wait times
Some callers of dmu_tx_assign() use the TXG_NOWAIT flag and call
dmu_tx_wait() themselves before retrying if the assignment fails.
The wait times for such callers are not accounted for in the
dmu_tx_assign kstat histogram, because the histogram only records
time spent in dmu_tx_assign(). This change moves the histogram
update to dmu_tx_wait() to properly account for all time spent there.
One downside of this approach is that it is possible to call
dmu_tx_wait() multiple times before successfully assigning a
transaction, in which case the cumulative wait time would not be
recorded. However, this case should not often arise in practice,
because most callers currently use one of these forms:
dmu_tx_assign(tx, TXG_WAIT);
dmu_tx_assign(tx, waited ? TXG_WAITED : TXG_NOWAIT);
The first form should make just one call to dmu_tx_delay() inside of
dmu_tx_assign(). The second form retries with TXG_WAITED if the first
assignment fails and incurs a delay, in which case no further waiting
is performed. Therefore transaction delays normally occur in one
call to dmu_tx_wait() so the histogram should be fairly accurate.
Another possible downside of this approach is that the histogram will
no longer record overhead outside of dmu_tx_wait() such as in
dmu_tx_try_assign(). While I'm not aware of any reason for concern on
this point, it is conceivable that lock contention, long list
traversal, etc. could cause assignment delays that would not be
reflected in the histogram. Therefore the histogram should strictly
be used for visibility in to the normal delay mechanisms and not as a
profiling tool for code performance.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1915
2014-03-01 03:07:00 +04:00
|
|
|
before = gethrtime();
|
|
|
|
|
|
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 (tx->tx_wait_dirty) {
|
|
|
|
|
uint64_t dirty;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* dmu_tx_try_assign() has determined that we need to wait
|
|
|
|
|
* because we've consumed much or all of the dirty buffer
|
|
|
|
|
* space.
|
|
|
|
|
*/
|
|
|
|
|
mutex_enter(&dp->dp_lock);
|
|
|
|
|
if (dp->dp_dirty_total >= zfs_dirty_data_max)
|
|
|
|
|
DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
|
|
|
|
|
while (dp->dp_dirty_total >= zfs_dirty_data_max)
|
|
|
|
|
cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
|
|
|
|
|
dirty = dp->dp_dirty_total;
|
|
|
|
|
mutex_exit(&dp->dp_lock);
|
|
|
|
|
|
|
|
|
|
dmu_tx_delay(tx, dirty);
|
|
|
|
|
|
|
|
|
|
tx->tx_wait_dirty = B_FALSE;
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Note: setting tx_waited only has effect if the caller
|
|
|
|
|
* used TX_WAIT. Otherwise they are going to destroy
|
|
|
|
|
* this tx and try again. The common case, zfs_write(),
|
|
|
|
|
* uses TX_WAIT.
|
|
|
|
|
*/
|
|
|
|
|
tx->tx_waited = B_TRUE;
|
|
|
|
|
} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
|
|
|
|
|
/*
|
|
|
|
|
* If the pool is suspended we need to wait until it
|
|
|
|
|
* is resumed. Note that it's possible that the pool
|
|
|
|
|
* has become active after this thread has tried to
|
|
|
|
|
* obtain a tx. If that's the case then tx_lasttried_txg
|
|
|
|
|
* would not have been set.
|
|
|
|
|
*/
|
|
|
|
|
txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
|
2008-11-20 23:01:55 +03:00
|
|
|
} else if (tx->tx_needassign_txh) {
|
|
|
|
|
dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
|
|
|
|
|
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
|
while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
|
|
|
|
|
cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
|
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
tx->tx_needassign_txh = NULL;
|
|
|
|
|
} else {
|
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
|
|
|
/*
|
|
|
|
|
* A dnode is assigned to the quiescing txg. Wait for its
|
|
|
|
|
* transaction to complete.
|
|
|
|
|
*/
|
2008-11-20 23:01:55 +03:00
|
|
|
txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
|
|
|
|
|
}
|
Improve reporting of tx assignment wait times
Some callers of dmu_tx_assign() use the TXG_NOWAIT flag and call
dmu_tx_wait() themselves before retrying if the assignment fails.
The wait times for such callers are not accounted for in the
dmu_tx_assign kstat histogram, because the histogram only records
time spent in dmu_tx_assign(). This change moves the histogram
update to dmu_tx_wait() to properly account for all time spent there.
One downside of this approach is that it is possible to call
dmu_tx_wait() multiple times before successfully assigning a
transaction, in which case the cumulative wait time would not be
recorded. However, this case should not often arise in practice,
because most callers currently use one of these forms:
dmu_tx_assign(tx, TXG_WAIT);
dmu_tx_assign(tx, waited ? TXG_WAITED : TXG_NOWAIT);
The first form should make just one call to dmu_tx_delay() inside of
dmu_tx_assign(). The second form retries with TXG_WAITED if the first
assignment fails and incurs a delay, in which case no further waiting
is performed. Therefore transaction delays normally occur in one
call to dmu_tx_wait() so the histogram should be fairly accurate.
Another possible downside of this approach is that the histogram will
no longer record overhead outside of dmu_tx_wait() such as in
dmu_tx_try_assign(). While I'm not aware of any reason for concern on
this point, it is conceivable that lock contention, long list
traversal, etc. could cause assignment delays that would not be
reflected in the histogram. Therefore the histogram should strictly
be used for visibility in to the normal delay mechanisms and not as a
profiling tool for code performance.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1915
2014-03-01 03:07:00 +04:00
|
|
|
|
|
|
|
|
spa_tx_assign_add_nsecs(spa, gethrtime() - before);
|
2008-11-20 23:01:55 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
|
|
|
|
|
{
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-11-20 23:01:55 +03:00
|
|
|
if (tx->tx_dir == NULL || delta == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (delta > 0) {
|
|
|
|
|
ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
|
|
|
|
|
tx->tx_space_towrite);
|
|
|
|
|
(void) refcount_add_many(&tx->tx_space_written, delta, NULL);
|
|
|
|
|
} else {
|
|
|
|
|
(void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_commit(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg != 0);
|
|
|
|
|
|
2013-06-11 21:12:34 +04:00
|
|
|
/*
|
|
|
|
|
* Go through the transaction's hold list and remove holds on
|
|
|
|
|
* associated dnodes, notifying waiters if no holds remain.
|
|
|
|
|
*/
|
2010-08-26 20:52:42 +04:00
|
|
|
while ((txh = list_head(&tx->tx_holds))) {
|
2008-11-20 23:01:55 +03:00
|
|
|
dnode_t *dn = txh->txh_dnode;
|
|
|
|
|
|
|
|
|
|
list_remove(&tx->tx_holds, txh);
|
|
|
|
|
kmem_free(txh, sizeof (dmu_tx_hold_t));
|
|
|
|
|
if (dn == NULL)
|
|
|
|
|
continue;
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
|
ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
|
|
|
|
|
|
|
|
|
|
if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
|
|
|
|
|
dn->dn_assigned_txg = 0;
|
|
|
|
|
cv_broadcast(&dn->dn_notxholds);
|
|
|
|
|
}
|
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
dnode_rele(dn, tx);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (tx->tx_tempreserve_cookie)
|
|
|
|
|
dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
|
|
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
if (!list_is_empty(&tx->tx_callbacks))
|
|
|
|
|
txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
|
|
|
|
|
|
2008-11-20 23:01:55 +03:00
|
|
|
if (tx->tx_anyobj == FALSE)
|
|
|
|
|
txg_rele_to_sync(&tx->tx_txgh);
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
list_destroy(&tx->tx_callbacks);
|
2008-11-20 23:01:55 +03:00
|
|
|
list_destroy(&tx->tx_holds);
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-11-20 23:01:55 +03:00
|
|
|
dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
|
|
|
|
|
tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
|
|
|
|
|
tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
|
|
|
|
|
refcount_destroy_many(&tx->tx_space_written,
|
|
|
|
|
refcount_count(&tx->tx_space_written));
|
|
|
|
|
refcount_destroy_many(&tx->tx_space_freed,
|
|
|
|
|
refcount_count(&tx->tx_space_freed));
|
|
|
|
|
#endif
|
|
|
|
|
kmem_free(tx, sizeof (dmu_tx_t));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_abort(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
|
2010-08-26 20:52:42 +04:00
|
|
|
while ((txh = list_head(&tx->tx_holds))) {
|
2008-11-20 23:01:55 +03:00
|
|
|
dnode_t *dn = txh->txh_dnode;
|
|
|
|
|
|
|
|
|
|
list_remove(&tx->tx_holds, txh);
|
|
|
|
|
kmem_free(txh, sizeof (dmu_tx_hold_t));
|
|
|
|
|
if (dn != NULL)
|
|
|
|
|
dnode_rele(dn, tx);
|
|
|
|
|
}
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Call any registered callbacks with an error code.
|
|
|
|
|
*/
|
|
|
|
|
if (!list_is_empty(&tx->tx_callbacks))
|
|
|
|
|
dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
|
|
|
|
|
|
|
|
|
|
list_destroy(&tx->tx_callbacks);
|
2008-11-20 23:01:55 +03:00
|
|
|
list_destroy(&tx->tx_holds);
|
2012-03-21 03:00:17 +04:00
|
|
|
#ifdef DEBUG_DMU_TX
|
2008-11-20 23:01:55 +03:00
|
|
|
refcount_destroy_many(&tx->tx_space_written,
|
|
|
|
|
refcount_count(&tx->tx_space_written));
|
|
|
|
|
refcount_destroy_many(&tx->tx_space_freed,
|
|
|
|
|
refcount_count(&tx->tx_space_freed));
|
|
|
|
|
#endif
|
|
|
|
|
kmem_free(tx, sizeof (dmu_tx_t));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
uint64_t
|
|
|
|
|
dmu_tx_get_txg(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
ASSERT(tx->tx_txg != 0);
|
|
|
|
|
return (tx->tx_txg);
|
|
|
|
|
}
|
2010-05-29 00:45:14 +04:00
|
|
|
|
2013-09-04 16:00:57 +04:00
|
|
|
dsl_pool_t *
|
|
|
|
|
dmu_tx_pool(dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
ASSERT(tx->tx_pool != NULL);
|
|
|
|
|
return (tx->tx_pool);
|
|
|
|
|
}
|
|
|
|
|
|
2010-05-29 00:45:14 +04:00
|
|
|
void
|
|
|
|
|
dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_callback_t *dcb;
|
|
|
|
|
|
2012-05-07 21:49:51 +04:00
|
|
|
dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_PUSHPAGE);
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
dcb->dcb_func = func;
|
|
|
|
|
dcb->dcb_data = data;
|
|
|
|
|
|
|
|
|
|
list_insert_tail(&tx->tx_callbacks, dcb);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Call all the commit callbacks on a list, with a given error code.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_do_callbacks(list_t *cb_list, int error)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_callback_t *dcb;
|
|
|
|
|
|
2010-08-26 20:52:42 +04:00
|
|
|
while ((dcb = list_head(cb_list))) {
|
2010-05-29 00:45:14 +04:00
|
|
|
list_remove(cb_list, dcb);
|
|
|
|
|
dcb->dcb_func(dcb->dcb_data, error);
|
|
|
|
|
kmem_free(dcb, sizeof (dmu_tx_callback_t));
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Interface to hold a bunch of attributes.
|
|
|
|
|
* used for creating new files.
|
|
|
|
|
* attrsize is the total size of all attributes
|
|
|
|
|
* to be added during object creation
|
|
|
|
|
*
|
|
|
|
|
* For updating/adding a single attribute dmu_tx_hold_sa() should be used.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* hold necessary attribute name for attribute registration.
|
|
|
|
|
* should be a very rare case where this is needed. If it does
|
|
|
|
|
* happen it would only happen on the first write to the file system.
|
|
|
|
|
*/
|
|
|
|
|
static void
|
|
|
|
|
dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
|
|
|
|
|
{
|
|
|
|
|
int i;
|
|
|
|
|
|
|
|
|
|
if (!sa->sa_need_attr_registration)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
for (i = 0; i != sa->sa_num_attrs; i++) {
|
|
|
|
|
if (!sa->sa_attr_table[i].sa_registered) {
|
|
|
|
|
if (sa->sa_reg_attr_obj)
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
|
|
|
|
|
B_TRUE, sa->sa_attr_table[i].sa_name);
|
|
|
|
|
else
|
|
|
|
|
dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
|
|
|
|
|
B_TRUE, sa->sa_attr_table[i].sa_name);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
|
|
|
|
|
{
|
|
|
|
|
dnode_t *dn;
|
|
|
|
|
dmu_tx_hold_t *txh;
|
|
|
|
|
|
|
|
|
|
txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
|
|
|
|
|
THT_SPILL, 0, 0);
|
2013-07-23 21:32:57 +04:00
|
|
|
if (txh == NULL)
|
|
|
|
|
return;
|
2010-05-29 00:45:14 +04:00
|
|
|
|
|
|
|
|
dn = txh->txh_dnode;
|
|
|
|
|
|
|
|
|
|
if (dn == NULL)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/* If blkptr doesn't exist then add space to towrite */
|
2012-04-08 21:10:49 +04:00
|
|
|
if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
|
2010-05-29 00:45:14 +04:00
|
|
|
txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
|
|
|
|
|
} else {
|
2012-04-08 21:10:49 +04:00
|
|
|
blkptr_t *bp;
|
|
|
|
|
|
|
|
|
|
bp = &dn->dn_phys->dn_spill;
|
2010-05-29 00:45:14 +04:00
|
|
|
if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
|
|
|
|
|
bp, bp->blk_birth))
|
|
|
|
|
txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
|
|
|
|
|
else
|
|
|
|
|
txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
|
2012-04-08 21:10:49 +04:00
|
|
|
if (!BP_IS_HOLE(bp))
|
2010-05-29 00:45:14 +04:00
|
|
|
txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
|
|
|
|
|
{
|
|
|
|
|
sa_os_t *sa = tx->tx_objset->os_sa;
|
|
|
|
|
|
|
|
|
|
dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
|
|
|
|
|
|
|
|
|
|
if (tx->tx_objset->os_sa->sa_master_obj == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (tx->tx_objset->os_sa->sa_layout_attr_obj)
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
|
|
|
|
|
else {
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
|
|
|
|
|
dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
|
|
|
|
|
dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dmu_tx_sa_registration_hold(sa, tx);
|
|
|
|
|
|
|
|
|
|
if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
|
|
|
|
|
THT_SPILL, 0, 0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* Hold SA attribute
|
|
|
|
|
*
|
|
|
|
|
* dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
|
|
|
|
|
*
|
|
|
|
|
* variable_size is the total size of all variable sized attributes
|
|
|
|
|
* passed to this function. It is not the total size of all
|
|
|
|
|
* variable size attributes that *may* exist on this object.
|
|
|
|
|
*/
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
|
|
|
|
|
{
|
|
|
|
|
uint64_t object;
|
|
|
|
|
sa_os_t *sa = tx->tx_objset->os_sa;
|
|
|
|
|
|
|
|
|
|
ASSERT(hdl != NULL);
|
|
|
|
|
|
|
|
|
|
object = sa_handle_object(hdl);
|
|
|
|
|
|
|
|
|
|
dmu_tx_hold_bonus(tx, object);
|
|
|
|
|
|
|
|
|
|
if (tx->tx_objset->os_sa->sa_master_obj == 0)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
|
|
|
|
|
tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
|
|
|
|
|
dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
|
|
|
|
|
dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dmu_tx_sa_registration_hold(sa, tx);
|
|
|
|
|
|
|
|
|
|
if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
|
|
|
|
|
dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
|
|
|
|
|
|
2010-08-27 01:24:34 +04:00
|
|
|
if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
|
2010-05-29 00:45:14 +04:00
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
dmu_tx_hold_spill(tx, object);
|
2010-08-27 01:24:34 +04:00
|
|
|
} else {
|
|
|
|
|
dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
|
|
|
|
|
dnode_t *dn;
|
|
|
|
|
|
|
|
|
|
DB_DNODE_ENTER(db);
|
|
|
|
|
dn = DB_DNODE(db);
|
|
|
|
|
if (dn->dn_have_spill) {
|
|
|
|
|
ASSERT(tx->tx_txg == 0);
|
|
|
|
|
dmu_tx_hold_spill(tx, object);
|
|
|
|
|
}
|
|
|
|
|
DB_DNODE_EXIT(db);
|
2010-05-29 00:45:14 +04:00
|
|
|
}
|
|
|
|
|
}
|
2010-08-26 22:49:16 +04:00
|
|
|
|
2012-01-20 22:58:57 +04:00
|
|
|
void
|
|
|
|
|
dmu_tx_init(void)
|
|
|
|
|
{
|
|
|
|
|
dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
|
|
|
|
|
KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
|
|
|
|
|
KSTAT_FLAG_VIRTUAL);
|
|
|
|
|
|
|
|
|
|
if (dmu_tx_ksp != NULL) {
|
|
|
|
|
dmu_tx_ksp->ks_data = &dmu_tx_stats;
|
|
|
|
|
kstat_install(dmu_tx_ksp);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void
|
|
|
|
|
dmu_tx_fini(void)
|
|
|
|
|
{
|
|
|
|
|
if (dmu_tx_ksp != NULL) {
|
|
|
|
|
kstat_delete(dmu_tx_ksp);
|
|
|
|
|
dmu_tx_ksp = NULL;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2010-08-26 22:49:16 +04:00
|
|
|
#if defined(_KERNEL) && defined(HAVE_SPL)
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_create);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_write);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_free);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_zap);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_bonus);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_abort);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_assign);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_wait);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_commit);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_get_txg);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_callback_register);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_do_callbacks);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_spill);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_sa_create);
|
|
|
|
|
EXPORT_SYMBOL(dmu_tx_hold_sa);
|
|
|
|
|
#endif
|