4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/zfs_context.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
37 #include <sys/vdev_indirect_mapping.h>
40 #define GANG_ALLOCATION(flags) \
41 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
43 uint64_t metaslab_aliquot
= 512ULL << 10;
44 uint64_t metaslab_force_ganging
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
47 * Since we can touch multiple metaslabs (and their respective space maps)
48 * with each transaction group, we benefit from having a smaller space map
49 * block size since it allows us to issue more I/O operations scattered
52 int zfs_metaslab_sm_blksz
= (1 << 12);
55 * The in-core space map representation is more compact than its on-disk form.
56 * The zfs_condense_pct determines how much more compact the in-core
57 * space map representation must be before we compact it on-disk.
58 * Values should be greater than or equal to 100.
60 int zfs_condense_pct
= 200;
63 * Condensing a metaslab is not guaranteed to actually reduce the amount of
64 * space used on disk. In particular, a space map uses data in increments of
65 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
66 * same number of blocks after condensing. Since the goal of condensing is to
67 * reduce the number of IOPs required to read the space map, we only want to
68 * condense when we can be sure we will reduce the number of blocks used by the
69 * space map. Unfortunately, we cannot precisely compute whether or not this is
70 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
71 * we apply the following heuristic: do not condense a spacemap unless the
72 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
75 int zfs_metaslab_condense_block_threshold
= 4;
78 * The zfs_mg_noalloc_threshold defines which metaslab groups should
79 * be eligible for allocation. The value is defined as a percentage of
80 * free space. Metaslab groups that have more free space than
81 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
82 * a metaslab group's free space is less than or equal to the
83 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
84 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
85 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
86 * groups are allowed to accept allocations. Gang blocks are always
87 * eligible to allocate on any metaslab group. The default value of 0 means
88 * no metaslab group will be excluded based on this criterion.
90 int zfs_mg_noalloc_threshold
= 0;
93 * Metaslab groups are considered eligible for allocations if their
94 * fragmenation metric (measured as a percentage) is less than or equal to
95 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
96 * then it will be skipped unless all metaslab groups within the metaslab
97 * class have also crossed this threshold.
99 int zfs_mg_fragmentation_threshold
= 85;
102 * Allow metaslabs to keep their active state as long as their fragmentation
103 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
104 * active metaslab that exceeds this threshold will no longer keep its active
105 * status allowing better metaslabs to be selected.
107 int zfs_metaslab_fragmentation_threshold
= 70;
110 * When set will load all metaslabs when pool is first opened.
112 int metaslab_debug_load
= 0;
115 * When set will prevent metaslabs from being unloaded.
117 int metaslab_debug_unload
= 0;
120 * Minimum size which forces the dynamic allocator to change
121 * it's allocation strategy. Once the space map cannot satisfy
122 * an allocation of this size then it switches to using more
123 * aggressive strategy (i.e search by size rather than offset).
125 uint64_t metaslab_df_alloc_threshold
= SPA_OLD_MAXBLOCKSIZE
;
128 * The minimum free space, in percent, which must be available
129 * in a space map to continue allocations in a first-fit fashion.
130 * Once the space map's free space drops below this level we dynamically
131 * switch to using best-fit allocations.
133 int metaslab_df_free_pct
= 4;
136 * A metaslab is considered "free" if it contains a contiguous
137 * segment which is greater than metaslab_min_alloc_size.
139 uint64_t metaslab_min_alloc_size
= DMU_MAX_ACCESS
;
142 * Percentage of all cpus that can be used by the metaslab taskq.
144 int metaslab_load_pct
= 50;
147 * Determines how many txgs a metaslab may remain loaded without having any
148 * allocations from it. As long as a metaslab continues to be used we will
151 int metaslab_unload_delay
= TXG_SIZE
* 2;
154 * Max number of metaslabs per group to preload.
156 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
159 * Enable/disable preloading of metaslab.
161 boolean_t metaslab_preload_enabled
= B_TRUE
;
164 * Enable/disable fragmentation weighting on metaslabs.
166 boolean_t metaslab_fragmentation_factor_enabled
= B_TRUE
;
169 * Enable/disable lba weighting (i.e. outer tracks are given preference).
171 boolean_t metaslab_lba_weighting_enabled
= B_TRUE
;
174 * Enable/disable metaslab group biasing.
176 boolean_t metaslab_bias_enabled
= B_TRUE
;
179 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
181 boolean_t zfs_remap_blkptr_enable
= B_TRUE
;
184 * Enable/disable segment-based metaslab selection.
186 boolean_t zfs_metaslab_segment_weight_enabled
= B_TRUE
;
189 * When using segment-based metaslab selection, we will continue
190 * allocating from the active metaslab until we have exhausted
191 * zfs_metaslab_switch_threshold of its buckets.
193 int zfs_metaslab_switch_threshold
= 2;
196 * Internal switch to enable/disable the metaslab allocation tracing
199 boolean_t metaslab_trace_enabled
= B_TRUE
;
202 * Maximum entries that the metaslab allocation tracing facility will keep
203 * in a given list when running in non-debug mode. We limit the number
204 * of entries in non-debug mode to prevent us from using up too much memory.
205 * The limit should be sufficiently large that we don't expect any allocation
206 * to every exceed this value. In debug mode, the system will panic if this
207 * limit is ever reached allowing for further investigation.
209 uint64_t metaslab_trace_max_entries
= 5000;
211 static uint64_t metaslab_weight(metaslab_t
*);
212 static void metaslab_set_fragmentation(metaslab_t
*);
213 static void metaslab_free_impl(vdev_t
*, uint64_t, uint64_t, boolean_t
);
214 static void metaslab_check_free_impl(vdev_t
*, uint64_t, uint64_t);
216 kmem_cache_t
*metaslab_alloc_trace_cache
;
219 * ==========================================================================
221 * ==========================================================================
224 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
226 metaslab_class_t
*mc
;
228 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
233 mutex_init(&mc
->mc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
234 refcount_create_tracked(&mc
->mc_alloc_slots
);
240 metaslab_class_destroy(metaslab_class_t
*mc
)
242 ASSERT(mc
->mc_rotor
== NULL
);
243 ASSERT(mc
->mc_alloc
== 0);
244 ASSERT(mc
->mc_deferred
== 0);
245 ASSERT(mc
->mc_space
== 0);
246 ASSERT(mc
->mc_dspace
== 0);
248 refcount_destroy(&mc
->mc_alloc_slots
);
249 mutex_destroy(&mc
->mc_lock
);
250 kmem_free(mc
, sizeof (metaslab_class_t
));
254 metaslab_class_validate(metaslab_class_t
*mc
)
256 metaslab_group_t
*mg
;
260 * Must hold one of the spa_config locks.
262 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
263 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
265 if ((mg
= mc
->mc_rotor
) == NULL
)
270 ASSERT(vd
->vdev_mg
!= NULL
);
271 ASSERT3P(vd
->vdev_top
, ==, vd
);
272 ASSERT3P(mg
->mg_class
, ==, mc
);
273 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
274 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
280 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
281 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
283 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
284 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
285 atomic_add_64(&mc
->mc_space
, space_delta
);
286 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
290 metaslab_class_get_alloc(metaslab_class_t
*mc
)
292 return (mc
->mc_alloc
);
296 metaslab_class_get_deferred(metaslab_class_t
*mc
)
298 return (mc
->mc_deferred
);
302 metaslab_class_get_space(metaslab_class_t
*mc
)
304 return (mc
->mc_space
);
308 metaslab_class_get_dspace(metaslab_class_t
*mc
)
310 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
314 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
316 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
320 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
323 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
326 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
327 vdev_t
*tvd
= rvd
->vdev_child
[c
];
328 metaslab_group_t
*mg
= tvd
->vdev_mg
;
331 * Skip any holes, uninitialized top-levels, or
332 * vdevs that are not in this metalab class.
334 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
335 mg
->mg_class
!= mc
) {
339 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
340 mc_hist
[i
] += mg
->mg_histogram
[i
];
343 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
344 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
346 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
350 * Calculate the metaslab class's fragmentation metric. The metric
351 * is weighted based on the space contribution of each metaslab group.
352 * The return value will be a number between 0 and 100 (inclusive), or
353 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
354 * zfs_frag_table for more information about the metric.
357 metaslab_class_fragmentation(metaslab_class_t
*mc
)
359 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
360 uint64_t fragmentation
= 0;
362 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
364 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
365 vdev_t
*tvd
= rvd
->vdev_child
[c
];
366 metaslab_group_t
*mg
= tvd
->vdev_mg
;
369 * Skip any holes, uninitialized top-levels,
370 * or vdevs that are not in this metalab class.
372 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
373 mg
->mg_class
!= mc
) {
378 * If a metaslab group does not contain a fragmentation
379 * metric then just bail out.
381 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
382 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
383 return (ZFS_FRAG_INVALID
);
387 * Determine how much this metaslab_group is contributing
388 * to the overall pool fragmentation metric.
390 fragmentation
+= mg
->mg_fragmentation
*
391 metaslab_group_get_space(mg
);
393 fragmentation
/= metaslab_class_get_space(mc
);
395 ASSERT3U(fragmentation
, <=, 100);
396 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
397 return (fragmentation
);
401 * Calculate the amount of expandable space that is available in
402 * this metaslab class. If a device is expanded then its expandable
403 * space will be the amount of allocatable space that is currently not
404 * part of this metaslab class.
407 metaslab_class_expandable_space(metaslab_class_t
*mc
)
409 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
412 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
413 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
415 vdev_t
*tvd
= rvd
->vdev_child
[c
];
416 metaslab_group_t
*mg
= tvd
->vdev_mg
;
418 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
419 mg
->mg_class
!= mc
) {
424 * Calculate if we have enough space to add additional
425 * metaslabs. We report the expandable space in terms
426 * of the metaslab size since that's the unit of expansion.
427 * Adjust by efi system partition size.
429 tspace
= tvd
->vdev_max_asize
- tvd
->vdev_asize
;
430 if (tspace
> mc
->mc_spa
->spa_bootsize
) {
431 tspace
-= mc
->mc_spa
->spa_bootsize
;
433 space
+= P2ALIGN(tspace
, 1ULL << tvd
->vdev_ms_shift
);
435 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
440 metaslab_compare(const void *x1
, const void *x2
)
442 const metaslab_t
*m1
= x1
;
443 const metaslab_t
*m2
= x2
;
445 if (m1
->ms_weight
< m2
->ms_weight
)
447 if (m1
->ms_weight
> m2
->ms_weight
)
451 * If the weights are identical, use the offset to force uniqueness.
453 if (m1
->ms_start
< m2
->ms_start
)
455 if (m1
->ms_start
> m2
->ms_start
)
458 ASSERT3P(m1
, ==, m2
);
464 * Verify that the space accounting on disk matches the in-core range_trees.
467 metaslab_verify_space(metaslab_t
*msp
, uint64_t txg
)
469 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
470 uint64_t allocated
= 0;
471 uint64_t sm_free_space
, msp_free_space
;
473 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
475 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
479 * We can only verify the metaslab space when we're called
480 * from syncing context with a loaded metaslab that has an allocated
481 * space map. Calling this in non-syncing context does not
482 * provide a consistent view of the metaslab since we're performing
483 * allocations in the future.
485 if (txg
!= spa_syncing_txg(spa
) || msp
->ms_sm
== NULL
||
489 sm_free_space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
) -
490 space_map_alloc_delta(msp
->ms_sm
);
493 * Account for future allocations since we would have already
494 * deducted that space from the ms_freetree.
496 for (int t
= 0; t
< TXG_CONCURRENT_STATES
; t
++) {
498 range_tree_space(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]);
501 msp_free_space
= range_tree_space(msp
->ms_allocatable
) + allocated
+
502 msp
->ms_deferspace
+ range_tree_space(msp
->ms_freed
);
504 VERIFY3U(sm_free_space
, ==, msp_free_space
);
508 * ==========================================================================
510 * ==========================================================================
513 * Update the allocatable flag and the metaslab group's capacity.
514 * The allocatable flag is set to true if the capacity is below
515 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
516 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
517 * transitions from allocatable to non-allocatable or vice versa then the
518 * metaslab group's class is updated to reflect the transition.
521 metaslab_group_alloc_update(metaslab_group_t
*mg
)
523 vdev_t
*vd
= mg
->mg_vd
;
524 metaslab_class_t
*mc
= mg
->mg_class
;
525 vdev_stat_t
*vs
= &vd
->vdev_stat
;
526 boolean_t was_allocatable
;
527 boolean_t was_initialized
;
529 ASSERT(vd
== vd
->vdev_top
);
530 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_READER
), ==,
533 mutex_enter(&mg
->mg_lock
);
534 was_allocatable
= mg
->mg_allocatable
;
535 was_initialized
= mg
->mg_initialized
;
537 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
540 mutex_enter(&mc
->mc_lock
);
543 * If the metaslab group was just added then it won't
544 * have any space until we finish syncing out this txg.
545 * At that point we will consider it initialized and available
546 * for allocations. We also don't consider non-activated
547 * metaslab groups (e.g. vdevs that are in the middle of being removed)
548 * to be initialized, because they can't be used for allocation.
550 mg
->mg_initialized
= metaslab_group_initialized(mg
);
551 if (!was_initialized
&& mg
->mg_initialized
) {
553 } else if (was_initialized
&& !mg
->mg_initialized
) {
554 ASSERT3U(mc
->mc_groups
, >, 0);
557 if (mg
->mg_initialized
)
558 mg
->mg_no_free_space
= B_FALSE
;
561 * A metaslab group is considered allocatable if it has plenty
562 * of free space or is not heavily fragmented. We only take
563 * fragmentation into account if the metaslab group has a valid
564 * fragmentation metric (i.e. a value between 0 and 100).
566 mg
->mg_allocatable
= (mg
->mg_activation_count
> 0 &&
567 mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
568 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
569 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
572 * The mc_alloc_groups maintains a count of the number of
573 * groups in this metaslab class that are still above the
574 * zfs_mg_noalloc_threshold. This is used by the allocating
575 * threads to determine if they should avoid allocations to
576 * a given group. The allocator will avoid allocations to a group
577 * if that group has reached or is below the zfs_mg_noalloc_threshold
578 * and there are still other groups that are above the threshold.
579 * When a group transitions from allocatable to non-allocatable or
580 * vice versa we update the metaslab class to reflect that change.
581 * When the mc_alloc_groups value drops to 0 that means that all
582 * groups have reached the zfs_mg_noalloc_threshold making all groups
583 * eligible for allocations. This effectively means that all devices
584 * are balanced again.
586 if (was_allocatable
&& !mg
->mg_allocatable
)
587 mc
->mc_alloc_groups
--;
588 else if (!was_allocatable
&& mg
->mg_allocatable
)
589 mc
->mc_alloc_groups
++;
590 mutex_exit(&mc
->mc_lock
);
592 mutex_exit(&mg
->mg_lock
);
596 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
598 metaslab_group_t
*mg
;
600 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
601 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
602 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
603 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
606 mg
->mg_activation_count
= 0;
607 mg
->mg_initialized
= B_FALSE
;
608 mg
->mg_no_free_space
= B_TRUE
;
609 refcount_create_tracked(&mg
->mg_alloc_queue_depth
);
611 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
612 minclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
);
618 metaslab_group_destroy(metaslab_group_t
*mg
)
620 ASSERT(mg
->mg_prev
== NULL
);
621 ASSERT(mg
->mg_next
== NULL
);
623 * We may have gone below zero with the activation count
624 * either because we never activated in the first place or
625 * because we're done, and possibly removing the vdev.
627 ASSERT(mg
->mg_activation_count
<= 0);
629 taskq_destroy(mg
->mg_taskq
);
630 avl_destroy(&mg
->mg_metaslab_tree
);
631 mutex_destroy(&mg
->mg_lock
);
632 refcount_destroy(&mg
->mg_alloc_queue_depth
);
633 kmem_free(mg
, sizeof (metaslab_group_t
));
637 metaslab_group_activate(metaslab_group_t
*mg
)
639 metaslab_class_t
*mc
= mg
->mg_class
;
640 metaslab_group_t
*mgprev
, *mgnext
;
642 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
), !=, 0);
644 ASSERT(mc
->mc_rotor
!= mg
);
645 ASSERT(mg
->mg_prev
== NULL
);
646 ASSERT(mg
->mg_next
== NULL
);
647 ASSERT(mg
->mg_activation_count
<= 0);
649 if (++mg
->mg_activation_count
<= 0)
652 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
653 metaslab_group_alloc_update(mg
);
655 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
659 mgnext
= mgprev
->mg_next
;
660 mg
->mg_prev
= mgprev
;
661 mg
->mg_next
= mgnext
;
662 mgprev
->mg_next
= mg
;
663 mgnext
->mg_prev
= mg
;
669 * Passivate a metaslab group and remove it from the allocation rotor.
670 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
671 * a metaslab group. This function will momentarily drop spa_config_locks
672 * that are lower than the SCL_ALLOC lock (see comment below).
675 metaslab_group_passivate(metaslab_group_t
*mg
)
677 metaslab_class_t
*mc
= mg
->mg_class
;
678 spa_t
*spa
= mc
->mc_spa
;
679 metaslab_group_t
*mgprev
, *mgnext
;
680 int locks
= spa_config_held(spa
, SCL_ALL
, RW_WRITER
);
682 ASSERT3U(spa_config_held(spa
, SCL_ALLOC
| SCL_ZIO
, RW_WRITER
), ==,
683 (SCL_ALLOC
| SCL_ZIO
));
685 if (--mg
->mg_activation_count
!= 0) {
686 ASSERT(mc
->mc_rotor
!= mg
);
687 ASSERT(mg
->mg_prev
== NULL
);
688 ASSERT(mg
->mg_next
== NULL
);
689 ASSERT(mg
->mg_activation_count
< 0);
694 * The spa_config_lock is an array of rwlocks, ordered as
695 * follows (from highest to lowest):
696 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
697 * SCL_ZIO > SCL_FREE > SCL_VDEV
698 * (For more information about the spa_config_lock see spa_misc.c)
699 * The higher the lock, the broader its coverage. When we passivate
700 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
701 * config locks. However, the metaslab group's taskq might be trying
702 * to preload metaslabs so we must drop the SCL_ZIO lock and any
703 * lower locks to allow the I/O to complete. At a minimum,
704 * we continue to hold the SCL_ALLOC lock, which prevents any future
705 * allocations from taking place and any changes to the vdev tree.
707 spa_config_exit(spa
, locks
& ~(SCL_ZIO
- 1), spa
);
708 taskq_wait(mg
->mg_taskq
);
709 spa_config_enter(spa
, locks
& ~(SCL_ZIO
- 1), spa
, RW_WRITER
);
710 metaslab_group_alloc_update(mg
);
712 mgprev
= mg
->mg_prev
;
713 mgnext
= mg
->mg_next
;
718 mc
->mc_rotor
= mgnext
;
719 mgprev
->mg_next
= mgnext
;
720 mgnext
->mg_prev
= mgprev
;
728 metaslab_group_initialized(metaslab_group_t
*mg
)
730 vdev_t
*vd
= mg
->mg_vd
;
731 vdev_stat_t
*vs
= &vd
->vdev_stat
;
733 return (vs
->vs_space
!= 0 && mg
->mg_activation_count
> 0);
737 metaslab_group_get_space(metaslab_group_t
*mg
)
739 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
743 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
746 vdev_t
*vd
= mg
->mg_vd
;
747 uint64_t ashift
= vd
->vdev_ashift
;
750 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
753 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
756 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
757 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
759 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
760 metaslab_t
*msp
= vd
->vdev_ms
[m
];
762 if (msp
->ms_sm
== NULL
)
765 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
766 mg_hist
[i
+ ashift
] +=
767 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
770 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
771 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
773 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
777 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
779 metaslab_class_t
*mc
= mg
->mg_class
;
780 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
782 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
783 if (msp
->ms_sm
== NULL
)
786 mutex_enter(&mg
->mg_lock
);
787 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
788 mg
->mg_histogram
[i
+ ashift
] +=
789 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
790 mc
->mc_histogram
[i
+ ashift
] +=
791 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
793 mutex_exit(&mg
->mg_lock
);
797 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
799 metaslab_class_t
*mc
= mg
->mg_class
;
800 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
802 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
803 if (msp
->ms_sm
== NULL
)
806 mutex_enter(&mg
->mg_lock
);
807 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
808 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
809 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
810 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
811 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
813 mg
->mg_histogram
[i
+ ashift
] -=
814 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
815 mc
->mc_histogram
[i
+ ashift
] -=
816 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
818 mutex_exit(&mg
->mg_lock
);
822 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
824 ASSERT(msp
->ms_group
== NULL
);
825 mutex_enter(&mg
->mg_lock
);
828 avl_add(&mg
->mg_metaslab_tree
, msp
);
829 mutex_exit(&mg
->mg_lock
);
831 mutex_enter(&msp
->ms_lock
);
832 metaslab_group_histogram_add(mg
, msp
);
833 mutex_exit(&msp
->ms_lock
);
837 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
839 mutex_enter(&msp
->ms_lock
);
840 metaslab_group_histogram_remove(mg
, msp
);
841 mutex_exit(&msp
->ms_lock
);
843 mutex_enter(&mg
->mg_lock
);
844 ASSERT(msp
->ms_group
== mg
);
845 avl_remove(&mg
->mg_metaslab_tree
, msp
);
846 msp
->ms_group
= NULL
;
847 mutex_exit(&mg
->mg_lock
);
851 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
854 * Although in principle the weight can be any value, in
855 * practice we do not use values in the range [1, 511].
857 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
858 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
860 mutex_enter(&mg
->mg_lock
);
861 ASSERT(msp
->ms_group
== mg
);
862 avl_remove(&mg
->mg_metaslab_tree
, msp
);
863 msp
->ms_weight
= weight
;
864 avl_add(&mg
->mg_metaslab_tree
, msp
);
865 mutex_exit(&mg
->mg_lock
);
869 * Calculate the fragmentation for a given metaslab group. We can use
870 * a simple average here since all metaslabs within the group must have
871 * the same size. The return value will be a value between 0 and 100
872 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
873 * group have a fragmentation metric.
876 metaslab_group_fragmentation(metaslab_group_t
*mg
)
878 vdev_t
*vd
= mg
->mg_vd
;
879 uint64_t fragmentation
= 0;
880 uint64_t valid_ms
= 0;
882 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
883 metaslab_t
*msp
= vd
->vdev_ms
[m
];
885 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
889 fragmentation
+= msp
->ms_fragmentation
;
892 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
893 return (ZFS_FRAG_INVALID
);
895 fragmentation
/= valid_ms
;
896 ASSERT3U(fragmentation
, <=, 100);
897 return (fragmentation
);
901 * Determine if a given metaslab group should skip allocations. A metaslab
902 * group should avoid allocations if its free capacity is less than the
903 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
904 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
905 * that can still handle allocations. If the allocation throttle is enabled
906 * then we skip allocations to devices that have reached their maximum
907 * allocation queue depth unless the selected metaslab group is the only
908 * eligible group remaining.
911 metaslab_group_allocatable(metaslab_group_t
*mg
, metaslab_group_t
*rotor
,
914 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
915 metaslab_class_t
*mc
= mg
->mg_class
;
918 * We can only consider skipping this metaslab group if it's
919 * in the normal metaslab class and there are other metaslab
920 * groups to select from. Otherwise, we always consider it eligible
923 if (mc
!= spa_normal_class(spa
) || mc
->mc_groups
<= 1)
927 * If the metaslab group's mg_allocatable flag is set (see comments
928 * in metaslab_group_alloc_update() for more information) and
929 * the allocation throttle is disabled then allow allocations to this
930 * device. However, if the allocation throttle is enabled then
931 * check if we have reached our allocation limit (mg_alloc_queue_depth)
932 * to determine if we should allow allocations to this metaslab group.
933 * If all metaslab groups are no longer considered allocatable
934 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
935 * gang block size then we allow allocations on this metaslab group
936 * regardless of the mg_allocatable or throttle settings.
938 if (mg
->mg_allocatable
) {
939 metaslab_group_t
*mgp
;
941 uint64_t qmax
= mg
->mg_max_alloc_queue_depth
;
943 if (!mc
->mc_alloc_throttle_enabled
)
947 * If this metaslab group does not have any free space, then
948 * there is no point in looking further.
950 if (mg
->mg_no_free_space
)
953 qdepth
= refcount_count(&mg
->mg_alloc_queue_depth
);
956 * If this metaslab group is below its qmax or it's
957 * the only allocatable metasable group, then attempt
958 * to allocate from it.
960 if (qdepth
< qmax
|| mc
->mc_alloc_groups
== 1)
962 ASSERT3U(mc
->mc_alloc_groups
, >, 1);
965 * Since this metaslab group is at or over its qmax, we
966 * need to determine if there are metaslab groups after this
967 * one that might be able to handle this allocation. This is
968 * racy since we can't hold the locks for all metaslab
969 * groups at the same time when we make this check.
971 for (mgp
= mg
->mg_next
; mgp
!= rotor
; mgp
= mgp
->mg_next
) {
972 qmax
= mgp
->mg_max_alloc_queue_depth
;
974 qdepth
= refcount_count(&mgp
->mg_alloc_queue_depth
);
977 * If there is another metaslab group that
978 * might be able to handle the allocation, then
979 * we return false so that we skip this group.
981 if (qdepth
< qmax
&& !mgp
->mg_no_free_space
)
986 * We didn't find another group to handle the allocation
987 * so we can't skip this metaslab group even though
988 * we are at or over our qmax.
992 } else if (mc
->mc_alloc_groups
== 0 || psize
== SPA_MINBLOCKSIZE
) {
999 * ==========================================================================
1000 * Range tree callbacks
1001 * ==========================================================================
1005 * Comparison function for the private size-ordered tree. Tree is sorted
1006 * by size, larger sizes at the end of the tree.
1009 metaslab_rangesize_compare(const void *x1
, const void *x2
)
1011 const range_seg_t
*r1
= x1
;
1012 const range_seg_t
*r2
= x2
;
1013 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
1014 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
1016 if (rs_size1
< rs_size2
)
1018 if (rs_size1
> rs_size2
)
1021 if (r1
->rs_start
< r2
->rs_start
)
1024 if (r1
->rs_start
> r2
->rs_start
)
1031 * Create any block allocator specific components. The current allocators
1032 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1035 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
1037 metaslab_t
*msp
= arg
;
1039 ASSERT3P(rt
->rt_arg
, ==, msp
);
1040 ASSERT(msp
->ms_allocatable
== NULL
);
1042 avl_create(&msp
->ms_allocatable_by_size
, metaslab_rangesize_compare
,
1043 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
1047 * Destroy the block allocator specific components.
1050 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
1052 metaslab_t
*msp
= arg
;
1054 ASSERT3P(rt
->rt_arg
, ==, msp
);
1055 ASSERT3P(msp
->ms_allocatable
, ==, rt
);
1056 ASSERT0(avl_numnodes(&msp
->ms_allocatable_by_size
));
1058 avl_destroy(&msp
->ms_allocatable_by_size
);
1062 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
1064 metaslab_t
*msp
= arg
;
1066 ASSERT3P(rt
->rt_arg
, ==, msp
);
1067 ASSERT3P(msp
->ms_allocatable
, ==, rt
);
1068 VERIFY(!msp
->ms_condensing
);
1069 avl_add(&msp
->ms_allocatable_by_size
, rs
);
1073 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
1075 metaslab_t
*msp
= arg
;
1077 ASSERT3P(rt
->rt_arg
, ==, msp
);
1078 ASSERT3P(msp
->ms_allocatable
, ==, rt
);
1079 VERIFY(!msp
->ms_condensing
);
1080 avl_remove(&msp
->ms_allocatable_by_size
, rs
);
1084 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
1086 metaslab_t
*msp
= arg
;
1088 ASSERT3P(rt
->rt_arg
, ==, msp
);
1089 ASSERT3P(msp
->ms_allocatable
, ==, rt
);
1092 * Normally one would walk the tree freeing nodes along the way.
1093 * Since the nodes are shared with the range trees we can avoid
1094 * walking all nodes and just reinitialize the avl tree. The nodes
1095 * will be freed by the range tree, so we don't want to free them here.
1097 avl_create(&msp
->ms_allocatable_by_size
, metaslab_rangesize_compare
,
1098 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
1101 static range_tree_ops_t metaslab_rt_ops
= {
1103 metaslab_rt_destroy
,
1110 * ==========================================================================
1111 * Common allocator routines
1112 * ==========================================================================
1116 * Return the maximum contiguous segment within the metaslab.
1119 metaslab_block_maxsize(metaslab_t
*msp
)
1121 avl_tree_t
*t
= &msp
->ms_allocatable_by_size
;
1124 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
1127 return (rs
->rs_end
- rs
->rs_start
);
1130 static range_seg_t
*
1131 metaslab_block_find(avl_tree_t
*t
, uint64_t start
, uint64_t size
)
1133 range_seg_t
*rs
, rsearch
;
1136 rsearch
.rs_start
= start
;
1137 rsearch
.rs_end
= start
+ size
;
1139 rs
= avl_find(t
, &rsearch
, &where
);
1141 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1148 * This is a helper function that can be used by the allocator to find
1149 * a suitable block to allocate. This will search the specified AVL
1150 * tree looking for a block that matches the specified criteria.
1153 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
1156 range_seg_t
*rs
= metaslab_block_find(t
, *cursor
, size
);
1158 while (rs
!= NULL
) {
1159 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
1161 if (offset
+ size
<= rs
->rs_end
) {
1162 *cursor
= offset
+ size
;
1165 rs
= AVL_NEXT(t
, rs
);
1169 * If we know we've searched the whole map (*cursor == 0), give up.
1170 * Otherwise, reset the cursor to the beginning and try again.
1176 return (metaslab_block_picker(t
, cursor
, size
, align
));
1180 * ==========================================================================
1181 * The first-fit block allocator
1182 * ==========================================================================
1185 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
1188 * Find the largest power of 2 block size that evenly divides the
1189 * requested size. This is used to try to allocate blocks with similar
1190 * alignment from the same area of the metaslab (i.e. same cursor
1191 * bucket) but it does not guarantee that other allocations sizes
1192 * may exist in the same region.
1194 uint64_t align
= size
& -size
;
1195 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1196 avl_tree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1198 return (metaslab_block_picker(t
, cursor
, size
, align
));
1201 static metaslab_ops_t metaslab_ff_ops
= {
1206 * ==========================================================================
1207 * Dynamic block allocator -
1208 * Uses the first fit allocation scheme until space get low and then
1209 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1210 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1211 * ==========================================================================
1214 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1217 * Find the largest power of 2 block size that evenly divides the
1218 * requested size. This is used to try to allocate blocks with similar
1219 * alignment from the same area of the metaslab (i.e. same cursor
1220 * bucket) but it does not guarantee that other allocations sizes
1221 * may exist in the same region.
1223 uint64_t align
= size
& -size
;
1224 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1225 range_tree_t
*rt
= msp
->ms_allocatable
;
1226 avl_tree_t
*t
= &rt
->rt_root
;
1227 uint64_t max_size
= metaslab_block_maxsize(msp
);
1228 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1230 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1231 ASSERT3U(avl_numnodes(t
), ==,
1232 avl_numnodes(&msp
->ms_allocatable_by_size
));
1234 if (max_size
< size
)
1238 * If we're running low on space switch to using the size
1239 * sorted AVL tree (best-fit).
1241 if (max_size
< metaslab_df_alloc_threshold
||
1242 free_pct
< metaslab_df_free_pct
) {
1243 t
= &msp
->ms_allocatable_by_size
;
1247 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1250 static metaslab_ops_t metaslab_df_ops
= {
1255 * ==========================================================================
1256 * Cursor fit block allocator -
1257 * Select the largest region in the metaslab, set the cursor to the beginning
1258 * of the range and the cursor_end to the end of the range. As allocations
1259 * are made advance the cursor. Continue allocating from the cursor until
1260 * the range is exhausted and then find a new range.
1261 * ==========================================================================
1264 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1266 range_tree_t
*rt
= msp
->ms_allocatable
;
1267 avl_tree_t
*t
= &msp
->ms_allocatable_by_size
;
1268 uint64_t *cursor
= &msp
->ms_lbas
[0];
1269 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1270 uint64_t offset
= 0;
1272 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1273 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1275 ASSERT3U(*cursor_end
, >=, *cursor
);
1277 if ((*cursor
+ size
) > *cursor_end
) {
1280 rs
= avl_last(&msp
->ms_allocatable_by_size
);
1281 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1284 *cursor
= rs
->rs_start
;
1285 *cursor_end
= rs
->rs_end
;
1294 static metaslab_ops_t metaslab_cf_ops
= {
1299 * ==========================================================================
1300 * New dynamic fit allocator -
1301 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1302 * contiguous blocks. If no region is found then just use the largest segment
1304 * ==========================================================================
1308 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1309 * to request from the allocator.
1311 uint64_t metaslab_ndf_clump_shift
= 4;
1314 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1316 avl_tree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1318 range_seg_t
*rs
, rsearch
;
1319 uint64_t hbit
= highbit64(size
);
1320 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1321 uint64_t max_size
= metaslab_block_maxsize(msp
);
1323 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1324 ASSERT3U(avl_numnodes(t
), ==,
1325 avl_numnodes(&msp
->ms_allocatable_by_size
));
1327 if (max_size
< size
)
1330 rsearch
.rs_start
= *cursor
;
1331 rsearch
.rs_end
= *cursor
+ size
;
1333 rs
= avl_find(t
, &rsearch
, &where
);
1334 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1335 t
= &msp
->ms_allocatable_by_size
;
1337 rsearch
.rs_start
= 0;
1338 rsearch
.rs_end
= MIN(max_size
,
1339 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1340 rs
= avl_find(t
, &rsearch
, &where
);
1342 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1346 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1347 *cursor
= rs
->rs_start
+ size
;
1348 return (rs
->rs_start
);
1353 static metaslab_ops_t metaslab_ndf_ops
= {
1357 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1360 * ==========================================================================
1362 * ==========================================================================
1366 * Wait for any in-progress metaslab loads to complete.
1369 metaslab_load_wait(metaslab_t
*msp
)
1371 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1373 while (msp
->ms_loading
) {
1374 ASSERT(!msp
->ms_loaded
);
1375 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1380 metaslab_load(metaslab_t
*msp
)
1383 boolean_t success
= B_FALSE
;
1385 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1386 ASSERT(!msp
->ms_loaded
);
1387 ASSERT(!msp
->ms_loading
);
1389 msp
->ms_loading
= B_TRUE
;
1391 * Nobody else can manipulate a loading metaslab, so it's now safe
1392 * to drop the lock. This way we don't have to hold the lock while
1393 * reading the spacemap from disk.
1395 mutex_exit(&msp
->ms_lock
);
1398 * If the space map has not been allocated yet, then treat
1399 * all the space in the metaslab as free and add it to ms_allocatable.
1401 if (msp
->ms_sm
!= NULL
) {
1402 error
= space_map_load(msp
->ms_sm
, msp
->ms_allocatable
,
1405 range_tree_add(msp
->ms_allocatable
,
1406 msp
->ms_start
, msp
->ms_size
);
1409 success
= (error
== 0);
1411 mutex_enter(&msp
->ms_lock
);
1412 msp
->ms_loading
= B_FALSE
;
1415 ASSERT3P(msp
->ms_group
, !=, NULL
);
1416 msp
->ms_loaded
= B_TRUE
;
1419 * If the metaslab already has a spacemap, then we need to
1420 * remove all segments from the defer tree; otherwise, the
1421 * metaslab is completely empty and we can skip this.
1423 if (msp
->ms_sm
!= NULL
) {
1424 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1425 range_tree_walk(msp
->ms_defer
[t
],
1426 range_tree_remove
, msp
->ms_allocatable
);
1429 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1431 cv_broadcast(&msp
->ms_load_cv
);
1436 metaslab_unload(metaslab_t
*msp
)
1438 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1439 range_tree_vacate(msp
->ms_allocatable
, NULL
, NULL
);
1440 msp
->ms_loaded
= B_FALSE
;
1441 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1442 msp
->ms_max_size
= 0;
1446 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1449 vdev_t
*vd
= mg
->mg_vd
;
1450 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1454 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1455 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1456 mutex_init(&ms
->ms_sync_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1457 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1459 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1460 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1463 * We only open space map objects that already exist. All others
1464 * will be opened when we finally allocate an object for it.
1467 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1468 ms
->ms_size
, vd
->vdev_ashift
);
1471 kmem_free(ms
, sizeof (metaslab_t
));
1475 ASSERT(ms
->ms_sm
!= NULL
);
1479 * We create the main range tree here, but we don't create the
1480 * other range trees until metaslab_sync_done(). This serves
1481 * two purposes: it allows metaslab_sync_done() to detect the
1482 * addition of new space; and for debugging, it ensures that we'd
1483 * data fault on any attempt to use this metaslab before it's ready.
1485 ms
->ms_allocatable
= range_tree_create(&metaslab_rt_ops
, ms
);
1486 metaslab_group_add(mg
, ms
);
1488 metaslab_set_fragmentation(ms
);
1491 * If we're opening an existing pool (txg == 0) or creating
1492 * a new one (txg == TXG_INITIAL), all space is available now.
1493 * If we're adding space to an existing pool, the new space
1494 * does not become available until after this txg has synced.
1495 * The metaslab's weight will also be initialized when we sync
1496 * out this txg. This ensures that we don't attempt to allocate
1497 * from it before we have initialized it completely.
1499 if (txg
<= TXG_INITIAL
)
1500 metaslab_sync_done(ms
, 0);
1503 * If metaslab_debug_load is set and we're initializing a metaslab
1504 * that has an allocated space map object then load the its space
1505 * map so that can verify frees.
1507 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1508 mutex_enter(&ms
->ms_lock
);
1509 VERIFY0(metaslab_load(ms
));
1510 mutex_exit(&ms
->ms_lock
);
1514 vdev_dirty(vd
, 0, NULL
, txg
);
1515 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1524 metaslab_fini(metaslab_t
*msp
)
1526 metaslab_group_t
*mg
= msp
->ms_group
;
1528 metaslab_group_remove(mg
, msp
);
1530 mutex_enter(&msp
->ms_lock
);
1531 VERIFY(msp
->ms_group
== NULL
);
1532 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1534 space_map_close(msp
->ms_sm
);
1536 metaslab_unload(msp
);
1537 range_tree_destroy(msp
->ms_allocatable
);
1538 range_tree_destroy(msp
->ms_freeing
);
1539 range_tree_destroy(msp
->ms_freed
);
1541 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1542 range_tree_destroy(msp
->ms_allocating
[t
]);
1545 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1546 range_tree_destroy(msp
->ms_defer
[t
]);
1548 ASSERT0(msp
->ms_deferspace
);
1550 range_tree_destroy(msp
->ms_checkpointing
);
1552 mutex_exit(&msp
->ms_lock
);
1553 cv_destroy(&msp
->ms_load_cv
);
1554 mutex_destroy(&msp
->ms_lock
);
1555 mutex_destroy(&msp
->ms_sync_lock
);
1557 kmem_free(msp
, sizeof (metaslab_t
));
1560 #define FRAGMENTATION_TABLE_SIZE 17
1563 * This table defines a segment size based fragmentation metric that will
1564 * allow each metaslab to derive its own fragmentation value. This is done
1565 * by calculating the space in each bucket of the spacemap histogram and
1566 * multiplying that by the fragmetation metric in this table. Doing
1567 * this for all buckets and dividing it by the total amount of free
1568 * space in this metaslab (i.e. the total free space in all buckets) gives
1569 * us the fragmentation metric. This means that a high fragmentation metric
1570 * equates to most of the free space being comprised of small segments.
1571 * Conversely, if the metric is low, then most of the free space is in
1572 * large segments. A 10% change in fragmentation equates to approximately
1573 * double the number of segments.
1575 * This table defines 0% fragmented space using 16MB segments. Testing has
1576 * shown that segments that are greater than or equal to 16MB do not suffer
1577 * from drastic performance problems. Using this value, we derive the rest
1578 * of the table. Since the fragmentation value is never stored on disk, it
1579 * is possible to change these calculations in the future.
1581 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1601 * Calclate the metaslab's fragmentation metric. A return value
1602 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1603 * not support this metric. Otherwise, the return value should be in the
1607 metaslab_set_fragmentation(metaslab_t
*msp
)
1609 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1610 uint64_t fragmentation
= 0;
1612 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1613 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1615 if (!feature_enabled
) {
1616 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1621 * A null space map means that the entire metaslab is free
1622 * and thus is not fragmented.
1624 if (msp
->ms_sm
== NULL
) {
1625 msp
->ms_fragmentation
= 0;
1630 * If this metaslab's space map has not been upgraded, flag it
1631 * so that we upgrade next time we encounter it.
1633 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1634 uint64_t txg
= spa_syncing_txg(spa
);
1635 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1638 * If we've reached the final dirty txg, then we must
1639 * be shutting down the pool. We don't want to dirty
1640 * any data past this point so skip setting the condense
1641 * flag. We can retry this action the next time the pool
1644 if (spa_writeable(spa
) && txg
< spa_final_dirty_txg(spa
)) {
1645 msp
->ms_condense_wanted
= B_TRUE
;
1646 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1647 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1648 "ms_id %llu, vdev_id %llu", txg
, msp
->ms_id
,
1651 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1655 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1657 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1659 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1660 FRAGMENTATION_TABLE_SIZE
- 1);
1662 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1665 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1668 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1669 fragmentation
+= space
* zfs_frag_table
[idx
];
1673 fragmentation
/= total
;
1674 ASSERT3U(fragmentation
, <=, 100);
1676 msp
->ms_fragmentation
= fragmentation
;
1680 * Compute a weight -- a selection preference value -- for the given metaslab.
1681 * This is based on the amount of free space, the level of fragmentation,
1682 * the LBA range, and whether the metaslab is loaded.
1685 metaslab_space_weight(metaslab_t
*msp
)
1687 metaslab_group_t
*mg
= msp
->ms_group
;
1688 vdev_t
*vd
= mg
->mg_vd
;
1689 uint64_t weight
, space
;
1691 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1692 ASSERT(!vd
->vdev_removing
);
1695 * The baseline weight is the metaslab's free space.
1697 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1699 if (metaslab_fragmentation_factor_enabled
&&
1700 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1702 * Use the fragmentation information to inversely scale
1703 * down the baseline weight. We need to ensure that we
1704 * don't exclude this metaslab completely when it's 100%
1705 * fragmented. To avoid this we reduce the fragmented value
1708 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1711 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1712 * this metaslab again. The fragmentation metric may have
1713 * decreased the space to something smaller than
1714 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1715 * so that we can consume any remaining space.
1717 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1718 space
= SPA_MINBLOCKSIZE
;
1723 * Modern disks have uniform bit density and constant angular velocity.
1724 * Therefore, the outer recording zones are faster (higher bandwidth)
1725 * than the inner zones by the ratio of outer to inner track diameter,
1726 * which is typically around 2:1. We account for this by assigning
1727 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1728 * In effect, this means that we'll select the metaslab with the most
1729 * free bandwidth rather than simply the one with the most free space.
1731 if (metaslab_lba_weighting_enabled
) {
1732 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1733 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1737 * If this metaslab is one we're actively using, adjust its
1738 * weight to make it preferable to any inactive metaslab so
1739 * we'll polish it off. If the fragmentation on this metaslab
1740 * has exceed our threshold, then don't mark it active.
1742 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1743 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1744 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1747 WEIGHT_SET_SPACEBASED(weight
);
1752 * Return the weight of the specified metaslab, according to the segment-based
1753 * weighting algorithm. The metaslab must be loaded. This function can
1754 * be called within a sync pass since it relies only on the metaslab's
1755 * range tree which is always accurate when the metaslab is loaded.
1758 metaslab_weight_from_range_tree(metaslab_t
*msp
)
1760 uint64_t weight
= 0;
1761 uint32_t segments
= 0;
1763 ASSERT(msp
->ms_loaded
);
1765 for (int i
= RANGE_TREE_HISTOGRAM_SIZE
- 1; i
>= SPA_MINBLOCKSHIFT
;
1767 uint8_t shift
= msp
->ms_group
->mg_vd
->vdev_ashift
;
1768 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1771 segments
+= msp
->ms_allocatable
->rt_histogram
[i
];
1774 * The range tree provides more precision than the space map
1775 * and must be downgraded so that all values fit within the
1776 * space map's histogram. This allows us to compare loaded
1777 * vs. unloaded metaslabs to determine which metaslab is
1778 * considered "best".
1783 if (segments
!= 0) {
1784 WEIGHT_SET_COUNT(weight
, segments
);
1785 WEIGHT_SET_INDEX(weight
, i
);
1786 WEIGHT_SET_ACTIVE(weight
, 0);
1794 * Calculate the weight based on the on-disk histogram. This should only
1795 * be called after a sync pass has completely finished since the on-disk
1796 * information is updated in metaslab_sync().
1799 metaslab_weight_from_spacemap(metaslab_t
*msp
)
1801 uint64_t weight
= 0;
1803 for (int i
= SPACE_MAP_HISTOGRAM_SIZE
- 1; i
>= 0; i
--) {
1804 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] != 0) {
1805 WEIGHT_SET_COUNT(weight
,
1806 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
1807 WEIGHT_SET_INDEX(weight
, i
+
1808 msp
->ms_sm
->sm_shift
);
1809 WEIGHT_SET_ACTIVE(weight
, 0);
1817 * Compute a segment-based weight for the specified metaslab. The weight
1818 * is determined by highest bucket in the histogram. The information
1819 * for the highest bucket is encoded into the weight value.
1822 metaslab_segment_weight(metaslab_t
*msp
)
1824 metaslab_group_t
*mg
= msp
->ms_group
;
1825 uint64_t weight
= 0;
1826 uint8_t shift
= mg
->mg_vd
->vdev_ashift
;
1828 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1831 * The metaslab is completely free.
1833 if (space_map_allocated(msp
->ms_sm
) == 0) {
1834 int idx
= highbit64(msp
->ms_size
) - 1;
1835 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1837 if (idx
< max_idx
) {
1838 WEIGHT_SET_COUNT(weight
, 1ULL);
1839 WEIGHT_SET_INDEX(weight
, idx
);
1841 WEIGHT_SET_COUNT(weight
, 1ULL << (idx
- max_idx
));
1842 WEIGHT_SET_INDEX(weight
, max_idx
);
1844 WEIGHT_SET_ACTIVE(weight
, 0);
1845 ASSERT(!WEIGHT_IS_SPACEBASED(weight
));
1850 ASSERT3U(msp
->ms_sm
->sm_dbuf
->db_size
, ==, sizeof (space_map_phys_t
));
1853 * If the metaslab is fully allocated then just make the weight 0.
1855 if (space_map_allocated(msp
->ms_sm
) == msp
->ms_size
)
1858 * If the metaslab is already loaded, then use the range tree to
1859 * determine the weight. Otherwise, we rely on the space map information
1860 * to generate the weight.
1862 if (msp
->ms_loaded
) {
1863 weight
= metaslab_weight_from_range_tree(msp
);
1865 weight
= metaslab_weight_from_spacemap(msp
);
1869 * If the metaslab was active the last time we calculated its weight
1870 * then keep it active. We want to consume the entire region that
1871 * is associated with this weight.
1873 if (msp
->ms_activation_weight
!= 0 && weight
!= 0)
1874 WEIGHT_SET_ACTIVE(weight
, WEIGHT_GET_ACTIVE(msp
->ms_weight
));
1879 * Determine if we should attempt to allocate from this metaslab. If the
1880 * metaslab has a maximum size then we can quickly determine if the desired
1881 * allocation size can be satisfied. Otherwise, if we're using segment-based
1882 * weighting then we can determine the maximum allocation that this metaslab
1883 * can accommodate based on the index encoded in the weight. If we're using
1884 * space-based weights then rely on the entire weight (excluding the weight
1888 metaslab_should_allocate(metaslab_t
*msp
, uint64_t asize
)
1890 boolean_t should_allocate
;
1892 if (msp
->ms_max_size
!= 0)
1893 return (msp
->ms_max_size
>= asize
);
1895 if (!WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
1897 * The metaslab segment weight indicates segments in the
1898 * range [2^i, 2^(i+1)), where i is the index in the weight.
1899 * Since the asize might be in the middle of the range, we
1900 * should attempt the allocation if asize < 2^(i+1).
1902 should_allocate
= (asize
<
1903 1ULL << (WEIGHT_GET_INDEX(msp
->ms_weight
) + 1));
1905 should_allocate
= (asize
<=
1906 (msp
->ms_weight
& ~METASLAB_WEIGHT_TYPE
));
1908 return (should_allocate
);
1912 metaslab_weight(metaslab_t
*msp
)
1914 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1915 spa_t
*spa
= vd
->vdev_spa
;
1918 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1921 * If this vdev is in the process of being removed, there is nothing
1922 * for us to do here.
1924 if (vd
->vdev_removing
)
1927 metaslab_set_fragmentation(msp
);
1930 * Update the maximum size if the metaslab is loaded. This will
1931 * ensure that we get an accurate maximum size if newly freed space
1932 * has been added back into the free tree.
1935 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1938 * Segment-based weighting requires space map histogram support.
1940 if (zfs_metaslab_segment_weight_enabled
&&
1941 spa_feature_is_enabled(spa
, SPA_FEATURE_SPACEMAP_HISTOGRAM
) &&
1942 (msp
->ms_sm
== NULL
|| msp
->ms_sm
->sm_dbuf
->db_size
==
1943 sizeof (space_map_phys_t
))) {
1944 weight
= metaslab_segment_weight(msp
);
1946 weight
= metaslab_space_weight(msp
);
1952 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1954 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1956 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1957 metaslab_load_wait(msp
);
1958 if (!msp
->ms_loaded
) {
1959 int error
= metaslab_load(msp
);
1961 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1966 msp
->ms_activation_weight
= msp
->ms_weight
;
1967 metaslab_group_sort(msp
->ms_group
, msp
,
1968 msp
->ms_weight
| activation_weight
);
1970 ASSERT(msp
->ms_loaded
);
1971 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1977 metaslab_passivate(metaslab_t
*msp
, uint64_t weight
)
1979 uint64_t size
= weight
& ~METASLAB_WEIGHT_TYPE
;
1982 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1983 * this metaslab again. In that case, it had better be empty,
1984 * or we would be leaving space on the table.
1986 ASSERT(size
>= SPA_MINBLOCKSIZE
||
1987 range_tree_is_empty(msp
->ms_allocatable
));
1988 ASSERT0(weight
& METASLAB_ACTIVE_MASK
);
1990 msp
->ms_activation_weight
= 0;
1991 metaslab_group_sort(msp
->ms_group
, msp
, weight
);
1992 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1996 * Segment-based metaslabs are activated once and remain active until
1997 * we either fail an allocation attempt (similar to space-based metaslabs)
1998 * or have exhausted the free space in zfs_metaslab_switch_threshold
1999 * buckets since the metaslab was activated. This function checks to see
2000 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
2001 * metaslab and passivates it proactively. This will allow us to select a
2002 * metaslabs with larger contiguous region if any remaining within this
2003 * metaslab group. If we're in sync pass > 1, then we continue using this
2004 * metaslab so that we don't dirty more block and cause more sync passes.
2007 metaslab_segment_may_passivate(metaslab_t
*msp
)
2009 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2011 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
) || spa_sync_pass(spa
) > 1)
2015 * Since we are in the middle of a sync pass, the most accurate
2016 * information that is accessible to us is the in-core range tree
2017 * histogram; calculate the new weight based on that information.
2019 uint64_t weight
= metaslab_weight_from_range_tree(msp
);
2020 int activation_idx
= WEIGHT_GET_INDEX(msp
->ms_activation_weight
);
2021 int current_idx
= WEIGHT_GET_INDEX(weight
);
2023 if (current_idx
<= activation_idx
- zfs_metaslab_switch_threshold
)
2024 metaslab_passivate(msp
, weight
);
2028 metaslab_preload(void *arg
)
2030 metaslab_t
*msp
= arg
;
2031 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2033 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
2035 mutex_enter(&msp
->ms_lock
);
2036 metaslab_load_wait(msp
);
2037 if (!msp
->ms_loaded
)
2038 (void) metaslab_load(msp
);
2039 msp
->ms_selected_txg
= spa_syncing_txg(spa
);
2040 mutex_exit(&msp
->ms_lock
);
2044 metaslab_group_preload(metaslab_group_t
*mg
)
2046 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2048 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2051 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
2052 taskq_wait(mg
->mg_taskq
);
2056 mutex_enter(&mg
->mg_lock
);
2059 * Load the next potential metaslabs
2061 for (msp
= avl_first(t
); msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
2062 ASSERT3P(msp
->ms_group
, ==, mg
);
2065 * We preload only the maximum number of metaslabs specified
2066 * by metaslab_preload_limit. If a metaslab is being forced
2067 * to condense then we preload it too. This will ensure
2068 * that force condensing happens in the next txg.
2070 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
2074 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
2075 msp
, TQ_SLEEP
) != (uintptr_t)NULL
);
2077 mutex_exit(&mg
->mg_lock
);
2081 * Determine if the space map's on-disk footprint is past our tolerance
2082 * for inefficiency. We would like to use the following criteria to make
2085 * 1. The size of the space map object should not dramatically increase as a
2086 * result of writing out the free space range tree.
2088 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2089 * times the size than the free space range tree representation
2090 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1MB).
2092 * 3. The on-disk size of the space map should actually decrease.
2094 * Unfortunately, we cannot compute the on-disk size of the space map in this
2095 * context because we cannot accurately compute the effects of compression, etc.
2096 * Instead, we apply the heuristic described in the block comment for
2097 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2098 * is greater than a threshold number of blocks.
2101 metaslab_should_condense(metaslab_t
*msp
)
2103 space_map_t
*sm
= msp
->ms_sm
;
2104 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
2105 uint64_t vdev_blocksize
= 1 << vd
->vdev_ashift
;
2106 uint64_t current_txg
= spa_syncing_txg(vd
->vdev_spa
);
2108 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2109 ASSERT(msp
->ms_loaded
);
2112 * Allocations and frees in early passes are generally more space
2113 * efficient (in terms of blocks described in space map entries)
2114 * than the ones in later passes (e.g. we don't compress after
2115 * sync pass 5) and condensing a metaslab multiple times in a txg
2116 * could degrade performance.
2118 * Thus we prefer condensing each metaslab at most once every txg at
2119 * the earliest sync pass possible. If a metaslab is eligible for
2120 * condensing again after being considered for condensing within the
2121 * same txg, it will hopefully be dirty in the next txg where it will
2122 * be condensed at an earlier pass.
2124 if (msp
->ms_condense_checked_txg
== current_txg
)
2126 msp
->ms_condense_checked_txg
= current_txg
;
2129 * We always condense metaslabs that are empty and metaslabs for
2130 * which a condense request has been made.
2132 if (avl_is_empty(&msp
->ms_allocatable_by_size
) ||
2133 msp
->ms_condense_wanted
)
2136 uint64_t object_size
= space_map_length(msp
->ms_sm
);
2137 uint64_t optimal_size
= space_map_estimate_optimal_size(sm
,
2138 msp
->ms_allocatable
, SM_NO_VDEVID
);
2140 dmu_object_info_t doi
;
2141 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
2142 uint64_t record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
2144 return (object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
2145 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
2149 * Condense the on-disk space map representation to its minimized form.
2150 * The minimized form consists of a small number of allocations followed by
2151 * the entries of the free range tree.
2154 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
2156 range_tree_t
*condense_tree
;
2157 space_map_t
*sm
= msp
->ms_sm
;
2159 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2160 ASSERT(msp
->ms_loaded
);
2162 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2163 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
2164 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
2165 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
2166 space_map_length(msp
->ms_sm
),
2167 avl_numnodes(&msp
->ms_allocatable
->rt_root
),
2168 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
2170 msp
->ms_condense_wanted
= B_FALSE
;
2173 * Create an range tree that is 100% allocated. We remove segments
2174 * that have been freed in this txg, any deferred frees that exist,
2175 * and any allocation in the future. Removing segments should be
2176 * a relatively inexpensive operation since we expect these trees to
2177 * have a small number of nodes.
2179 condense_tree
= range_tree_create(NULL
, NULL
);
2180 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
2182 range_tree_walk(msp
->ms_freeing
, range_tree_remove
, condense_tree
);
2183 range_tree_walk(msp
->ms_freed
, range_tree_remove
, condense_tree
);
2185 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2186 range_tree_walk(msp
->ms_defer
[t
],
2187 range_tree_remove
, condense_tree
);
2190 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2191 range_tree_walk(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
],
2192 range_tree_remove
, condense_tree
);
2196 * We're about to drop the metaslab's lock thus allowing
2197 * other consumers to change it's content. Set the
2198 * metaslab's ms_condensing flag to ensure that
2199 * allocations on this metaslab do not occur while we're
2200 * in the middle of committing it to disk. This is only critical
2201 * for ms_allocatable as all other range trees use per txg
2202 * views of their content.
2204 msp
->ms_condensing
= B_TRUE
;
2206 mutex_exit(&msp
->ms_lock
);
2207 space_map_truncate(sm
, zfs_metaslab_sm_blksz
, tx
);
2210 * While we would ideally like to create a space map representation
2211 * that consists only of allocation records, doing so can be
2212 * prohibitively expensive because the in-core free tree can be
2213 * large, and therefore computationally expensive to subtract
2214 * from the condense_tree. Instead we sync out two trees, a cheap
2215 * allocation only tree followed by the in-core free tree. While not
2216 * optimal, this is typically close to optimal, and much cheaper to
2219 space_map_write(sm
, condense_tree
, SM_ALLOC
, SM_NO_VDEVID
, tx
);
2220 range_tree_vacate(condense_tree
, NULL
, NULL
);
2221 range_tree_destroy(condense_tree
);
2223 space_map_write(sm
, msp
->ms_allocatable
, SM_FREE
, SM_NO_VDEVID
, tx
);
2224 mutex_enter(&msp
->ms_lock
);
2225 msp
->ms_condensing
= B_FALSE
;
2229 * Write a metaslab to disk in the context of the specified transaction group.
2232 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
2234 metaslab_group_t
*mg
= msp
->ms_group
;
2235 vdev_t
*vd
= mg
->mg_vd
;
2236 spa_t
*spa
= vd
->vdev_spa
;
2237 objset_t
*mos
= spa_meta_objset(spa
);
2238 range_tree_t
*alloctree
= msp
->ms_allocating
[txg
& TXG_MASK
];
2240 uint64_t object
= space_map_object(msp
->ms_sm
);
2242 ASSERT(!vd
->vdev_ishole
);
2245 * This metaslab has just been added so there's no work to do now.
2247 if (msp
->ms_freeing
== NULL
) {
2248 ASSERT3P(alloctree
, ==, NULL
);
2252 ASSERT3P(alloctree
, !=, NULL
);
2253 ASSERT3P(msp
->ms_freeing
, !=, NULL
);
2254 ASSERT3P(msp
->ms_freed
, !=, NULL
);
2255 ASSERT3P(msp
->ms_checkpointing
, !=, NULL
);
2258 * Normally, we don't want to process a metaslab if there are no
2259 * allocations or frees to perform. However, if the metaslab is being
2260 * forced to condense and it's loaded, we need to let it through.
2262 if (range_tree_is_empty(alloctree
) &&
2263 range_tree_is_empty(msp
->ms_freeing
) &&
2264 range_tree_is_empty(msp
->ms_checkpointing
) &&
2265 !(msp
->ms_loaded
&& msp
->ms_condense_wanted
))
2269 VERIFY(txg
<= spa_final_dirty_txg(spa
));
2272 * The only state that can actually be changing concurrently with
2273 * metaslab_sync() is the metaslab's ms_allocatable. No other
2274 * thread can be modifying this txg's alloc, freeing,
2275 * freed, or space_map_phys_t. We drop ms_lock whenever we
2276 * could call into the DMU, because the DMU can call down to us
2277 * (e.g. via zio_free()) at any time.
2279 * The spa_vdev_remove_thread() can be reading metaslab state
2280 * concurrently, and it is locked out by the ms_sync_lock. Note
2281 * that the ms_lock is insufficient for this, because it is dropped
2282 * by space_map_write().
2284 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
2286 if (msp
->ms_sm
== NULL
) {
2287 uint64_t new_object
;
2289 new_object
= space_map_alloc(mos
, zfs_metaslab_sm_blksz
, tx
);
2290 VERIFY3U(new_object
, !=, 0);
2292 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
2293 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
));
2294 ASSERT(msp
->ms_sm
!= NULL
);
2297 if (!range_tree_is_empty(msp
->ms_checkpointing
) &&
2298 vd
->vdev_checkpoint_sm
== NULL
) {
2299 ASSERT(spa_has_checkpoint(spa
));
2301 uint64_t new_object
= space_map_alloc(mos
,
2302 vdev_standard_sm_blksz
, tx
);
2303 VERIFY3U(new_object
, !=, 0);
2305 VERIFY0(space_map_open(&vd
->vdev_checkpoint_sm
,
2306 mos
, new_object
, 0, vd
->vdev_asize
, vd
->vdev_ashift
));
2307 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
2310 * We save the space map object as an entry in vdev_top_zap
2311 * so it can be retrieved when the pool is reopened after an
2312 * export or through zdb.
2314 VERIFY0(zap_add(vd
->vdev_spa
->spa_meta_objset
,
2315 vd
->vdev_top_zap
, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM
,
2316 sizeof (new_object
), 1, &new_object
, tx
));
2319 mutex_enter(&msp
->ms_sync_lock
);
2320 mutex_enter(&msp
->ms_lock
);
2323 * Note: metaslab_condense() clears the space map's histogram.
2324 * Therefore we must verify and remove this histogram before
2327 metaslab_group_histogram_verify(mg
);
2328 metaslab_class_histogram_verify(mg
->mg_class
);
2329 metaslab_group_histogram_remove(mg
, msp
);
2331 if (msp
->ms_loaded
&& metaslab_should_condense(msp
)) {
2332 metaslab_condense(msp
, txg
, tx
);
2334 mutex_exit(&msp
->ms_lock
);
2335 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
,
2337 space_map_write(msp
->ms_sm
, msp
->ms_freeing
, SM_FREE
,
2339 mutex_enter(&msp
->ms_lock
);
2342 if (!range_tree_is_empty(msp
->ms_checkpointing
)) {
2343 ASSERT(spa_has_checkpoint(spa
));
2344 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
2347 * Since we are doing writes to disk and the ms_checkpointing
2348 * tree won't be changing during that time, we drop the
2349 * ms_lock while writing to the checkpoint space map.
2351 mutex_exit(&msp
->ms_lock
);
2352 space_map_write(vd
->vdev_checkpoint_sm
,
2353 msp
->ms_checkpointing
, SM_FREE
, SM_NO_VDEVID
, tx
);
2354 mutex_enter(&msp
->ms_lock
);
2355 space_map_update(vd
->vdev_checkpoint_sm
);
2357 spa
->spa_checkpoint_info
.sci_dspace
+=
2358 range_tree_space(msp
->ms_checkpointing
);
2359 vd
->vdev_stat
.vs_checkpoint_space
+=
2360 range_tree_space(msp
->ms_checkpointing
);
2361 ASSERT3U(vd
->vdev_stat
.vs_checkpoint_space
, ==,
2362 -vd
->vdev_checkpoint_sm
->sm_alloc
);
2364 range_tree_vacate(msp
->ms_checkpointing
, NULL
, NULL
);
2367 if (msp
->ms_loaded
) {
2369 * When the space map is loaded, we have an accurate
2370 * histogram in the range tree. This gives us an opportunity
2371 * to bring the space map's histogram up-to-date so we clear
2372 * it first before updating it.
2374 space_map_histogram_clear(msp
->ms_sm
);
2375 space_map_histogram_add(msp
->ms_sm
, msp
->ms_allocatable
, tx
);
2378 * Since we've cleared the histogram we need to add back
2379 * any free space that has already been processed, plus
2380 * any deferred space. This allows the on-disk histogram
2381 * to accurately reflect all free space even if some space
2382 * is not yet available for allocation (i.e. deferred).
2384 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freed
, tx
);
2387 * Add back any deferred free space that has not been
2388 * added back into the in-core free tree yet. This will
2389 * ensure that we don't end up with a space map histogram
2390 * that is completely empty unless the metaslab is fully
2393 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2394 space_map_histogram_add(msp
->ms_sm
,
2395 msp
->ms_defer
[t
], tx
);
2400 * Always add the free space from this sync pass to the space
2401 * map histogram. We want to make sure that the on-disk histogram
2402 * accounts for all free space. If the space map is not loaded,
2403 * then we will lose some accuracy but will correct it the next
2404 * time we load the space map.
2406 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freeing
, tx
);
2408 metaslab_group_histogram_add(mg
, msp
);
2409 metaslab_group_histogram_verify(mg
);
2410 metaslab_class_histogram_verify(mg
->mg_class
);
2413 * For sync pass 1, we avoid traversing this txg's free range tree
2414 * and instead will just swap the pointers for freeing and
2415 * freed. We can safely do this since the freed_tree is
2416 * guaranteed to be empty on the initial pass.
2418 if (spa_sync_pass(spa
) == 1) {
2419 range_tree_swap(&msp
->ms_freeing
, &msp
->ms_freed
);
2421 range_tree_vacate(msp
->ms_freeing
,
2422 range_tree_add
, msp
->ms_freed
);
2424 range_tree_vacate(alloctree
, NULL
, NULL
);
2426 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
2427 ASSERT0(range_tree_space(msp
->ms_allocating
[TXG_CLEAN(txg
)
2429 ASSERT0(range_tree_space(msp
->ms_freeing
));
2430 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2432 mutex_exit(&msp
->ms_lock
);
2434 if (object
!= space_map_object(msp
->ms_sm
)) {
2435 object
= space_map_object(msp
->ms_sm
);
2436 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
2437 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
2439 mutex_exit(&msp
->ms_sync_lock
);
2444 * Called after a transaction group has completely synced to mark
2445 * all of the metaslab's free space as usable.
2448 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
2450 metaslab_group_t
*mg
= msp
->ms_group
;
2451 vdev_t
*vd
= mg
->mg_vd
;
2452 spa_t
*spa
= vd
->vdev_spa
;
2453 range_tree_t
**defer_tree
;
2454 int64_t alloc_delta
, defer_delta
;
2455 boolean_t defer_allowed
= B_TRUE
;
2457 ASSERT(!vd
->vdev_ishole
);
2459 mutex_enter(&msp
->ms_lock
);
2462 * If this metaslab is just becoming available, initialize its
2463 * range trees and add its capacity to the vdev.
2465 if (msp
->ms_freed
== NULL
) {
2466 for (int t
= 0; t
< TXG_SIZE
; t
++) {
2467 ASSERT(msp
->ms_allocating
[t
] == NULL
);
2469 msp
->ms_allocating
[t
] = range_tree_create(NULL
, NULL
);
2472 ASSERT3P(msp
->ms_freeing
, ==, NULL
);
2473 msp
->ms_freeing
= range_tree_create(NULL
, NULL
);
2475 ASSERT3P(msp
->ms_freed
, ==, NULL
);
2476 msp
->ms_freed
= range_tree_create(NULL
, NULL
);
2478 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2479 ASSERT(msp
->ms_defer
[t
] == NULL
);
2481 msp
->ms_defer
[t
] = range_tree_create(NULL
, NULL
);
2484 ASSERT3P(msp
->ms_checkpointing
, ==, NULL
);
2485 msp
->ms_checkpointing
= range_tree_create(NULL
, NULL
);
2487 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
2489 ASSERT0(range_tree_space(msp
->ms_freeing
));
2490 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2492 defer_tree
= &msp
->ms_defer
[txg
% TXG_DEFER_SIZE
];
2494 uint64_t free_space
= metaslab_class_get_space(spa_normal_class(spa
)) -
2495 metaslab_class_get_alloc(spa_normal_class(spa
));
2496 if (free_space
<= spa_get_slop_space(spa
) || vd
->vdev_removing
) {
2497 defer_allowed
= B_FALSE
;
2501 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
2502 if (defer_allowed
) {
2503 defer_delta
= range_tree_space(msp
->ms_freed
) -
2504 range_tree_space(*defer_tree
);
2506 defer_delta
-= range_tree_space(*defer_tree
);
2509 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
2512 * If there's a metaslab_load() in progress, wait for it to complete
2513 * so that we have a consistent view of the in-core space map.
2515 metaslab_load_wait(msp
);
2518 * Move the frees from the defer_tree back to the free
2519 * range tree (if it's loaded). Swap the freed_tree and
2520 * the defer_tree -- this is safe to do because we've
2521 * just emptied out the defer_tree.
2523 range_tree_vacate(*defer_tree
,
2524 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_allocatable
);
2525 if (defer_allowed
) {
2526 range_tree_swap(&msp
->ms_freed
, defer_tree
);
2528 range_tree_vacate(msp
->ms_freed
,
2529 msp
->ms_loaded
? range_tree_add
: NULL
,
2530 msp
->ms_allocatable
);
2532 space_map_update(msp
->ms_sm
);
2534 msp
->ms_deferspace
+= defer_delta
;
2535 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2536 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2537 if (msp
->ms_deferspace
!= 0) {
2539 * Keep syncing this metaslab until all deferred frees
2540 * are back in circulation.
2542 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2546 * Calculate the new weights before unloading any metaslabs.
2547 * This will give us the most accurate weighting.
2549 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
2552 * If the metaslab is loaded and we've not tried to load or allocate
2553 * from it in 'metaslab_unload_delay' txgs, then unload it.
2555 if (msp
->ms_loaded
&&
2556 msp
->ms_selected_txg
+ metaslab_unload_delay
< txg
) {
2557 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2558 VERIFY0(range_tree_space(
2559 msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]));
2562 if (!metaslab_debug_unload
)
2563 metaslab_unload(msp
);
2566 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
2567 ASSERT0(range_tree_space(msp
->ms_freeing
));
2568 ASSERT0(range_tree_space(msp
->ms_freed
));
2569 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2571 mutex_exit(&msp
->ms_lock
);
2575 metaslab_sync_reassess(metaslab_group_t
*mg
)
2577 spa_t
*spa
= mg
->mg_class
->mc_spa
;
2579 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2580 metaslab_group_alloc_update(mg
);
2581 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2584 * Preload the next potential metaslabs but only on active
2585 * metaslab groups. We can get into a state where the metaslab
2586 * is no longer active since we dirty metaslabs as we remove a
2587 * a device, thus potentially making the metaslab group eligible
2590 if (mg
->mg_activation_count
> 0) {
2591 metaslab_group_preload(mg
);
2593 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2597 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2599 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2600 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2601 uint64_t start
= msp
->ms_id
;
2603 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2604 return (1ULL << 63);
2607 return ((start
- offset
) << ms_shift
);
2609 return ((offset
- start
) << ms_shift
);
2614 * ==========================================================================
2615 * Metaslab allocation tracing facility
2616 * ==========================================================================
2618 kstat_t
*metaslab_trace_ksp
;
2619 kstat_named_t metaslab_trace_over_limit
;
2622 metaslab_alloc_trace_init(void)
2624 ASSERT(metaslab_alloc_trace_cache
== NULL
);
2625 metaslab_alloc_trace_cache
= kmem_cache_create(
2626 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t
),
2627 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
2628 metaslab_trace_ksp
= kstat_create("zfs", 0, "metaslab_trace_stats",
2629 "misc", KSTAT_TYPE_NAMED
, 1, KSTAT_FLAG_VIRTUAL
);
2630 if (metaslab_trace_ksp
!= NULL
) {
2631 metaslab_trace_ksp
->ks_data
= &metaslab_trace_over_limit
;
2632 kstat_named_init(&metaslab_trace_over_limit
,
2633 "metaslab_trace_over_limit", KSTAT_DATA_UINT64
);
2634 kstat_install(metaslab_trace_ksp
);
2639 metaslab_alloc_trace_fini(void)
2641 if (metaslab_trace_ksp
!= NULL
) {
2642 kstat_delete(metaslab_trace_ksp
);
2643 metaslab_trace_ksp
= NULL
;
2645 kmem_cache_destroy(metaslab_alloc_trace_cache
);
2646 metaslab_alloc_trace_cache
= NULL
;
2650 * Add an allocation trace element to the allocation tracing list.
2653 metaslab_trace_add(zio_alloc_list_t
*zal
, metaslab_group_t
*mg
,
2654 metaslab_t
*msp
, uint64_t psize
, uint32_t dva_id
, uint64_t offset
)
2656 if (!metaslab_trace_enabled
)
2660 * When the tracing list reaches its maximum we remove
2661 * the second element in the list before adding a new one.
2662 * By removing the second element we preserve the original
2663 * entry as a clue to what allocations steps have already been
2666 if (zal
->zal_size
== metaslab_trace_max_entries
) {
2667 metaslab_alloc_trace_t
*mat_next
;
2669 panic("too many entries in allocation list");
2671 atomic_inc_64(&metaslab_trace_over_limit
.value
.ui64
);
2673 mat_next
= list_next(&zal
->zal_list
, list_head(&zal
->zal_list
));
2674 list_remove(&zal
->zal_list
, mat_next
);
2675 kmem_cache_free(metaslab_alloc_trace_cache
, mat_next
);
2678 metaslab_alloc_trace_t
*mat
=
2679 kmem_cache_alloc(metaslab_alloc_trace_cache
, KM_SLEEP
);
2680 list_link_init(&mat
->mat_list_node
);
2683 mat
->mat_size
= psize
;
2684 mat
->mat_dva_id
= dva_id
;
2685 mat
->mat_offset
= offset
;
2686 mat
->mat_weight
= 0;
2689 mat
->mat_weight
= msp
->ms_weight
;
2692 * The list is part of the zio so locking is not required. Only
2693 * a single thread will perform allocations for a given zio.
2695 list_insert_tail(&zal
->zal_list
, mat
);
2698 ASSERT3U(zal
->zal_size
, <=, metaslab_trace_max_entries
);
2702 metaslab_trace_init(zio_alloc_list_t
*zal
)
2704 list_create(&zal
->zal_list
, sizeof (metaslab_alloc_trace_t
),
2705 offsetof(metaslab_alloc_trace_t
, mat_list_node
));
2710 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2712 metaslab_alloc_trace_t
*mat
;
2714 while ((mat
= list_remove_head(&zal
->zal_list
)) != NULL
)
2715 kmem_cache_free(metaslab_alloc_trace_cache
, mat
);
2716 list_destroy(&zal
->zal_list
);
2721 * ==========================================================================
2722 * Metaslab block operations
2723 * ==========================================================================
2727 metaslab_group_alloc_increment(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
)
2729 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2730 flags
& METASLAB_DONT_THROTTLE
)
2733 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2734 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2737 (void) refcount_add(&mg
->mg_alloc_queue_depth
, tag
);
2741 metaslab_group_alloc_decrement(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
)
2743 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2744 flags
& METASLAB_DONT_THROTTLE
)
2747 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2748 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2751 (void) refcount_remove(&mg
->mg_alloc_queue_depth
, tag
);
2755 metaslab_group_alloc_verify(spa_t
*spa
, const blkptr_t
*bp
, void *tag
)
2758 const dva_t
*dva
= bp
->blk_dva
;
2759 int ndvas
= BP_GET_NDVAS(bp
);
2761 for (int d
= 0; d
< ndvas
; d
++) {
2762 uint64_t vdev
= DVA_GET_VDEV(&dva
[d
]);
2763 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2764 VERIFY(refcount_not_held(&mg
->mg_alloc_queue_depth
, tag
));
2770 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
, uint64_t txg
)
2773 range_tree_t
*rt
= msp
->ms_allocatable
;
2774 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
2776 VERIFY(!msp
->ms_condensing
);
2778 start
= mc
->mc_ops
->msop_alloc(msp
, size
);
2779 if (start
!= -1ULL) {
2780 metaslab_group_t
*mg
= msp
->ms_group
;
2781 vdev_t
*vd
= mg
->mg_vd
;
2783 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
2784 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2785 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
2786 range_tree_remove(rt
, start
, size
);
2788 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
2789 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2791 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
], start
, size
);
2793 /* Track the last successful allocation */
2794 msp
->ms_alloc_txg
= txg
;
2795 metaslab_verify_space(msp
, txg
);
2799 * Now that we've attempted the allocation we need to update the
2800 * metaslab's maximum block size since it may have changed.
2802 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
2807 metaslab_group_alloc_normal(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
2808 uint64_t asize
, uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2810 metaslab_t
*msp
= NULL
;
2811 uint64_t offset
= -1ULL;
2812 uint64_t activation_weight
;
2813 uint64_t target_distance
;
2816 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2817 for (i
= 0; i
< d
; i
++) {
2818 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2819 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2824 metaslab_t
*search
= kmem_alloc(sizeof (*search
), KM_SLEEP
);
2825 search
->ms_weight
= UINT64_MAX
;
2826 search
->ms_start
= 0;
2828 boolean_t was_active
;
2829 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2832 mutex_enter(&mg
->mg_lock
);
2835 * Find the metaslab with the highest weight that is less
2836 * than what we've already tried. In the common case, this
2837 * means that we will examine each metaslab at most once.
2838 * Note that concurrent callers could reorder metaslabs
2839 * by activation/passivation once we have dropped the mg_lock.
2840 * If a metaslab is activated by another thread, and we fail
2841 * to allocate from the metaslab we have selected, we may
2842 * not try the newly-activated metaslab, and instead activate
2843 * another metaslab. This is not optimal, but generally
2844 * does not cause any problems (a possible exception being
2845 * if every metaslab is completely full except for the
2846 * the newly-activated metaslab which we fail to examine).
2848 msp
= avl_find(t
, search
, &idx
);
2850 msp
= avl_nearest(t
, idx
, AVL_AFTER
);
2851 for (; msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
2853 if (!metaslab_should_allocate(msp
, asize
)) {
2854 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
2860 * If the selected metaslab is condensing, skip it.
2862 if (msp
->ms_condensing
)
2865 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2866 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2869 target_distance
= min_distance
+
2870 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2873 for (i
= 0; i
< d
; i
++) {
2874 if (metaslab_distance(msp
, &dva
[i
]) <
2881 mutex_exit(&mg
->mg_lock
);
2883 kmem_free(search
, sizeof (*search
));
2886 search
->ms_weight
= msp
->ms_weight
;
2887 search
->ms_start
= msp
->ms_start
+ 1;
2889 mutex_enter(&msp
->ms_lock
);
2892 * Ensure that the metaslab we have selected is still
2893 * capable of handling our request. It's possible that
2894 * another thread may have changed the weight while we
2895 * were blocked on the metaslab lock. We check the
2896 * active status first to see if we need to reselect
2899 if (was_active
&& !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
)) {
2900 mutex_exit(&msp
->ms_lock
);
2904 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2905 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2906 metaslab_passivate(msp
,
2907 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2908 mutex_exit(&msp
->ms_lock
);
2912 if (metaslab_activate(msp
, activation_weight
) != 0) {
2913 mutex_exit(&msp
->ms_lock
);
2916 msp
->ms_selected_txg
= txg
;
2919 * Now that we have the lock, recheck to see if we should
2920 * continue to use this metaslab for this allocation. The
2921 * the metaslab is now loaded so metaslab_should_allocate() can
2922 * accurately determine if the allocation attempt should
2925 if (!metaslab_should_allocate(msp
, asize
)) {
2926 /* Passivate this metaslab and select a new one. */
2927 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
2933 * If this metaslab is currently condensing then pick again as
2934 * we can't manipulate this metaslab until it's committed
2937 if (msp
->ms_condensing
) {
2938 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
2940 mutex_exit(&msp
->ms_lock
);
2944 offset
= metaslab_block_alloc(msp
, asize
, txg
);
2945 metaslab_trace_add(zal
, mg
, msp
, asize
, d
, offset
);
2947 if (offset
!= -1ULL) {
2948 /* Proactively passivate the metaslab, if needed */
2949 metaslab_segment_may_passivate(msp
);
2953 ASSERT(msp
->ms_loaded
);
2956 * We were unable to allocate from this metaslab so determine
2957 * a new weight for this metaslab. Now that we have loaded
2958 * the metaslab we can provide a better hint to the metaslab
2961 * For space-based metaslabs, we use the maximum block size.
2962 * This information is only available when the metaslab
2963 * is loaded and is more accurate than the generic free
2964 * space weight that was calculated by metaslab_weight().
2965 * This information allows us to quickly compare the maximum
2966 * available allocation in the metaslab to the allocation
2967 * size being requested.
2969 * For segment-based metaslabs, determine the new weight
2970 * based on the highest bucket in the range tree. We
2971 * explicitly use the loaded segment weight (i.e. the range
2972 * tree histogram) since it contains the space that is
2973 * currently available for allocation and is accurate
2974 * even within a sync pass.
2976 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
2977 uint64_t weight
= metaslab_block_maxsize(msp
);
2978 WEIGHT_SET_SPACEBASED(weight
);
2979 metaslab_passivate(msp
, weight
);
2981 metaslab_passivate(msp
,
2982 metaslab_weight_from_range_tree(msp
));
2986 * We have just failed an allocation attempt, check
2987 * that metaslab_should_allocate() agrees. Otherwise,
2988 * we may end up in an infinite loop retrying the same
2991 ASSERT(!metaslab_should_allocate(msp
, asize
));
2992 mutex_exit(&msp
->ms_lock
);
2994 mutex_exit(&msp
->ms_lock
);
2995 kmem_free(search
, sizeof (*search
));
3000 metaslab_group_alloc(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
3001 uint64_t asize
, uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
3004 ASSERT(mg
->mg_initialized
);
3006 offset
= metaslab_group_alloc_normal(mg
, zal
, asize
, txg
,
3007 min_distance
, dva
, d
);
3009 mutex_enter(&mg
->mg_lock
);
3010 if (offset
== -1ULL) {
3011 mg
->mg_failed_allocations
++;
3012 metaslab_trace_add(zal
, mg
, NULL
, asize
, d
,
3013 TRACE_GROUP_FAILURE
);
3014 if (asize
== SPA_GANGBLOCKSIZE
) {
3016 * This metaslab group was unable to allocate
3017 * the minimum gang block size so it must be out of
3018 * space. We must notify the allocation throttle
3019 * to start skipping allocation attempts to this
3020 * metaslab group until more space becomes available.
3021 * Note: this failure cannot be caused by the
3022 * allocation throttle since the allocation throttle
3023 * is only responsible for skipping devices and
3024 * not failing block allocations.
3026 mg
->mg_no_free_space
= B_TRUE
;
3029 mg
->mg_allocations
++;
3030 mutex_exit(&mg
->mg_lock
);
3035 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
3036 * on the same vdev as an existing DVA of this BP, then try to allocate it
3037 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
3040 int ditto_same_vdev_distance_shift
= 3;
3043 * Allocate a block for the specified i/o.
3046 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
3047 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
,
3048 zio_alloc_list_t
*zal
)
3050 metaslab_group_t
*mg
, *rotor
;
3052 boolean_t try_hard
= B_FALSE
;
3054 ASSERT(!DVA_IS_VALID(&dva
[d
]));
3057 * For testing, make some blocks above a certain size be gang blocks.
3059 if (psize
>= metaslab_force_ganging
&& (ddi_get_lbolt() & 3) == 0) {
3060 metaslab_trace_add(zal
, NULL
, NULL
, psize
, d
, TRACE_FORCE_GANG
);
3061 return (SET_ERROR(ENOSPC
));
3065 * Start at the rotor and loop through all mgs until we find something.
3066 * Note that there's no locking on mc_rotor or mc_aliquot because
3067 * nothing actually breaks if we miss a few updates -- we just won't
3068 * allocate quite as evenly. It all balances out over time.
3070 * If we are doing ditto or log blocks, try to spread them across
3071 * consecutive vdevs. If we're forced to reuse a vdev before we've
3072 * allocated all of our ditto blocks, then try and spread them out on
3073 * that vdev as much as possible. If it turns out to not be possible,
3074 * gradually lower our standards until anything becomes acceptable.
3075 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3076 * gives us hope of containing our fault domains to something we're
3077 * able to reason about. Otherwise, any two top-level vdev failures
3078 * will guarantee the loss of data. With consecutive allocation,
3079 * only two adjacent top-level vdev failures will result in data loss.
3081 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3082 * ourselves on the same vdev as our gang block header. That
3083 * way, we can hope for locality in vdev_cache, plus it makes our
3084 * fault domains something tractable.
3087 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
3090 * It's possible the vdev we're using as the hint no
3091 * longer exists or its mg has been closed (e.g. by
3092 * device removal). Consult the rotor when
3095 if (vd
!= NULL
&& vd
->vdev_mg
!= NULL
) {
3098 if (flags
& METASLAB_HINTBP_AVOID
&&
3099 mg
->mg_next
!= NULL
)
3104 } else if (d
!= 0) {
3105 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
3106 mg
= vd
->vdev_mg
->mg_next
;
3112 * If the hint put us into the wrong metaslab class, or into a
3113 * metaslab group that has been passivated, just follow the rotor.
3115 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
3121 boolean_t allocatable
;
3123 ASSERT(mg
->mg_activation_count
== 1);
3127 * Don't allocate from faulted devices.
3130 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
3131 allocatable
= vdev_allocatable(vd
);
3132 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
3134 allocatable
= vdev_allocatable(vd
);
3138 * Determine if the selected metaslab group is eligible
3139 * for allocations. If we're ganging then don't allow
3140 * this metaslab group to skip allocations since that would
3141 * inadvertently return ENOSPC and suspend the pool
3142 * even though space is still available.
3144 if (allocatable
&& !GANG_ALLOCATION(flags
) && !try_hard
) {
3145 allocatable
= metaslab_group_allocatable(mg
, rotor
,
3150 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3151 TRACE_NOT_ALLOCATABLE
);
3155 ASSERT(mg
->mg_initialized
);
3158 * Avoid writing single-copy data to a failing,
3159 * non-redundant vdev, unless we've already tried all
3162 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
3163 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
3164 d
== 0 && !try_hard
&& vd
->vdev_children
== 0) {
3165 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3170 ASSERT(mg
->mg_class
== mc
);
3173 * If we don't need to try hard, then require that the
3174 * block be 1/8th of the device away from any other DVAs
3175 * in this BP. If we are trying hard, allow any offset
3176 * to be used (distance=0).
3178 uint64_t distance
= 0;
3180 distance
= vd
->vdev_asize
>>
3181 ditto_same_vdev_distance_shift
;
3182 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
3186 uint64_t asize
= vdev_psize_to_asize(vd
, psize
);
3187 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
3189 uint64_t offset
= metaslab_group_alloc(mg
, zal
, asize
, txg
,
3192 if (offset
!= -1ULL) {
3194 * If we've just selected this metaslab group,
3195 * figure out whether the corresponding vdev is
3196 * over- or under-used relative to the pool,
3197 * and set an allocation bias to even it out.
3199 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
3200 vdev_stat_t
*vs
= &vd
->vdev_stat
;
3203 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
3204 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
3207 * Calculate how much more or less we should
3208 * try to allocate from this device during
3209 * this iteration around the rotor.
3210 * For example, if a device is 80% full
3211 * and the pool is 20% full then we should
3212 * reduce allocations by 60% on this device.
3214 * mg_bias = (20 - 80) * 512K / 100 = -307K
3216 * This reduces allocations by 307K for this
3219 mg
->mg_bias
= ((cu
- vu
) *
3220 (int64_t)mg
->mg_aliquot
) / 100;
3221 } else if (!metaslab_bias_enabled
) {
3225 if (atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
3226 mg
->mg_aliquot
+ mg
->mg_bias
) {
3227 mc
->mc_rotor
= mg
->mg_next
;
3231 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
3232 DVA_SET_OFFSET(&dva
[d
], offset
);
3233 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
3234 DVA_SET_ASIZE(&dva
[d
], asize
);
3239 mc
->mc_rotor
= mg
->mg_next
;
3241 } while ((mg
= mg
->mg_next
) != rotor
);
3244 * If we haven't tried hard, do so now.
3251 bzero(&dva
[d
], sizeof (dva_t
));
3253 metaslab_trace_add(zal
, rotor
, NULL
, psize
, d
, TRACE_ENOSPC
);
3254 return (SET_ERROR(ENOSPC
));
3258 metaslab_free_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t asize
,
3259 boolean_t checkpoint
)
3262 spa_t
*spa
= vd
->vdev_spa
;
3264 ASSERT(vdev_is_concrete(vd
));
3265 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3266 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
3268 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3270 VERIFY(!msp
->ms_condensing
);
3271 VERIFY3U(offset
, >=, msp
->ms_start
);
3272 VERIFY3U(offset
+ asize
, <=, msp
->ms_start
+ msp
->ms_size
);
3273 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3274 VERIFY0(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
));
3276 metaslab_check_free_impl(vd
, offset
, asize
);
3278 mutex_enter(&msp
->ms_lock
);
3279 if (range_tree_is_empty(msp
->ms_freeing
) &&
3280 range_tree_is_empty(msp
->ms_checkpointing
)) {
3281 vdev_dirty(vd
, VDD_METASLAB
, msp
, spa_syncing_txg(spa
));
3285 ASSERT(spa_has_checkpoint(spa
));
3286 range_tree_add(msp
->ms_checkpointing
, offset
, asize
);
3288 range_tree_add(msp
->ms_freeing
, offset
, asize
);
3290 mutex_exit(&msp
->ms_lock
);
3295 metaslab_free_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3296 uint64_t size
, void *arg
)
3298 boolean_t
*checkpoint
= arg
;
3300 ASSERT3P(checkpoint
, !=, NULL
);
3302 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
)
3303 vdev_indirect_mark_obsolete(vd
, offset
, size
);
3305 metaslab_free_impl(vd
, offset
, size
, *checkpoint
);
3309 metaslab_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
3310 boolean_t checkpoint
)
3312 spa_t
*spa
= vd
->vdev_spa
;
3314 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3316 if (spa_syncing_txg(spa
) > spa_freeze_txg(spa
))
3319 if (spa
->spa_vdev_removal
!= NULL
&&
3320 spa
->spa_vdev_removal
->svr_vdev
== vd
&&
3321 vdev_is_concrete(vd
)) {
3323 * Note: we check if the vdev is concrete because when
3324 * we complete the removal, we first change the vdev to be
3325 * an indirect vdev (in open context), and then (in syncing
3326 * context) clear spa_vdev_removal.
3328 free_from_removing_vdev(vd
, offset
, size
);
3329 } else if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
3330 vdev_indirect_mark_obsolete(vd
, offset
, size
);
3331 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
3332 metaslab_free_impl_cb
, &checkpoint
);
3334 metaslab_free_concrete(vd
, offset
, size
, checkpoint
);
3338 typedef struct remap_blkptr_cb_arg
{
3340 spa_remap_cb_t rbca_cb
;
3341 vdev_t
*rbca_remap_vd
;
3342 uint64_t rbca_remap_offset
;
3344 } remap_blkptr_cb_arg_t
;
3347 remap_blkptr_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3348 uint64_t size
, void *arg
)
3350 remap_blkptr_cb_arg_t
*rbca
= arg
;
3351 blkptr_t
*bp
= rbca
->rbca_bp
;
3353 /* We can not remap split blocks. */
3354 if (size
!= DVA_GET_ASIZE(&bp
->blk_dva
[0]))
3356 ASSERT0(inner_offset
);
3358 if (rbca
->rbca_cb
!= NULL
) {
3360 * At this point we know that we are not handling split
3361 * blocks and we invoke the callback on the previous
3362 * vdev which must be indirect.
3364 ASSERT3P(rbca
->rbca_remap_vd
->vdev_ops
, ==, &vdev_indirect_ops
);
3366 rbca
->rbca_cb(rbca
->rbca_remap_vd
->vdev_id
,
3367 rbca
->rbca_remap_offset
, size
, rbca
->rbca_cb_arg
);
3369 /* set up remap_blkptr_cb_arg for the next call */
3370 rbca
->rbca_remap_vd
= vd
;
3371 rbca
->rbca_remap_offset
= offset
;
3375 * The phys birth time is that of dva[0]. This ensures that we know
3376 * when each dva was written, so that resilver can determine which
3377 * blocks need to be scrubbed (i.e. those written during the time
3378 * the vdev was offline). It also ensures that the key used in
3379 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
3380 * we didn't change the phys_birth, a lookup in the ARC for a
3381 * remapped BP could find the data that was previously stored at
3382 * this vdev + offset.
3384 vdev_t
*oldvd
= vdev_lookup_top(vd
->vdev_spa
,
3385 DVA_GET_VDEV(&bp
->blk_dva
[0]));
3386 vdev_indirect_births_t
*vib
= oldvd
->vdev_indirect_births
;
3387 bp
->blk_phys_birth
= vdev_indirect_births_physbirth(vib
,
3388 DVA_GET_OFFSET(&bp
->blk_dva
[0]), DVA_GET_ASIZE(&bp
->blk_dva
[0]));
3390 DVA_SET_VDEV(&bp
->blk_dva
[0], vd
->vdev_id
);
3391 DVA_SET_OFFSET(&bp
->blk_dva
[0], offset
);
3395 * If the block pointer contains any indirect DVAs, modify them to refer to
3396 * concrete DVAs. Note that this will sometimes not be possible, leaving
3397 * the indirect DVA in place. This happens if the indirect DVA spans multiple
3398 * segments in the mapping (i.e. it is a "split block").
3400 * If the BP was remapped, calls the callback on the original dva (note the
3401 * callback can be called multiple times if the original indirect DVA refers
3402 * to another indirect DVA, etc).
3404 * Returns TRUE if the BP was remapped.
3407 spa_remap_blkptr(spa_t
*spa
, blkptr_t
*bp
, spa_remap_cb_t callback
, void *arg
)
3409 remap_blkptr_cb_arg_t rbca
;
3411 if (!zfs_remap_blkptr_enable
)
3414 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
))
3418 * Dedup BP's can not be remapped, because ddt_phys_select() depends
3419 * on DVA[0] being the same in the BP as in the DDT (dedup table).
3421 if (BP_GET_DEDUP(bp
))
3425 * Gang blocks can not be remapped, because
3426 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
3427 * the BP used to read the gang block header (GBH) being the same
3428 * as the DVA[0] that we allocated for the GBH.
3434 * Embedded BP's have no DVA to remap.
3436 if (BP_GET_NDVAS(bp
) < 1)
3440 * Note: we only remap dva[0]. If we remapped other dvas, we
3441 * would no longer know what their phys birth txg is.
3443 dva_t
*dva
= &bp
->blk_dva
[0];
3445 uint64_t offset
= DVA_GET_OFFSET(dva
);
3446 uint64_t size
= DVA_GET_ASIZE(dva
);
3447 vdev_t
*vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
3449 if (vd
->vdev_ops
->vdev_op_remap
== NULL
)
3453 rbca
.rbca_cb
= callback
;
3454 rbca
.rbca_remap_vd
= vd
;
3455 rbca
.rbca_remap_offset
= offset
;
3456 rbca
.rbca_cb_arg
= arg
;
3459 * remap_blkptr_cb() will be called in order for each level of
3460 * indirection, until a concrete vdev is reached or a split block is
3461 * encountered. old_vd and old_offset are updated within the callback
3462 * as we go from the one indirect vdev to the next one (either concrete
3463 * or indirect again) in that order.
3465 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
, remap_blkptr_cb
, &rbca
);
3467 /* Check if the DVA wasn't remapped because it is a split block */
3468 if (DVA_GET_VDEV(&rbca
.rbca_bp
->blk_dva
[0]) == vd
->vdev_id
)
3475 * Undo the allocation of a DVA which happened in the given transaction group.
3478 metaslab_unalloc_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
3482 uint64_t vdev
= DVA_GET_VDEV(dva
);
3483 uint64_t offset
= DVA_GET_OFFSET(dva
);
3484 uint64_t size
= DVA_GET_ASIZE(dva
);
3486 ASSERT(DVA_IS_VALID(dva
));
3487 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3489 if (txg
> spa_freeze_txg(spa
))
3492 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
3493 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
3494 cmn_err(CE_WARN
, "metaslab_free_dva(): bad DVA %llu:%llu",
3495 (u_longlong_t
)vdev
, (u_longlong_t
)offset
);
3500 ASSERT(!vd
->vdev_removing
);
3501 ASSERT(vdev_is_concrete(vd
));
3502 ASSERT0(vd
->vdev_indirect_config
.vic_mapping_object
);
3503 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
3505 if (DVA_GET_GANG(dva
))
3506 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3508 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3510 mutex_enter(&msp
->ms_lock
);
3511 range_tree_remove(msp
->ms_allocating
[txg
& TXG_MASK
],
3514 VERIFY(!msp
->ms_condensing
);
3515 VERIFY3U(offset
, >=, msp
->ms_start
);
3516 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
3517 VERIFY3U(range_tree_space(msp
->ms_allocatable
) + size
, <=,
3519 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3520 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3521 range_tree_add(msp
->ms_allocatable
, offset
, size
);
3522 mutex_exit(&msp
->ms_lock
);
3526 * Free the block represented by the given DVA.
3529 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, boolean_t checkpoint
)
3531 uint64_t vdev
= DVA_GET_VDEV(dva
);
3532 uint64_t offset
= DVA_GET_OFFSET(dva
);
3533 uint64_t size
= DVA_GET_ASIZE(dva
);
3534 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
3536 ASSERT(DVA_IS_VALID(dva
));
3537 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3539 if (DVA_GET_GANG(dva
)) {
3540 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3543 metaslab_free_impl(vd
, offset
, size
, checkpoint
);
3547 * Reserve some allocation slots. The reservation system must be called
3548 * before we call into the allocator. If there aren't any available slots
3549 * then the I/O will be throttled until an I/O completes and its slots are
3550 * freed up. The function returns true if it was successful in placing
3554 metaslab_class_throttle_reserve(metaslab_class_t
*mc
, int slots
, zio_t
*zio
,
3557 uint64_t available_slots
= 0;
3558 boolean_t slot_reserved
= B_FALSE
;
3560 ASSERT(mc
->mc_alloc_throttle_enabled
);
3561 mutex_enter(&mc
->mc_lock
);
3563 uint64_t reserved_slots
= refcount_count(&mc
->mc_alloc_slots
);
3564 if (reserved_slots
< mc
->mc_alloc_max_slots
)
3565 available_slots
= mc
->mc_alloc_max_slots
- reserved_slots
;
3567 if (slots
<= available_slots
|| GANG_ALLOCATION(flags
)) {
3569 * We reserve the slots individually so that we can unreserve
3570 * them individually when an I/O completes.
3572 for (int d
= 0; d
< slots
; d
++) {
3573 reserved_slots
= refcount_add(&mc
->mc_alloc_slots
, zio
);
3575 zio
->io_flags
|= ZIO_FLAG_IO_ALLOCATING
;
3576 slot_reserved
= B_TRUE
;
3579 mutex_exit(&mc
->mc_lock
);
3580 return (slot_reserved
);
3584 metaslab_class_throttle_unreserve(metaslab_class_t
*mc
, int slots
, zio_t
*zio
)
3586 ASSERT(mc
->mc_alloc_throttle_enabled
);
3587 mutex_enter(&mc
->mc_lock
);
3588 for (int d
= 0; d
< slots
; d
++) {
3589 (void) refcount_remove(&mc
->mc_alloc_slots
, zio
);
3591 mutex_exit(&mc
->mc_lock
);
3595 metaslab_claim_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
3599 spa_t
*spa
= vd
->vdev_spa
;
3602 if (offset
>> vd
->vdev_ms_shift
>= vd
->vdev_ms_count
)
3605 ASSERT3P(vd
->vdev_ms
, !=, NULL
);
3606 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3608 mutex_enter(&msp
->ms_lock
);
3610 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
3611 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
3614 !range_tree_contains(msp
->ms_allocatable
, offset
, size
))
3615 error
= SET_ERROR(ENOENT
);
3617 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
3618 mutex_exit(&msp
->ms_lock
);
3622 VERIFY(!msp
->ms_condensing
);
3623 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3624 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3625 VERIFY3U(range_tree_space(msp
->ms_allocatable
) - size
, <=,
3627 range_tree_remove(msp
->ms_allocatable
, offset
, size
);
3629 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(8) */
3630 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
3631 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
3632 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
],
3636 mutex_exit(&msp
->ms_lock
);
3641 typedef struct metaslab_claim_cb_arg_t
{
3644 } metaslab_claim_cb_arg_t
;
3648 metaslab_claim_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3649 uint64_t size
, void *arg
)
3651 metaslab_claim_cb_arg_t
*mcca_arg
= arg
;
3653 if (mcca_arg
->mcca_error
== 0) {
3654 mcca_arg
->mcca_error
= metaslab_claim_concrete(vd
, offset
,
3655 size
, mcca_arg
->mcca_txg
);
3660 metaslab_claim_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
, uint64_t txg
)
3662 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
3663 metaslab_claim_cb_arg_t arg
;
3666 * Only zdb(8) can claim on indirect vdevs. This is used
3667 * to detect leaks of mapped space (that are not accounted
3668 * for in the obsolete counts, spacemap, or bpobj).
3670 ASSERT(!spa_writeable(vd
->vdev_spa
));
3674 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
3675 metaslab_claim_impl_cb
, &arg
);
3677 if (arg
.mcca_error
== 0) {
3678 arg
.mcca_error
= metaslab_claim_concrete(vd
,
3681 return (arg
.mcca_error
);
3683 return (metaslab_claim_concrete(vd
, offset
, size
, txg
));
3688 * Intent log support: upon opening the pool after a crash, notify the SPA
3689 * of blocks that the intent log has allocated for immediate write, but
3690 * which are still considered free by the SPA because the last transaction
3691 * group didn't commit yet.
3694 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
3696 uint64_t vdev
= DVA_GET_VDEV(dva
);
3697 uint64_t offset
= DVA_GET_OFFSET(dva
);
3698 uint64_t size
= DVA_GET_ASIZE(dva
);
3701 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
) {
3702 return (SET_ERROR(ENXIO
));
3705 ASSERT(DVA_IS_VALID(dva
));
3707 if (DVA_GET_GANG(dva
))
3708 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3710 return (metaslab_claim_impl(vd
, offset
, size
, txg
));
3714 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
3715 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
,
3716 zio_alloc_list_t
*zal
, zio_t
*zio
)
3718 dva_t
*dva
= bp
->blk_dva
;
3719 dva_t
*hintdva
= hintbp
->blk_dva
;
3722 ASSERT(bp
->blk_birth
== 0);
3723 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
3725 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
3727 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
3728 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3729 return (SET_ERROR(ENOSPC
));
3732 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
3733 ASSERT(BP_GET_NDVAS(bp
) == 0);
3734 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
3735 ASSERT3P(zal
, !=, NULL
);
3737 for (int d
= 0; d
< ndvas
; d
++) {
3738 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
3741 for (d
--; d
>= 0; d
--) {
3742 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
3743 metaslab_group_alloc_decrement(spa
,
3744 DVA_GET_VDEV(&dva
[d
]), zio
, flags
);
3745 bzero(&dva
[d
], sizeof (dva_t
));
3747 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3751 * Update the metaslab group's queue depth
3752 * based on the newly allocated dva.
3754 metaslab_group_alloc_increment(spa
,
3755 DVA_GET_VDEV(&dva
[d
]), zio
, flags
);
3760 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
3762 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3764 BP_SET_BIRTH(bp
, txg
, txg
);
3770 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
3772 const dva_t
*dva
= bp
->blk_dva
;
3773 int ndvas
= BP_GET_NDVAS(bp
);
3775 ASSERT(!BP_IS_HOLE(bp
));
3776 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
3779 * If we have a checkpoint for the pool we need to make sure that
3780 * the blocks that we free that are part of the checkpoint won't be
3781 * reused until the checkpoint is discarded or we revert to it.
3783 * The checkpoint flag is passed down the metaslab_free code path
3784 * and is set whenever we want to add a block to the checkpoint's
3785 * accounting. That is, we "checkpoint" blocks that existed at the
3786 * time the checkpoint was created and are therefore referenced by
3787 * the checkpointed uberblock.
3789 * Note that, we don't checkpoint any blocks if the current
3790 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
3791 * normally as they will be referenced by the checkpointed uberblock.
3793 boolean_t checkpoint
= B_FALSE
;
3794 if (bp
->blk_birth
<= spa
->spa_checkpoint_txg
&&
3795 spa_syncing_txg(spa
) > spa
->spa_checkpoint_txg
) {
3797 * At this point, if the block is part of the checkpoint
3798 * there is no way it was created in the current txg.
3801 ASSERT3U(spa_syncing_txg(spa
), ==, txg
);
3802 checkpoint
= B_TRUE
;
3805 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
3807 for (int d
= 0; d
< ndvas
; d
++) {
3809 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
3811 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
3812 metaslab_free_dva(spa
, &dva
[d
], checkpoint
);
3816 spa_config_exit(spa
, SCL_FREE
, FTAG
);
3820 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
3822 const dva_t
*dva
= bp
->blk_dva
;
3823 int ndvas
= BP_GET_NDVAS(bp
);
3826 ASSERT(!BP_IS_HOLE(bp
));
3830 * First do a dry run to make sure all DVAs are claimable,
3831 * so we don't have to unwind from partial failures below.
3833 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
3837 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
3839 for (int d
= 0; d
< ndvas
; d
++)
3840 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
3843 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3845 ASSERT(error
== 0 || txg
== 0);
3852 metaslab_check_free_impl_cb(uint64_t inner
, vdev_t
*vd
, uint64_t offset
,
3853 uint64_t size
, void *arg
)
3855 if (vd
->vdev_ops
== &vdev_indirect_ops
)
3858 metaslab_check_free_impl(vd
, offset
, size
);
3862 metaslab_check_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
3865 spa_t
*spa
= vd
->vdev_spa
;
3867 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
3870 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
3871 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
3872 metaslab_check_free_impl_cb
, NULL
);
3876 ASSERT(vdev_is_concrete(vd
));
3877 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
3878 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3880 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3882 mutex_enter(&msp
->ms_lock
);
3884 range_tree_verify(msp
->ms_allocatable
, offset
, size
);
3886 range_tree_verify(msp
->ms_freeing
, offset
, size
);
3887 range_tree_verify(msp
->ms_checkpointing
, offset
, size
);
3888 range_tree_verify(msp
->ms_freed
, offset
, size
);
3889 for (int j
= 0; j
< TXG_DEFER_SIZE
; j
++)
3890 range_tree_verify(msp
->ms_defer
[j
], offset
, size
);
3891 mutex_exit(&msp
->ms_lock
);
3895 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
3897 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
3900 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
3901 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3902 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
3903 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
3904 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
3905 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
3907 if (DVA_GET_GANG(&bp
->blk_dva
[i
]))
3908 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3910 ASSERT3P(vd
, !=, NULL
);
3912 metaslab_check_free_impl(vd
, offset
, size
);
3914 spa_config_exit(spa
, SCL_VDEV
, FTAG
);