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, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
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 WITH_DF_BLOCK_ALLOCATOR
42 #define GANG_ALLOCATION(flags) \
43 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
46 * Metaslab granularity, in bytes. This is roughly similar to what would be
47 * referred to as the "stripe size" in traditional RAID arrays. In normal
48 * operation, we will try to write this amount of data to a top-level vdev
49 * before moving on to the next one.
51 unsigned long metaslab_aliquot
= 512 << 10;
54 * For testing, make some blocks above a certain size be gang blocks.
56 unsigned long metaslab_force_ganging
= SPA_MAXBLOCKSIZE
+ 1;
59 * Since we can touch multiple metaslabs (and their respective space maps)
60 * with each transaction group, we benefit from having a smaller space map
61 * block size since it allows us to issue more I/O operations scattered
64 int zfs_metaslab_sm_blksz
= (1 << 12);
67 * The in-core space map representation is more compact than its on-disk form.
68 * The zfs_condense_pct determines how much more compact the in-core
69 * space map representation must be before we compact it on-disk.
70 * Values should be greater than or equal to 100.
72 int zfs_condense_pct
= 200;
75 * Condensing a metaslab is not guaranteed to actually reduce the amount of
76 * space used on disk. In particular, a space map uses data in increments of
77 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
78 * same number of blocks after condensing. Since the goal of condensing is to
79 * reduce the number of IOPs required to read the space map, we only want to
80 * condense when we can be sure we will reduce the number of blocks used by the
81 * space map. Unfortunately, we cannot precisely compute whether or not this is
82 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
83 * we apply the following heuristic: do not condense a spacemap unless the
84 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
87 int zfs_metaslab_condense_block_threshold
= 4;
90 * The zfs_mg_noalloc_threshold defines which metaslab groups should
91 * be eligible for allocation. The value is defined as a percentage of
92 * free space. Metaslab groups that have more free space than
93 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
94 * a metaslab group's free space is less than or equal to the
95 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
96 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
97 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
98 * groups are allowed to accept allocations. Gang blocks are always
99 * eligible to allocate on any metaslab group. The default value of 0 means
100 * no metaslab group will be excluded based on this criterion.
102 int zfs_mg_noalloc_threshold
= 0;
105 * Metaslab groups are considered eligible for allocations if their
106 * fragmenation metric (measured as a percentage) is less than or equal to
107 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
108 * then it will be skipped unless all metaslab groups within the metaslab
109 * class have also crossed this threshold.
111 int zfs_mg_fragmentation_threshold
= 85;
114 * Allow metaslabs to keep their active state as long as their fragmentation
115 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
116 * active metaslab that exceeds this threshold will no longer keep its active
117 * status allowing better metaslabs to be selected.
119 int zfs_metaslab_fragmentation_threshold
= 70;
122 * When set will load all metaslabs when pool is first opened.
124 int metaslab_debug_load
= 0;
127 * When set will prevent metaslabs from being unloaded.
129 int metaslab_debug_unload
= 0;
132 * Minimum size which forces the dynamic allocator to change
133 * it's allocation strategy. Once the space map cannot satisfy
134 * an allocation of this size then it switches to using more
135 * aggressive strategy (i.e search by size rather than offset).
137 uint64_t metaslab_df_alloc_threshold
= SPA_OLD_MAXBLOCKSIZE
;
140 * The minimum free space, in percent, which must be available
141 * in a space map to continue allocations in a first-fit fashion.
142 * Once the space map's free space drops below this level we dynamically
143 * switch to using best-fit allocations.
145 int metaslab_df_free_pct
= 4;
148 * Percentage of all cpus that can be used by the metaslab taskq.
150 int metaslab_load_pct
= 50;
153 * Determines how many txgs a metaslab may remain loaded without having any
154 * allocations from it. As long as a metaslab continues to be used we will
157 int metaslab_unload_delay
= TXG_SIZE
* 2;
160 * Max number of metaslabs per group to preload.
162 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
165 * Enable/disable preloading of metaslab.
167 int metaslab_preload_enabled
= B_TRUE
;
170 * Enable/disable fragmentation weighting on metaslabs.
172 int metaslab_fragmentation_factor_enabled
= B_TRUE
;
175 * Enable/disable lba weighting (i.e. outer tracks are given preference).
177 int metaslab_lba_weighting_enabled
= B_TRUE
;
180 * Enable/disable metaslab group biasing.
182 int metaslab_bias_enabled
= B_TRUE
;
186 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
188 boolean_t zfs_remap_blkptr_enable
= B_TRUE
;
191 * Enable/disable segment-based metaslab selection.
193 int zfs_metaslab_segment_weight_enabled
= B_TRUE
;
196 * When using segment-based metaslab selection, we will continue
197 * allocating from the active metaslab until we have exhausted
198 * zfs_metaslab_switch_threshold of its buckets.
200 int zfs_metaslab_switch_threshold
= 2;
203 * Internal switch to enable/disable the metaslab allocation tracing
206 #ifdef _METASLAB_TRACING
207 boolean_t metaslab_trace_enabled
= B_TRUE
;
211 * Maximum entries that the metaslab allocation tracing facility will keep
212 * in a given list when running in non-debug mode. We limit the number
213 * of entries in non-debug mode to prevent us from using up too much memory.
214 * The limit should be sufficiently large that we don't expect any allocation
215 * to every exceed this value. In debug mode, the system will panic if this
216 * limit is ever reached allowing for further investigation.
218 #ifdef _METASLAB_TRACING
219 uint64_t metaslab_trace_max_entries
= 5000;
222 static uint64_t metaslab_weight(metaslab_t
*);
223 static void metaslab_set_fragmentation(metaslab_t
*);
224 static void metaslab_free_impl(vdev_t
*, uint64_t, uint64_t, boolean_t
);
225 static void metaslab_check_free_impl(vdev_t
*, uint64_t, uint64_t);
227 static void metaslab_passivate(metaslab_t
*msp
, uint64_t weight
);
228 static uint64_t metaslab_weight_from_range_tree(metaslab_t
*msp
);
229 #ifdef _METASLAB_TRACING
230 kmem_cache_t
*metaslab_alloc_trace_cache
;
234 * ==========================================================================
236 * ==========================================================================
239 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
241 metaslab_class_t
*mc
;
243 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
248 mutex_init(&mc
->mc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
249 mc
->mc_alloc_slots
= kmem_zalloc(spa
->spa_alloc_count
*
250 sizeof (zfs_refcount_t
), KM_SLEEP
);
251 mc
->mc_alloc_max_slots
= kmem_zalloc(spa
->spa_alloc_count
*
252 sizeof (uint64_t), KM_SLEEP
);
253 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++)
254 refcount_create_tracked(&mc
->mc_alloc_slots
[i
]);
260 metaslab_class_destroy(metaslab_class_t
*mc
)
262 ASSERT(mc
->mc_rotor
== NULL
);
263 ASSERT(mc
->mc_alloc
== 0);
264 ASSERT(mc
->mc_deferred
== 0);
265 ASSERT(mc
->mc_space
== 0);
266 ASSERT(mc
->mc_dspace
== 0);
268 for (int i
= 0; i
< mc
->mc_spa
->spa_alloc_count
; i
++)
269 refcount_destroy(&mc
->mc_alloc_slots
[i
]);
270 kmem_free(mc
->mc_alloc_slots
, mc
->mc_spa
->spa_alloc_count
*
271 sizeof (zfs_refcount_t
));
272 kmem_free(mc
->mc_alloc_max_slots
, mc
->mc_spa
->spa_alloc_count
*
274 mutex_destroy(&mc
->mc_lock
);
275 kmem_free(mc
, sizeof (metaslab_class_t
));
279 metaslab_class_validate(metaslab_class_t
*mc
)
281 metaslab_group_t
*mg
;
285 * Must hold one of the spa_config locks.
287 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
288 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
290 if ((mg
= mc
->mc_rotor
) == NULL
)
295 ASSERT(vd
->vdev_mg
!= NULL
);
296 ASSERT3P(vd
->vdev_top
, ==, vd
);
297 ASSERT3P(mg
->mg_class
, ==, mc
);
298 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
299 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
305 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
306 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
308 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
309 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
310 atomic_add_64(&mc
->mc_space
, space_delta
);
311 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
315 metaslab_class_get_alloc(metaslab_class_t
*mc
)
317 return (mc
->mc_alloc
);
321 metaslab_class_get_deferred(metaslab_class_t
*mc
)
323 return (mc
->mc_deferred
);
327 metaslab_class_get_space(metaslab_class_t
*mc
)
329 return (mc
->mc_space
);
333 metaslab_class_get_dspace(metaslab_class_t
*mc
)
335 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
339 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
341 spa_t
*spa
= mc
->mc_spa
;
342 vdev_t
*rvd
= spa
->spa_root_vdev
;
346 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
349 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
352 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
353 vdev_t
*tvd
= rvd
->vdev_child
[c
];
354 metaslab_group_t
*mg
= tvd
->vdev_mg
;
357 * Skip any holes, uninitialized top-levels, or
358 * vdevs that are not in this metalab class.
360 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
361 mg
->mg_class
!= mc
) {
365 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
366 mc_hist
[i
] += mg
->mg_histogram
[i
];
369 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
370 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
372 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
376 * Calculate the metaslab class's fragmentation metric. The metric
377 * is weighted based on the space contribution of each metaslab group.
378 * The return value will be a number between 0 and 100 (inclusive), or
379 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
380 * zfs_frag_table for more information about the metric.
383 metaslab_class_fragmentation(metaslab_class_t
*mc
)
385 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
386 uint64_t fragmentation
= 0;
388 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
390 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
391 vdev_t
*tvd
= rvd
->vdev_child
[c
];
392 metaslab_group_t
*mg
= tvd
->vdev_mg
;
395 * Skip any holes, uninitialized top-levels,
396 * or vdevs that are not in this metalab class.
398 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
399 mg
->mg_class
!= mc
) {
404 * If a metaslab group does not contain a fragmentation
405 * metric then just bail out.
407 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
408 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
409 return (ZFS_FRAG_INVALID
);
413 * Determine how much this metaslab_group is contributing
414 * to the overall pool fragmentation metric.
416 fragmentation
+= mg
->mg_fragmentation
*
417 metaslab_group_get_space(mg
);
419 fragmentation
/= metaslab_class_get_space(mc
);
421 ASSERT3U(fragmentation
, <=, 100);
422 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
423 return (fragmentation
);
427 * Calculate the amount of expandable space that is available in
428 * this metaslab class. If a device is expanded then its expandable
429 * space will be the amount of allocatable space that is currently not
430 * part of this metaslab class.
433 metaslab_class_expandable_space(metaslab_class_t
*mc
)
435 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
438 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
439 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
440 vdev_t
*tvd
= rvd
->vdev_child
[c
];
441 metaslab_group_t
*mg
= tvd
->vdev_mg
;
443 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
444 mg
->mg_class
!= mc
) {
449 * Calculate if we have enough space to add additional
450 * metaslabs. We report the expandable space in terms
451 * of the metaslab size since that's the unit of expansion.
453 space
+= P2ALIGN(tvd
->vdev_max_asize
- tvd
->vdev_asize
,
454 1ULL << tvd
->vdev_ms_shift
);
456 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
461 metaslab_compare(const void *x1
, const void *x2
)
463 const metaslab_t
*m1
= (const metaslab_t
*)x1
;
464 const metaslab_t
*m2
= (const metaslab_t
*)x2
;
468 if (m1
->ms_allocator
!= -1 && m1
->ms_primary
)
470 else if (m1
->ms_allocator
!= -1 && !m1
->ms_primary
)
472 if (m2
->ms_allocator
!= -1 && m2
->ms_primary
)
474 else if (m2
->ms_allocator
!= -1 && !m2
->ms_primary
)
478 * Sort inactive metaslabs first, then primaries, then secondaries. When
479 * selecting a metaslab to allocate from, an allocator first tries its
480 * primary, then secondary active metaslab. If it doesn't have active
481 * metaslabs, or can't allocate from them, it searches for an inactive
482 * metaslab to activate. If it can't find a suitable one, it will steal
483 * a primary or secondary metaslab from another allocator.
490 int cmp
= AVL_CMP(m2
->ms_weight
, m1
->ms_weight
);
494 IMPLY(AVL_CMP(m1
->ms_start
, m2
->ms_start
) == 0, m1
== m2
);
496 return (AVL_CMP(m1
->ms_start
, m2
->ms_start
));
500 * Verify that the space accounting on disk matches the in-core range_trees.
503 metaslab_verify_space(metaslab_t
*msp
, uint64_t txg
)
505 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
506 uint64_t allocated
= 0;
507 uint64_t sm_free_space
, msp_free_space
;
509 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
511 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
515 * We can only verify the metaslab space when we're called
516 * from syncing context with a loaded metaslab that has an allocated
517 * space map. Calling this in non-syncing context does not
518 * provide a consistent view of the metaslab since we're performing
519 * allocations in the future.
521 if (txg
!= spa_syncing_txg(spa
) || msp
->ms_sm
== NULL
||
525 sm_free_space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
) -
526 space_map_alloc_delta(msp
->ms_sm
);
529 * Account for future allocations since we would have already
530 * deducted that space from the ms_freetree.
532 for (int t
= 0; t
< TXG_CONCURRENT_STATES
; t
++) {
534 range_tree_space(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]);
537 msp_free_space
= range_tree_space(msp
->ms_allocatable
) + allocated
+
538 msp
->ms_deferspace
+ range_tree_space(msp
->ms_freed
);
540 VERIFY3U(sm_free_space
, ==, msp_free_space
);
544 * ==========================================================================
546 * ==========================================================================
549 * Update the allocatable flag and the metaslab group's capacity.
550 * The allocatable flag is set to true if the capacity is below
551 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
552 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
553 * transitions from allocatable to non-allocatable or vice versa then the
554 * metaslab group's class is updated to reflect the transition.
557 metaslab_group_alloc_update(metaslab_group_t
*mg
)
559 vdev_t
*vd
= mg
->mg_vd
;
560 metaslab_class_t
*mc
= mg
->mg_class
;
561 vdev_stat_t
*vs
= &vd
->vdev_stat
;
562 boolean_t was_allocatable
;
563 boolean_t was_initialized
;
565 ASSERT(vd
== vd
->vdev_top
);
566 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_READER
), ==,
569 mutex_enter(&mg
->mg_lock
);
570 was_allocatable
= mg
->mg_allocatable
;
571 was_initialized
= mg
->mg_initialized
;
573 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
576 mutex_enter(&mc
->mc_lock
);
579 * If the metaslab group was just added then it won't
580 * have any space until we finish syncing out this txg.
581 * At that point we will consider it initialized and available
582 * for allocations. We also don't consider non-activated
583 * metaslab groups (e.g. vdevs that are in the middle of being removed)
584 * to be initialized, because they can't be used for allocation.
586 mg
->mg_initialized
= metaslab_group_initialized(mg
);
587 if (!was_initialized
&& mg
->mg_initialized
) {
589 } else if (was_initialized
&& !mg
->mg_initialized
) {
590 ASSERT3U(mc
->mc_groups
, >, 0);
593 if (mg
->mg_initialized
)
594 mg
->mg_no_free_space
= B_FALSE
;
597 * A metaslab group is considered allocatable if it has plenty
598 * of free space or is not heavily fragmented. We only take
599 * fragmentation into account if the metaslab group has a valid
600 * fragmentation metric (i.e. a value between 0 and 100).
602 mg
->mg_allocatable
= (mg
->mg_activation_count
> 0 &&
603 mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
604 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
605 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
608 * The mc_alloc_groups maintains a count of the number of
609 * groups in this metaslab class that are still above the
610 * zfs_mg_noalloc_threshold. This is used by the allocating
611 * threads to determine if they should avoid allocations to
612 * a given group. The allocator will avoid allocations to a group
613 * if that group has reached or is below the zfs_mg_noalloc_threshold
614 * and there are still other groups that are above the threshold.
615 * When a group transitions from allocatable to non-allocatable or
616 * vice versa we update the metaslab class to reflect that change.
617 * When the mc_alloc_groups value drops to 0 that means that all
618 * groups have reached the zfs_mg_noalloc_threshold making all groups
619 * eligible for allocations. This effectively means that all devices
620 * are balanced again.
622 if (was_allocatable
&& !mg
->mg_allocatable
)
623 mc
->mc_alloc_groups
--;
624 else if (!was_allocatable
&& mg
->mg_allocatable
)
625 mc
->mc_alloc_groups
++;
626 mutex_exit(&mc
->mc_lock
);
628 mutex_exit(&mg
->mg_lock
);
632 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
, int allocators
)
634 metaslab_group_t
*mg
;
636 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
637 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
638 mg
->mg_primaries
= kmem_zalloc(allocators
* sizeof (metaslab_t
*),
640 mg
->mg_secondaries
= kmem_zalloc(allocators
* sizeof (metaslab_t
*),
642 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
643 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
646 mg
->mg_activation_count
= 0;
647 mg
->mg_initialized
= B_FALSE
;
648 mg
->mg_no_free_space
= B_TRUE
;
649 mg
->mg_allocators
= allocators
;
651 mg
->mg_alloc_queue_depth
= kmem_zalloc(allocators
*
652 sizeof (zfs_refcount_t
), KM_SLEEP
);
653 mg
->mg_cur_max_alloc_queue_depth
= kmem_zalloc(allocators
*
654 sizeof (uint64_t), KM_SLEEP
);
655 for (int i
= 0; i
< allocators
; i
++) {
656 refcount_create_tracked(&mg
->mg_alloc_queue_depth
[i
]);
657 mg
->mg_cur_max_alloc_queue_depth
[i
] = 0;
660 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
661 maxclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
| TASKQ_DYNAMIC
);
667 metaslab_group_destroy(metaslab_group_t
*mg
)
669 ASSERT(mg
->mg_prev
== NULL
);
670 ASSERT(mg
->mg_next
== NULL
);
672 * We may have gone below zero with the activation count
673 * either because we never activated in the first place or
674 * because we're done, and possibly removing the vdev.
676 ASSERT(mg
->mg_activation_count
<= 0);
678 taskq_destroy(mg
->mg_taskq
);
679 avl_destroy(&mg
->mg_metaslab_tree
);
680 kmem_free(mg
->mg_primaries
, mg
->mg_allocators
* sizeof (metaslab_t
*));
681 kmem_free(mg
->mg_secondaries
, mg
->mg_allocators
*
682 sizeof (metaslab_t
*));
683 mutex_destroy(&mg
->mg_lock
);
685 for (int i
= 0; i
< mg
->mg_allocators
; i
++) {
686 refcount_destroy(&mg
->mg_alloc_queue_depth
[i
]);
687 mg
->mg_cur_max_alloc_queue_depth
[i
] = 0;
689 kmem_free(mg
->mg_alloc_queue_depth
, mg
->mg_allocators
*
690 sizeof (zfs_refcount_t
));
691 kmem_free(mg
->mg_cur_max_alloc_queue_depth
, mg
->mg_allocators
*
694 kmem_free(mg
, sizeof (metaslab_group_t
));
698 metaslab_group_activate(metaslab_group_t
*mg
)
700 metaslab_class_t
*mc
= mg
->mg_class
;
701 metaslab_group_t
*mgprev
, *mgnext
;
703 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
), !=, 0);
705 ASSERT(mc
->mc_rotor
!= mg
);
706 ASSERT(mg
->mg_prev
== NULL
);
707 ASSERT(mg
->mg_next
== NULL
);
708 ASSERT(mg
->mg_activation_count
<= 0);
710 if (++mg
->mg_activation_count
<= 0)
713 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
714 metaslab_group_alloc_update(mg
);
716 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
720 mgnext
= mgprev
->mg_next
;
721 mg
->mg_prev
= mgprev
;
722 mg
->mg_next
= mgnext
;
723 mgprev
->mg_next
= mg
;
724 mgnext
->mg_prev
= mg
;
730 * Passivate a metaslab group and remove it from the allocation rotor.
731 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
732 * a metaslab group. This function will momentarily drop spa_config_locks
733 * that are lower than the SCL_ALLOC lock (see comment below).
736 metaslab_group_passivate(metaslab_group_t
*mg
)
738 metaslab_class_t
*mc
= mg
->mg_class
;
739 spa_t
*spa
= mc
->mc_spa
;
740 metaslab_group_t
*mgprev
, *mgnext
;
741 int locks
= spa_config_held(spa
, SCL_ALL
, RW_WRITER
);
743 ASSERT3U(spa_config_held(spa
, SCL_ALLOC
| SCL_ZIO
, RW_WRITER
), ==,
744 (SCL_ALLOC
| SCL_ZIO
));
746 if (--mg
->mg_activation_count
!= 0) {
747 ASSERT(mc
->mc_rotor
!= mg
);
748 ASSERT(mg
->mg_prev
== NULL
);
749 ASSERT(mg
->mg_next
== NULL
);
750 ASSERT(mg
->mg_activation_count
< 0);
755 * The spa_config_lock is an array of rwlocks, ordered as
756 * follows (from highest to lowest):
757 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
758 * SCL_ZIO > SCL_FREE > SCL_VDEV
759 * (For more information about the spa_config_lock see spa_misc.c)
760 * The higher the lock, the broader its coverage. When we passivate
761 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
762 * config locks. However, the metaslab group's taskq might be trying
763 * to preload metaslabs so we must drop the SCL_ZIO lock and any
764 * lower locks to allow the I/O to complete. At a minimum,
765 * we continue to hold the SCL_ALLOC lock, which prevents any future
766 * allocations from taking place and any changes to the vdev tree.
768 spa_config_exit(spa
, locks
& ~(SCL_ZIO
- 1), spa
);
769 taskq_wait_outstanding(mg
->mg_taskq
, 0);
770 spa_config_enter(spa
, locks
& ~(SCL_ZIO
- 1), spa
, RW_WRITER
);
771 metaslab_group_alloc_update(mg
);
772 for (int i
= 0; i
< mg
->mg_allocators
; i
++) {
773 metaslab_t
*msp
= mg
->mg_primaries
[i
];
775 mutex_enter(&msp
->ms_lock
);
776 metaslab_passivate(msp
,
777 metaslab_weight_from_range_tree(msp
));
778 mutex_exit(&msp
->ms_lock
);
780 msp
= mg
->mg_secondaries
[i
];
782 mutex_enter(&msp
->ms_lock
);
783 metaslab_passivate(msp
,
784 metaslab_weight_from_range_tree(msp
));
785 mutex_exit(&msp
->ms_lock
);
789 mgprev
= mg
->mg_prev
;
790 mgnext
= mg
->mg_next
;
795 mc
->mc_rotor
= mgnext
;
796 mgprev
->mg_next
= mgnext
;
797 mgnext
->mg_prev
= mgprev
;
805 metaslab_group_initialized(metaslab_group_t
*mg
)
807 vdev_t
*vd
= mg
->mg_vd
;
808 vdev_stat_t
*vs
= &vd
->vdev_stat
;
810 return (vs
->vs_space
!= 0 && mg
->mg_activation_count
> 0);
814 metaslab_group_get_space(metaslab_group_t
*mg
)
816 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
820 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
823 vdev_t
*vd
= mg
->mg_vd
;
824 uint64_t ashift
= vd
->vdev_ashift
;
827 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
830 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
833 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
834 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
836 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
837 metaslab_t
*msp
= vd
->vdev_ms
[m
];
839 /* skip if not active or not a member */
840 if (msp
->ms_sm
== NULL
|| msp
->ms_group
!= mg
)
843 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
844 mg_hist
[i
+ ashift
] +=
845 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
848 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
849 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
851 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
855 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
857 metaslab_class_t
*mc
= mg
->mg_class
;
858 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
860 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
861 if (msp
->ms_sm
== NULL
)
864 mutex_enter(&mg
->mg_lock
);
865 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
866 mg
->mg_histogram
[i
+ ashift
] +=
867 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
868 mc
->mc_histogram
[i
+ ashift
] +=
869 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
871 mutex_exit(&mg
->mg_lock
);
875 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
877 metaslab_class_t
*mc
= mg
->mg_class
;
878 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
880 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
881 if (msp
->ms_sm
== NULL
)
884 mutex_enter(&mg
->mg_lock
);
885 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
886 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
887 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
888 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
889 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
891 mg
->mg_histogram
[i
+ ashift
] -=
892 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
893 mc
->mc_histogram
[i
+ ashift
] -=
894 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
896 mutex_exit(&mg
->mg_lock
);
900 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
902 ASSERT(msp
->ms_group
== NULL
);
903 mutex_enter(&mg
->mg_lock
);
906 avl_add(&mg
->mg_metaslab_tree
, msp
);
907 mutex_exit(&mg
->mg_lock
);
909 mutex_enter(&msp
->ms_lock
);
910 metaslab_group_histogram_add(mg
, msp
);
911 mutex_exit(&msp
->ms_lock
);
915 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
917 mutex_enter(&msp
->ms_lock
);
918 metaslab_group_histogram_remove(mg
, msp
);
919 mutex_exit(&msp
->ms_lock
);
921 mutex_enter(&mg
->mg_lock
);
922 ASSERT(msp
->ms_group
== mg
);
923 avl_remove(&mg
->mg_metaslab_tree
, msp
);
924 msp
->ms_group
= NULL
;
925 mutex_exit(&mg
->mg_lock
);
929 metaslab_group_sort_impl(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
931 ASSERT(MUTEX_HELD(&mg
->mg_lock
));
932 ASSERT(msp
->ms_group
== mg
);
933 avl_remove(&mg
->mg_metaslab_tree
, msp
);
934 msp
->ms_weight
= weight
;
935 avl_add(&mg
->mg_metaslab_tree
, msp
);
940 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
943 * Although in principle the weight can be any value, in
944 * practice we do not use values in the range [1, 511].
946 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
947 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
949 mutex_enter(&mg
->mg_lock
);
950 metaslab_group_sort_impl(mg
, msp
, weight
);
951 mutex_exit(&mg
->mg_lock
);
955 * Calculate the fragmentation for a given metaslab group. We can use
956 * a simple average here since all metaslabs within the group must have
957 * the same size. The return value will be a value between 0 and 100
958 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
959 * group have a fragmentation metric.
962 metaslab_group_fragmentation(metaslab_group_t
*mg
)
964 vdev_t
*vd
= mg
->mg_vd
;
965 uint64_t fragmentation
= 0;
966 uint64_t valid_ms
= 0;
968 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
969 metaslab_t
*msp
= vd
->vdev_ms
[m
];
971 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
973 if (msp
->ms_group
!= mg
)
977 fragmentation
+= msp
->ms_fragmentation
;
980 if (valid_ms
<= mg
->mg_vd
->vdev_ms_count
/ 2)
981 return (ZFS_FRAG_INVALID
);
983 fragmentation
/= valid_ms
;
984 ASSERT3U(fragmentation
, <=, 100);
985 return (fragmentation
);
989 * Determine if a given metaslab group should skip allocations. A metaslab
990 * group should avoid allocations if its free capacity is less than the
991 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
992 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
993 * that can still handle allocations. If the allocation throttle is enabled
994 * then we skip allocations to devices that have reached their maximum
995 * allocation queue depth unless the selected metaslab group is the only
996 * eligible group remaining.
999 metaslab_group_allocatable(metaslab_group_t
*mg
, metaslab_group_t
*rotor
,
1000 uint64_t psize
, int allocator
, int d
)
1002 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1003 metaslab_class_t
*mc
= mg
->mg_class
;
1006 * We can only consider skipping this metaslab group if it's
1007 * in the normal metaslab class and there are other metaslab
1008 * groups to select from. Otherwise, we always consider it eligible
1011 if ((mc
!= spa_normal_class(spa
) &&
1012 mc
!= spa_special_class(spa
) &&
1013 mc
!= spa_dedup_class(spa
)) ||
1018 * If the metaslab group's mg_allocatable flag is set (see comments
1019 * in metaslab_group_alloc_update() for more information) and
1020 * the allocation throttle is disabled then allow allocations to this
1021 * device. However, if the allocation throttle is enabled then
1022 * check if we have reached our allocation limit (mg_alloc_queue_depth)
1023 * to determine if we should allow allocations to this metaslab group.
1024 * If all metaslab groups are no longer considered allocatable
1025 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1026 * gang block size then we allow allocations on this metaslab group
1027 * regardless of the mg_allocatable or throttle settings.
1029 if (mg
->mg_allocatable
) {
1030 metaslab_group_t
*mgp
;
1032 uint64_t qmax
= mg
->mg_cur_max_alloc_queue_depth
[allocator
];
1034 if (!mc
->mc_alloc_throttle_enabled
)
1038 * If this metaslab group does not have any free space, then
1039 * there is no point in looking further.
1041 if (mg
->mg_no_free_space
)
1045 * Relax allocation throttling for ditto blocks. Due to
1046 * random imbalances in allocation it tends to push copies
1047 * to one vdev, that looks a bit better at the moment.
1049 qmax
= qmax
* (4 + d
) / 4;
1051 qdepth
= refcount_count(&mg
->mg_alloc_queue_depth
[allocator
]);
1054 * If this metaslab group is below its qmax or it's
1055 * the only allocatable metasable group, then attempt
1056 * to allocate from it.
1058 if (qdepth
< qmax
|| mc
->mc_alloc_groups
== 1)
1060 ASSERT3U(mc
->mc_alloc_groups
, >, 1);
1063 * Since this metaslab group is at or over its qmax, we
1064 * need to determine if there are metaslab groups after this
1065 * one that might be able to handle this allocation. This is
1066 * racy since we can't hold the locks for all metaslab
1067 * groups at the same time when we make this check.
1069 for (mgp
= mg
->mg_next
; mgp
!= rotor
; mgp
= mgp
->mg_next
) {
1070 qmax
= mgp
->mg_cur_max_alloc_queue_depth
[allocator
];
1071 qmax
= qmax
* (4 + d
) / 4;
1072 qdepth
= refcount_count(
1073 &mgp
->mg_alloc_queue_depth
[allocator
]);
1076 * If there is another metaslab group that
1077 * might be able to handle the allocation, then
1078 * we return false so that we skip this group.
1080 if (qdepth
< qmax
&& !mgp
->mg_no_free_space
)
1085 * We didn't find another group to handle the allocation
1086 * so we can't skip this metaslab group even though
1087 * we are at or over our qmax.
1091 } else if (mc
->mc_alloc_groups
== 0 || psize
== SPA_MINBLOCKSIZE
) {
1098 * ==========================================================================
1099 * Range tree callbacks
1100 * ==========================================================================
1104 * Comparison function for the private size-ordered tree. Tree is sorted
1105 * by size, larger sizes at the end of the tree.
1108 metaslab_rangesize_compare(const void *x1
, const void *x2
)
1110 const range_seg_t
*r1
= x1
;
1111 const range_seg_t
*r2
= x2
;
1112 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
1113 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
1115 int cmp
= AVL_CMP(rs_size1
, rs_size2
);
1119 return (AVL_CMP(r1
->rs_start
, r2
->rs_start
));
1123 * ==========================================================================
1124 * Common allocator routines
1125 * ==========================================================================
1129 * Return the maximum contiguous segment within the metaslab.
1132 metaslab_block_maxsize(metaslab_t
*msp
)
1134 avl_tree_t
*t
= &msp
->ms_allocatable_by_size
;
1137 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
1140 return (rs
->rs_end
- rs
->rs_start
);
1143 static range_seg_t
*
1144 metaslab_block_find(avl_tree_t
*t
, uint64_t start
, uint64_t size
)
1146 range_seg_t
*rs
, rsearch
;
1149 rsearch
.rs_start
= start
;
1150 rsearch
.rs_end
= start
+ size
;
1152 rs
= avl_find(t
, &rsearch
, &where
);
1154 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1160 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
1161 defined(WITH_DF_BLOCK_ALLOCATOR) || \
1162 defined(WITH_CF_BLOCK_ALLOCATOR)
1164 * This is a helper function that can be used by the allocator to find
1165 * a suitable block to allocate. This will search the specified AVL
1166 * tree looking for a block that matches the specified criteria.
1169 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
1172 range_seg_t
*rs
= metaslab_block_find(t
, *cursor
, size
);
1174 while (rs
!= NULL
) {
1175 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
1177 if (offset
+ size
<= rs
->rs_end
) {
1178 *cursor
= offset
+ size
;
1181 rs
= AVL_NEXT(t
, rs
);
1185 * If we know we've searched the whole map (*cursor == 0), give up.
1186 * Otherwise, reset the cursor to the beginning and try again.
1192 return (metaslab_block_picker(t
, cursor
, size
, align
));
1194 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
1196 #if defined(WITH_FF_BLOCK_ALLOCATOR)
1198 * ==========================================================================
1199 * The first-fit block allocator
1200 * ==========================================================================
1203 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
1206 * Find the largest power of 2 block size that evenly divides the
1207 * requested size. This is used to try to allocate blocks with similar
1208 * alignment from the same area of the metaslab (i.e. same cursor
1209 * bucket) but it does not guarantee that other allocations sizes
1210 * may exist in the same region.
1212 uint64_t align
= size
& -size
;
1213 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1214 avl_tree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1216 return (metaslab_block_picker(t
, cursor
, size
, align
));
1219 static metaslab_ops_t metaslab_ff_ops
= {
1223 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
1224 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1226 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1228 * ==========================================================================
1229 * Dynamic block allocator -
1230 * Uses the first fit allocation scheme until space get low and then
1231 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1232 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1233 * ==========================================================================
1236 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1239 * Find the largest power of 2 block size that evenly divides the
1240 * requested size. This is used to try to allocate blocks with similar
1241 * alignment from the same area of the metaslab (i.e. same cursor
1242 * bucket) but it does not guarantee that other allocations sizes
1243 * may exist in the same region.
1245 uint64_t align
= size
& -size
;
1246 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1247 range_tree_t
*rt
= msp
->ms_allocatable
;
1248 avl_tree_t
*t
= &rt
->rt_root
;
1249 uint64_t max_size
= metaslab_block_maxsize(msp
);
1250 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1252 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1253 ASSERT3U(avl_numnodes(t
), ==,
1254 avl_numnodes(&msp
->ms_allocatable_by_size
));
1256 if (max_size
< size
)
1260 * If we're running low on space switch to using the size
1261 * sorted AVL tree (best-fit).
1263 if (max_size
< metaslab_df_alloc_threshold
||
1264 free_pct
< metaslab_df_free_pct
) {
1265 t
= &msp
->ms_allocatable_by_size
;
1269 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1272 static metaslab_ops_t metaslab_df_ops
= {
1276 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1277 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1279 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1281 * ==========================================================================
1282 * Cursor fit block allocator -
1283 * Select the largest region in the metaslab, set the cursor to the beginning
1284 * of the range and the cursor_end to the end of the range. As allocations
1285 * are made advance the cursor. Continue allocating from the cursor until
1286 * the range is exhausted and then find a new range.
1287 * ==========================================================================
1290 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1292 range_tree_t
*rt
= msp
->ms_allocatable
;
1293 avl_tree_t
*t
= &msp
->ms_allocatable_by_size
;
1294 uint64_t *cursor
= &msp
->ms_lbas
[0];
1295 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1296 uint64_t offset
= 0;
1298 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1299 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1301 ASSERT3U(*cursor_end
, >=, *cursor
);
1303 if ((*cursor
+ size
) > *cursor_end
) {
1306 rs
= avl_last(&msp
->ms_allocatable_by_size
);
1307 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1310 *cursor
= rs
->rs_start
;
1311 *cursor_end
= rs
->rs_end
;
1320 static metaslab_ops_t metaslab_cf_ops
= {
1324 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1325 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1327 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1329 * ==========================================================================
1330 * New dynamic fit allocator -
1331 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1332 * contiguous blocks. If no region is found then just use the largest segment
1334 * ==========================================================================
1338 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1339 * to request from the allocator.
1341 uint64_t metaslab_ndf_clump_shift
= 4;
1344 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1346 avl_tree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1348 range_seg_t
*rs
, rsearch
;
1349 uint64_t hbit
= highbit64(size
);
1350 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1351 uint64_t max_size
= metaslab_block_maxsize(msp
);
1353 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1354 ASSERT3U(avl_numnodes(t
), ==,
1355 avl_numnodes(&msp
->ms_allocatable_by_size
));
1357 if (max_size
< size
)
1360 rsearch
.rs_start
= *cursor
;
1361 rsearch
.rs_end
= *cursor
+ size
;
1363 rs
= avl_find(t
, &rsearch
, &where
);
1364 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1365 t
= &msp
->ms_allocatable_by_size
;
1367 rsearch
.rs_start
= 0;
1368 rsearch
.rs_end
= MIN(max_size
,
1369 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1370 rs
= avl_find(t
, &rsearch
, &where
);
1372 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1376 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1377 *cursor
= rs
->rs_start
+ size
;
1378 return (rs
->rs_start
);
1383 static metaslab_ops_t metaslab_ndf_ops
= {
1387 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1388 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1392 * ==========================================================================
1394 * ==========================================================================
1398 * Wait for any in-progress metaslab loads to complete.
1401 metaslab_load_wait(metaslab_t
*msp
)
1403 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1405 while (msp
->ms_loading
) {
1406 ASSERT(!msp
->ms_loaded
);
1407 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1412 metaslab_load(metaslab_t
*msp
)
1415 boolean_t success
= B_FALSE
;
1417 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1418 ASSERT(!msp
->ms_loaded
);
1419 ASSERT(!msp
->ms_loading
);
1421 msp
->ms_loading
= B_TRUE
;
1423 * Nobody else can manipulate a loading metaslab, so it's now safe
1424 * to drop the lock. This way we don't have to hold the lock while
1425 * reading the spacemap from disk.
1427 mutex_exit(&msp
->ms_lock
);
1430 * If the space map has not been allocated yet, then treat
1431 * all the space in the metaslab as free and add it to ms_allocatable.
1433 if (msp
->ms_sm
!= NULL
) {
1434 error
= space_map_load(msp
->ms_sm
, msp
->ms_allocatable
,
1437 range_tree_add(msp
->ms_allocatable
,
1438 msp
->ms_start
, msp
->ms_size
);
1441 success
= (error
== 0);
1443 mutex_enter(&msp
->ms_lock
);
1444 msp
->ms_loading
= B_FALSE
;
1447 ASSERT3P(msp
->ms_group
, !=, NULL
);
1448 msp
->ms_loaded
= B_TRUE
;
1451 * If the metaslab already has a spacemap, then we need to
1452 * remove all segments from the defer tree; otherwise, the
1453 * metaslab is completely empty and we can skip this.
1455 if (msp
->ms_sm
!= NULL
) {
1456 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1457 range_tree_walk(msp
->ms_defer
[t
],
1458 range_tree_remove
, msp
->ms_allocatable
);
1461 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1463 cv_broadcast(&msp
->ms_load_cv
);
1468 metaslab_unload(metaslab_t
*msp
)
1470 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1471 range_tree_vacate(msp
->ms_allocatable
, NULL
, NULL
);
1472 msp
->ms_loaded
= B_FALSE
;
1473 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1474 msp
->ms_max_size
= 0;
1478 metaslab_space_update(vdev_t
*vd
, metaslab_class_t
*mc
, int64_t alloc_delta
,
1479 int64_t defer_delta
, int64_t space_delta
)
1481 vdev_space_update(vd
, alloc_delta
, defer_delta
, space_delta
);
1483 ASSERT3P(vd
->vdev_spa
->spa_root_vdev
, ==, vd
->vdev_parent
);
1484 ASSERT(vd
->vdev_ms_count
!= 0);
1486 metaslab_class_space_update(mc
, alloc_delta
, defer_delta
, space_delta
,
1487 vdev_deflated_space(vd
, space_delta
));
1491 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1494 vdev_t
*vd
= mg
->mg_vd
;
1495 spa_t
*spa
= vd
->vdev_spa
;
1496 objset_t
*mos
= spa
->spa_meta_objset
;
1500 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1501 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1502 mutex_init(&ms
->ms_sync_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1503 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1505 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1506 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1507 ms
->ms_allocator
= -1;
1508 ms
->ms_new
= B_TRUE
;
1511 * We only open space map objects that already exist. All others
1512 * will be opened when we finally allocate an object for it.
1515 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1516 ms
->ms_size
, vd
->vdev_ashift
);
1519 kmem_free(ms
, sizeof (metaslab_t
));
1523 ASSERT(ms
->ms_sm
!= NULL
);
1527 * We create the main range tree here, but we don't create the
1528 * other range trees until metaslab_sync_done(). This serves
1529 * two purposes: it allows metaslab_sync_done() to detect the
1530 * addition of new space; and for debugging, it ensures that we'd
1531 * data fault on any attempt to use this metaslab before it's ready.
1533 ms
->ms_allocatable
= range_tree_create_impl(&rt_avl_ops
,
1534 &ms
->ms_allocatable_by_size
, metaslab_rangesize_compare
, 0);
1535 metaslab_group_add(mg
, ms
);
1537 metaslab_set_fragmentation(ms
);
1540 * If we're opening an existing pool (txg == 0) or creating
1541 * a new one (txg == TXG_INITIAL), all space is available now.
1542 * If we're adding space to an existing pool, the new space
1543 * does not become available until after this txg has synced.
1544 * The metaslab's weight will also be initialized when we sync
1545 * out this txg. This ensures that we don't attempt to allocate
1546 * from it before we have initialized it completely.
1548 if (txg
<= TXG_INITIAL
)
1549 metaslab_sync_done(ms
, 0);
1552 * If metaslab_debug_load is set and we're initializing a metaslab
1553 * that has an allocated space map object then load the space map
1554 * so that we can verify frees.
1556 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1557 mutex_enter(&ms
->ms_lock
);
1558 VERIFY0(metaslab_load(ms
));
1559 mutex_exit(&ms
->ms_lock
);
1563 vdev_dirty(vd
, 0, NULL
, txg
);
1564 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1573 metaslab_fini(metaslab_t
*msp
)
1575 metaslab_group_t
*mg
= msp
->ms_group
;
1576 vdev_t
*vd
= mg
->mg_vd
;
1578 metaslab_group_remove(mg
, msp
);
1580 mutex_enter(&msp
->ms_lock
);
1581 VERIFY(msp
->ms_group
== NULL
);
1582 metaslab_space_update(vd
, mg
->mg_class
,
1583 -space_map_allocated(msp
->ms_sm
), 0, -msp
->ms_size
);
1585 space_map_close(msp
->ms_sm
);
1587 metaslab_unload(msp
);
1589 range_tree_destroy(msp
->ms_allocatable
);
1590 range_tree_destroy(msp
->ms_freeing
);
1591 range_tree_destroy(msp
->ms_freed
);
1593 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1594 range_tree_destroy(msp
->ms_allocating
[t
]);
1597 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1598 range_tree_destroy(msp
->ms_defer
[t
]);
1600 ASSERT0(msp
->ms_deferspace
);
1602 range_tree_destroy(msp
->ms_checkpointing
);
1604 mutex_exit(&msp
->ms_lock
);
1605 cv_destroy(&msp
->ms_load_cv
);
1606 mutex_destroy(&msp
->ms_lock
);
1607 mutex_destroy(&msp
->ms_sync_lock
);
1608 ASSERT3U(msp
->ms_allocator
, ==, -1);
1610 kmem_free(msp
, sizeof (metaslab_t
));
1613 #define FRAGMENTATION_TABLE_SIZE 17
1616 * This table defines a segment size based fragmentation metric that will
1617 * allow each metaslab to derive its own fragmentation value. This is done
1618 * by calculating the space in each bucket of the spacemap histogram and
1619 * multiplying that by the fragmetation metric in this table. Doing
1620 * this for all buckets and dividing it by the total amount of free
1621 * space in this metaslab (i.e. the total free space in all buckets) gives
1622 * us the fragmentation metric. This means that a high fragmentation metric
1623 * equates to most of the free space being comprised of small segments.
1624 * Conversely, if the metric is low, then most of the free space is in
1625 * large segments. A 10% change in fragmentation equates to approximately
1626 * double the number of segments.
1628 * This table defines 0% fragmented space using 16MB segments. Testing has
1629 * shown that segments that are greater than or equal to 16MB do not suffer
1630 * from drastic performance problems. Using this value, we derive the rest
1631 * of the table. Since the fragmentation value is never stored on disk, it
1632 * is possible to change these calculations in the future.
1634 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1654 * Calclate the metaslab's fragmentation metric. A return value
1655 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1656 * not support this metric. Otherwise, the return value should be in the
1660 metaslab_set_fragmentation(metaslab_t
*msp
)
1662 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1663 uint64_t fragmentation
= 0;
1665 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1666 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1668 if (!feature_enabled
) {
1669 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1674 * A null space map means that the entire metaslab is free
1675 * and thus is not fragmented.
1677 if (msp
->ms_sm
== NULL
) {
1678 msp
->ms_fragmentation
= 0;
1683 * If this metaslab's space map has not been upgraded, flag it
1684 * so that we upgrade next time we encounter it.
1686 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1687 uint64_t txg
= spa_syncing_txg(spa
);
1688 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1691 * If we've reached the final dirty txg, then we must
1692 * be shutting down the pool. We don't want to dirty
1693 * any data past this point so skip setting the condense
1694 * flag. We can retry this action the next time the pool
1697 if (spa_writeable(spa
) && txg
< spa_final_dirty_txg(spa
)) {
1698 msp
->ms_condense_wanted
= B_TRUE
;
1699 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1700 zfs_dbgmsg("txg %llu, requesting force condense: "
1701 "ms_id %llu, vdev_id %llu", txg
, msp
->ms_id
,
1704 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1708 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1710 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1712 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1713 FRAGMENTATION_TABLE_SIZE
- 1);
1715 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1718 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1721 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1722 fragmentation
+= space
* zfs_frag_table
[idx
];
1726 fragmentation
/= total
;
1727 ASSERT3U(fragmentation
, <=, 100);
1729 msp
->ms_fragmentation
= fragmentation
;
1733 * Compute a weight -- a selection preference value -- for the given metaslab.
1734 * This is based on the amount of free space, the level of fragmentation,
1735 * the LBA range, and whether the metaslab is loaded.
1738 metaslab_space_weight(metaslab_t
*msp
)
1740 metaslab_group_t
*mg
= msp
->ms_group
;
1741 vdev_t
*vd
= mg
->mg_vd
;
1742 uint64_t weight
, space
;
1744 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1745 ASSERT(!vd
->vdev_removing
);
1748 * The baseline weight is the metaslab's free space.
1750 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1752 if (metaslab_fragmentation_factor_enabled
&&
1753 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1755 * Use the fragmentation information to inversely scale
1756 * down the baseline weight. We need to ensure that we
1757 * don't exclude this metaslab completely when it's 100%
1758 * fragmented. To avoid this we reduce the fragmented value
1761 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1764 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1765 * this metaslab again. The fragmentation metric may have
1766 * decreased the space to something smaller than
1767 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1768 * so that we can consume any remaining space.
1770 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1771 space
= SPA_MINBLOCKSIZE
;
1776 * Modern disks have uniform bit density and constant angular velocity.
1777 * Therefore, the outer recording zones are faster (higher bandwidth)
1778 * than the inner zones by the ratio of outer to inner track diameter,
1779 * which is typically around 2:1. We account for this by assigning
1780 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1781 * In effect, this means that we'll select the metaslab with the most
1782 * free bandwidth rather than simply the one with the most free space.
1784 if (!vd
->vdev_nonrot
&& metaslab_lba_weighting_enabled
) {
1785 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1786 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1790 * If this metaslab is one we're actively using, adjust its
1791 * weight to make it preferable to any inactive metaslab so
1792 * we'll polish it off. If the fragmentation on this metaslab
1793 * has exceed our threshold, then don't mark it active.
1795 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1796 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1797 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1800 WEIGHT_SET_SPACEBASED(weight
);
1805 * Return the weight of the specified metaslab, according to the segment-based
1806 * weighting algorithm. The metaslab must be loaded. This function can
1807 * be called within a sync pass since it relies only on the metaslab's
1808 * range tree which is always accurate when the metaslab is loaded.
1811 metaslab_weight_from_range_tree(metaslab_t
*msp
)
1813 uint64_t weight
= 0;
1814 uint32_t segments
= 0;
1816 ASSERT(msp
->ms_loaded
);
1818 for (int i
= RANGE_TREE_HISTOGRAM_SIZE
- 1; i
>= SPA_MINBLOCKSHIFT
;
1820 uint8_t shift
= msp
->ms_group
->mg_vd
->vdev_ashift
;
1821 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1824 segments
+= msp
->ms_allocatable
->rt_histogram
[i
];
1827 * The range tree provides more precision than the space map
1828 * and must be downgraded so that all values fit within the
1829 * space map's histogram. This allows us to compare loaded
1830 * vs. unloaded metaslabs to determine which metaslab is
1831 * considered "best".
1836 if (segments
!= 0) {
1837 WEIGHT_SET_COUNT(weight
, segments
);
1838 WEIGHT_SET_INDEX(weight
, i
);
1839 WEIGHT_SET_ACTIVE(weight
, 0);
1847 * Calculate the weight based on the on-disk histogram. This should only
1848 * be called after a sync pass has completely finished since the on-disk
1849 * information is updated in metaslab_sync().
1852 metaslab_weight_from_spacemap(metaslab_t
*msp
)
1854 uint64_t weight
= 0;
1856 for (int i
= SPACE_MAP_HISTOGRAM_SIZE
- 1; i
>= 0; i
--) {
1857 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] != 0) {
1858 WEIGHT_SET_COUNT(weight
,
1859 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
1860 WEIGHT_SET_INDEX(weight
, i
+
1861 msp
->ms_sm
->sm_shift
);
1862 WEIGHT_SET_ACTIVE(weight
, 0);
1870 * Compute a segment-based weight for the specified metaslab. The weight
1871 * is determined by highest bucket in the histogram. The information
1872 * for the highest bucket is encoded into the weight value.
1875 metaslab_segment_weight(metaslab_t
*msp
)
1877 metaslab_group_t
*mg
= msp
->ms_group
;
1878 uint64_t weight
= 0;
1879 uint8_t shift
= mg
->mg_vd
->vdev_ashift
;
1881 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1884 * The metaslab is completely free.
1886 if (space_map_allocated(msp
->ms_sm
) == 0) {
1887 int idx
= highbit64(msp
->ms_size
) - 1;
1888 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1890 if (idx
< max_idx
) {
1891 WEIGHT_SET_COUNT(weight
, 1ULL);
1892 WEIGHT_SET_INDEX(weight
, idx
);
1894 WEIGHT_SET_COUNT(weight
, 1ULL << (idx
- max_idx
));
1895 WEIGHT_SET_INDEX(weight
, max_idx
);
1897 WEIGHT_SET_ACTIVE(weight
, 0);
1898 ASSERT(!WEIGHT_IS_SPACEBASED(weight
));
1903 ASSERT3U(msp
->ms_sm
->sm_dbuf
->db_size
, ==, sizeof (space_map_phys_t
));
1906 * If the metaslab is fully allocated then just make the weight 0.
1908 if (space_map_allocated(msp
->ms_sm
) == msp
->ms_size
)
1911 * If the metaslab is already loaded, then use the range tree to
1912 * determine the weight. Otherwise, we rely on the space map information
1913 * to generate the weight.
1915 if (msp
->ms_loaded
) {
1916 weight
= metaslab_weight_from_range_tree(msp
);
1918 weight
= metaslab_weight_from_spacemap(msp
);
1922 * If the metaslab was active the last time we calculated its weight
1923 * then keep it active. We want to consume the entire region that
1924 * is associated with this weight.
1926 if (msp
->ms_activation_weight
!= 0 && weight
!= 0)
1927 WEIGHT_SET_ACTIVE(weight
, WEIGHT_GET_ACTIVE(msp
->ms_weight
));
1932 * Determine if we should attempt to allocate from this metaslab. If the
1933 * metaslab has a maximum size then we can quickly determine if the desired
1934 * allocation size can be satisfied. Otherwise, if we're using segment-based
1935 * weighting then we can determine the maximum allocation that this metaslab
1936 * can accommodate based on the index encoded in the weight. If we're using
1937 * space-based weights then rely on the entire weight (excluding the weight
1941 metaslab_should_allocate(metaslab_t
*msp
, uint64_t asize
)
1943 boolean_t should_allocate
;
1945 if (msp
->ms_max_size
!= 0)
1946 return (msp
->ms_max_size
>= asize
);
1948 if (!WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
1950 * The metaslab segment weight indicates segments in the
1951 * range [2^i, 2^(i+1)), where i is the index in the weight.
1952 * Since the asize might be in the middle of the range, we
1953 * should attempt the allocation if asize < 2^(i+1).
1955 should_allocate
= (asize
<
1956 1ULL << (WEIGHT_GET_INDEX(msp
->ms_weight
) + 1));
1958 should_allocate
= (asize
<=
1959 (msp
->ms_weight
& ~METASLAB_WEIGHT_TYPE
));
1961 return (should_allocate
);
1964 metaslab_weight(metaslab_t
*msp
)
1966 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1967 spa_t
*spa
= vd
->vdev_spa
;
1970 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1973 * If this vdev is in the process of being removed, there is nothing
1974 * for us to do here.
1976 if (vd
->vdev_removing
)
1979 metaslab_set_fragmentation(msp
);
1982 * Update the maximum size if the metaslab is loaded. This will
1983 * ensure that we get an accurate maximum size if newly freed space
1984 * has been added back into the free tree.
1987 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1990 * Segment-based weighting requires space map histogram support.
1992 if (zfs_metaslab_segment_weight_enabled
&&
1993 spa_feature_is_enabled(spa
, SPA_FEATURE_SPACEMAP_HISTOGRAM
) &&
1994 (msp
->ms_sm
== NULL
|| msp
->ms_sm
->sm_dbuf
->db_size
==
1995 sizeof (space_map_phys_t
))) {
1996 weight
= metaslab_segment_weight(msp
);
1998 weight
= metaslab_space_weight(msp
);
2004 metaslab_activate_allocator(metaslab_group_t
*mg
, metaslab_t
*msp
,
2005 int allocator
, uint64_t activation_weight
)
2008 * If we're activating for the claim code, we don't want to actually
2009 * set the metaslab up for a specific allocator.
2011 if (activation_weight
== METASLAB_WEIGHT_CLAIM
)
2013 metaslab_t
**arr
= (activation_weight
== METASLAB_WEIGHT_PRIMARY
?
2014 mg
->mg_primaries
: mg
->mg_secondaries
);
2016 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2017 mutex_enter(&mg
->mg_lock
);
2018 if (arr
[allocator
] != NULL
) {
2019 mutex_exit(&mg
->mg_lock
);
2023 arr
[allocator
] = msp
;
2024 ASSERT3S(msp
->ms_allocator
, ==, -1);
2025 msp
->ms_allocator
= allocator
;
2026 msp
->ms_primary
= (activation_weight
== METASLAB_WEIGHT_PRIMARY
);
2027 mutex_exit(&mg
->mg_lock
);
2033 metaslab_activate(metaslab_t
*msp
, int allocator
, uint64_t activation_weight
)
2035 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2037 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
2039 metaslab_load_wait(msp
);
2040 if (!msp
->ms_loaded
) {
2041 if ((error
= metaslab_load(msp
)) != 0) {
2042 metaslab_group_sort(msp
->ms_group
, msp
, 0);
2046 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) != 0) {
2048 * The metaslab was activated for another allocator
2049 * while we were waiting, we should reselect.
2053 if ((error
= metaslab_activate_allocator(msp
->ms_group
, msp
,
2054 allocator
, activation_weight
)) != 0) {
2058 msp
->ms_activation_weight
= msp
->ms_weight
;
2059 metaslab_group_sort(msp
->ms_group
, msp
,
2060 msp
->ms_weight
| activation_weight
);
2062 ASSERT(msp
->ms_loaded
);
2063 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
2069 metaslab_passivate_allocator(metaslab_group_t
*mg
, metaslab_t
*msp
,
2072 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2073 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
) {
2074 metaslab_group_sort(mg
, msp
, weight
);
2078 mutex_enter(&mg
->mg_lock
);
2079 ASSERT3P(msp
->ms_group
, ==, mg
);
2080 if (msp
->ms_primary
) {
2081 ASSERT3U(0, <=, msp
->ms_allocator
);
2082 ASSERT3U(msp
->ms_allocator
, <, mg
->mg_allocators
);
2083 ASSERT3P(mg
->mg_primaries
[msp
->ms_allocator
], ==, msp
);
2084 ASSERT(msp
->ms_weight
& METASLAB_WEIGHT_PRIMARY
);
2085 mg
->mg_primaries
[msp
->ms_allocator
] = NULL
;
2087 ASSERT(msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
);
2088 ASSERT3P(mg
->mg_secondaries
[msp
->ms_allocator
], ==, msp
);
2089 mg
->mg_secondaries
[msp
->ms_allocator
] = NULL
;
2091 msp
->ms_allocator
= -1;
2092 metaslab_group_sort_impl(mg
, msp
, weight
);
2093 mutex_exit(&mg
->mg_lock
);
2097 metaslab_passivate(metaslab_t
*msp
, uint64_t weight
)
2099 ASSERTV(uint64_t size
= weight
& ~METASLAB_WEIGHT_TYPE
);
2102 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
2103 * this metaslab again. In that case, it had better be empty,
2104 * or we would be leaving space on the table.
2106 ASSERT(!WEIGHT_IS_SPACEBASED(msp
->ms_weight
) ||
2107 size
>= SPA_MINBLOCKSIZE
||
2108 range_tree_space(msp
->ms_allocatable
) == 0);
2109 ASSERT0(weight
& METASLAB_ACTIVE_MASK
);
2111 msp
->ms_activation_weight
= 0;
2112 metaslab_passivate_allocator(msp
->ms_group
, msp
, weight
);
2113 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
2117 * Segment-based metaslabs are activated once and remain active until
2118 * we either fail an allocation attempt (similar to space-based metaslabs)
2119 * or have exhausted the free space in zfs_metaslab_switch_threshold
2120 * buckets since the metaslab was activated. This function checks to see
2121 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
2122 * metaslab and passivates it proactively. This will allow us to select a
2123 * metaslab with a larger contiguous region, if any, remaining within this
2124 * metaslab group. If we're in sync pass > 1, then we continue using this
2125 * metaslab so that we don't dirty more block and cause more sync passes.
2128 metaslab_segment_may_passivate(metaslab_t
*msp
)
2130 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2132 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
) || spa_sync_pass(spa
) > 1)
2136 * Since we are in the middle of a sync pass, the most accurate
2137 * information that is accessible to us is the in-core range tree
2138 * histogram; calculate the new weight based on that information.
2140 uint64_t weight
= metaslab_weight_from_range_tree(msp
);
2141 int activation_idx
= WEIGHT_GET_INDEX(msp
->ms_activation_weight
);
2142 int current_idx
= WEIGHT_GET_INDEX(weight
);
2144 if (current_idx
<= activation_idx
- zfs_metaslab_switch_threshold
)
2145 metaslab_passivate(msp
, weight
);
2149 metaslab_preload(void *arg
)
2151 metaslab_t
*msp
= arg
;
2152 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2153 fstrans_cookie_t cookie
= spl_fstrans_mark();
2155 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
2157 mutex_enter(&msp
->ms_lock
);
2158 metaslab_load_wait(msp
);
2159 if (!msp
->ms_loaded
)
2160 (void) metaslab_load(msp
);
2161 msp
->ms_selected_txg
= spa_syncing_txg(spa
);
2162 mutex_exit(&msp
->ms_lock
);
2163 spl_fstrans_unmark(cookie
);
2167 metaslab_group_preload(metaslab_group_t
*mg
)
2169 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2171 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2174 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
2175 taskq_wait_outstanding(mg
->mg_taskq
, 0);
2179 mutex_enter(&mg
->mg_lock
);
2182 * Load the next potential metaslabs
2184 for (msp
= avl_first(t
); msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
2185 ASSERT3P(msp
->ms_group
, ==, mg
);
2188 * We preload only the maximum number of metaslabs specified
2189 * by metaslab_preload_limit. If a metaslab is being forced
2190 * to condense then we preload it too. This will ensure
2191 * that force condensing happens in the next txg.
2193 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
2197 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
2198 msp
, TQ_SLEEP
) != TASKQID_INVALID
);
2200 mutex_exit(&mg
->mg_lock
);
2204 * Determine if the space map's on-disk footprint is past our tolerance
2205 * for inefficiency. We would like to use the following criteria to make
2208 * 1. The size of the space map object should not dramatically increase as a
2209 * result of writing out the free space range tree.
2211 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2212 * times the size than the free space range tree representation
2213 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1MB).
2215 * 3. The on-disk size of the space map should actually decrease.
2217 * Unfortunately, we cannot compute the on-disk size of the space map in this
2218 * context because we cannot accurately compute the effects of compression, etc.
2219 * Instead, we apply the heuristic described in the block comment for
2220 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2221 * is greater than a threshold number of blocks.
2224 metaslab_should_condense(metaslab_t
*msp
)
2226 space_map_t
*sm
= msp
->ms_sm
;
2227 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
2228 uint64_t vdev_blocksize
= 1 << vd
->vdev_ashift
;
2229 uint64_t current_txg
= spa_syncing_txg(vd
->vdev_spa
);
2231 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2232 ASSERT(msp
->ms_loaded
);
2235 * Allocations and frees in early passes are generally more space
2236 * efficient (in terms of blocks described in space map entries)
2237 * than the ones in later passes (e.g. we don't compress after
2238 * sync pass 5) and condensing a metaslab multiple times in a txg
2239 * could degrade performance.
2241 * Thus we prefer condensing each metaslab at most once every txg at
2242 * the earliest sync pass possible. If a metaslab is eligible for
2243 * condensing again after being considered for condensing within the
2244 * same txg, it will hopefully be dirty in the next txg where it will
2245 * be condensed at an earlier pass.
2247 if (msp
->ms_condense_checked_txg
== current_txg
)
2249 msp
->ms_condense_checked_txg
= current_txg
;
2252 * We always condense metaslabs that are empty and metaslabs for
2253 * which a condense request has been made.
2255 if (avl_is_empty(&msp
->ms_allocatable_by_size
) ||
2256 msp
->ms_condense_wanted
)
2259 uint64_t object_size
= space_map_length(msp
->ms_sm
);
2260 uint64_t optimal_size
= space_map_estimate_optimal_size(sm
,
2261 msp
->ms_allocatable
, SM_NO_VDEVID
);
2263 dmu_object_info_t doi
;
2264 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
2265 uint64_t record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
2267 return (object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
2268 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
2272 * Condense the on-disk space map representation to its minimized form.
2273 * The minimized form consists of a small number of allocations followed by
2274 * the entries of the free range tree.
2277 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
2279 range_tree_t
*condense_tree
;
2280 space_map_t
*sm
= msp
->ms_sm
;
2282 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2283 ASSERT(msp
->ms_loaded
);
2286 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2287 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
2288 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
2289 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
2290 space_map_length(msp
->ms_sm
),
2291 avl_numnodes(&msp
->ms_allocatable
->rt_root
),
2292 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
2294 msp
->ms_condense_wanted
= B_FALSE
;
2297 * Create an range tree that is 100% allocated. We remove segments
2298 * that have been freed in this txg, any deferred frees that exist,
2299 * and any allocation in the future. Removing segments should be
2300 * a relatively inexpensive operation since we expect these trees to
2301 * have a small number of nodes.
2303 condense_tree
= range_tree_create(NULL
, NULL
);
2304 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
2306 range_tree_walk(msp
->ms_freeing
, range_tree_remove
, condense_tree
);
2307 range_tree_walk(msp
->ms_freed
, range_tree_remove
, condense_tree
);
2309 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2310 range_tree_walk(msp
->ms_defer
[t
],
2311 range_tree_remove
, condense_tree
);
2314 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2315 range_tree_walk(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
],
2316 range_tree_remove
, condense_tree
);
2320 * We're about to drop the metaslab's lock thus allowing
2321 * other consumers to change it's content. Set the
2322 * metaslab's ms_condensing flag to ensure that
2323 * allocations on this metaslab do not occur while we're
2324 * in the middle of committing it to disk. This is only critical
2325 * for ms_allocatable as all other range trees use per txg
2326 * views of their content.
2328 msp
->ms_condensing
= B_TRUE
;
2330 mutex_exit(&msp
->ms_lock
);
2331 space_map_truncate(sm
, zfs_metaslab_sm_blksz
, tx
);
2334 * While we would ideally like to create a space map representation
2335 * that consists only of allocation records, doing so can be
2336 * prohibitively expensive because the in-core free tree can be
2337 * large, and therefore computationally expensive to subtract
2338 * from the condense_tree. Instead we sync out two trees, a cheap
2339 * allocation only tree followed by the in-core free tree. While not
2340 * optimal, this is typically close to optimal, and much cheaper to
2343 space_map_write(sm
, condense_tree
, SM_ALLOC
, SM_NO_VDEVID
, tx
);
2344 range_tree_vacate(condense_tree
, NULL
, NULL
);
2345 range_tree_destroy(condense_tree
);
2347 space_map_write(sm
, msp
->ms_allocatable
, SM_FREE
, SM_NO_VDEVID
, tx
);
2348 mutex_enter(&msp
->ms_lock
);
2349 msp
->ms_condensing
= B_FALSE
;
2353 * Write a metaslab to disk in the context of the specified transaction group.
2356 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
2358 metaslab_group_t
*mg
= msp
->ms_group
;
2359 vdev_t
*vd
= mg
->mg_vd
;
2360 spa_t
*spa
= vd
->vdev_spa
;
2361 objset_t
*mos
= spa_meta_objset(spa
);
2362 range_tree_t
*alloctree
= msp
->ms_allocating
[txg
& TXG_MASK
];
2364 uint64_t object
= space_map_object(msp
->ms_sm
);
2366 ASSERT(!vd
->vdev_ishole
);
2369 * This metaslab has just been added so there's no work to do now.
2371 if (msp
->ms_freeing
== NULL
) {
2372 ASSERT3P(alloctree
, ==, NULL
);
2376 ASSERT3P(alloctree
, !=, NULL
);
2377 ASSERT3P(msp
->ms_freeing
, !=, NULL
);
2378 ASSERT3P(msp
->ms_freed
, !=, NULL
);
2379 ASSERT3P(msp
->ms_checkpointing
, !=, NULL
);
2382 * Normally, we don't want to process a metaslab if there are no
2383 * allocations or frees to perform. However, if the metaslab is being
2384 * forced to condense and it's loaded, we need to let it through.
2386 if (range_tree_is_empty(alloctree
) &&
2387 range_tree_is_empty(msp
->ms_freeing
) &&
2388 range_tree_is_empty(msp
->ms_checkpointing
) &&
2389 !(msp
->ms_loaded
&& msp
->ms_condense_wanted
))
2393 VERIFY(txg
<= spa_final_dirty_txg(spa
));
2396 * The only state that can actually be changing concurrently with
2397 * metaslab_sync() is the metaslab's ms_allocatable. No other
2398 * thread can be modifying this txg's alloc, freeing,
2399 * freed, or space_map_phys_t. We drop ms_lock whenever we
2400 * could call into the DMU, because the DMU can call down to us
2401 * (e.g. via zio_free()) at any time.
2403 * The spa_vdev_remove_thread() can be reading metaslab state
2404 * concurrently, and it is locked out by the ms_sync_lock. Note
2405 * that the ms_lock is insufficient for this, because it is dropped
2406 * by space_map_write().
2408 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
2410 if (msp
->ms_sm
== NULL
) {
2411 uint64_t new_object
;
2413 new_object
= space_map_alloc(mos
, zfs_metaslab_sm_blksz
, tx
);
2414 VERIFY3U(new_object
, !=, 0);
2416 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
2417 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
));
2418 ASSERT(msp
->ms_sm
!= NULL
);
2421 if (!range_tree_is_empty(msp
->ms_checkpointing
) &&
2422 vd
->vdev_checkpoint_sm
== NULL
) {
2423 ASSERT(spa_has_checkpoint(spa
));
2425 uint64_t new_object
= space_map_alloc(mos
,
2426 vdev_standard_sm_blksz
, tx
);
2427 VERIFY3U(new_object
, !=, 0);
2429 VERIFY0(space_map_open(&vd
->vdev_checkpoint_sm
,
2430 mos
, new_object
, 0, vd
->vdev_asize
, vd
->vdev_ashift
));
2431 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
2434 * We save the space map object as an entry in vdev_top_zap
2435 * so it can be retrieved when the pool is reopened after an
2436 * export or through zdb.
2438 VERIFY0(zap_add(vd
->vdev_spa
->spa_meta_objset
,
2439 vd
->vdev_top_zap
, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM
,
2440 sizeof (new_object
), 1, &new_object
, tx
));
2443 mutex_enter(&msp
->ms_sync_lock
);
2444 mutex_enter(&msp
->ms_lock
);
2447 * Note: metaslab_condense() clears the space map's histogram.
2448 * Therefore we must verify and remove this histogram before
2451 metaslab_group_histogram_verify(mg
);
2452 metaslab_class_histogram_verify(mg
->mg_class
);
2453 metaslab_group_histogram_remove(mg
, msp
);
2455 if (msp
->ms_loaded
&& metaslab_should_condense(msp
)) {
2456 metaslab_condense(msp
, txg
, tx
);
2458 mutex_exit(&msp
->ms_lock
);
2459 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
,
2461 space_map_write(msp
->ms_sm
, msp
->ms_freeing
, SM_FREE
,
2463 mutex_enter(&msp
->ms_lock
);
2466 if (!range_tree_is_empty(msp
->ms_checkpointing
)) {
2467 ASSERT(spa_has_checkpoint(spa
));
2468 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
2471 * Since we are doing writes to disk and the ms_checkpointing
2472 * tree won't be changing during that time, we drop the
2473 * ms_lock while writing to the checkpoint space map.
2475 mutex_exit(&msp
->ms_lock
);
2476 space_map_write(vd
->vdev_checkpoint_sm
,
2477 msp
->ms_checkpointing
, SM_FREE
, SM_NO_VDEVID
, tx
);
2478 mutex_enter(&msp
->ms_lock
);
2479 space_map_update(vd
->vdev_checkpoint_sm
);
2481 spa
->spa_checkpoint_info
.sci_dspace
+=
2482 range_tree_space(msp
->ms_checkpointing
);
2483 vd
->vdev_stat
.vs_checkpoint_space
+=
2484 range_tree_space(msp
->ms_checkpointing
);
2485 ASSERT3U(vd
->vdev_stat
.vs_checkpoint_space
, ==,
2486 -vd
->vdev_checkpoint_sm
->sm_alloc
);
2488 range_tree_vacate(msp
->ms_checkpointing
, NULL
, NULL
);
2491 if (msp
->ms_loaded
) {
2493 * When the space map is loaded, we have an accurate
2494 * histogram in the range tree. This gives us an opportunity
2495 * to bring the space map's histogram up-to-date so we clear
2496 * it first before updating it.
2498 space_map_histogram_clear(msp
->ms_sm
);
2499 space_map_histogram_add(msp
->ms_sm
, msp
->ms_allocatable
, tx
);
2502 * Since we've cleared the histogram we need to add back
2503 * any free space that has already been processed, plus
2504 * any deferred space. This allows the on-disk histogram
2505 * to accurately reflect all free space even if some space
2506 * is not yet available for allocation (i.e. deferred).
2508 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freed
, tx
);
2511 * Add back any deferred free space that has not been
2512 * added back into the in-core free tree yet. This will
2513 * ensure that we don't end up with a space map histogram
2514 * that is completely empty unless the metaslab is fully
2517 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2518 space_map_histogram_add(msp
->ms_sm
,
2519 msp
->ms_defer
[t
], tx
);
2524 * Always add the free space from this sync pass to the space
2525 * map histogram. We want to make sure that the on-disk histogram
2526 * accounts for all free space. If the space map is not loaded,
2527 * then we will lose some accuracy but will correct it the next
2528 * time we load the space map.
2530 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freeing
, tx
);
2532 metaslab_group_histogram_add(mg
, msp
);
2533 metaslab_group_histogram_verify(mg
);
2534 metaslab_class_histogram_verify(mg
->mg_class
);
2537 * For sync pass 1, we avoid traversing this txg's free range tree
2538 * and instead will just swap the pointers for freeing and
2539 * freed. We can safely do this since the freed_tree is
2540 * guaranteed to be empty on the initial pass.
2542 if (spa_sync_pass(spa
) == 1) {
2543 range_tree_swap(&msp
->ms_freeing
, &msp
->ms_freed
);
2545 range_tree_vacate(msp
->ms_freeing
,
2546 range_tree_add
, msp
->ms_freed
);
2548 range_tree_vacate(alloctree
, NULL
, NULL
);
2550 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
2551 ASSERT0(range_tree_space(msp
->ms_allocating
[TXG_CLEAN(txg
)
2553 ASSERT0(range_tree_space(msp
->ms_freeing
));
2554 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2556 mutex_exit(&msp
->ms_lock
);
2558 if (object
!= space_map_object(msp
->ms_sm
)) {
2559 object
= space_map_object(msp
->ms_sm
);
2560 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
2561 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
2563 mutex_exit(&msp
->ms_sync_lock
);
2568 * Called after a transaction group has completely synced to mark
2569 * all of the metaslab's free space as usable.
2572 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
2574 metaslab_group_t
*mg
= msp
->ms_group
;
2575 vdev_t
*vd
= mg
->mg_vd
;
2576 spa_t
*spa
= vd
->vdev_spa
;
2577 range_tree_t
**defer_tree
;
2578 int64_t alloc_delta
, defer_delta
;
2579 boolean_t defer_allowed
= B_TRUE
;
2581 ASSERT(!vd
->vdev_ishole
);
2583 mutex_enter(&msp
->ms_lock
);
2586 * If this metaslab is just becoming available, initialize its
2587 * range trees and add its capacity to the vdev.
2589 if (msp
->ms_freed
== NULL
) {
2590 for (int t
= 0; t
< TXG_SIZE
; t
++) {
2591 ASSERT(msp
->ms_allocating
[t
] == NULL
);
2593 msp
->ms_allocating
[t
] = range_tree_create(NULL
, NULL
);
2596 ASSERT3P(msp
->ms_freeing
, ==, NULL
);
2597 msp
->ms_freeing
= range_tree_create(NULL
, NULL
);
2599 ASSERT3P(msp
->ms_freed
, ==, NULL
);
2600 msp
->ms_freed
= range_tree_create(NULL
, NULL
);
2602 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2603 ASSERT(msp
->ms_defer
[t
] == NULL
);
2605 msp
->ms_defer
[t
] = range_tree_create(NULL
, NULL
);
2608 ASSERT3P(msp
->ms_checkpointing
, ==, NULL
);
2609 msp
->ms_checkpointing
= range_tree_create(NULL
, NULL
);
2611 metaslab_space_update(vd
, mg
->mg_class
, 0, 0, msp
->ms_size
);
2613 ASSERT0(range_tree_space(msp
->ms_freeing
));
2614 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2616 defer_tree
= &msp
->ms_defer
[txg
% TXG_DEFER_SIZE
];
2618 uint64_t free_space
= metaslab_class_get_space(spa_normal_class(spa
)) -
2619 metaslab_class_get_alloc(spa_normal_class(spa
));
2620 if (free_space
<= spa_get_slop_space(spa
) || vd
->vdev_removing
) {
2621 defer_allowed
= B_FALSE
;
2625 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
2626 if (defer_allowed
) {
2627 defer_delta
= range_tree_space(msp
->ms_freed
) -
2628 range_tree_space(*defer_tree
);
2630 defer_delta
-= range_tree_space(*defer_tree
);
2633 metaslab_space_update(vd
, mg
->mg_class
, alloc_delta
+ defer_delta
,
2637 * If there's a metaslab_load() in progress, wait for it to complete
2638 * so that we have a consistent view of the in-core space map.
2640 metaslab_load_wait(msp
);
2643 * Move the frees from the defer_tree back to the free
2644 * range tree (if it's loaded). Swap the freed_tree and
2645 * the defer_tree -- this is safe to do because we've
2646 * just emptied out the defer_tree.
2648 range_tree_vacate(*defer_tree
,
2649 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_allocatable
);
2650 if (defer_allowed
) {
2651 range_tree_swap(&msp
->ms_freed
, defer_tree
);
2653 range_tree_vacate(msp
->ms_freed
,
2654 msp
->ms_loaded
? range_tree_add
: NULL
,
2655 msp
->ms_allocatable
);
2657 space_map_update(msp
->ms_sm
);
2659 msp
->ms_deferspace
+= defer_delta
;
2660 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2661 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2662 if (msp
->ms_deferspace
!= 0) {
2664 * Keep syncing this metaslab until all deferred frees
2665 * are back in circulation.
2667 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2671 msp
->ms_new
= B_FALSE
;
2672 mutex_enter(&mg
->mg_lock
);
2674 mutex_exit(&mg
->mg_lock
);
2677 * Calculate the new weights before unloading any metaslabs.
2678 * This will give us the most accurate weighting.
2680 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
) |
2681 (msp
->ms_weight
& METASLAB_ACTIVE_MASK
));
2684 * If the metaslab is loaded and we've not tried to load or allocate
2685 * from it in 'metaslab_unload_delay' txgs, then unload it.
2687 if (msp
->ms_loaded
&&
2688 msp
->ms_selected_txg
+ metaslab_unload_delay
< txg
) {
2690 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2691 VERIFY0(range_tree_space(
2692 msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]));
2694 if (msp
->ms_allocator
!= -1) {
2695 metaslab_passivate(msp
, msp
->ms_weight
&
2696 ~METASLAB_ACTIVE_MASK
);
2699 if (!metaslab_debug_unload
)
2700 metaslab_unload(msp
);
2703 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
2704 ASSERT0(range_tree_space(msp
->ms_freeing
));
2705 ASSERT0(range_tree_space(msp
->ms_freed
));
2706 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2708 mutex_exit(&msp
->ms_lock
);
2712 metaslab_sync_reassess(metaslab_group_t
*mg
)
2714 spa_t
*spa
= mg
->mg_class
->mc_spa
;
2716 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2717 metaslab_group_alloc_update(mg
);
2718 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2721 * Preload the next potential metaslabs but only on active
2722 * metaslab groups. We can get into a state where the metaslab
2723 * is no longer active since we dirty metaslabs as we remove a
2724 * a device, thus potentially making the metaslab group eligible
2727 if (mg
->mg_activation_count
> 0) {
2728 metaslab_group_preload(mg
);
2730 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2734 * When writing a ditto block (i.e. more than one DVA for a given BP) on
2735 * the same vdev as an existing DVA of this BP, then try to allocate it
2736 * on a different metaslab than existing DVAs (i.e. a unique metaslab).
2739 metaslab_is_unique(metaslab_t
*msp
, dva_t
*dva
)
2743 if (DVA_GET_ASIZE(dva
) == 0)
2746 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2749 dva_ms_id
= DVA_GET_OFFSET(dva
) >> msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2751 return (msp
->ms_id
!= dva_ms_id
);
2755 * ==========================================================================
2756 * Metaslab allocation tracing facility
2757 * ==========================================================================
2759 #ifdef _METASLAB_TRACING
2760 kstat_t
*metaslab_trace_ksp
;
2761 kstat_named_t metaslab_trace_over_limit
;
2764 metaslab_alloc_trace_init(void)
2766 ASSERT(metaslab_alloc_trace_cache
== NULL
);
2767 metaslab_alloc_trace_cache
= kmem_cache_create(
2768 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t
),
2769 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
2770 metaslab_trace_ksp
= kstat_create("zfs", 0, "metaslab_trace_stats",
2771 "misc", KSTAT_TYPE_NAMED
, 1, KSTAT_FLAG_VIRTUAL
);
2772 if (metaslab_trace_ksp
!= NULL
) {
2773 metaslab_trace_ksp
->ks_data
= &metaslab_trace_over_limit
;
2774 kstat_named_init(&metaslab_trace_over_limit
,
2775 "metaslab_trace_over_limit", KSTAT_DATA_UINT64
);
2776 kstat_install(metaslab_trace_ksp
);
2781 metaslab_alloc_trace_fini(void)
2783 if (metaslab_trace_ksp
!= NULL
) {
2784 kstat_delete(metaslab_trace_ksp
);
2785 metaslab_trace_ksp
= NULL
;
2787 kmem_cache_destroy(metaslab_alloc_trace_cache
);
2788 metaslab_alloc_trace_cache
= NULL
;
2792 * Add an allocation trace element to the allocation tracing list.
2795 metaslab_trace_add(zio_alloc_list_t
*zal
, metaslab_group_t
*mg
,
2796 metaslab_t
*msp
, uint64_t psize
, uint32_t dva_id
, uint64_t offset
,
2799 metaslab_alloc_trace_t
*mat
;
2801 if (!metaslab_trace_enabled
)
2805 * When the tracing list reaches its maximum we remove
2806 * the second element in the list before adding a new one.
2807 * By removing the second element we preserve the original
2808 * entry as a clue to what allocations steps have already been
2811 if (zal
->zal_size
== metaslab_trace_max_entries
) {
2812 metaslab_alloc_trace_t
*mat_next
;
2814 panic("too many entries in allocation list");
2816 atomic_inc_64(&metaslab_trace_over_limit
.value
.ui64
);
2818 mat_next
= list_next(&zal
->zal_list
, list_head(&zal
->zal_list
));
2819 list_remove(&zal
->zal_list
, mat_next
);
2820 kmem_cache_free(metaslab_alloc_trace_cache
, mat_next
);
2823 mat
= kmem_cache_alloc(metaslab_alloc_trace_cache
, KM_SLEEP
);
2824 list_link_init(&mat
->mat_list_node
);
2827 mat
->mat_size
= psize
;
2828 mat
->mat_dva_id
= dva_id
;
2829 mat
->mat_offset
= offset
;
2830 mat
->mat_weight
= 0;
2831 mat
->mat_allocator
= allocator
;
2834 mat
->mat_weight
= msp
->ms_weight
;
2837 * The list is part of the zio so locking is not required. Only
2838 * a single thread will perform allocations for a given zio.
2840 list_insert_tail(&zal
->zal_list
, mat
);
2843 ASSERT3U(zal
->zal_size
, <=, metaslab_trace_max_entries
);
2847 metaslab_trace_init(zio_alloc_list_t
*zal
)
2849 list_create(&zal
->zal_list
, sizeof (metaslab_alloc_trace_t
),
2850 offsetof(metaslab_alloc_trace_t
, mat_list_node
));
2855 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2857 metaslab_alloc_trace_t
*mat
;
2859 while ((mat
= list_remove_head(&zal
->zal_list
)) != NULL
)
2860 kmem_cache_free(metaslab_alloc_trace_cache
, mat
);
2861 list_destroy(&zal
->zal_list
);
2866 #define metaslab_trace_add(zal, mg, msp, psize, id, off, alloc)
2869 metaslab_alloc_trace_init(void)
2874 metaslab_alloc_trace_fini(void)
2879 metaslab_trace_init(zio_alloc_list_t
*zal
)
2884 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2888 #endif /* _METASLAB_TRACING */
2891 * ==========================================================================
2892 * Metaslab block operations
2893 * ==========================================================================
2897 metaslab_group_alloc_increment(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
,
2900 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2901 (flags
& METASLAB_DONT_THROTTLE
))
2904 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2905 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2908 (void) zfs_refcount_add(&mg
->mg_alloc_queue_depth
[allocator
], tag
);
2912 metaslab_group_increment_qdepth(metaslab_group_t
*mg
, int allocator
)
2914 uint64_t max
= mg
->mg_max_alloc_queue_depth
;
2915 uint64_t cur
= mg
->mg_cur_max_alloc_queue_depth
[allocator
];
2917 if (atomic_cas_64(&mg
->mg_cur_max_alloc_queue_depth
[allocator
],
2918 cur
, cur
+ 1) == cur
) {
2920 &mg
->mg_class
->mc_alloc_max_slots
[allocator
]);
2923 cur
= mg
->mg_cur_max_alloc_queue_depth
[allocator
];
2928 metaslab_group_alloc_decrement(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
,
2929 int allocator
, boolean_t io_complete
)
2931 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2932 (flags
& METASLAB_DONT_THROTTLE
))
2935 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2936 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2939 (void) refcount_remove(&mg
->mg_alloc_queue_depth
[allocator
], tag
);
2941 metaslab_group_increment_qdepth(mg
, allocator
);
2945 metaslab_group_alloc_verify(spa_t
*spa
, const blkptr_t
*bp
, void *tag
,
2949 const dva_t
*dva
= bp
->blk_dva
;
2950 int ndvas
= BP_GET_NDVAS(bp
);
2952 for (int d
= 0; d
< ndvas
; d
++) {
2953 uint64_t vdev
= DVA_GET_VDEV(&dva
[d
]);
2954 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2955 VERIFY(refcount_not_held(&mg
->mg_alloc_queue_depth
[allocator
],
2962 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
, uint64_t txg
)
2965 range_tree_t
*rt
= msp
->ms_allocatable
;
2966 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
2968 VERIFY(!msp
->ms_condensing
);
2970 start
= mc
->mc_ops
->msop_alloc(msp
, size
);
2971 if (start
!= -1ULL) {
2972 metaslab_group_t
*mg
= msp
->ms_group
;
2973 vdev_t
*vd
= mg
->mg_vd
;
2975 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
2976 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2977 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
2978 range_tree_remove(rt
, start
, size
);
2980 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
2981 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2983 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
], start
, size
);
2985 /* Track the last successful allocation */
2986 msp
->ms_alloc_txg
= txg
;
2987 metaslab_verify_space(msp
, txg
);
2991 * Now that we've attempted the allocation we need to update the
2992 * metaslab's maximum block size since it may have changed.
2994 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
2999 * Find the metaslab with the highest weight that is less than what we've
3000 * already tried. In the common case, this means that we will examine each
3001 * metaslab at most once. Note that concurrent callers could reorder metaslabs
3002 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
3003 * activated by another thread, and we fail to allocate from the metaslab we
3004 * have selected, we may not try the newly-activated metaslab, and instead
3005 * activate another metaslab. This is not optimal, but generally does not cause
3006 * any problems (a possible exception being if every metaslab is completely full
3007 * except for the the newly-activated metaslab which we fail to examine).
3010 find_valid_metaslab(metaslab_group_t
*mg
, uint64_t activation_weight
,
3011 dva_t
*dva
, int d
, boolean_t want_unique
, uint64_t asize
, int allocator
,
3012 zio_alloc_list_t
*zal
, metaslab_t
*search
, boolean_t
*was_active
)
3015 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
3016 metaslab_t
*msp
= avl_find(t
, search
, &idx
);
3018 msp
= avl_nearest(t
, idx
, AVL_AFTER
);
3020 for (; msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
3022 if (!metaslab_should_allocate(msp
, asize
)) {
3023 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3024 TRACE_TOO_SMALL
, allocator
);
3029 * If the selected metaslab is condensing, skip it.
3031 if (msp
->ms_condensing
)
3034 *was_active
= msp
->ms_allocator
!= -1;
3036 * If we're activating as primary, this is our first allocation
3037 * from this disk, so we don't need to check how close we are.
3038 * If the metaslab under consideration was already active,
3039 * we're getting desperate enough to steal another allocator's
3040 * metaslab, so we still don't care about distances.
3042 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
|| *was_active
)
3045 for (i
= 0; i
< d
; i
++) {
3047 !metaslab_is_unique(msp
, &dva
[i
]))
3048 break; /* try another metaslab */
3055 search
->ms_weight
= msp
->ms_weight
;
3056 search
->ms_start
= msp
->ms_start
+ 1;
3057 search
->ms_allocator
= msp
->ms_allocator
;
3058 search
->ms_primary
= msp
->ms_primary
;
3065 metaslab_group_alloc_normal(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
3066 uint64_t asize
, uint64_t txg
, boolean_t want_unique
, dva_t
*dva
,
3067 int d
, int allocator
)
3069 metaslab_t
*msp
= NULL
;
3070 uint64_t offset
= -1ULL;
3071 uint64_t activation_weight
;
3073 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
3074 for (int i
= 0; i
< d
; i
++) {
3075 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
&&
3076 DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
3077 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
3078 } else if (activation_weight
== METASLAB_WEIGHT_SECONDARY
&&
3079 DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
3080 activation_weight
= METASLAB_WEIGHT_CLAIM
;
3086 * If we don't have enough metaslabs active to fill the entire array, we
3087 * just use the 0th slot.
3089 if (mg
->mg_ms_ready
< mg
->mg_allocators
* 3)
3092 ASSERT3U(mg
->mg_vd
->vdev_ms_count
, >=, 2);
3094 metaslab_t
*search
= kmem_alloc(sizeof (*search
), KM_SLEEP
);
3095 search
->ms_weight
= UINT64_MAX
;
3096 search
->ms_start
= 0;
3098 * At the end of the metaslab tree are the already-active metaslabs,
3099 * first the primaries, then the secondaries. When we resume searching
3100 * through the tree, we need to consider ms_allocator and ms_primary so
3101 * we start in the location right after where we left off, and don't
3102 * accidentally loop forever considering the same metaslabs.
3104 search
->ms_allocator
= -1;
3105 search
->ms_primary
= B_TRUE
;
3107 boolean_t was_active
= B_FALSE
;
3109 mutex_enter(&mg
->mg_lock
);
3111 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
&&
3112 mg
->mg_primaries
[allocator
] != NULL
) {
3113 msp
= mg
->mg_primaries
[allocator
];
3114 was_active
= B_TRUE
;
3115 } else if (activation_weight
== METASLAB_WEIGHT_SECONDARY
&&
3116 mg
->mg_secondaries
[allocator
] != NULL
) {
3117 msp
= mg
->mg_secondaries
[allocator
];
3118 was_active
= B_TRUE
;
3120 msp
= find_valid_metaslab(mg
, activation_weight
, dva
, d
,
3121 want_unique
, asize
, allocator
, zal
, search
,
3125 mutex_exit(&mg
->mg_lock
);
3127 kmem_free(search
, sizeof (*search
));
3131 mutex_enter(&msp
->ms_lock
);
3133 * Ensure that the metaslab we have selected is still
3134 * capable of handling our request. It's possible that
3135 * another thread may have changed the weight while we
3136 * were blocked on the metaslab lock. We check the
3137 * active status first to see if we need to reselect
3140 if (was_active
&& !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
)) {
3141 mutex_exit(&msp
->ms_lock
);
3146 * If the metaslab is freshly activated for an allocator that
3147 * isn't the one we're allocating from, or if it's a primary and
3148 * we're seeking a secondary (or vice versa), we go back and
3149 * select a new metaslab.
3151 if (!was_active
&& (msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
3152 (msp
->ms_allocator
!= -1) &&
3153 (msp
->ms_allocator
!= allocator
|| ((activation_weight
==
3154 METASLAB_WEIGHT_PRIMARY
) != msp
->ms_primary
))) {
3155 mutex_exit(&msp
->ms_lock
);
3159 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
&&
3160 activation_weight
!= METASLAB_WEIGHT_CLAIM
) {
3161 metaslab_passivate(msp
, msp
->ms_weight
&
3162 ~METASLAB_WEIGHT_CLAIM
);
3163 mutex_exit(&msp
->ms_lock
);
3167 if (metaslab_activate(msp
, allocator
, activation_weight
) != 0) {
3168 mutex_exit(&msp
->ms_lock
);
3172 msp
->ms_selected_txg
= txg
;
3175 * Now that we have the lock, recheck to see if we should
3176 * continue to use this metaslab for this allocation. The
3177 * the metaslab is now loaded so metaslab_should_allocate() can
3178 * accurately determine if the allocation attempt should
3181 if (!metaslab_should_allocate(msp
, asize
)) {
3182 /* Passivate this metaslab and select a new one. */
3183 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3184 TRACE_TOO_SMALL
, allocator
);
3190 * If this metaslab is currently condensing then pick again as
3191 * we can't manipulate this metaslab until it's committed
3194 if (msp
->ms_condensing
) {
3195 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3196 TRACE_CONDENSING
, allocator
);
3197 metaslab_passivate(msp
, msp
->ms_weight
&
3198 ~METASLAB_ACTIVE_MASK
);
3199 mutex_exit(&msp
->ms_lock
);
3203 offset
= metaslab_block_alloc(msp
, asize
, txg
);
3204 metaslab_trace_add(zal
, mg
, msp
, asize
, d
, offset
, allocator
);
3206 if (offset
!= -1ULL) {
3207 /* Proactively passivate the metaslab, if needed */
3208 metaslab_segment_may_passivate(msp
);
3212 ASSERT(msp
->ms_loaded
);
3215 * We were unable to allocate from this metaslab so determine
3216 * a new weight for this metaslab. Now that we have loaded
3217 * the metaslab we can provide a better hint to the metaslab
3220 * For space-based metaslabs, we use the maximum block size.
3221 * This information is only available when the metaslab
3222 * is loaded and is more accurate than the generic free
3223 * space weight that was calculated by metaslab_weight().
3224 * This information allows us to quickly compare the maximum
3225 * available allocation in the metaslab to the allocation
3226 * size being requested.
3228 * For segment-based metaslabs, determine the new weight
3229 * based on the highest bucket in the range tree. We
3230 * explicitly use the loaded segment weight (i.e. the range
3231 * tree histogram) since it contains the space that is
3232 * currently available for allocation and is accurate
3233 * even within a sync pass.
3235 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
3236 uint64_t weight
= metaslab_block_maxsize(msp
);
3237 WEIGHT_SET_SPACEBASED(weight
);
3238 metaslab_passivate(msp
, weight
);
3240 metaslab_passivate(msp
,
3241 metaslab_weight_from_range_tree(msp
));
3245 * We have just failed an allocation attempt, check
3246 * that metaslab_should_allocate() agrees. Otherwise,
3247 * we may end up in an infinite loop retrying the same
3250 ASSERT(!metaslab_should_allocate(msp
, asize
));
3252 mutex_exit(&msp
->ms_lock
);
3254 mutex_exit(&msp
->ms_lock
);
3255 kmem_free(search
, sizeof (*search
));
3260 metaslab_group_alloc(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
3261 uint64_t asize
, uint64_t txg
, boolean_t want_unique
, dva_t
*dva
,
3262 int d
, int allocator
)
3265 ASSERT(mg
->mg_initialized
);
3267 offset
= metaslab_group_alloc_normal(mg
, zal
, asize
, txg
, want_unique
,
3270 mutex_enter(&mg
->mg_lock
);
3271 if (offset
== -1ULL) {
3272 mg
->mg_failed_allocations
++;
3273 metaslab_trace_add(zal
, mg
, NULL
, asize
, d
,
3274 TRACE_GROUP_FAILURE
, allocator
);
3275 if (asize
== SPA_GANGBLOCKSIZE
) {
3277 * This metaslab group was unable to allocate
3278 * the minimum gang block size so it must be out of
3279 * space. We must notify the allocation throttle
3280 * to start skipping allocation attempts to this
3281 * metaslab group until more space becomes available.
3282 * Note: this failure cannot be caused by the
3283 * allocation throttle since the allocation throttle
3284 * is only responsible for skipping devices and
3285 * not failing block allocations.
3287 mg
->mg_no_free_space
= B_TRUE
;
3290 mg
->mg_allocations
++;
3291 mutex_exit(&mg
->mg_lock
);
3296 * Allocate a block for the specified i/o.
3299 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
3300 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
,
3301 zio_alloc_list_t
*zal
, int allocator
)
3303 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
3305 boolean_t try_hard
= B_FALSE
;
3307 ASSERT(!DVA_IS_VALID(&dva
[d
]));
3310 * For testing, make some blocks above a certain size be gang blocks.
3311 * This will also test spilling from special to normal.
3313 if (psize
>= metaslab_force_ganging
&& (ddi_get_lbolt() & 3) == 0) {
3314 metaslab_trace_add(zal
, NULL
, NULL
, psize
, d
, TRACE_FORCE_GANG
,
3316 return (SET_ERROR(ENOSPC
));
3320 * Start at the rotor and loop through all mgs until we find something.
3321 * Note that there's no locking on mc_rotor or mc_aliquot because
3322 * nothing actually breaks if we miss a few updates -- we just won't
3323 * allocate quite as evenly. It all balances out over time.
3325 * If we are doing ditto or log blocks, try to spread them across
3326 * consecutive vdevs. If we're forced to reuse a vdev before we've
3327 * allocated all of our ditto blocks, then try and spread them out on
3328 * that vdev as much as possible. If it turns out to not be possible,
3329 * gradually lower our standards until anything becomes acceptable.
3330 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3331 * gives us hope of containing our fault domains to something we're
3332 * able to reason about. Otherwise, any two top-level vdev failures
3333 * will guarantee the loss of data. With consecutive allocation,
3334 * only two adjacent top-level vdev failures will result in data loss.
3336 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3337 * ourselves on the same vdev as our gang block header. That
3338 * way, we can hope for locality in vdev_cache, plus it makes our
3339 * fault domains something tractable.
3342 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
3345 * It's possible the vdev we're using as the hint no
3346 * longer exists or its mg has been closed (e.g. by
3347 * device removal). Consult the rotor when
3350 if (vd
!= NULL
&& vd
->vdev_mg
!= NULL
) {
3353 if (flags
& METASLAB_HINTBP_AVOID
&&
3354 mg
->mg_next
!= NULL
)
3359 } else if (d
!= 0) {
3360 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
3361 mg
= vd
->vdev_mg
->mg_next
;
3362 } else if (flags
& METASLAB_FASTWRITE
) {
3363 mg
= fast_mg
= mc
->mc_rotor
;
3366 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
3367 mg
->mg_vd
->vdev_pending_fastwrite
)
3369 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
3372 ASSERT(mc
->mc_rotor
!= NULL
);
3377 * If the hint put us into the wrong metaslab class, or into a
3378 * metaslab group that has been passivated, just follow the rotor.
3380 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
3386 boolean_t allocatable
;
3388 ASSERT(mg
->mg_activation_count
== 1);
3392 * Don't allocate from faulted devices.
3395 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
3396 allocatable
= vdev_allocatable(vd
);
3397 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
3399 allocatable
= vdev_allocatable(vd
);
3403 * Determine if the selected metaslab group is eligible
3404 * for allocations. If we're ganging then don't allow
3405 * this metaslab group to skip allocations since that would
3406 * inadvertently return ENOSPC and suspend the pool
3407 * even though space is still available.
3409 if (allocatable
&& !GANG_ALLOCATION(flags
) && !try_hard
) {
3410 allocatable
= metaslab_group_allocatable(mg
, rotor
,
3411 psize
, allocator
, d
);
3415 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3416 TRACE_NOT_ALLOCATABLE
, allocator
);
3420 ASSERT(mg
->mg_initialized
);
3423 * Avoid writing single-copy data to a failing,
3424 * non-redundant vdev, unless we've already tried all
3427 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
3428 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
3429 d
== 0 && !try_hard
&& vd
->vdev_children
== 0) {
3430 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3431 TRACE_VDEV_ERROR
, allocator
);
3435 ASSERT(mg
->mg_class
== mc
);
3437 uint64_t asize
= vdev_psize_to_asize(vd
, psize
);
3438 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
3441 * If we don't need to try hard, then require that the
3442 * block be on an different metaslab from any other DVAs
3443 * in this BP (unique=true). If we are trying hard, then
3444 * allow any metaslab to be used (unique=false).
3446 uint64_t offset
= metaslab_group_alloc(mg
, zal
, asize
, txg
,
3447 !try_hard
, dva
, d
, allocator
);
3449 if (offset
!= -1ULL) {
3451 * If we've just selected this metaslab group,
3452 * figure out whether the corresponding vdev is
3453 * over- or under-used relative to the pool,
3454 * and set an allocation bias to even it out.
3456 * Bias is also used to compensate for unequally
3457 * sized vdevs so that space is allocated fairly.
3459 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
3460 vdev_stat_t
*vs
= &vd
->vdev_stat
;
3461 int64_t vs_free
= vs
->vs_space
- vs
->vs_alloc
;
3462 int64_t mc_free
= mc
->mc_space
- mc
->mc_alloc
;
3466 * Calculate how much more or less we should
3467 * try to allocate from this device during
3468 * this iteration around the rotor.
3470 * This basically introduces a zero-centered
3471 * bias towards the devices with the most
3472 * free space, while compensating for vdev
3476 * vdev V1 = 16M/128M
3477 * vdev V2 = 16M/128M
3478 * ratio(V1) = 100% ratio(V2) = 100%
3480 * vdev V1 = 16M/128M
3481 * vdev V2 = 64M/128M
3482 * ratio(V1) = 127% ratio(V2) = 72%
3484 * vdev V1 = 16M/128M
3485 * vdev V2 = 64M/512M
3486 * ratio(V1) = 40% ratio(V2) = 160%
3488 ratio
= (vs_free
* mc
->mc_alloc_groups
* 100) /
3490 mg
->mg_bias
= ((ratio
- 100) *
3491 (int64_t)mg
->mg_aliquot
) / 100;
3492 } else if (!metaslab_bias_enabled
) {
3496 if ((flags
& METASLAB_FASTWRITE
) ||
3497 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
3498 mg
->mg_aliquot
+ mg
->mg_bias
) {
3499 mc
->mc_rotor
= mg
->mg_next
;
3503 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
3504 DVA_SET_OFFSET(&dva
[d
], offset
);
3505 DVA_SET_GANG(&dva
[d
],
3506 ((flags
& METASLAB_GANG_HEADER
) ? 1 : 0));
3507 DVA_SET_ASIZE(&dva
[d
], asize
);
3509 if (flags
& METASLAB_FASTWRITE
) {
3510 atomic_add_64(&vd
->vdev_pending_fastwrite
,
3517 mc
->mc_rotor
= mg
->mg_next
;
3519 } while ((mg
= mg
->mg_next
) != rotor
);
3522 * If we haven't tried hard, do so now.
3529 bzero(&dva
[d
], sizeof (dva_t
));
3531 metaslab_trace_add(zal
, rotor
, NULL
, psize
, d
, TRACE_ENOSPC
, allocator
);
3532 return (SET_ERROR(ENOSPC
));
3536 metaslab_free_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t asize
,
3537 boolean_t checkpoint
)
3540 spa_t
*spa
= vd
->vdev_spa
;
3542 ASSERT(vdev_is_concrete(vd
));
3543 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3544 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
3546 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3548 VERIFY(!msp
->ms_condensing
);
3549 VERIFY3U(offset
, >=, msp
->ms_start
);
3550 VERIFY3U(offset
+ asize
, <=, msp
->ms_start
+ msp
->ms_size
);
3551 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3552 VERIFY0(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
));
3554 metaslab_check_free_impl(vd
, offset
, asize
);
3556 mutex_enter(&msp
->ms_lock
);
3557 if (range_tree_is_empty(msp
->ms_freeing
) &&
3558 range_tree_is_empty(msp
->ms_checkpointing
)) {
3559 vdev_dirty(vd
, VDD_METASLAB
, msp
, spa_syncing_txg(spa
));
3563 ASSERT(spa_has_checkpoint(spa
));
3564 range_tree_add(msp
->ms_checkpointing
, offset
, asize
);
3566 range_tree_add(msp
->ms_freeing
, offset
, asize
);
3568 mutex_exit(&msp
->ms_lock
);
3573 metaslab_free_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3574 uint64_t size
, void *arg
)
3576 boolean_t
*checkpoint
= arg
;
3578 ASSERT3P(checkpoint
, !=, NULL
);
3580 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
)
3581 vdev_indirect_mark_obsolete(vd
, offset
, size
);
3583 metaslab_free_impl(vd
, offset
, size
, *checkpoint
);
3587 metaslab_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
3588 boolean_t checkpoint
)
3590 spa_t
*spa
= vd
->vdev_spa
;
3592 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3594 if (spa_syncing_txg(spa
) > spa_freeze_txg(spa
))
3597 if (spa
->spa_vdev_removal
!= NULL
&&
3598 spa
->spa_vdev_removal
->svr_vdev_id
== vd
->vdev_id
&&
3599 vdev_is_concrete(vd
)) {
3601 * Note: we check if the vdev is concrete because when
3602 * we complete the removal, we first change the vdev to be
3603 * an indirect vdev (in open context), and then (in syncing
3604 * context) clear spa_vdev_removal.
3606 free_from_removing_vdev(vd
, offset
, size
);
3607 } else if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
3608 vdev_indirect_mark_obsolete(vd
, offset
, size
);
3609 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
3610 metaslab_free_impl_cb
, &checkpoint
);
3612 metaslab_free_concrete(vd
, offset
, size
, checkpoint
);
3616 typedef struct remap_blkptr_cb_arg
{
3618 spa_remap_cb_t rbca_cb
;
3619 vdev_t
*rbca_remap_vd
;
3620 uint64_t rbca_remap_offset
;
3622 } remap_blkptr_cb_arg_t
;
3625 remap_blkptr_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3626 uint64_t size
, void *arg
)
3628 remap_blkptr_cb_arg_t
*rbca
= arg
;
3629 blkptr_t
*bp
= rbca
->rbca_bp
;
3631 /* We can not remap split blocks. */
3632 if (size
!= DVA_GET_ASIZE(&bp
->blk_dva
[0]))
3634 ASSERT0(inner_offset
);
3636 if (rbca
->rbca_cb
!= NULL
) {
3638 * At this point we know that we are not handling split
3639 * blocks and we invoke the callback on the previous
3640 * vdev which must be indirect.
3642 ASSERT3P(rbca
->rbca_remap_vd
->vdev_ops
, ==, &vdev_indirect_ops
);
3644 rbca
->rbca_cb(rbca
->rbca_remap_vd
->vdev_id
,
3645 rbca
->rbca_remap_offset
, size
, rbca
->rbca_cb_arg
);
3647 /* set up remap_blkptr_cb_arg for the next call */
3648 rbca
->rbca_remap_vd
= vd
;
3649 rbca
->rbca_remap_offset
= offset
;
3653 * The phys birth time is that of dva[0]. This ensures that we know
3654 * when each dva was written, so that resilver can determine which
3655 * blocks need to be scrubbed (i.e. those written during the time
3656 * the vdev was offline). It also ensures that the key used in
3657 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
3658 * we didn't change the phys_birth, a lookup in the ARC for a
3659 * remapped BP could find the data that was previously stored at
3660 * this vdev + offset.
3662 vdev_t
*oldvd
= vdev_lookup_top(vd
->vdev_spa
,
3663 DVA_GET_VDEV(&bp
->blk_dva
[0]));
3664 vdev_indirect_births_t
*vib
= oldvd
->vdev_indirect_births
;
3665 bp
->blk_phys_birth
= vdev_indirect_births_physbirth(vib
,
3666 DVA_GET_OFFSET(&bp
->blk_dva
[0]), DVA_GET_ASIZE(&bp
->blk_dva
[0]));
3668 DVA_SET_VDEV(&bp
->blk_dva
[0], vd
->vdev_id
);
3669 DVA_SET_OFFSET(&bp
->blk_dva
[0], offset
);
3673 * If the block pointer contains any indirect DVAs, modify them to refer to
3674 * concrete DVAs. Note that this will sometimes not be possible, leaving
3675 * the indirect DVA in place. This happens if the indirect DVA spans multiple
3676 * segments in the mapping (i.e. it is a "split block").
3678 * If the BP was remapped, calls the callback on the original dva (note the
3679 * callback can be called multiple times if the original indirect DVA refers
3680 * to another indirect DVA, etc).
3682 * Returns TRUE if the BP was remapped.
3685 spa_remap_blkptr(spa_t
*spa
, blkptr_t
*bp
, spa_remap_cb_t callback
, void *arg
)
3687 remap_blkptr_cb_arg_t rbca
;
3689 if (!zfs_remap_blkptr_enable
)
3692 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
))
3696 * Dedup BP's can not be remapped, because ddt_phys_select() depends
3697 * on DVA[0] being the same in the BP as in the DDT (dedup table).
3699 if (BP_GET_DEDUP(bp
))
3703 * Gang blocks can not be remapped, because
3704 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
3705 * the BP used to read the gang block header (GBH) being the same
3706 * as the DVA[0] that we allocated for the GBH.
3712 * Embedded BP's have no DVA to remap.
3714 if (BP_GET_NDVAS(bp
) < 1)
3718 * Note: we only remap dva[0]. If we remapped other dvas, we
3719 * would no longer know what their phys birth txg is.
3721 dva_t
*dva
= &bp
->blk_dva
[0];
3723 uint64_t offset
= DVA_GET_OFFSET(dva
);
3724 uint64_t size
= DVA_GET_ASIZE(dva
);
3725 vdev_t
*vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
3727 if (vd
->vdev_ops
->vdev_op_remap
== NULL
)
3731 rbca
.rbca_cb
= callback
;
3732 rbca
.rbca_remap_vd
= vd
;
3733 rbca
.rbca_remap_offset
= offset
;
3734 rbca
.rbca_cb_arg
= arg
;
3737 * remap_blkptr_cb() will be called in order for each level of
3738 * indirection, until a concrete vdev is reached or a split block is
3739 * encountered. old_vd and old_offset are updated within the callback
3740 * as we go from the one indirect vdev to the next one (either concrete
3741 * or indirect again) in that order.
3743 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
, remap_blkptr_cb
, &rbca
);
3745 /* Check if the DVA wasn't remapped because it is a split block */
3746 if (DVA_GET_VDEV(&rbca
.rbca_bp
->blk_dva
[0]) == vd
->vdev_id
)
3753 * Undo the allocation of a DVA which happened in the given transaction group.
3756 metaslab_unalloc_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
3760 uint64_t vdev
= DVA_GET_VDEV(dva
);
3761 uint64_t offset
= DVA_GET_OFFSET(dva
);
3762 uint64_t size
= DVA_GET_ASIZE(dva
);
3764 ASSERT(DVA_IS_VALID(dva
));
3765 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3767 if (txg
> spa_freeze_txg(spa
))
3770 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
3771 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
3772 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
3773 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
3774 (u_longlong_t
)size
);
3778 ASSERT(!vd
->vdev_removing
);
3779 ASSERT(vdev_is_concrete(vd
));
3780 ASSERT0(vd
->vdev_indirect_config
.vic_mapping_object
);
3781 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
3783 if (DVA_GET_GANG(dva
))
3784 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3786 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3788 mutex_enter(&msp
->ms_lock
);
3789 range_tree_remove(msp
->ms_allocating
[txg
& TXG_MASK
],
3792 VERIFY(!msp
->ms_condensing
);
3793 VERIFY3U(offset
, >=, msp
->ms_start
);
3794 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
3795 VERIFY3U(range_tree_space(msp
->ms_allocatable
) + size
, <=,
3797 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3798 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3799 range_tree_add(msp
->ms_allocatable
, offset
, size
);
3800 mutex_exit(&msp
->ms_lock
);
3804 * Free the block represented by the given DVA.
3807 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, boolean_t checkpoint
)
3809 uint64_t vdev
= DVA_GET_VDEV(dva
);
3810 uint64_t offset
= DVA_GET_OFFSET(dva
);
3811 uint64_t size
= DVA_GET_ASIZE(dva
);
3812 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
3814 ASSERT(DVA_IS_VALID(dva
));
3815 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3817 if (DVA_GET_GANG(dva
)) {
3818 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3821 metaslab_free_impl(vd
, offset
, size
, checkpoint
);
3825 * Reserve some allocation slots. The reservation system must be called
3826 * before we call into the allocator. If there aren't any available slots
3827 * then the I/O will be throttled until an I/O completes and its slots are
3828 * freed up. The function returns true if it was successful in placing
3832 metaslab_class_throttle_reserve(metaslab_class_t
*mc
, int slots
, int allocator
,
3833 zio_t
*zio
, int flags
)
3835 uint64_t available_slots
= 0;
3836 boolean_t slot_reserved
= B_FALSE
;
3837 uint64_t max
= mc
->mc_alloc_max_slots
[allocator
];
3839 ASSERT(mc
->mc_alloc_throttle_enabled
);
3840 mutex_enter(&mc
->mc_lock
);
3842 uint64_t reserved_slots
=
3843 refcount_count(&mc
->mc_alloc_slots
[allocator
]);
3844 if (reserved_slots
< max
)
3845 available_slots
= max
- reserved_slots
;
3847 if (slots
<= available_slots
|| GANG_ALLOCATION(flags
) ||
3848 flags
& METASLAB_MUST_RESERVE
) {
3850 * We reserve the slots individually so that we can unreserve
3851 * them individually when an I/O completes.
3853 for (int d
= 0; d
< slots
; d
++) {
3855 zfs_refcount_add(&mc
->mc_alloc_slots
[allocator
],
3858 zio
->io_flags
|= ZIO_FLAG_IO_ALLOCATING
;
3859 slot_reserved
= B_TRUE
;
3862 mutex_exit(&mc
->mc_lock
);
3863 return (slot_reserved
);
3867 metaslab_class_throttle_unreserve(metaslab_class_t
*mc
, int slots
,
3868 int allocator
, zio_t
*zio
)
3870 ASSERT(mc
->mc_alloc_throttle_enabled
);
3871 mutex_enter(&mc
->mc_lock
);
3872 for (int d
= 0; d
< slots
; d
++) {
3873 (void) refcount_remove(&mc
->mc_alloc_slots
[allocator
],
3876 mutex_exit(&mc
->mc_lock
);
3880 metaslab_claim_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
3884 spa_t
*spa
= vd
->vdev_spa
;
3887 if (offset
>> vd
->vdev_ms_shift
>= vd
->vdev_ms_count
)
3890 ASSERT3P(vd
->vdev_ms
, !=, NULL
);
3891 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3893 mutex_enter(&msp
->ms_lock
);
3895 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
3896 error
= metaslab_activate(msp
, 0, METASLAB_WEIGHT_CLAIM
);
3899 !range_tree_contains(msp
->ms_allocatable
, offset
, size
))
3900 error
= SET_ERROR(ENOENT
);
3902 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
3903 mutex_exit(&msp
->ms_lock
);
3907 VERIFY(!msp
->ms_condensing
);
3908 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3909 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3910 VERIFY3U(range_tree_space(msp
->ms_allocatable
) - size
, <=,
3912 range_tree_remove(msp
->ms_allocatable
, offset
, size
);
3914 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
3915 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
3916 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
3917 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
],
3921 mutex_exit(&msp
->ms_lock
);
3926 typedef struct metaslab_claim_cb_arg_t
{
3929 } metaslab_claim_cb_arg_t
;
3933 metaslab_claim_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3934 uint64_t size
, void *arg
)
3936 metaslab_claim_cb_arg_t
*mcca_arg
= arg
;
3938 if (mcca_arg
->mcca_error
== 0) {
3939 mcca_arg
->mcca_error
= metaslab_claim_concrete(vd
, offset
,
3940 size
, mcca_arg
->mcca_txg
);
3945 metaslab_claim_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
, uint64_t txg
)
3947 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
3948 metaslab_claim_cb_arg_t arg
;
3951 * Only zdb(1M) can claim on indirect vdevs. This is used
3952 * to detect leaks of mapped space (that are not accounted
3953 * for in the obsolete counts, spacemap, or bpobj).
3955 ASSERT(!spa_writeable(vd
->vdev_spa
));
3959 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
3960 metaslab_claim_impl_cb
, &arg
);
3962 if (arg
.mcca_error
== 0) {
3963 arg
.mcca_error
= metaslab_claim_concrete(vd
,
3966 return (arg
.mcca_error
);
3968 return (metaslab_claim_concrete(vd
, offset
, size
, txg
));
3973 * Intent log support: upon opening the pool after a crash, notify the SPA
3974 * of blocks that the intent log has allocated for immediate write, but
3975 * which are still considered free by the SPA because the last transaction
3976 * group didn't commit yet.
3979 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
3981 uint64_t vdev
= DVA_GET_VDEV(dva
);
3982 uint64_t offset
= DVA_GET_OFFSET(dva
);
3983 uint64_t size
= DVA_GET_ASIZE(dva
);
3986 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
) {
3987 return (SET_ERROR(ENXIO
));
3990 ASSERT(DVA_IS_VALID(dva
));
3992 if (DVA_GET_GANG(dva
))
3993 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3995 return (metaslab_claim_impl(vd
, offset
, size
, txg
));
3999 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
4000 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
,
4001 zio_alloc_list_t
*zal
, zio_t
*zio
, int allocator
)
4003 dva_t
*dva
= bp
->blk_dva
;
4004 dva_t
*hintdva
= hintbp
->blk_dva
;
4007 ASSERT(bp
->blk_birth
== 0);
4008 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
4010 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
4012 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
4013 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4014 return (SET_ERROR(ENOSPC
));
4017 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
4018 ASSERT(BP_GET_NDVAS(bp
) == 0);
4019 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
4020 ASSERT3P(zal
, !=, NULL
);
4022 for (int d
= 0; d
< ndvas
; d
++) {
4023 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
4024 txg
, flags
, zal
, allocator
);
4026 for (d
--; d
>= 0; d
--) {
4027 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
4028 metaslab_group_alloc_decrement(spa
,
4029 DVA_GET_VDEV(&dva
[d
]), zio
, flags
,
4030 allocator
, B_FALSE
);
4031 bzero(&dva
[d
], sizeof (dva_t
));
4033 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4037 * Update the metaslab group's queue depth
4038 * based on the newly allocated dva.
4040 metaslab_group_alloc_increment(spa
,
4041 DVA_GET_VDEV(&dva
[d
]), zio
, flags
, allocator
);
4046 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
4048 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4050 BP_SET_BIRTH(bp
, txg
, 0);
4056 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
4058 const dva_t
*dva
= bp
->blk_dva
;
4059 int ndvas
= BP_GET_NDVAS(bp
);
4061 ASSERT(!BP_IS_HOLE(bp
));
4062 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
4065 * If we have a checkpoint for the pool we need to make sure that
4066 * the blocks that we free that are part of the checkpoint won't be
4067 * reused until the checkpoint is discarded or we revert to it.
4069 * The checkpoint flag is passed down the metaslab_free code path
4070 * and is set whenever we want to add a block to the checkpoint's
4071 * accounting. That is, we "checkpoint" blocks that existed at the
4072 * time the checkpoint was created and are therefore referenced by
4073 * the checkpointed uberblock.
4075 * Note that, we don't checkpoint any blocks if the current
4076 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
4077 * normally as they will be referenced by the checkpointed uberblock.
4079 boolean_t checkpoint
= B_FALSE
;
4080 if (bp
->blk_birth
<= spa
->spa_checkpoint_txg
&&
4081 spa_syncing_txg(spa
) > spa
->spa_checkpoint_txg
) {
4083 * At this point, if the block is part of the checkpoint
4084 * there is no way it was created in the current txg.
4087 ASSERT3U(spa_syncing_txg(spa
), ==, txg
);
4088 checkpoint
= B_TRUE
;
4091 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
4093 for (int d
= 0; d
< ndvas
; d
++) {
4095 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
4097 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
4098 metaslab_free_dva(spa
, &dva
[d
], checkpoint
);
4102 spa_config_exit(spa
, SCL_FREE
, FTAG
);
4106 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
4108 const dva_t
*dva
= bp
->blk_dva
;
4109 int ndvas
= BP_GET_NDVAS(bp
);
4112 ASSERT(!BP_IS_HOLE(bp
));
4116 * First do a dry run to make sure all DVAs are claimable,
4117 * so we don't have to unwind from partial failures below.
4119 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
4123 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
4125 for (int d
= 0; d
< ndvas
; d
++) {
4126 error
= metaslab_claim_dva(spa
, &dva
[d
], txg
);
4131 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4133 ASSERT(error
== 0 || txg
== 0);
4139 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
4141 const dva_t
*dva
= bp
->blk_dva
;
4142 int ndvas
= BP_GET_NDVAS(bp
);
4143 uint64_t psize
= BP_GET_PSIZE(bp
);
4147 ASSERT(!BP_IS_HOLE(bp
));
4148 ASSERT(!BP_IS_EMBEDDED(bp
));
4151 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4153 for (d
= 0; d
< ndvas
; d
++) {
4154 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
4156 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
4159 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4163 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
4165 const dva_t
*dva
= bp
->blk_dva
;
4166 int ndvas
= BP_GET_NDVAS(bp
);
4167 uint64_t psize
= BP_GET_PSIZE(bp
);
4171 ASSERT(!BP_IS_HOLE(bp
));
4172 ASSERT(!BP_IS_EMBEDDED(bp
));
4175 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4177 for (d
= 0; d
< ndvas
; d
++) {
4178 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
4180 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
4181 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
4184 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4189 metaslab_check_free_impl_cb(uint64_t inner
, vdev_t
*vd
, uint64_t offset
,
4190 uint64_t size
, void *arg
)
4192 if (vd
->vdev_ops
== &vdev_indirect_ops
)
4195 metaslab_check_free_impl(vd
, offset
, size
);
4199 metaslab_check_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
4202 ASSERTV(spa_t
*spa
= vd
->vdev_spa
);
4204 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
4207 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
4208 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
4209 metaslab_check_free_impl_cb
, NULL
);
4213 ASSERT(vdev_is_concrete(vd
));
4214 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
4215 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
4217 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
4219 mutex_enter(&msp
->ms_lock
);
4221 range_tree_verify(msp
->ms_allocatable
, offset
, size
);
4223 range_tree_verify(msp
->ms_freeing
, offset
, size
);
4224 range_tree_verify(msp
->ms_checkpointing
, offset
, size
);
4225 range_tree_verify(msp
->ms_freed
, offset
, size
);
4226 for (int j
= 0; j
< TXG_DEFER_SIZE
; j
++)
4227 range_tree_verify(msp
->ms_defer
[j
], offset
, size
);
4228 mutex_exit(&msp
->ms_lock
);
4232 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
4234 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
4237 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4238 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
4239 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
4240 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
4241 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
4242 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
4244 if (DVA_GET_GANG(&bp
->blk_dva
[i
]))
4245 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
4247 ASSERT3P(vd
, !=, NULL
);
4249 metaslab_check_free_impl(vd
, offset
, size
);
4251 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4254 #if defined(_KERNEL)
4256 module_param(metaslab_aliquot
, ulong
, 0644);
4257 MODULE_PARM_DESC(metaslab_aliquot
,
4258 "allocation granularity (a.k.a. stripe size)");
4260 module_param(metaslab_debug_load
, int, 0644);
4261 MODULE_PARM_DESC(metaslab_debug_load
,
4262 "load all metaslabs when pool is first opened");
4264 module_param(metaslab_debug_unload
, int, 0644);
4265 MODULE_PARM_DESC(metaslab_debug_unload
,
4266 "prevent metaslabs from being unloaded");
4268 module_param(metaslab_preload_enabled
, int, 0644);
4269 MODULE_PARM_DESC(metaslab_preload_enabled
,
4270 "preload potential metaslabs during reassessment");
4272 module_param(zfs_mg_noalloc_threshold
, int, 0644);
4273 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
4274 "percentage of free space for metaslab group to allow allocation");
4276 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
4277 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
4278 "fragmentation for metaslab group to allow allocation");
4280 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
4281 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
4282 "fragmentation for metaslab to allow allocation");
4284 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
4285 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
4286 "use the fragmentation metric to prefer less fragmented metaslabs");
4288 module_param(metaslab_lba_weighting_enabled
, int, 0644);
4289 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
4290 "prefer metaslabs with lower LBAs");
4292 module_param(metaslab_bias_enabled
, int, 0644);
4293 MODULE_PARM_DESC(metaslab_bias_enabled
,
4294 "enable metaslab group biasing");
4296 module_param(zfs_metaslab_segment_weight_enabled
, int, 0644);
4297 MODULE_PARM_DESC(zfs_metaslab_segment_weight_enabled
,
4298 "enable segment-based metaslab selection");
4300 module_param(zfs_metaslab_switch_threshold
, int, 0644);
4301 MODULE_PARM_DESC(zfs_metaslab_switch_threshold
,
4302 "segment-based metaslab selection maximum buckets before switching");
4304 module_param(metaslab_force_ganging
, ulong
, 0644);
4305 MODULE_PARM_DESC(metaslab_force_ganging
,
4306 "blocks larger than this size are forced to be gang blocks");