mirror_zfs/module/zfs/metaslab.c

1982 lines
52 KiB
C
Raw Normal View History

2008-11-20 23:01:55 +03:00
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2012 by Delphix. All rights reserved.
2008-11-20 23:01:55 +03:00
*/
#include <sys/zfs_context.h>
#include <sys/dmu.h>
#include <sys/dmu_tx.h>
#include <sys/space_map.h>
#include <sys/metaslab_impl.h>
#include <sys/vdev_impl.h>
#include <sys/zio.h>
#define WITH_DF_BLOCK_ALLOCATOR
/*
* Allow allocations to switch to gang blocks quickly. We do this to
* avoid having to load lots of space_maps in a given txg. There are,
* however, some cases where we want to avoid "fast" ganging and instead
* we want to do an exhaustive search of all metaslabs on this device.
* Currently we don't allow any gang, zil, or dump device related allocations
* to "fast" gang.
*/
#define CAN_FASTGANG(flags) \
(!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
METASLAB_GANG_AVOID)))
2008-11-20 23:01:55 +03:00
uint64_t metaslab_aliquot = 512ULL << 10;
uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
/*
* The in-core space map representation is more compact than its on-disk form.
* The zfs_condense_pct determines how much more compact the in-core
* space_map representation must be before we compact it on-disk.
* Values should be greater than or equal to 100.
*/
int zfs_condense_pct = 200;
/*
* This value defines the number of allowed allocation failures per vdev.
* If a device reaches this threshold in a given txg then we consider skipping
* allocations on that device.
*/
int zfs_mg_alloc_failures;
/*
* Metaslab debugging: when set, keeps all space maps in core to verify frees.
*/
int metaslab_debug = 0;
2009-07-03 02:44:48 +04:00
/*
* Minimum size which forces the dynamic allocator to change
* it's allocation strategy. Once the space map cannot satisfy
2009-07-03 02:44:48 +04:00
* an allocation of this size then it switches to using more
* aggressive strategy (i.e search by size rather than offset).
*/
uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
/*
* The minimum free space, in percent, which must be available
* in a space map to continue allocations in a first-fit fashion.
* Once the space_map's free space drops below this level we dynamically
* switch to using best-fit allocations.
*/
int metaslab_df_free_pct = 4;
/*
* A metaslab is considered "free" if it contains a contiguous
* segment which is greater than metaslab_min_alloc_size.
*/
uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
/*
* Max number of space_maps to prefetch.
*/
int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
/*
* Percentage bonus multiplier for metaslabs that are in the bonus area.
*/
int metaslab_smo_bonus_pct = 150;
2009-07-03 02:44:48 +04:00
2008-11-20 23:01:55 +03:00
/*
* ==========================================================================
* Metaslab classes
* ==========================================================================
*/
metaslab_class_t *
metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
2008-11-20 23:01:55 +03:00
{
metaslab_class_t *mc;
mc = kmem_zalloc(sizeof (metaslab_class_t), KM_PUSHPAGE);
2008-11-20 23:01:55 +03:00
mc->mc_spa = spa;
2008-11-20 23:01:55 +03:00
mc->mc_rotor = NULL;
2009-07-03 02:44:48 +04:00
mc->mc_ops = ops;
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
mutex_init(&mc->mc_fastwrite_lock, NULL, MUTEX_DEFAULT, NULL);
2008-11-20 23:01:55 +03:00
return (mc);
}
void
metaslab_class_destroy(metaslab_class_t *mc)
{
ASSERT(mc->mc_rotor == NULL);
ASSERT(mc->mc_alloc == 0);
ASSERT(mc->mc_deferred == 0);
ASSERT(mc->mc_space == 0);
ASSERT(mc->mc_dspace == 0);
2008-11-20 23:01:55 +03:00
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
mutex_destroy(&mc->mc_fastwrite_lock);
2008-11-20 23:01:55 +03:00
kmem_free(mc, sizeof (metaslab_class_t));
}
int
metaslab_class_validate(metaslab_class_t *mc)
2008-11-20 23:01:55 +03:00
{
metaslab_group_t *mg;
vdev_t *vd;
2008-11-20 23:01:55 +03:00
/*
* Must hold one of the spa_config locks.
*/
ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
2008-11-20 23:01:55 +03:00
if ((mg = mc->mc_rotor) == NULL)
return (0);
do {
vd = mg->mg_vd;
ASSERT(vd->vdev_mg != NULL);
ASSERT3P(vd->vdev_top, ==, vd);
ASSERT3P(mg->mg_class, ==, mc);
ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
} while ((mg = mg->mg_next) != mc->mc_rotor);
return (0);
2008-11-20 23:01:55 +03:00
}
void
metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
2008-11-20 23:01:55 +03:00
{
atomic_add_64(&mc->mc_alloc, alloc_delta);
atomic_add_64(&mc->mc_deferred, defer_delta);
atomic_add_64(&mc->mc_space, space_delta);
atomic_add_64(&mc->mc_dspace, dspace_delta);
}
2008-11-20 23:01:55 +03:00
uint64_t
metaslab_class_get_alloc(metaslab_class_t *mc)
{
return (mc->mc_alloc);
}
2008-11-20 23:01:55 +03:00
uint64_t
metaslab_class_get_deferred(metaslab_class_t *mc)
{
return (mc->mc_deferred);
}
2008-11-20 23:01:55 +03:00
uint64_t
metaslab_class_get_space(metaslab_class_t *mc)
{
return (mc->mc_space);
}
2008-11-20 23:01:55 +03:00
uint64_t
metaslab_class_get_dspace(metaslab_class_t *mc)
{
return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
2008-11-20 23:01:55 +03:00
}
/*
* ==========================================================================
* Metaslab groups
* ==========================================================================
*/
static int
metaslab_compare(const void *x1, const void *x2)
{
const metaslab_t *m1 = x1;
const metaslab_t *m2 = x2;
if (m1->ms_weight < m2->ms_weight)
return (1);
if (m1->ms_weight > m2->ms_weight)
return (-1);
/*
* If the weights are identical, use the offset to force uniqueness.
*/
if (m1->ms_map->sm_start < m2->ms_map->sm_start)
2008-11-20 23:01:55 +03:00
return (-1);
if (m1->ms_map->sm_start > m2->ms_map->sm_start)
2008-11-20 23:01:55 +03:00
return (1);
ASSERT3P(m1, ==, m2);
return (0);
}
metaslab_group_t *
metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
{
metaslab_group_t *mg;
mg = kmem_zalloc(sizeof (metaslab_group_t), KM_PUSHPAGE);
2008-11-20 23:01:55 +03:00
mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
avl_create(&mg->mg_metaslab_tree, metaslab_compare,
sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
mg->mg_vd = vd;
mg->mg_class = mc;
mg->mg_activation_count = 0;
2008-11-20 23:01:55 +03:00
return (mg);
}
void
metaslab_group_destroy(metaslab_group_t *mg)
{
ASSERT(mg->mg_prev == NULL);
ASSERT(mg->mg_next == NULL);
/*
* We may have gone below zero with the activation count
* either because we never activated in the first place or
* because we're done, and possibly removing the vdev.
*/
ASSERT(mg->mg_activation_count <= 0);
2008-11-20 23:01:55 +03:00
avl_destroy(&mg->mg_metaslab_tree);
mutex_destroy(&mg->mg_lock);
kmem_free(mg, sizeof (metaslab_group_t));
}
void
metaslab_group_activate(metaslab_group_t *mg)
{
metaslab_class_t *mc = mg->mg_class;
metaslab_group_t *mgprev, *mgnext;
ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
ASSERT(mc->mc_rotor != mg);
ASSERT(mg->mg_prev == NULL);
ASSERT(mg->mg_next == NULL);
ASSERT(mg->mg_activation_count <= 0);
if (++mg->mg_activation_count <= 0)
return;
mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
if ((mgprev = mc->mc_rotor) == NULL) {
mg->mg_prev = mg;
mg->mg_next = mg;
} else {
mgnext = mgprev->mg_next;
mg->mg_prev = mgprev;
mg->mg_next = mgnext;
mgprev->mg_next = mg;
mgnext->mg_prev = mg;
}
mc->mc_rotor = mg;
}
void
metaslab_group_passivate(metaslab_group_t *mg)
{
metaslab_class_t *mc = mg->mg_class;
metaslab_group_t *mgprev, *mgnext;
ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
if (--mg->mg_activation_count != 0) {
ASSERT(mc->mc_rotor != mg);
ASSERT(mg->mg_prev == NULL);
ASSERT(mg->mg_next == NULL);
ASSERT(mg->mg_activation_count < 0);
return;
}
mgprev = mg->mg_prev;
mgnext = mg->mg_next;
if (mg == mgnext) {
mc->mc_rotor = NULL;
} else {
mc->mc_rotor = mgnext;
mgprev->mg_next = mgnext;
mgnext->mg_prev = mgprev;
}
mg->mg_prev = NULL;
mg->mg_next = NULL;
}
2008-11-20 23:01:55 +03:00
static void
metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
{
mutex_enter(&mg->mg_lock);
ASSERT(msp->ms_group == NULL);
msp->ms_group = mg;
msp->ms_weight = 0;
avl_add(&mg->mg_metaslab_tree, msp);
mutex_exit(&mg->mg_lock);
}
static void
metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
{
mutex_enter(&mg->mg_lock);
ASSERT(msp->ms_group == mg);
avl_remove(&mg->mg_metaslab_tree, msp);
msp->ms_group = NULL;
mutex_exit(&mg->mg_lock);
}
static void
metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
{
/*
* Although in principle the weight can be any value, in
* practice we do not use values in the range [1, 510].
*/
ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
ASSERT(MUTEX_HELD(&msp->ms_lock));
mutex_enter(&mg->mg_lock);
ASSERT(msp->ms_group == mg);
avl_remove(&mg->mg_metaslab_tree, msp);
msp->ms_weight = weight;
avl_add(&mg->mg_metaslab_tree, msp);
mutex_exit(&mg->mg_lock);
}
/*
* ==========================================================================
* Common allocator routines
* ==========================================================================
*/
static int
metaslab_segsize_compare(const void *x1, const void *x2)
{
const space_seg_t *s1 = x1;
const space_seg_t *s2 = x2;
uint64_t ss_size1 = s1->ss_end - s1->ss_start;
uint64_t ss_size2 = s2->ss_end - s2->ss_start;
if (ss_size1 < ss_size2)
return (-1);
if (ss_size1 > ss_size2)
return (1);
if (s1->ss_start < s2->ss_start)
return (-1);
if (s1->ss_start > s2->ss_start)
return (1);
return (0);
}
#if defined(WITH_FF_BLOCK_ALLOCATOR) || \
defined(WITH_DF_BLOCK_ALLOCATOR) || \
defined(WITH_CDF_BLOCK_ALLOCATOR)
2008-11-20 23:01:55 +03:00
/*
2009-07-03 02:44:48 +04:00
* This is a helper function that can be used by the allocator to find
* a suitable block to allocate. This will search the specified AVL
* tree looking for a block that matches the specified criteria.
2008-11-20 23:01:55 +03:00
*/
static uint64_t
2009-07-03 02:44:48 +04:00
metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
uint64_t align)
2008-11-20 23:01:55 +03:00
{
space_seg_t *ss, ssearch;
avl_index_t where;
ssearch.ss_start = *cursor;
ssearch.ss_end = *cursor + size;
ss = avl_find(t, &ssearch, &where);
if (ss == NULL)
ss = avl_nearest(t, where, AVL_AFTER);
while (ss != NULL) {
uint64_t offset = P2ROUNDUP(ss->ss_start, align);
if (offset + size <= ss->ss_end) {
*cursor = offset + size;
return (offset);
}
ss = AVL_NEXT(t, ss);
}
/*
* If we know we've searched the whole map (*cursor == 0), give up.
* Otherwise, reset the cursor to the beginning and try again.
*/
if (*cursor == 0)
return (-1ULL);
*cursor = 0;
2009-07-03 02:44:48 +04:00
return (metaslab_block_picker(t, cursor, size, align));
}
#endif /* WITH_FF/DF/CDF_BLOCK_ALLOCATOR */
2009-07-03 02:44:48 +04:00
static void
metaslab_pp_load(space_map_t *sm)
2009-07-03 02:44:48 +04:00
{
space_seg_t *ss;
2009-07-03 02:44:48 +04:00
ASSERT(sm->sm_ppd == NULL);
sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_PUSHPAGE);
sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_PUSHPAGE);
avl_create(sm->sm_pp_root, metaslab_segsize_compare,
sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
avl_add(sm->sm_pp_root, ss);
2009-07-03 02:44:48 +04:00
}
static void
metaslab_pp_unload(space_map_t *sm)
2009-07-03 02:44:48 +04:00
{
void *cookie = NULL;
2009-07-03 02:44:48 +04:00
kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
sm->sm_ppd = NULL;
while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
/* tear down the tree */
}
2009-07-03 02:44:48 +04:00
avl_destroy(sm->sm_pp_root);
kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
sm->sm_pp_root = NULL;
2008-11-20 23:01:55 +03:00
}
/* ARGSUSED */
static void
metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
2008-11-20 23:01:55 +03:00
{
/* No need to update cursor */
}
/* ARGSUSED */
static void
metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
2008-11-20 23:01:55 +03:00
{
/* No need to update cursor */
}
2009-07-03 02:44:48 +04:00
/*
* Return the maximum contiguous segment within the metaslab.
2009-07-03 02:44:48 +04:00
*/
uint64_t
metaslab_pp_maxsize(space_map_t *sm)
2009-07-03 02:44:48 +04:00
{
avl_tree_t *t = sm->sm_pp_root;
space_seg_t *ss;
if (t == NULL || (ss = avl_last(t)) == NULL)
return (0ULL);
return (ss->ss_end - ss->ss_start);
}
#if defined(WITH_FF_BLOCK_ALLOCATOR)
/*
* ==========================================================================
* The first-fit block allocator
* ==========================================================================
*/
static uint64_t
metaslab_ff_alloc(space_map_t *sm, uint64_t size)
2009-07-03 02:44:48 +04:00
{
avl_tree_t *t = &sm->sm_root;
uint64_t align = size & -size;
uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
2009-07-03 02:44:48 +04:00
return (metaslab_block_picker(t, cursor, size, align));
2009-07-03 02:44:48 +04:00
}
/* ARGSUSED */
boolean_t
metaslab_ff_fragmented(space_map_t *sm)
2009-07-03 02:44:48 +04:00
{
return (B_TRUE);
2009-07-03 02:44:48 +04:00
}
static space_map_ops_t metaslab_ff_ops = {
metaslab_pp_load,
metaslab_pp_unload,
metaslab_ff_alloc,
metaslab_pp_claim,
metaslab_pp_free,
metaslab_pp_maxsize,
metaslab_ff_fragmented
};
2009-07-03 02:44:48 +04:00
space_map_ops_t *zfs_metaslab_ops = &metaslab_ff_ops;
#endif /* WITH_FF_BLOCK_ALLOCATOR */
#if defined(WITH_DF_BLOCK_ALLOCATOR)
/*
* ==========================================================================
* Dynamic block allocator -
* Uses the first fit allocation scheme until space get low and then
* adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
* and metaslab_df_free_pct to determine when to switch the allocation scheme.
* ==========================================================================
*/
2009-07-03 02:44:48 +04:00
static uint64_t
metaslab_df_alloc(space_map_t *sm, uint64_t size)
{
avl_tree_t *t = &sm->sm_root;
uint64_t align = size & -size;
uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
uint64_t max_size = metaslab_pp_maxsize(sm);
2009-07-03 02:44:48 +04:00
int free_pct = sm->sm_space * 100 / sm->sm_size;
ASSERT(MUTEX_HELD(sm->sm_lock));
ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
if (max_size < size)
return (-1ULL);
/*
* If we're running low on space switch to using the size
* sorted AVL tree (best-fit).
*/
if (max_size < metaslab_df_alloc_threshold ||
free_pct < metaslab_df_free_pct) {
t = sm->sm_pp_root;
*cursor = 0;
}
return (metaslab_block_picker(t, cursor, size, 1ULL));
}
static boolean_t
metaslab_df_fragmented(space_map_t *sm)
2009-07-03 02:44:48 +04:00
{
uint64_t max_size = metaslab_pp_maxsize(sm);
int free_pct = sm->sm_space * 100 / sm->sm_size;
2009-07-03 02:44:48 +04:00
if (max_size >= metaslab_df_alloc_threshold &&
free_pct >= metaslab_df_free_pct)
return (B_FALSE);
return (B_TRUE);
2009-07-03 02:44:48 +04:00
}
static space_map_ops_t metaslab_df_ops = {
metaslab_pp_load,
metaslab_pp_unload,
2009-07-03 02:44:48 +04:00
metaslab_df_alloc,
metaslab_pp_claim,
metaslab_pp_free,
metaslab_pp_maxsize,
metaslab_df_fragmented
2008-11-20 23:01:55 +03:00
};
space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
#endif /* WITH_DF_BLOCK_ALLOCATOR */
/*
* ==========================================================================
* Other experimental allocators
* ==========================================================================
*/
#if defined(WITH_CDF_BLOCK_ALLOCATOR)
static uint64_t
metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
{
avl_tree_t *t = &sm->sm_root;
uint64_t *cursor = (uint64_t *)sm->sm_ppd;
uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
uint64_t max_size = metaslab_pp_maxsize(sm);
uint64_t rsize = size;
uint64_t offset = 0;
ASSERT(MUTEX_HELD(sm->sm_lock));
ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
if (max_size < size)
return (-1ULL);
ASSERT3U(*extent_end, >=, *cursor);
/*
* If we're running low on space switch to using the size
* sorted AVL tree (best-fit).
*/
if ((*cursor + size) > *extent_end) {
t = sm->sm_pp_root;
*cursor = *extent_end = 0;
if (max_size > 2 * SPA_MAXBLOCKSIZE)
rsize = MIN(metaslab_min_alloc_size, max_size);
offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
if (offset != -1)
*cursor = offset + size;
} else {
offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
}
ASSERT3U(*cursor, <=, *extent_end);
return (offset);
}
static boolean_t
metaslab_cdf_fragmented(space_map_t *sm)
{
uint64_t max_size = metaslab_pp_maxsize(sm);
if (max_size > (metaslab_min_alloc_size * 10))
return (B_FALSE);
return (B_TRUE);
}
static space_map_ops_t metaslab_cdf_ops = {
metaslab_pp_load,
metaslab_pp_unload,
metaslab_cdf_alloc,
metaslab_pp_claim,
metaslab_pp_free,
metaslab_pp_maxsize,
metaslab_cdf_fragmented
};
space_map_ops_t *zfs_metaslab_ops = &metaslab_cdf_ops;
#endif /* WITH_CDF_BLOCK_ALLOCATOR */
#if defined(WITH_NDF_BLOCK_ALLOCATOR)
uint64_t metaslab_ndf_clump_shift = 4;
static uint64_t
metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
{
avl_tree_t *t = &sm->sm_root;
avl_index_t where;
space_seg_t *ss, ssearch;
uint64_t hbit = highbit(size);
uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
uint64_t max_size = metaslab_pp_maxsize(sm);
ASSERT(MUTEX_HELD(sm->sm_lock));
ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
if (max_size < size)
return (-1ULL);
ssearch.ss_start = *cursor;
ssearch.ss_end = *cursor + size;
ss = avl_find(t, &ssearch, &where);
if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
t = sm->sm_pp_root;
ssearch.ss_start = 0;
ssearch.ss_end = MIN(max_size,
1ULL << (hbit + metaslab_ndf_clump_shift));
ss = avl_find(t, &ssearch, &where);
if (ss == NULL)
ss = avl_nearest(t, where, AVL_AFTER);
ASSERT(ss != NULL);
}
if (ss != NULL) {
if (ss->ss_start + size <= ss->ss_end) {
*cursor = ss->ss_start + size;
return (ss->ss_start);
}
}
return (-1ULL);
}
static boolean_t
metaslab_ndf_fragmented(space_map_t *sm)
{
uint64_t max_size = metaslab_pp_maxsize(sm);
if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
return (B_FALSE);
return (B_TRUE);
}
static space_map_ops_t metaslab_ndf_ops = {
metaslab_pp_load,
metaslab_pp_unload,
metaslab_ndf_alloc,
metaslab_pp_claim,
metaslab_pp_free,
metaslab_pp_maxsize,
metaslab_ndf_fragmented
};
space_map_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops;
#endif /* WITH_NDF_BLOCK_ALLOCATOR */
2009-07-03 02:44:48 +04:00
2008-11-20 23:01:55 +03:00
/*
* ==========================================================================
* Metaslabs
* ==========================================================================
*/
metaslab_t *
metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
uint64_t start, uint64_t size, uint64_t txg)
{
vdev_t *vd = mg->mg_vd;
metaslab_t *msp;
msp = kmem_zalloc(sizeof (metaslab_t), KM_PUSHPAGE);
2008-11-20 23:01:55 +03:00
mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
msp->ms_smo_syncing = *smo;
/*
* We create the main space map here, but we don't create the
* allocmaps and freemaps until metaslab_sync_done(). This serves
* two purposes: it allows metaslab_sync_done() to detect the
* addition of new space; and for debugging, it ensures that we'd
* data fault on any attempt to use this metaslab before it's ready.
*/
msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_PUSHPAGE);
space_map_create(msp->ms_map, start, size,
2008-11-20 23:01:55 +03:00
vd->vdev_ashift, &msp->ms_lock);
metaslab_group_add(mg, msp);
if (metaslab_debug && smo->smo_object != 0) {
mutex_enter(&msp->ms_lock);
VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops,
SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
mutex_exit(&msp->ms_lock);
}
2008-11-20 23:01:55 +03:00
/*
* If we're opening an existing pool (txg == 0) or creating
* a new one (txg == TXG_INITIAL), all space is available now.
* If we're adding space to an existing pool, the new space
* does not become available until after this txg has synced.
*/
if (txg <= TXG_INITIAL)
metaslab_sync_done(msp, 0);
if (txg != 0) {
vdev_dirty(vd, 0, NULL, txg);
vdev_dirty(vd, VDD_METASLAB, msp, txg);
2008-11-20 23:01:55 +03:00
}
return (msp);
}
void
metaslab_fini(metaslab_t *msp)
{
metaslab_group_t *mg = msp->ms_group;
int t;
2008-11-20 23:01:55 +03:00
vdev_space_update(mg->mg_vd,
-msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size);
2008-11-20 23:01:55 +03:00
metaslab_group_remove(mg, msp);
mutex_enter(&msp->ms_lock);
space_map_unload(msp->ms_map);
space_map_destroy(msp->ms_map);
kmem_free(msp->ms_map, sizeof (*msp->ms_map));
2008-11-20 23:01:55 +03:00
for (t = 0; t < TXG_SIZE; t++) {
space_map_destroy(msp->ms_allocmap[t]);
space_map_destroy(msp->ms_freemap[t]);
kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t]));
kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t]));
2008-11-20 23:01:55 +03:00
}
for (t = 0; t < TXG_DEFER_SIZE; t++) {
space_map_destroy(msp->ms_defermap[t]);
kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t]));
}
ASSERT0(msp->ms_deferspace);
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
mutex_destroy(&msp->ms_lock);
kmem_free(msp, sizeof (metaslab_t));
}
#define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
#define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
#define METASLAB_ACTIVE_MASK \
(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
static uint64_t
metaslab_weight(metaslab_t *msp)
{
metaslab_group_t *mg = msp->ms_group;
space_map_t *sm = msp->ms_map;
2008-11-20 23:01:55 +03:00
space_map_obj_t *smo = &msp->ms_smo;
vdev_t *vd = mg->mg_vd;
uint64_t weight, space;
ASSERT(MUTEX_HELD(&msp->ms_lock));
/*
* The baseline weight is the metaslab's free space.
*/
space = sm->sm_size - smo->smo_alloc;
weight = space;
/*
* Modern disks have uniform bit density and constant angular velocity.
* Therefore, the outer recording zones are faster (higher bandwidth)
* than the inner zones by the ratio of outer to inner track diameter,
* which is typically around 2:1. We account for this by assigning
* higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
* In effect, this means that we'll select the metaslab with the most
* free bandwidth rather than simply the one with the most free space.
*/
weight = 2 * weight -
((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
ASSERT(weight >= space && weight <= 2 * space);
/*
* For locality, assign higher weight to metaslabs which have
* a lower offset than what we've already activated.
2008-11-20 23:01:55 +03:00
*/
if (sm->sm_start <= mg->mg_bonus_area)
weight *= (metaslab_smo_bonus_pct / 100);
2008-11-20 23:01:55 +03:00
ASSERT(weight >= space &&
weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
/*
* If this metaslab is one we're actively using, adjust its
* weight to make it preferable to any inactive metaslab so
* we'll polish it off.
*/
weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
}
return (weight);
}
static void
metaslab_prefetch(metaslab_group_t *mg)
{
spa_t *spa = mg->mg_vd->vdev_spa;
metaslab_t *msp;
avl_tree_t *t = &mg->mg_metaslab_tree;
int m;
mutex_enter(&mg->mg_lock);
2008-11-20 23:01:55 +03:00
/*
* Prefetch the next potential metaslabs
2008-11-20 23:01:55 +03:00
*/
for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
space_map_t *sm = msp->ms_map;
space_map_obj_t *smo = &msp->ms_smo;
2008-11-20 23:01:55 +03:00
/* If we have reached our prefetch limit then we're done */
if (m >= metaslab_prefetch_limit)
break;
if (!sm->sm_loaded && smo->smo_object != 0) {
mutex_exit(&mg->mg_lock);
dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
0ULL, smo->smo_objsize);
mutex_enter(&mg->mg_lock);
}
}
mutex_exit(&mg->mg_lock);
2008-11-20 23:01:55 +03:00
}
static int
metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
2008-11-20 23:01:55 +03:00
{
metaslab_group_t *mg = msp->ms_group;
space_map_t *sm = msp->ms_map;
2009-07-03 02:44:48 +04:00
space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
int t;
2008-11-20 23:01:55 +03:00
ASSERT(MUTEX_HELD(&msp->ms_lock));
if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
space_map_load_wait(sm);
if (!sm->sm_loaded) {
space_map_obj_t *smo = &msp->ms_smo;
int error = space_map_load(sm, sm_ops, SM_FREE, smo,
spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
if (error) {
metaslab_group_sort(msp->ms_group, msp, 0);
return (error);
}
for (t = 0; t < TXG_DEFER_SIZE; t++)
space_map_walk(msp->ms_defermap[t],
space_map_claim, sm);
}
/*
* Track the bonus area as we activate new metaslabs.
*/
if (sm->sm_start > mg->mg_bonus_area) {
mutex_enter(&mg->mg_lock);
mg->mg_bonus_area = sm->sm_start;
mutex_exit(&mg->mg_lock);
2008-11-20 23:01:55 +03:00
}
2009-07-03 02:44:48 +04:00
2008-11-20 23:01:55 +03:00
metaslab_group_sort(msp->ms_group, msp,
msp->ms_weight | activation_weight);
}
ASSERT(sm->sm_loaded);
ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
return (0);
}
static void
metaslab_passivate(metaslab_t *msp, uint64_t size)
{
/*
* If size < SPA_MINBLOCKSIZE, then we will not allocate from
* this metaslab again. In that case, it had better be empty,
* or we would be leaving space on the table.
*/
ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0);
2008-11-20 23:01:55 +03:00
metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
}
/*
* Determine if the in-core space map representation can be condensed on-disk.
* We would like to use the following criteria to make our decision:
*
* 1. The size of the space map object should not dramatically increase as a
* result of writing out our in-core free map.
*
* 2. The minimal on-disk space map representation is zfs_condense_pct/100
* times the size than the in-core representation (i.e. zfs_condense_pct = 110
* and in-core = 1MB, minimal = 1.1.MB).
*
* Checking the first condition is tricky since we don't want to walk
* the entire AVL tree calculating the estimated on-disk size. Instead we
* use the size-ordered AVL tree in the space map and calculate the
* size required for the largest segment in our in-core free map. If the
* size required to represent that segment on disk is larger than the space
* map object then we avoid condensing this map.
*
* To determine the second criterion we use a best-case estimate and assume
* each segment can be represented on-disk as a single 64-bit entry. We refer
* to this best-case estimate as the space map's minimal form.
*/
static boolean_t
metaslab_should_condense(metaslab_t *msp)
{
space_map_t *sm = msp->ms_map;
space_map_obj_t *smo = &msp->ms_smo_syncing;
space_seg_t *ss;
uint64_t size, entries, segsz;
ASSERT(MUTEX_HELD(&msp->ms_lock));
ASSERT(sm->sm_loaded);
/*
* Use the sm_pp_root AVL tree, which is ordered by size, to obtain
* the largest segment in the in-core free map. If the tree is
* empty then we should condense the map.
*/
ss = avl_last(sm->sm_pp_root);
if (ss == NULL)
return (B_TRUE);
/*
* Calculate the number of 64-bit entries this segment would
* require when written to disk. If this single segment would be
* larger on-disk than the entire current on-disk structure, then
* clearly condensing will increase the on-disk structure size.
*/
size = (ss->ss_end - ss->ss_start) >> sm->sm_shift;
entries = size / (MIN(size, SM_RUN_MAX));
segsz = entries * sizeof (uint64_t);
return (segsz <= smo->smo_objsize &&
smo->smo_objsize >= (zfs_condense_pct *
sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100);
}
/*
* Condense the on-disk space map representation to its minimized form.
* The minimized form consists of a small number of allocations followed by
* the in-core free map.
*/
static void
metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
{
spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK];
space_map_t condense_map;
space_map_t *sm = msp->ms_map;
objset_t *mos = spa_meta_objset(spa);
space_map_obj_t *smo = &msp->ms_smo_syncing;
int t;
ASSERT(MUTEX_HELD(&msp->ms_lock));
ASSERT3U(spa_sync_pass(spa), ==, 1);
ASSERT(sm->sm_loaded);
spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
"smo size %llu, segments %lu", txg,
(msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
smo->smo_objsize, avl_numnodes(&sm->sm_root));
/*
* Create an map that is a 100% allocated map. We remove segments
* that have been freed in this txg, any deferred frees that exist,
* and any allocation in the future. Removing segments should be
* a relatively inexpensive operation since we expect these maps to
* a small number of nodes.
*/
space_map_create(&condense_map, sm->sm_start, sm->sm_size,
sm->sm_shift, sm->sm_lock);
space_map_add(&condense_map, condense_map.sm_start,
condense_map.sm_size);
/*
* Remove what's been freed in this txg from the condense_map.
* Since we're in sync_pass 1, we know that all the frees from
* this txg are in the freemap.
*/
space_map_walk(freemap, space_map_remove, &condense_map);
for (t = 0; t < TXG_DEFER_SIZE; t++)
space_map_walk(msp->ms_defermap[t],
space_map_remove, &condense_map);
for (t = 1; t < TXG_CONCURRENT_STATES; t++)
space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK],
space_map_remove, &condense_map);
/*
* We're about to drop the metaslab's lock thus allowing
* other consumers to change it's content. Set the
* space_map's sm_condensing flag to ensure that
* allocations on this metaslab do not occur while we're
* in the middle of committing it to disk. This is only critical
* for the ms_map as all other space_maps use per txg
* views of their content.
*/
sm->sm_condensing = B_TRUE;
mutex_exit(&msp->ms_lock);
space_map_truncate(smo, mos, tx);
mutex_enter(&msp->ms_lock);
/*
* While we would ideally like to create a space_map representation
* that consists only of allocation records, doing so can be
* prohibitively expensive because the in-core free map can be
* large, and therefore computationally expensive to subtract
* from the condense_map. Instead we sync out two maps, a cheap
* allocation only map followed by the in-core free map. While not
* optimal, this is typically close to optimal, and much cheaper to
* compute.
*/
space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx);
space_map_vacate(&condense_map, NULL, NULL);
space_map_destroy(&condense_map);
space_map_sync(sm, SM_FREE, smo, mos, tx);
sm->sm_condensing = B_FALSE;
spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, "
"smo size %llu", txg,
(msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
smo->smo_objsize);
}
2008-11-20 23:01:55 +03:00
/*
* Write a metaslab to disk in the context of the specified transaction group.
*/
void
metaslab_sync(metaslab_t *msp, uint64_t txg)
{
vdev_t *vd = msp->ms_group->mg_vd;
spa_t *spa = vd->vdev_spa;
objset_t *mos = spa_meta_objset(spa);
space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK];
space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK];
space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
space_map_t *sm = msp->ms_map;
2008-11-20 23:01:55 +03:00
space_map_obj_t *smo = &msp->ms_smo_syncing;
dmu_buf_t *db;
dmu_tx_t *tx;
ASSERT(!vd->vdev_ishole);
/*
* This metaslab has just been added so there's no work to do now.
*/
if (*freemap == NULL) {
ASSERT3P(allocmap, ==, NULL);
return;
}
ASSERT3P(allocmap, !=, NULL);
ASSERT3P(*freemap, !=, NULL);
ASSERT3P(*freed_map, !=, NULL);
if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0)
return;
2008-11-20 23:01:55 +03:00
/*
* The only state that can actually be changing concurrently with
* metaslab_sync() is the metaslab's ms_map. No other thread can
* be modifying this txg's allocmap, freemap, freed_map, or smo.
* Therefore, we only hold ms_lock to satify space_map ASSERTs.
* We drop it whenever we call into the DMU, because the DMU
* can call down to us (e.g. via zio_free()) at any time.
*/
tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2008-11-20 23:01:55 +03:00
if (smo->smo_object == 0) {
ASSERT(smo->smo_objsize == 0);
ASSERT(smo->smo_alloc == 0);
smo->smo_object = dmu_object_alloc(mos,
DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
ASSERT(smo->smo_object != 0);
dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
(sm->sm_start >> vd->vdev_ms_shift),
sizeof (uint64_t), &smo->smo_object, tx);
}
mutex_enter(&msp->ms_lock);
if (sm->sm_loaded && spa_sync_pass(spa) == 1 &&
metaslab_should_condense(msp)) {
metaslab_condense(msp, txg, tx);
} else {
space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
space_map_sync(*freemap, SM_FREE, smo, mos, tx);
}
space_map_vacate(allocmap, NULL, NULL);
2008-11-20 23:01:55 +03:00
/*
* For sync pass 1, we avoid walking the entire space map and
* instead will just swap the pointers for freemap and
* freed_map. We can safely do this since the freed_map is
* guaranteed to be empty on the initial pass.
*/
if (spa_sync_pass(spa) == 1) {
ASSERT0((*freed_map)->sm_space);
ASSERT0(avl_numnodes(&(*freed_map)->sm_root));
space_map_swap(freemap, freed_map);
} else {
space_map_vacate(*freemap, space_map_add, *freed_map);
2008-11-20 23:01:55 +03:00
}
ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space);
ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space);
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
2008-11-20 23:01:55 +03:00
dmu_buf_will_dirty(db, tx);
ASSERT3U(db->db_size, >=, sizeof (*smo));
bcopy(smo, db->db_data, sizeof (*smo));
dmu_buf_rele(db, FTAG);
dmu_tx_commit(tx);
}
/*
* Called after a transaction group has completely synced to mark
* all of the metaslab's free space as usable.
*/
void
metaslab_sync_done(metaslab_t *msp, uint64_t txg)
{
space_map_obj_t *smo = &msp->ms_smo;
space_map_obj_t *smosync = &msp->ms_smo_syncing;
space_map_t *sm = msp->ms_map;
space_map_t *freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
space_map_t *defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE];
2008-11-20 23:01:55 +03:00
metaslab_group_t *mg = msp->ms_group;
vdev_t *vd = mg->mg_vd;
int64_t alloc_delta, defer_delta;
int t;
ASSERT(!vd->vdev_ishole);
2008-11-20 23:01:55 +03:00
mutex_enter(&msp->ms_lock);
/*
* If this metaslab is just becoming available, initialize its
* allocmaps, freemaps, and defermap and add its capacity to the vdev.
2008-11-20 23:01:55 +03:00
*/
if (freed_map == NULL) {
ASSERT(defer_map == NULL);
for (t = 0; t < TXG_SIZE; t++) {
msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t),
KM_PUSHPAGE);
space_map_create(msp->ms_allocmap[t], sm->sm_start,
2008-11-20 23:01:55 +03:00
sm->sm_size, sm->sm_shift, sm->sm_lock);
msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t),
KM_PUSHPAGE);
space_map_create(msp->ms_freemap[t], sm->sm_start,
2008-11-20 23:01:55 +03:00
sm->sm_size, sm->sm_shift, sm->sm_lock);
}
for (t = 0; t < TXG_DEFER_SIZE; t++) {
msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t),
KM_PUSHPAGE);
space_map_create(msp->ms_defermap[t], sm->sm_start,
sm->sm_size, sm->sm_shift, sm->sm_lock);
}
freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE];
vdev_space_update(vd, 0, 0, sm->sm_size);
2008-11-20 23:01:55 +03:00
}
alloc_delta = smosync->smo_alloc - smo->smo_alloc;
defer_delta = freed_map->sm_space - defer_map->sm_space;
vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2008-11-20 23:01:55 +03:00
ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0);
ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0);
2008-11-20 23:01:55 +03:00
/*
* If there's a space_map_load() in progress, wait for it to complete
* so that we have a consistent view of the in-core space map.
* Then, add defer_map (oldest deferred frees) to this map and
* transfer freed_map (this txg's frees) to defer_map.
2008-11-20 23:01:55 +03:00
*/
space_map_load_wait(sm);
space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
space_map_vacate(freed_map, space_map_add, defer_map);
2008-11-20 23:01:55 +03:00
*smo = *smosync;
msp->ms_deferspace += defer_delta;
ASSERT3S(msp->ms_deferspace, >=, 0);
ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
if (msp->ms_deferspace != 0) {
/*
* Keep syncing this metaslab until all deferred frees
* are back in circulation.
*/
vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
}
2008-11-20 23:01:55 +03:00
/*
* If the map is loaded but no longer active, evict it as soon as all
* future allocations have synced. (If we unloaded it now and then
* loaded a moment later, the map wouldn't reflect those allocations.)
*/
if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
int evictable = 1;
for (t = 1; t < TXG_CONCURRENT_STATES; t++)
if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space)
2008-11-20 23:01:55 +03:00
evictable = 0;
if (evictable && !metaslab_debug)
2008-11-20 23:01:55 +03:00
space_map_unload(sm);
}
metaslab_group_sort(mg, msp, metaslab_weight(msp));
mutex_exit(&msp->ms_lock);
}
void
metaslab_sync_reassess(metaslab_group_t *mg)
{
vdev_t *vd = mg->mg_vd;
int64_t failures = mg->mg_alloc_failures;
int m;
/*
* Re-evaluate all metaslabs which have lower offsets than the
* bonus area.
*/
for (m = 0; m < vd->vdev_ms_count; m++) {
metaslab_t *msp = vd->vdev_ms[m];
if (msp->ms_map->sm_start > mg->mg_bonus_area)
break;
mutex_enter(&msp->ms_lock);
metaslab_group_sort(mg, msp, metaslab_weight(msp));
mutex_exit(&msp->ms_lock);
}
atomic_add_64(&mg->mg_alloc_failures, -failures);
/*
* Prefetch the next potential metaslabs
*/
metaslab_prefetch(mg);
}
2008-11-20 23:01:55 +03:00
static uint64_t
metaslab_distance(metaslab_t *msp, dva_t *dva)
{
uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
uint64_t start = msp->ms_map->sm_start >> ms_shift;
2008-11-20 23:01:55 +03:00
if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
return (1ULL << 63);
if (offset < start)
return ((start - offset) << ms_shift);
if (offset > start)
return ((offset - start) << ms_shift);
return (0);
}
static uint64_t
metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
2008-11-20 23:01:55 +03:00
{
spa_t *spa = mg->mg_vd->vdev_spa;
2008-11-20 23:01:55 +03:00
metaslab_t *msp = NULL;
uint64_t offset = -1ULL;
avl_tree_t *t = &mg->mg_metaslab_tree;
uint64_t activation_weight;
uint64_t target_distance;
int i;
activation_weight = METASLAB_WEIGHT_PRIMARY;
2009-07-03 02:44:48 +04:00
for (i = 0; i < d; i++) {
if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2008-11-20 23:01:55 +03:00
activation_weight = METASLAB_WEIGHT_SECONDARY;
2009-07-03 02:44:48 +04:00
break;
}
}
2008-11-20 23:01:55 +03:00
for (;;) {
2009-07-03 02:44:48 +04:00
boolean_t was_active;
2008-11-20 23:01:55 +03:00
mutex_enter(&mg->mg_lock);
for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
if (msp->ms_weight < asize) {
spa_dbgmsg(spa, "%s: failed to meet weight "
"requirement: vdev %llu, txg %llu, mg %p, "
"msp %p, psize %llu, asize %llu, "
"failures %llu, weight %llu",
spa_name(spa), mg->mg_vd->vdev_id, txg,
mg, msp, psize, asize,
mg->mg_alloc_failures, msp->ms_weight);
2008-11-20 23:01:55 +03:00
mutex_exit(&mg->mg_lock);
return (-1ULL);
}
/*
* If the selected metaslab is condensing, skip it.
*/
if (msp->ms_map->sm_condensing)
continue;
2009-07-03 02:44:48 +04:00
was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2008-11-20 23:01:55 +03:00
if (activation_weight == METASLAB_WEIGHT_PRIMARY)
break;
target_distance = min_distance +
(msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
for (i = 0; i < d; i++)
if (metaslab_distance(msp, &dva[i]) <
target_distance)
break;
if (i == d)
break;
}
mutex_exit(&mg->mg_lock);
if (msp == NULL)
return (-1ULL);
/*
* If we've already reached the allowable number of failed
* allocation attempts on this metaslab group then we
* consider skipping it. We skip it only if we're allowed
* to "fast" gang, the physical size is larger than
* a gang block, and we're attempting to allocate from
* the primary metaslab.
*/
if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
activation_weight == METASLAB_WEIGHT_PRIMARY) {
spa_dbgmsg(spa, "%s: skipping metaslab group: "
"vdev %llu, txg %llu, mg %p, psize %llu, "
"asize %llu, failures %llu", spa_name(spa),
mg->mg_vd->vdev_id, txg, mg, psize, asize,
mg->mg_alloc_failures);
return (-1ULL);
}
2008-11-20 23:01:55 +03:00
mutex_enter(&msp->ms_lock);
/*
* Ensure that the metaslab we have selected is still
* capable of handling our request. It's possible that
* another thread may have changed the weight while we
* were blocked on the metaslab lock.
*/
if (msp->ms_weight < asize || (was_active &&
2009-07-03 02:44:48 +04:00
!(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
activation_weight == METASLAB_WEIGHT_PRIMARY)) {
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
continue;
}
if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
activation_weight == METASLAB_WEIGHT_PRIMARY) {
metaslab_passivate(msp,
msp->ms_weight & ~METASLAB_ACTIVE_MASK);
mutex_exit(&msp->ms_lock);
continue;
}
if (metaslab_activate(msp, activation_weight) != 0) {
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
continue;
}
/*
* If this metaslab is currently condensing then pick again as
* we can't manipulate this metaslab until it's committed
* to disk.
*/
if (msp->ms_map->sm_condensing) {
mutex_exit(&msp->ms_lock);
continue;
}
if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL)
2008-11-20 23:01:55 +03:00
break;
atomic_inc_64(&mg->mg_alloc_failures);
metaslab_passivate(msp, space_map_maxsize(msp->ms_map));
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
}
if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
2008-11-20 23:01:55 +03:00
vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize);
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
return (offset);
}
/*
* Allocate a block for the specified i/o.
*/
static int
metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
2008-11-20 23:01:55 +03:00
{
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
metaslab_group_t *mg, *fast_mg, *rotor;
2008-11-20 23:01:55 +03:00
vdev_t *vd;
int dshift = 3;
int all_zero;
2009-01-16 00:59:39 +03:00
int zio_lock = B_FALSE;
boolean_t allocatable;
2008-11-20 23:01:55 +03:00
uint64_t offset = -1ULL;
uint64_t asize;
uint64_t distance;
ASSERT(!DVA_IS_VALID(&dva[d]));
/*
* For testing, make some blocks above a certain size be gang blocks.
*/
if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
2008-11-20 23:01:55 +03:00
return (ENOSPC);
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
if (flags & METASLAB_FASTWRITE)
mutex_enter(&mc->mc_fastwrite_lock);
2008-11-20 23:01:55 +03:00
/*
* Start at the rotor and loop through all mgs until we find something.
* Note that there's no locking on mc_rotor or mc_aliquot because
2008-11-20 23:01:55 +03:00
* nothing actually breaks if we miss a few updates -- we just won't
* allocate quite as evenly. It all balances out over time.
*
* If we are doing ditto or log blocks, try to spread them across
* consecutive vdevs. If we're forced to reuse a vdev before we've
* allocated all of our ditto blocks, then try and spread them out on
* that vdev as much as possible. If it turns out to not be possible,
* gradually lower our standards until anything becomes acceptable.
* Also, allocating on consecutive vdevs (as opposed to random vdevs)
* gives us hope of containing our fault domains to something we're
* able to reason about. Otherwise, any two top-level vdev failures
* will guarantee the loss of data. With consecutive allocation,
* only two adjacent top-level vdev failures will result in data loss.
*
* If we are doing gang blocks (hintdva is non-NULL), try to keep
* ourselves on the same vdev as our gang block header. That
* way, we can hope for locality in vdev_cache, plus it makes our
* fault domains something tractable.
*/
if (hintdva) {
vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
/*
* It's possible the vdev we're using as the hint no
* longer exists (i.e. removed). Consult the rotor when
* all else fails.
*/
if (vd != NULL) {
2008-11-20 23:01:55 +03:00
mg = vd->vdev_mg;
if (flags & METASLAB_HINTBP_AVOID &&
mg->mg_next != NULL)
mg = mg->mg_next;
} else {
mg = mc->mc_rotor;
}
2008-11-20 23:01:55 +03:00
} else if (d != 0) {
vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
mg = vd->vdev_mg->mg_next;
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
} else if (flags & METASLAB_FASTWRITE) {
mg = fast_mg = mc->mc_rotor;
do {
if (fast_mg->mg_vd->vdev_pending_fastwrite <
mg->mg_vd->vdev_pending_fastwrite)
mg = fast_mg;
} while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor);
2008-11-20 23:01:55 +03:00
} else {
mg = mc->mc_rotor;
}
/*
* If the hint put us into the wrong metaslab class, or into a
* metaslab group that has been passivated, just follow the rotor.
2008-11-20 23:01:55 +03:00
*/
if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2008-11-20 23:01:55 +03:00
mg = mc->mc_rotor;
rotor = mg;
top:
all_zero = B_TRUE;
do {
ASSERT(mg->mg_activation_count == 1);
2008-11-20 23:01:55 +03:00
vd = mg->mg_vd;
2009-01-16 00:59:39 +03:00
2008-11-20 23:01:55 +03:00
/*
* Don't allocate from faulted devices.
2008-11-20 23:01:55 +03:00
*/
2009-01-16 00:59:39 +03:00
if (zio_lock) {
spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
allocatable = vdev_allocatable(vd);
spa_config_exit(spa, SCL_ZIO, FTAG);
} else {
allocatable = vdev_allocatable(vd);
}
if (!allocatable)
2008-11-20 23:01:55 +03:00
goto next;
2009-01-16 00:59:39 +03:00
2008-11-20 23:01:55 +03:00
/*
* Avoid writing single-copy data to a failing vdev
*/
if ((vd->vdev_stat.vs_write_errors > 0 ||
vd->vdev_state < VDEV_STATE_HEALTHY) &&
d == 0 && dshift == 3) {
all_zero = B_FALSE;
goto next;
}
ASSERT(mg->mg_class == mc);
distance = vd->vdev_asize >> dshift;
if (distance <= (1ULL << vd->vdev_ms_shift))
distance = 0;
else
all_zero = B_FALSE;
asize = vdev_psize_to_asize(vd, psize);
ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
dva, d, flags);
2008-11-20 23:01:55 +03:00
if (offset != -1ULL) {
/*
* If we've just selected this metaslab group,
* figure out whether the corresponding vdev is
* over- or under-used relative to the pool,
* and set an allocation bias to even it out.
*/
if (mc->mc_aliquot == 0) {
2008-11-20 23:01:55 +03:00
vdev_stat_t *vs = &vd->vdev_stat;
int64_t vu, cu;
2008-11-20 23:01:55 +03:00
vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2008-11-20 23:01:55 +03:00
/*
* Calculate how much more or less we should
* try to allocate from this device during
* this iteration around the rotor.
* For example, if a device is 80% full
* and the pool is 20% full then we should
* reduce allocations by 60% on this device.
*
* mg_bias = (20 - 80) * 512K / 100 = -307K
*
* This reduces allocations by 307K for this
* iteration.
2008-11-20 23:01:55 +03:00
*/
mg->mg_bias = ((cu - vu) *
(int64_t)mg->mg_aliquot) / 100;
2008-11-20 23:01:55 +03:00
}
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
if ((flags & METASLAB_FASTWRITE) ||
atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2008-11-20 23:01:55 +03:00
mg->mg_aliquot + mg->mg_bias) {
mc->mc_rotor = mg->mg_next;
mc->mc_aliquot = 0;
2008-11-20 23:01:55 +03:00
}
DVA_SET_VDEV(&dva[d], vd->vdev_id);
DVA_SET_OFFSET(&dva[d], offset);
DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2008-11-20 23:01:55 +03:00
DVA_SET_ASIZE(&dva[d], asize);
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
if (flags & METASLAB_FASTWRITE) {
atomic_add_64(&vd->vdev_pending_fastwrite,
psize);
mutex_exit(&mc->mc_fastwrite_lock);
}
2008-11-20 23:01:55 +03:00
return (0);
}
next:
mc->mc_rotor = mg->mg_next;
mc->mc_aliquot = 0;
2008-11-20 23:01:55 +03:00
} while ((mg = mg->mg_next) != rotor);
if (!all_zero) {
dshift++;
ASSERT(dshift < 64);
goto top;
}
2009-07-03 02:44:48 +04:00
if (!allocatable && !zio_lock) {
2009-01-16 00:59:39 +03:00
dshift = 3;
zio_lock = B_TRUE;
goto top;
}
2008-11-20 23:01:55 +03:00
bzero(&dva[d], sizeof (dva_t));
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
if (flags & METASLAB_FASTWRITE)
mutex_exit(&mc->mc_fastwrite_lock);
2008-11-20 23:01:55 +03:00
return (ENOSPC);
}
/*
* Free the block represented by DVA in the context of the specified
* transaction group.
*/
static void
metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
{
uint64_t vdev = DVA_GET_VDEV(dva);
uint64_t offset = DVA_GET_OFFSET(dva);
uint64_t size = DVA_GET_ASIZE(dva);
vdev_t *vd;
metaslab_t *msp;
ASSERT(DVA_IS_VALID(dva));
if (txg > spa_freeze_txg(spa))
return;
if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
(offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
(u_longlong_t)vdev, (u_longlong_t)offset);
ASSERT(0);
return;
}
msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
if (DVA_GET_GANG(dva))
size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
mutex_enter(&msp->ms_lock);
if (now) {
space_map_remove(msp->ms_allocmap[txg & TXG_MASK],
2008-11-20 23:01:55 +03:00
offset, size);
space_map_free(msp->ms_map, offset, size);
2008-11-20 23:01:55 +03:00
} else {
if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0)
2008-11-20 23:01:55 +03:00
vdev_dirty(vd, VDD_METASLAB, msp, txg);
space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size);
2008-11-20 23:01:55 +03:00
}
mutex_exit(&msp->ms_lock);
}
/*
* Intent log support: upon opening the pool after a crash, notify the SPA
* of blocks that the intent log has allocated for immediate write, but
* which are still considered free by the SPA because the last transaction
* group didn't commit yet.
*/
static int
metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
{
uint64_t vdev = DVA_GET_VDEV(dva);
uint64_t offset = DVA_GET_OFFSET(dva);
uint64_t size = DVA_GET_ASIZE(dva);
vdev_t *vd;
metaslab_t *msp;
int error = 0;
2008-11-20 23:01:55 +03:00
ASSERT(DVA_IS_VALID(dva));
if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
(offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
return (ENXIO);
msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
if (DVA_GET_GANG(dva))
size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
mutex_enter(&msp->ms_lock);
if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded)
error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
if (error == 0 && !space_map_contains(msp->ms_map, offset, size))
error = ENOENT;
if (error || txg == 0) { /* txg == 0 indicates dry run */
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
return (error);
}
space_map_claim(msp->ms_map, offset, size);
2009-01-16 00:59:39 +03:00
if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
vdev_dirty(vd, VDD_METASLAB, msp, txg);
space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size);
}
2008-11-20 23:01:55 +03:00
mutex_exit(&msp->ms_lock);
return (0);
}
int
metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2008-11-20 23:01:55 +03:00
{
dva_t *dva = bp->blk_dva;
dva_t *hintdva = hintbp->blk_dva;
int d, error = 0;
2008-11-20 23:01:55 +03:00
ASSERT(bp->blk_birth == 0);
ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
if (mc->mc_rotor == NULL) { /* no vdevs in this class */
spa_config_exit(spa, SCL_ALLOC, FTAG);
2008-11-20 23:01:55 +03:00
return (ENOSPC);
}
2008-11-20 23:01:55 +03:00
ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
ASSERT(BP_GET_NDVAS(bp) == 0);
ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
for (d = 0; d < ndvas; d++) {
2008-11-20 23:01:55 +03:00
error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
txg, flags);
2008-11-20 23:01:55 +03:00
if (error) {
for (d--; d >= 0; d--) {
metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
bzero(&dva[d], sizeof (dva_t));
}
spa_config_exit(spa, SCL_ALLOC, FTAG);
2008-11-20 23:01:55 +03:00
return (error);
}
}
ASSERT(error == 0);
ASSERT(BP_GET_NDVAS(bp) == ndvas);
spa_config_exit(spa, SCL_ALLOC, FTAG);
BP_SET_BIRTH(bp, txg, txg);
2008-11-20 23:01:55 +03:00
return (0);
}
void
metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
{
const dva_t *dva = bp->blk_dva;
int d, ndvas = BP_GET_NDVAS(bp);
2008-11-20 23:01:55 +03:00
ASSERT(!BP_IS_HOLE(bp));
ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2008-11-20 23:01:55 +03:00
for (d = 0; d < ndvas; d++)
2008-11-20 23:01:55 +03:00
metaslab_free_dva(spa, &dva[d], txg, now);
spa_config_exit(spa, SCL_FREE, FTAG);
2008-11-20 23:01:55 +03:00
}
int
metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
{
const dva_t *dva = bp->blk_dva;
int ndvas = BP_GET_NDVAS(bp);
int d, error = 0;
2008-11-20 23:01:55 +03:00
ASSERT(!BP_IS_HOLE(bp));
if (txg != 0) {
/*
* First do a dry run to make sure all DVAs are claimable,
* so we don't have to unwind from partial failures below.
*/
if ((error = metaslab_claim(spa, bp, 0)) != 0)
return (error);
}
spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
for (d = 0; d < ndvas; d++)
2008-11-20 23:01:55 +03:00
if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
break;
spa_config_exit(spa, SCL_ALLOC, FTAG);
ASSERT(error == 0 || txg == 0);
2008-11-20 23:01:55 +03:00
return (error);
2008-11-20 23:01:55 +03:00
}
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 17:20:20 +04:00
void metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp)
{
const dva_t *dva = bp->blk_dva;
int ndvas = BP_GET_NDVAS(bp);
uint64_t psize = BP_GET_PSIZE(bp);
int d;
vdev_t *vd;
ASSERT(!BP_IS_HOLE(bp));
ASSERT(psize > 0);
spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
for (d = 0; d < ndvas; d++) {
if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
continue;
atomic_add_64(&vd->vdev_pending_fastwrite, psize);
}
spa_config_exit(spa, SCL_VDEV, FTAG);
}
void metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp)
{
const dva_t *dva = bp->blk_dva;
int ndvas = BP_GET_NDVAS(bp);
uint64_t psize = BP_GET_PSIZE(bp);
int d;
vdev_t *vd;
ASSERT(!BP_IS_HOLE(bp));
ASSERT(psize > 0);
spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
for (d = 0; d < ndvas; d++) {
if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
continue;
ASSERT3U(vd->vdev_pending_fastwrite, >=, psize);
atomic_sub_64(&vd->vdev_pending_fastwrite, psize);
}
spa_config_exit(spa, SCL_VDEV, FTAG);
}
static void
checkmap(space_map_t *sm, uint64_t off, uint64_t size)
{
space_seg_t *ss;
avl_index_t where;
mutex_enter(sm->sm_lock);
ss = space_map_find(sm, off, size, &where);
if (ss != NULL)
panic("freeing free block; ss=%p", (void *)ss);
mutex_exit(sm->sm_lock);
}
void
metaslab_check_free(spa_t *spa, const blkptr_t *bp)
{
int i, j;
if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
return;
spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
for (i = 0; i < BP_GET_NDVAS(bp); i++) {
uint64_t vdid = DVA_GET_VDEV(&bp->blk_dva[i]);
vdev_t *vd = vdev_lookup_top(spa, vdid);
uint64_t off = DVA_GET_OFFSET(&bp->blk_dva[i]);
uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
metaslab_t *ms = vd->vdev_ms[off >> vd->vdev_ms_shift];
if (ms->ms_map->sm_loaded)
checkmap(ms->ms_map, off, size);
for (j = 0; j < TXG_SIZE; j++)
checkmap(ms->ms_freemap[j], off, size);
for (j = 0; j < TXG_DEFER_SIZE; j++)
checkmap(ms->ms_defermap[j], off, size);
}
spa_config_exit(spa, SCL_VDEV, FTAG);
}
#if defined(_KERNEL) && defined(HAVE_SPL)
module_param(metaslab_debug, int, 0644);
MODULE_PARM_DESC(metaslab_debug, "keep space maps in core to verify frees");
#endif /* _KERNEL && HAVE_SPL */