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 https://opensource.org/licenses/CDDL-1.0.
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, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2015, Nexenta Systems, Inc. All rights reserved.
26 * Copyright (c) 2017, Intel Corporation.
29 #include <sys/zfs_context.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/space_map.h>
33 #include <sys/metaslab_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/vdev_draid.h>
37 #include <sys/spa_impl.h>
38 #include <sys/zfeature.h>
39 #include <sys/vdev_indirect_mapping.h>
41 #include <sys/btree.h>
43 #define GANG_ALLOCATION(flags) \
44 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
47 * Metaslab granularity, in bytes. This is roughly similar to what would be
48 * referred to as the "stripe size" in traditional RAID arrays. In normal
49 * operation, we will try to write this amount of data to each disk before
50 * moving on to the next top-level vdev.
52 static uint64_t metaslab_aliquot = 1024 * 1024;
55 * For testing, make some blocks above a certain size be gang blocks.
57 uint64_t metaslab_force_ganging = SPA_MAXBLOCKSIZE + 1;
60 * Of blocks of size >= metaslab_force_ganging, actually gang them this often.
62 uint_t metaslab_force_ganging_pct = 3;
65 * In pools where the log space map feature is not enabled we touch
66 * multiple metaslabs (and their respective space maps) with each
67 * transaction group. Thus, we benefit from having a small space map
68 * block size since it allows us to issue more I/O operations scattered
69 * around the disk. So a sane default for the space map block size
72 int zfs_metaslab_sm_blksz_no_log = (1 << 14);
75 * When the log space map feature is enabled, we accumulate a lot of
76 * changes per metaslab that are flushed once in a while so we benefit
77 * from a bigger block size like 128K for the metaslab space maps.
79 int zfs_metaslab_sm_blksz_with_log = (1 << 17);
82 * The in-core space map representation is more compact than its on-disk form.
83 * The zfs_condense_pct determines how much more compact the in-core
84 * space map representation must be before we compact it on-disk.
85 * Values should be greater than or equal to 100.
87 uint_t zfs_condense_pct = 200;
90 * Condensing a metaslab is not guaranteed to actually reduce the amount of
91 * space used on disk. In particular, a space map uses data in increments of
92 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
93 * same number of blocks after condensing. Since the goal of condensing is to
94 * reduce the number of IOPs required to read the space map, we only want to
95 * condense when we can be sure we will reduce the number of blocks used by the
96 * space map. Unfortunately, we cannot precisely compute whether or not this is
97 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
98 * we apply the following heuristic: do not condense a spacemap unless the
99 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
102 static const int zfs_metaslab_condense_block_threshold = 4;
105 * The zfs_mg_noalloc_threshold defines which metaslab groups should
106 * be eligible for allocation. The value is defined as a percentage of
107 * free space. Metaslab groups that have more free space than
108 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
109 * a metaslab group's free space is less than or equal to the
110 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
111 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
112 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
113 * groups are allowed to accept allocations. Gang blocks are always
114 * eligible to allocate on any metaslab group. The default value of 0 means
115 * no metaslab group will be excluded based on this criterion.
117 static uint_t zfs_mg_noalloc_threshold = 0;
120 * Metaslab groups are considered eligible for allocations if their
121 * fragmentation metric (measured as a percentage) is less than or
122 * equal to zfs_mg_fragmentation_threshold. If a metaslab group
123 * exceeds this threshold then it will be skipped unless all metaslab
124 * groups within the metaslab class have also crossed this threshold.
126 * This tunable was introduced to avoid edge cases where we continue
127 * allocating from very fragmented disks in our pool while other, less
128 * fragmented disks, exists. On the other hand, if all disks in the
129 * pool are uniformly approaching the threshold, the threshold can
130 * be a speed bump in performance, where we keep switching the disks
131 * that we allocate from (e.g. we allocate some segments from disk A
132 * making it bypassing the threshold while freeing segments from disk
133 * B getting its fragmentation below the threshold).
135 * Empirically, we've seen that our vdev selection for allocations is
136 * good enough that fragmentation increases uniformly across all vdevs
137 * the majority of the time. Thus we set the threshold percentage high
138 * enough to avoid hitting the speed bump on pools that are being pushed
141 static uint_t zfs_mg_fragmentation_threshold = 95;
144 * Allow metaslabs to keep their active state as long as their fragmentation
145 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
146 * active metaslab that exceeds this threshold will no longer keep its active
147 * status allowing better metaslabs to be selected.
149 static uint_t zfs_metaslab_fragmentation_threshold = 70;
152 * When set will load all metaslabs when pool is first opened.
154 int metaslab_debug_load = B_FALSE;
157 * When set will prevent metaslabs from being unloaded.
159 static int metaslab_debug_unload = B_FALSE;
162 * Minimum size which forces the dynamic allocator to change
163 * it's allocation strategy. Once the space map cannot satisfy
164 * an allocation of this size then it switches to using more
165 * aggressive strategy (i.e search by size rather than offset).
167 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
170 * The minimum free space, in percent, which must be available
171 * in a space map to continue allocations in a first-fit fashion.
172 * Once the space map's free space drops below this level we dynamically
173 * switch to using best-fit allocations.
175 uint_t metaslab_df_free_pct = 4;
178 * Maximum distance to search forward from the last offset. Without this
179 * limit, fragmented pools can see >100,000 iterations and
180 * metaslab_block_picker() becomes the performance limiting factor on
181 * high-performance storage.
183 * With the default setting of 16MB, we typically see less than 500
184 * iterations, even with very fragmented, ashift=9 pools. The maximum number
185 * of iterations possible is:
186 * metaslab_df_max_search / (2 * (1<<ashift))
187 * With the default setting of 16MB this is 16*1024 (with ashift=9) or
188 * 2048 (with ashift=12).
190 static uint_t metaslab_df_max_search = 16 * 1024 * 1024;
193 * Forces the metaslab_block_picker function to search for at least this many
194 * segments forwards until giving up on finding a segment that the allocation
197 static const uint32_t metaslab_min_search_count = 100;
200 * If we are not searching forward (due to metaslab_df_max_search,
201 * metaslab_df_free_pct, or metaslab_df_alloc_threshold), this tunable
202 * controls what segment is used. If it is set, we will use the largest free
203 * segment. If it is not set, we will use a segment of exactly the requested
206 static int metaslab_df_use_largest_segment = B_FALSE;
209 * Percentage of all cpus that can be used by the metaslab taskq.
211 int metaslab_load_pct = 50;
214 * These tunables control how long a metaslab will remain loaded after the
215 * last allocation from it. A metaslab can't be unloaded until at least
216 * metaslab_unload_delay TXG's and metaslab_unload_delay_ms milliseconds
217 * have elapsed. However, zfs_metaslab_mem_limit may cause it to be
218 * unloaded sooner. These settings are intended to be generous -- to keep
219 * metaslabs loaded for a long time, reducing the rate of metaslab loading.
221 static uint_t metaslab_unload_delay = 32;
222 static uint_t metaslab_unload_delay_ms = 10 * 60 * 1000; /* ten minutes */
225 * Max number of metaslabs per group to preload.
227 uint_t metaslab_preload_limit = 10;
230 * Enable/disable preloading of metaslab.
232 static int metaslab_preload_enabled = B_TRUE;
235 * Enable/disable fragmentation weighting on metaslabs.
237 static int metaslab_fragmentation_factor_enabled = B_TRUE;
240 * Enable/disable lba weighting (i.e. outer tracks are given preference).
242 static int metaslab_lba_weighting_enabled = B_TRUE;
245 * Enable/disable metaslab group biasing.
247 static int metaslab_bias_enabled = B_TRUE;
250 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
252 static const boolean_t zfs_remap_blkptr_enable = B_TRUE;
255 * Enable/disable segment-based metaslab selection.
257 static int zfs_metaslab_segment_weight_enabled = B_TRUE;
260 * When using segment-based metaslab selection, we will continue
261 * allocating from the active metaslab until we have exhausted
262 * zfs_metaslab_switch_threshold of its buckets.
264 static int zfs_metaslab_switch_threshold = 2;
267 * Internal switch to enable/disable the metaslab allocation tracing
270 static const boolean_t metaslab_trace_enabled = B_FALSE;
273 * Maximum entries that the metaslab allocation tracing facility will keep
274 * in a given list when running in non-debug mode. We limit the number
275 * of entries in non-debug mode to prevent us from using up too much memory.
276 * The limit should be sufficiently large that we don't expect any allocation
277 * to every exceed this value. In debug mode, the system will panic if this
278 * limit is ever reached allowing for further investigation.
280 static const uint64_t metaslab_trace_max_entries = 5000;
283 * Maximum number of metaslabs per group that can be disabled
286 static const int max_disabled_ms = 3;
289 * Time (in seconds) to respect ms_max_size when the metaslab is not loaded.
290 * To avoid 64-bit overflow, don't set above UINT32_MAX.
292 static uint64_t zfs_metaslab_max_size_cache_sec = 1 * 60 * 60; /* 1 hour */
295 * Maximum percentage of memory to use on storing loaded metaslabs. If loading
296 * a metaslab would take it over this percentage, the oldest selected metaslab
297 * is automatically unloaded.
299 static uint_t zfs_metaslab_mem_limit = 25;
302 * Force the per-metaslab range trees to use 64-bit integers to store
303 * segments. Used for debugging purposes.
305 static const boolean_t zfs_metaslab_force_large_segs = B_FALSE;
308 * By default we only store segments over a certain size in the size-sorted
309 * metaslab trees (ms_allocatable_by_size and
310 * ms_unflushed_frees_by_size). This dramatically reduces memory usage and
311 * improves load and unload times at the cost of causing us to use slightly
312 * larger segments than we would otherwise in some cases.
314 static const uint32_t metaslab_by_size_min_shift = 14;
317 * If not set, we will first try normal allocation. If that fails then
318 * we will do a gang allocation. If that fails then we will do a "try hard"
319 * gang allocation. If that fails then we will have a multi-layer gang
322 * If set, we will first try normal allocation. If that fails then
323 * we will do a "try hard" allocation. If that fails we will do a gang
324 * allocation. If that fails we will do a "try hard" gang allocation. If
325 * that fails then we will have a multi-layer gang block.
327 static int zfs_metaslab_try_hard_before_gang = B_FALSE;
330 * When not trying hard, we only consider the best zfs_metaslab_find_max_tries
331 * metaslabs. This improves performance, especially when there are many
332 * metaslabs per vdev and the allocation can't actually be satisfied (so we
333 * would otherwise iterate all the metaslabs). If there is a metaslab with a
334 * worse weight but it can actually satisfy the allocation, we won't find it
335 * until trying hard. This may happen if the worse metaslab is not loaded
336 * (and the true weight is better than we have calculated), or due to weight
337 * bucketization. E.g. we are looking for a 60K segment, and the best
338 * metaslabs all have free segments in the 32-63K bucket, but the best
339 * zfs_metaslab_find_max_tries metaslabs have ms_max_size <60KB, and a
340 * subsequent metaslab has ms_max_size >60KB (but fewer segments in this
341 * bucket, and therefore a lower weight).
343 static uint_t zfs_metaslab_find_max_tries = 100;
345 static uint64_t metaslab_weight(metaslab_t *, boolean_t);
346 static void metaslab_set_fragmentation(metaslab_t *, boolean_t);
347 static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t);
348 static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t);
350 static void metaslab_passivate(metaslab_t *msp, uint64_t weight);
351 static uint64_t metaslab_weight_from_range_tree(metaslab_t *msp);
352 static void metaslab_flush_update(metaslab_t *, dmu_tx_t *);
353 static unsigned int metaslab_idx_func(multilist_t *, void *);
354 static void metaslab_evict(metaslab_t *, uint64_t);
355 static void metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg);
356 kmem_cache_t *metaslab_alloc_trace_cache;
358 typedef struct metaslab_stats {
359 kstat_named_t metaslabstat_trace_over_limit;
360 kstat_named_t metaslabstat_reload_tree;
361 kstat_named_t metaslabstat_too_many_tries;
362 kstat_named_t metaslabstat_try_hard;
365 static metaslab_stats_t metaslab_stats = {
366 { "trace_over_limit", KSTAT_DATA_UINT64 },
367 { "reload_tree", KSTAT_DATA_UINT64 },
368 { "too_many_tries", KSTAT_DATA_UINT64 },
369 { "try_hard", KSTAT_DATA_UINT64 },
372 #define METASLABSTAT_BUMP(stat) \
373 atomic_inc_64(&metaslab_stats.stat.value.ui64);
376 static kstat_t *metaslab_ksp;
379 metaslab_stat_init(void)
381 ASSERT(metaslab_alloc_trace_cache == NULL);
382 metaslab_alloc_trace_cache = kmem_cache_create(
383 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
384 0, NULL, NULL, NULL, NULL, NULL, 0);
385 metaslab_ksp = kstat_create("zfs", 0, "metaslab_stats",
386 "misc", KSTAT_TYPE_NAMED, sizeof (metaslab_stats) /
387 sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
388 if (metaslab_ksp != NULL) {
389 metaslab_ksp->ks_data = &metaslab_stats;
390 kstat_install(metaslab_ksp);
395 metaslab_stat_fini(void)
397 if (metaslab_ksp != NULL) {
398 kstat_delete(metaslab_ksp);
402 kmem_cache_destroy(metaslab_alloc_trace_cache);
403 metaslab_alloc_trace_cache = NULL;
407 * ==========================================================================
409 * ==========================================================================
412 metaslab_class_create(spa_t *spa, const metaslab_ops_t *ops)
414 metaslab_class_t *mc;
416 mc = kmem_zalloc(offsetof(metaslab_class_t,
417 mc_allocator[spa->spa_alloc_count]), KM_SLEEP);
421 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
422 multilist_create(&mc->mc_metaslab_txg_list, sizeof (metaslab_t),
423 offsetof(metaslab_t, ms_class_txg_node), metaslab_idx_func);
424 for (int i = 0; i < spa->spa_alloc_count; i++) {
425 metaslab_class_allocator_t *mca = &mc->mc_allocator[i];
426 mca->mca_rotor = NULL;
427 zfs_refcount_create_tracked(&mca->mca_alloc_slots);
434 metaslab_class_destroy(metaslab_class_t *mc)
436 spa_t *spa = mc->mc_spa;
438 ASSERT(mc->mc_alloc == 0);
439 ASSERT(mc->mc_deferred == 0);
440 ASSERT(mc->mc_space == 0);
441 ASSERT(mc->mc_dspace == 0);
443 for (int i = 0; i < spa->spa_alloc_count; i++) {
444 metaslab_class_allocator_t *mca = &mc->mc_allocator[i];
445 ASSERT(mca->mca_rotor == NULL);
446 zfs_refcount_destroy(&mca->mca_alloc_slots);
448 mutex_destroy(&mc->mc_lock);
449 multilist_destroy(&mc->mc_metaslab_txg_list);
450 kmem_free(mc, offsetof(metaslab_class_t,
451 mc_allocator[spa->spa_alloc_count]));
455 metaslab_class_validate(metaslab_class_t *mc)
457 metaslab_group_t *mg;
461 * Must hold one of the spa_config locks.
463 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
464 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
466 if ((mg = mc->mc_allocator[0].mca_rotor) == NULL)
471 ASSERT(vd->vdev_mg != NULL);
472 ASSERT3P(vd->vdev_top, ==, vd);
473 ASSERT3P(mg->mg_class, ==, mc);
474 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
475 } while ((mg = mg->mg_next) != mc->mc_allocator[0].mca_rotor);
481 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
482 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
484 atomic_add_64(&mc->mc_alloc, alloc_delta);
485 atomic_add_64(&mc->mc_deferred, defer_delta);
486 atomic_add_64(&mc->mc_space, space_delta);
487 atomic_add_64(&mc->mc_dspace, dspace_delta);
491 metaslab_class_get_alloc(metaslab_class_t *mc)
493 return (mc->mc_alloc);
497 metaslab_class_get_deferred(metaslab_class_t *mc)
499 return (mc->mc_deferred);
503 metaslab_class_get_space(metaslab_class_t *mc)
505 return (mc->mc_space);
509 metaslab_class_get_dspace(metaslab_class_t *mc)
511 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
515 metaslab_class_histogram_verify(metaslab_class_t *mc)
517 spa_t *spa = mc->mc_spa;
518 vdev_t *rvd = spa->spa_root_vdev;
522 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
525 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
528 mutex_enter(&mc->mc_lock);
529 for (int c = 0; c < rvd->vdev_children; c++) {
530 vdev_t *tvd = rvd->vdev_child[c];
531 metaslab_group_t *mg = vdev_get_mg(tvd, mc);
534 * Skip any holes, uninitialized top-levels, or
535 * vdevs that are not in this metalab class.
537 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
538 mg->mg_class != mc) {
542 IMPLY(mg == mg->mg_vd->vdev_log_mg,
543 mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
545 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
546 mc_hist[i] += mg->mg_histogram[i];
549 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
550 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
553 mutex_exit(&mc->mc_lock);
554 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
558 * Calculate the metaslab class's fragmentation metric. The metric
559 * is weighted based on the space contribution of each metaslab group.
560 * The return value will be a number between 0 and 100 (inclusive), or
561 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
562 * zfs_frag_table for more information about the metric.
565 metaslab_class_fragmentation(metaslab_class_t *mc)
567 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
568 uint64_t fragmentation = 0;
570 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
572 for (int c = 0; c < rvd->vdev_children; c++) {
573 vdev_t *tvd = rvd->vdev_child[c];
574 metaslab_group_t *mg = tvd->vdev_mg;
577 * Skip any holes, uninitialized top-levels,
578 * or vdevs that are not in this metalab class.
580 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
581 mg->mg_class != mc) {
586 * If a metaslab group does not contain a fragmentation
587 * metric then just bail out.
589 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
590 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
591 return (ZFS_FRAG_INVALID);
595 * Determine how much this metaslab_group is contributing
596 * to the overall pool fragmentation metric.
598 fragmentation += mg->mg_fragmentation *
599 metaslab_group_get_space(mg);
601 fragmentation /= metaslab_class_get_space(mc);
603 ASSERT3U(fragmentation, <=, 100);
604 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
605 return (fragmentation);
609 * Calculate the amount of expandable space that is available in
610 * this metaslab class. If a device is expanded then its expandable
611 * space will be the amount of allocatable space that is currently not
612 * part of this metaslab class.
615 metaslab_class_expandable_space(metaslab_class_t *mc)
617 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
620 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
621 for (int c = 0; c < rvd->vdev_children; c++) {
622 vdev_t *tvd = rvd->vdev_child[c];
623 metaslab_group_t *mg = tvd->vdev_mg;
625 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
626 mg->mg_class != mc) {
631 * Calculate if we have enough space to add additional
632 * metaslabs. We report the expandable space in terms
633 * of the metaslab size since that's the unit of expansion.
635 space += P2ALIGN(tvd->vdev_max_asize - tvd->vdev_asize,
636 1ULL << tvd->vdev_ms_shift);
638 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
643 metaslab_class_evict_old(metaslab_class_t *mc, uint64_t txg)
645 multilist_t *ml = &mc->mc_metaslab_txg_list;
646 for (int i = 0; i < multilist_get_num_sublists(ml); i++) {
647 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
648 metaslab_t *msp = multilist_sublist_head(mls);
649 multilist_sublist_unlock(mls);
650 while (msp != NULL) {
651 mutex_enter(&msp->ms_lock);
654 * If the metaslab has been removed from the list
655 * (which could happen if we were at the memory limit
656 * and it was evicted during this loop), then we can't
657 * proceed and we should restart the sublist.
659 if (!multilist_link_active(&msp->ms_class_txg_node)) {
660 mutex_exit(&msp->ms_lock);
664 mls = multilist_sublist_lock(ml, i);
665 metaslab_t *next_msp = multilist_sublist_next(mls, msp);
666 multilist_sublist_unlock(mls);
668 msp->ms_selected_txg + metaslab_unload_delay &&
669 gethrtime() > msp->ms_selected_time +
670 (uint64_t)MSEC2NSEC(metaslab_unload_delay_ms)) {
671 metaslab_evict(msp, txg);
674 * Once we've hit a metaslab selected too
675 * recently to evict, we're done evicting for
678 mutex_exit(&msp->ms_lock);
681 mutex_exit(&msp->ms_lock);
688 metaslab_compare(const void *x1, const void *x2)
690 const metaslab_t *m1 = (const metaslab_t *)x1;
691 const metaslab_t *m2 = (const metaslab_t *)x2;
695 if (m1->ms_allocator != -1 && m1->ms_primary)
697 else if (m1->ms_allocator != -1 && !m1->ms_primary)
699 if (m2->ms_allocator != -1 && m2->ms_primary)
701 else if (m2->ms_allocator != -1 && !m2->ms_primary)
705 * Sort inactive metaslabs first, then primaries, then secondaries. When
706 * selecting a metaslab to allocate from, an allocator first tries its
707 * primary, then secondary active metaslab. If it doesn't have active
708 * metaslabs, or can't allocate from them, it searches for an inactive
709 * metaslab to activate. If it can't find a suitable one, it will steal
710 * a primary or secondary metaslab from another allocator.
717 int cmp = TREE_CMP(m2->ms_weight, m1->ms_weight);
721 IMPLY(TREE_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2);
723 return (TREE_CMP(m1->ms_start, m2->ms_start));
727 * ==========================================================================
729 * ==========================================================================
732 * Update the allocatable flag and the metaslab group's capacity.
733 * The allocatable flag is set to true if the capacity is below
734 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
735 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
736 * transitions from allocatable to non-allocatable or vice versa then the
737 * metaslab group's class is updated to reflect the transition.
740 metaslab_group_alloc_update(metaslab_group_t *mg)
742 vdev_t *vd = mg->mg_vd;
743 metaslab_class_t *mc = mg->mg_class;
744 vdev_stat_t *vs = &vd->vdev_stat;
745 boolean_t was_allocatable;
746 boolean_t was_initialized;
748 ASSERT(vd == vd->vdev_top);
749 ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==,
752 mutex_enter(&mg->mg_lock);
753 was_allocatable = mg->mg_allocatable;
754 was_initialized = mg->mg_initialized;
756 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
759 mutex_enter(&mc->mc_lock);
762 * If the metaslab group was just added then it won't
763 * have any space until we finish syncing out this txg.
764 * At that point we will consider it initialized and available
765 * for allocations. We also don't consider non-activated
766 * metaslab groups (e.g. vdevs that are in the middle of being removed)
767 * to be initialized, because they can't be used for allocation.
769 mg->mg_initialized = metaslab_group_initialized(mg);
770 if (!was_initialized && mg->mg_initialized) {
772 } else if (was_initialized && !mg->mg_initialized) {
773 ASSERT3U(mc->mc_groups, >, 0);
776 if (mg->mg_initialized)
777 mg->mg_no_free_space = B_FALSE;
780 * A metaslab group is considered allocatable if it has plenty
781 * of free space or is not heavily fragmented. We only take
782 * fragmentation into account if the metaslab group has a valid
783 * fragmentation metric (i.e. a value between 0 and 100).
785 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
786 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
787 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
788 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
791 * The mc_alloc_groups maintains a count of the number of
792 * groups in this metaslab class that are still above the
793 * zfs_mg_noalloc_threshold. This is used by the allocating
794 * threads to determine if they should avoid allocations to
795 * a given group. The allocator will avoid allocations to a group
796 * if that group has reached or is below the zfs_mg_noalloc_threshold
797 * and there are still other groups that are above the threshold.
798 * When a group transitions from allocatable to non-allocatable or
799 * vice versa we update the metaslab class to reflect that change.
800 * When the mc_alloc_groups value drops to 0 that means that all
801 * groups have reached the zfs_mg_noalloc_threshold making all groups
802 * eligible for allocations. This effectively means that all devices
803 * are balanced again.
805 if (was_allocatable && !mg->mg_allocatable)
806 mc->mc_alloc_groups--;
807 else if (!was_allocatable && mg->mg_allocatable)
808 mc->mc_alloc_groups++;
809 mutex_exit(&mc->mc_lock);
811 mutex_exit(&mg->mg_lock);
815 metaslab_sort_by_flushed(const void *va, const void *vb)
817 const metaslab_t *a = va;
818 const metaslab_t *b = vb;
820 int cmp = TREE_CMP(a->ms_unflushed_txg, b->ms_unflushed_txg);
824 uint64_t a_vdev_id = a->ms_group->mg_vd->vdev_id;
825 uint64_t b_vdev_id = b->ms_group->mg_vd->vdev_id;
826 cmp = TREE_CMP(a_vdev_id, b_vdev_id);
830 return (TREE_CMP(a->ms_id, b->ms_id));
834 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
836 metaslab_group_t *mg;
838 mg = kmem_zalloc(offsetof(metaslab_group_t,
839 mg_allocator[allocators]), KM_SLEEP);
840 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
841 mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL);
842 cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL);
843 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
844 sizeof (metaslab_t), offsetof(metaslab_t, ms_group_node));
847 mg->mg_activation_count = 0;
848 mg->mg_initialized = B_FALSE;
849 mg->mg_no_free_space = B_TRUE;
850 mg->mg_allocators = allocators;
852 for (int i = 0; i < allocators; i++) {
853 metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
854 zfs_refcount_create_tracked(&mga->mga_alloc_queue_depth);
857 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
858 maxclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT | TASKQ_DYNAMIC);
864 metaslab_group_destroy(metaslab_group_t *mg)
866 ASSERT(mg->mg_prev == NULL);
867 ASSERT(mg->mg_next == NULL);
869 * We may have gone below zero with the activation count
870 * either because we never activated in the first place or
871 * because we're done, and possibly removing the vdev.
873 ASSERT(mg->mg_activation_count <= 0);
875 taskq_destroy(mg->mg_taskq);
876 avl_destroy(&mg->mg_metaslab_tree);
877 mutex_destroy(&mg->mg_lock);
878 mutex_destroy(&mg->mg_ms_disabled_lock);
879 cv_destroy(&mg->mg_ms_disabled_cv);
881 for (int i = 0; i < mg->mg_allocators; i++) {
882 metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
883 zfs_refcount_destroy(&mga->mga_alloc_queue_depth);
885 kmem_free(mg, offsetof(metaslab_group_t,
886 mg_allocator[mg->mg_allocators]));
890 metaslab_group_activate(metaslab_group_t *mg)
892 metaslab_class_t *mc = mg->mg_class;
893 spa_t *spa = mc->mc_spa;
894 metaslab_group_t *mgprev, *mgnext;
896 ASSERT3U(spa_config_held(spa, SCL_ALLOC, RW_WRITER), !=, 0);
898 ASSERT(mg->mg_prev == NULL);
899 ASSERT(mg->mg_next == NULL);
900 ASSERT(mg->mg_activation_count <= 0);
902 if (++mg->mg_activation_count <= 0)
905 mg->mg_aliquot = metaslab_aliquot * MAX(1,
906 vdev_get_ndisks(mg->mg_vd) - vdev_get_nparity(mg->mg_vd));
907 metaslab_group_alloc_update(mg);
909 if ((mgprev = mc->mc_allocator[0].mca_rotor) == NULL) {
913 mgnext = mgprev->mg_next;
914 mg->mg_prev = mgprev;
915 mg->mg_next = mgnext;
916 mgprev->mg_next = mg;
917 mgnext->mg_prev = mg;
919 for (int i = 0; i < spa->spa_alloc_count; i++) {
920 mc->mc_allocator[i].mca_rotor = mg;
926 * Passivate a metaslab group and remove it from the allocation rotor.
927 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
928 * a metaslab group. This function will momentarily drop spa_config_locks
929 * that are lower than the SCL_ALLOC lock (see comment below).
932 metaslab_group_passivate(metaslab_group_t *mg)
934 metaslab_class_t *mc = mg->mg_class;
935 spa_t *spa = mc->mc_spa;
936 metaslab_group_t *mgprev, *mgnext;
937 int locks = spa_config_held(spa, SCL_ALL, RW_WRITER);
939 ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==,
940 (SCL_ALLOC | SCL_ZIO));
942 if (--mg->mg_activation_count != 0) {
943 for (int i = 0; i < spa->spa_alloc_count; i++)
944 ASSERT(mc->mc_allocator[i].mca_rotor != mg);
945 ASSERT(mg->mg_prev == NULL);
946 ASSERT(mg->mg_next == NULL);
947 ASSERT(mg->mg_activation_count < 0);
952 * The spa_config_lock is an array of rwlocks, ordered as
953 * follows (from highest to lowest):
954 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
955 * SCL_ZIO > SCL_FREE > SCL_VDEV
956 * (For more information about the spa_config_lock see spa_misc.c)
957 * The higher the lock, the broader its coverage. When we passivate
958 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
959 * config locks. However, the metaslab group's taskq might be trying
960 * to preload metaslabs so we must drop the SCL_ZIO lock and any
961 * lower locks to allow the I/O to complete. At a minimum,
962 * we continue to hold the SCL_ALLOC lock, which prevents any future
963 * allocations from taking place and any changes to the vdev tree.
965 spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa);
966 taskq_wait_outstanding(mg->mg_taskq, 0);
967 spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
968 metaslab_group_alloc_update(mg);
969 for (int i = 0; i < mg->mg_allocators; i++) {
970 metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
971 metaslab_t *msp = mga->mga_primary;
973 mutex_enter(&msp->ms_lock);
974 metaslab_passivate(msp,
975 metaslab_weight_from_range_tree(msp));
976 mutex_exit(&msp->ms_lock);
978 msp = mga->mga_secondary;
980 mutex_enter(&msp->ms_lock);
981 metaslab_passivate(msp,
982 metaslab_weight_from_range_tree(msp));
983 mutex_exit(&msp->ms_lock);
987 mgprev = mg->mg_prev;
988 mgnext = mg->mg_next;
993 mgprev->mg_next = mgnext;
994 mgnext->mg_prev = mgprev;
996 for (int i = 0; i < spa->spa_alloc_count; i++) {
997 if (mc->mc_allocator[i].mca_rotor == mg)
998 mc->mc_allocator[i].mca_rotor = mgnext;
1006 metaslab_group_initialized(metaslab_group_t *mg)
1008 vdev_t *vd = mg->mg_vd;
1009 vdev_stat_t *vs = &vd->vdev_stat;
1011 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
1015 metaslab_group_get_space(metaslab_group_t *mg)
1018 * Note that the number of nodes in mg_metaslab_tree may be one less
1019 * than vdev_ms_count, due to the embedded log metaslab.
1021 mutex_enter(&mg->mg_lock);
1022 uint64_t ms_count = avl_numnodes(&mg->mg_metaslab_tree);
1023 mutex_exit(&mg->mg_lock);
1024 return ((1ULL << mg->mg_vd->vdev_ms_shift) * ms_count);
1028 metaslab_group_histogram_verify(metaslab_group_t *mg)
1031 avl_tree_t *t = &mg->mg_metaslab_tree;
1032 uint64_t ashift = mg->mg_vd->vdev_ashift;
1034 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
1037 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
1040 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
1041 SPACE_MAP_HISTOGRAM_SIZE + ashift);
1043 mutex_enter(&mg->mg_lock);
1044 for (metaslab_t *msp = avl_first(t);
1045 msp != NULL; msp = AVL_NEXT(t, msp)) {
1046 VERIFY3P(msp->ms_group, ==, mg);
1047 /* skip if not active */
1048 if (msp->ms_sm == NULL)
1051 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1052 mg_hist[i + ashift] +=
1053 msp->ms_sm->sm_phys->smp_histogram[i];
1057 for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
1058 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
1060 mutex_exit(&mg->mg_lock);
1062 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
1066 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
1068 metaslab_class_t *mc = mg->mg_class;
1069 uint64_t ashift = mg->mg_vd->vdev_ashift;
1071 ASSERT(MUTEX_HELD(&msp->ms_lock));
1072 if (msp->ms_sm == NULL)
1075 mutex_enter(&mg->mg_lock);
1076 mutex_enter(&mc->mc_lock);
1077 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1078 IMPLY(mg == mg->mg_vd->vdev_log_mg,
1079 mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
1080 mg->mg_histogram[i + ashift] +=
1081 msp->ms_sm->sm_phys->smp_histogram[i];
1082 mc->mc_histogram[i + ashift] +=
1083 msp->ms_sm->sm_phys->smp_histogram[i];
1085 mutex_exit(&mc->mc_lock);
1086 mutex_exit(&mg->mg_lock);
1090 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
1092 metaslab_class_t *mc = mg->mg_class;
1093 uint64_t ashift = mg->mg_vd->vdev_ashift;
1095 ASSERT(MUTEX_HELD(&msp->ms_lock));
1096 if (msp->ms_sm == NULL)
1099 mutex_enter(&mg->mg_lock);
1100 mutex_enter(&mc->mc_lock);
1101 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1102 ASSERT3U(mg->mg_histogram[i + ashift], >=,
1103 msp->ms_sm->sm_phys->smp_histogram[i]);
1104 ASSERT3U(mc->mc_histogram[i + ashift], >=,
1105 msp->ms_sm->sm_phys->smp_histogram[i]);
1106 IMPLY(mg == mg->mg_vd->vdev_log_mg,
1107 mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
1109 mg->mg_histogram[i + ashift] -=
1110 msp->ms_sm->sm_phys->smp_histogram[i];
1111 mc->mc_histogram[i + ashift] -=
1112 msp->ms_sm->sm_phys->smp_histogram[i];
1114 mutex_exit(&mc->mc_lock);
1115 mutex_exit(&mg->mg_lock);
1119 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
1121 ASSERT(msp->ms_group == NULL);
1122 mutex_enter(&mg->mg_lock);
1125 avl_add(&mg->mg_metaslab_tree, msp);
1126 mutex_exit(&mg->mg_lock);
1128 mutex_enter(&msp->ms_lock);
1129 metaslab_group_histogram_add(mg, msp);
1130 mutex_exit(&msp->ms_lock);
1134 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
1136 mutex_enter(&msp->ms_lock);
1137 metaslab_group_histogram_remove(mg, msp);
1138 mutex_exit(&msp->ms_lock);
1140 mutex_enter(&mg->mg_lock);
1141 ASSERT(msp->ms_group == mg);
1142 avl_remove(&mg->mg_metaslab_tree, msp);
1144 metaslab_class_t *mc = msp->ms_group->mg_class;
1145 multilist_sublist_t *mls =
1146 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
1147 if (multilist_link_active(&msp->ms_class_txg_node))
1148 multilist_sublist_remove(mls, msp);
1149 multilist_sublist_unlock(mls);
1151 msp->ms_group = NULL;
1152 mutex_exit(&mg->mg_lock);
1156 metaslab_group_sort_impl(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
1158 ASSERT(MUTEX_HELD(&msp->ms_lock));
1159 ASSERT(MUTEX_HELD(&mg->mg_lock));
1160 ASSERT(msp->ms_group == mg);
1162 avl_remove(&mg->mg_metaslab_tree, msp);
1163 msp->ms_weight = weight;
1164 avl_add(&mg->mg_metaslab_tree, msp);
1169 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
1172 * Although in principle the weight can be any value, in
1173 * practice we do not use values in the range [1, 511].
1175 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
1176 ASSERT(MUTEX_HELD(&msp->ms_lock));
1178 mutex_enter(&mg->mg_lock);
1179 metaslab_group_sort_impl(mg, msp, weight);
1180 mutex_exit(&mg->mg_lock);
1184 * Calculate the fragmentation for a given metaslab group. We can use
1185 * a simple average here since all metaslabs within the group must have
1186 * the same size. The return value will be a value between 0 and 100
1187 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
1188 * group have a fragmentation metric.
1191 metaslab_group_fragmentation(metaslab_group_t *mg)
1193 vdev_t *vd = mg->mg_vd;
1194 uint64_t fragmentation = 0;
1195 uint64_t valid_ms = 0;
1197 for (int m = 0; m < vd->vdev_ms_count; m++) {
1198 metaslab_t *msp = vd->vdev_ms[m];
1200 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
1202 if (msp->ms_group != mg)
1206 fragmentation += msp->ms_fragmentation;
1209 if (valid_ms <= mg->mg_vd->vdev_ms_count / 2)
1210 return (ZFS_FRAG_INVALID);
1212 fragmentation /= valid_ms;
1213 ASSERT3U(fragmentation, <=, 100);
1214 return (fragmentation);
1218 * Determine if a given metaslab group should skip allocations. A metaslab
1219 * group should avoid allocations if its free capacity is less than the
1220 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
1221 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
1222 * that can still handle allocations. If the allocation throttle is enabled
1223 * then we skip allocations to devices that have reached their maximum
1224 * allocation queue depth unless the selected metaslab group is the only
1225 * eligible group remaining.
1228 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
1229 int flags, uint64_t psize, int allocator, int d)
1231 spa_t *spa = mg->mg_vd->vdev_spa;
1232 metaslab_class_t *mc = mg->mg_class;
1235 * We can only consider skipping this metaslab group if it's
1236 * in the normal metaslab class and there are other metaslab
1237 * groups to select from. Otherwise, we always consider it eligible
1240 if ((mc != spa_normal_class(spa) &&
1241 mc != spa_special_class(spa) &&
1242 mc != spa_dedup_class(spa)) ||
1247 * If the metaslab group's mg_allocatable flag is set (see comments
1248 * in metaslab_group_alloc_update() for more information) and
1249 * the allocation throttle is disabled then allow allocations to this
1250 * device. However, if the allocation throttle is enabled then
1251 * check if we have reached our allocation limit (mga_alloc_queue_depth)
1252 * to determine if we should allow allocations to this metaslab group.
1253 * If all metaslab groups are no longer considered allocatable
1254 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1255 * gang block size then we allow allocations on this metaslab group
1256 * regardless of the mg_allocatable or throttle settings.
1258 if (mg->mg_allocatable) {
1259 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
1261 uint64_t qmax = mga->mga_cur_max_alloc_queue_depth;
1263 if (!mc->mc_alloc_throttle_enabled)
1267 * If this metaslab group does not have any free space, then
1268 * there is no point in looking further.
1270 if (mg->mg_no_free_space)
1274 * Some allocations (e.g., those coming from device removal
1275 * where the * allocations are not even counted in the
1276 * metaslab * allocation queues) are allowed to bypass
1279 if (flags & METASLAB_DONT_THROTTLE)
1283 * Relax allocation throttling for ditto blocks. Due to
1284 * random imbalances in allocation it tends to push copies
1285 * to one vdev, that looks a bit better at the moment.
1287 qmax = qmax * (4 + d) / 4;
1289 qdepth = zfs_refcount_count(&mga->mga_alloc_queue_depth);
1292 * If this metaslab group is below its qmax or it's
1293 * the only allocatable metaslab group, then attempt
1294 * to allocate from it.
1296 if (qdepth < qmax || mc->mc_alloc_groups == 1)
1298 ASSERT3U(mc->mc_alloc_groups, >, 1);
1301 * Since this metaslab group is at or over its qmax, we
1302 * need to determine if there are metaslab groups after this
1303 * one that might be able to handle this allocation. This is
1304 * racy since we can't hold the locks for all metaslab
1305 * groups at the same time when we make this check.
1307 for (metaslab_group_t *mgp = mg->mg_next;
1308 mgp != rotor; mgp = mgp->mg_next) {
1309 metaslab_group_allocator_t *mgap =
1310 &mgp->mg_allocator[allocator];
1311 qmax = mgap->mga_cur_max_alloc_queue_depth;
1312 qmax = qmax * (4 + d) / 4;
1314 zfs_refcount_count(&mgap->mga_alloc_queue_depth);
1317 * If there is another metaslab group that
1318 * might be able to handle the allocation, then
1319 * we return false so that we skip this group.
1321 if (qdepth < qmax && !mgp->mg_no_free_space)
1326 * We didn't find another group to handle the allocation
1327 * so we can't skip this metaslab group even though
1328 * we are at or over our qmax.
1332 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
1339 * ==========================================================================
1340 * Range tree callbacks
1341 * ==========================================================================
1345 * Comparison function for the private size-ordered tree using 32-bit
1346 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1348 __attribute__((always_inline)) inline
1350 metaslab_rangesize32_compare(const void *x1, const void *x2)
1352 const range_seg32_t *r1 = x1;
1353 const range_seg32_t *r2 = x2;
1355 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1356 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1358 int cmp = TREE_CMP(rs_size1, rs_size2);
1360 return (cmp + !cmp * TREE_CMP(r1->rs_start, r2->rs_start));
1364 * Comparison function for the private size-ordered tree using 64-bit
1365 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1367 __attribute__((always_inline)) inline
1369 metaslab_rangesize64_compare(const void *x1, const void *x2)
1371 const range_seg64_t *r1 = x1;
1372 const range_seg64_t *r2 = x2;
1374 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1375 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1377 int cmp = TREE_CMP(rs_size1, rs_size2);
1379 return (cmp + !cmp * TREE_CMP(r1->rs_start, r2->rs_start));
1382 typedef struct metaslab_rt_arg {
1383 zfs_btree_t *mra_bt;
1384 uint32_t mra_floor_shift;
1385 } metaslab_rt_arg_t;
1389 metaslab_rt_arg_t *mra;
1393 metaslab_size_sorted_add(void *arg, uint64_t start, uint64_t size)
1395 struct mssa_arg *mssap = arg;
1396 range_tree_t *rt = mssap->rt;
1397 metaslab_rt_arg_t *mrap = mssap->mra;
1398 range_seg_max_t seg = {0};
1399 rs_set_start(&seg, rt, start);
1400 rs_set_end(&seg, rt, start + size);
1401 metaslab_rt_add(rt, &seg, mrap);
1405 metaslab_size_tree_full_load(range_tree_t *rt)
1407 metaslab_rt_arg_t *mrap = rt->rt_arg;
1408 METASLABSTAT_BUMP(metaslabstat_reload_tree);
1409 ASSERT0(zfs_btree_numnodes(mrap->mra_bt));
1410 mrap->mra_floor_shift = 0;
1411 struct mssa_arg arg = {0};
1414 range_tree_walk(rt, metaslab_size_sorted_add, &arg);
1418 ZFS_BTREE_FIND_IN_BUF_FUNC(metaslab_rt_find_rangesize32_in_buf,
1419 range_seg32_t, metaslab_rangesize32_compare)
1421 ZFS_BTREE_FIND_IN_BUF_FUNC(metaslab_rt_find_rangesize64_in_buf,
1422 range_seg64_t, metaslab_rangesize64_compare)
1425 * Create any block allocator specific components. The current allocators
1426 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1429 metaslab_rt_create(range_tree_t *rt, void *arg)
1431 metaslab_rt_arg_t *mrap = arg;
1432 zfs_btree_t *size_tree = mrap->mra_bt;
1435 int (*compare) (const void *, const void *);
1436 bt_find_in_buf_f bt_find;
1437 switch (rt->rt_type) {
1439 size = sizeof (range_seg32_t);
1440 compare = metaslab_rangesize32_compare;
1441 bt_find = metaslab_rt_find_rangesize32_in_buf;
1444 size = sizeof (range_seg64_t);
1445 compare = metaslab_rangesize64_compare;
1446 bt_find = metaslab_rt_find_rangesize64_in_buf;
1449 panic("Invalid range seg type %d", rt->rt_type);
1451 zfs_btree_create(size_tree, compare, bt_find, size);
1452 mrap->mra_floor_shift = metaslab_by_size_min_shift;
1456 metaslab_rt_destroy(range_tree_t *rt, void *arg)
1459 metaslab_rt_arg_t *mrap = arg;
1460 zfs_btree_t *size_tree = mrap->mra_bt;
1462 zfs_btree_destroy(size_tree);
1463 kmem_free(mrap, sizeof (*mrap));
1467 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1469 metaslab_rt_arg_t *mrap = arg;
1470 zfs_btree_t *size_tree = mrap->mra_bt;
1472 if (rs_get_end(rs, rt) - rs_get_start(rs, rt) <
1473 (1ULL << mrap->mra_floor_shift))
1476 zfs_btree_add(size_tree, rs);
1480 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1482 metaslab_rt_arg_t *mrap = arg;
1483 zfs_btree_t *size_tree = mrap->mra_bt;
1485 if (rs_get_end(rs, rt) - rs_get_start(rs, rt) < (1ULL <<
1486 mrap->mra_floor_shift))
1489 zfs_btree_remove(size_tree, rs);
1493 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1495 metaslab_rt_arg_t *mrap = arg;
1496 zfs_btree_t *size_tree = mrap->mra_bt;
1497 zfs_btree_clear(size_tree);
1498 zfs_btree_destroy(size_tree);
1500 metaslab_rt_create(rt, arg);
1503 static const range_tree_ops_t metaslab_rt_ops = {
1504 .rtop_create = metaslab_rt_create,
1505 .rtop_destroy = metaslab_rt_destroy,
1506 .rtop_add = metaslab_rt_add,
1507 .rtop_remove = metaslab_rt_remove,
1508 .rtop_vacate = metaslab_rt_vacate
1512 * ==========================================================================
1513 * Common allocator routines
1514 * ==========================================================================
1518 * Return the maximum contiguous segment within the metaslab.
1521 metaslab_largest_allocatable(metaslab_t *msp)
1523 zfs_btree_t *t = &msp->ms_allocatable_by_size;
1528 if (zfs_btree_numnodes(t) == 0)
1529 metaslab_size_tree_full_load(msp->ms_allocatable);
1531 rs = zfs_btree_last(t, NULL);
1535 return (rs_get_end(rs, msp->ms_allocatable) - rs_get_start(rs,
1536 msp->ms_allocatable));
1540 * Return the maximum contiguous segment within the unflushed frees of this
1544 metaslab_largest_unflushed_free(metaslab_t *msp)
1546 ASSERT(MUTEX_HELD(&msp->ms_lock));
1548 if (msp->ms_unflushed_frees == NULL)
1551 if (zfs_btree_numnodes(&msp->ms_unflushed_frees_by_size) == 0)
1552 metaslab_size_tree_full_load(msp->ms_unflushed_frees);
1553 range_seg_t *rs = zfs_btree_last(&msp->ms_unflushed_frees_by_size,
1559 * When a range is freed from the metaslab, that range is added to
1560 * both the unflushed frees and the deferred frees. While the block
1561 * will eventually be usable, if the metaslab were loaded the range
1562 * would not be added to the ms_allocatable tree until TXG_DEFER_SIZE
1563 * txgs had passed. As a result, when attempting to estimate an upper
1564 * bound for the largest currently-usable free segment in the
1565 * metaslab, we need to not consider any ranges currently in the defer
1566 * trees. This algorithm approximates the largest available chunk in
1567 * the largest range in the unflushed_frees tree by taking the first
1568 * chunk. While this may be a poor estimate, it should only remain so
1569 * briefly and should eventually self-correct as frees are no longer
1570 * deferred. Similar logic applies to the ms_freed tree. See
1571 * metaslab_load() for more details.
1573 * There are two primary sources of inaccuracy in this estimate. Both
1574 * are tolerated for performance reasons. The first source is that we
1575 * only check the largest segment for overlaps. Smaller segments may
1576 * have more favorable overlaps with the other trees, resulting in
1577 * larger usable chunks. Second, we only look at the first chunk in
1578 * the largest segment; there may be other usable chunks in the
1579 * largest segment, but we ignore them.
1581 uint64_t rstart = rs_get_start(rs, msp->ms_unflushed_frees);
1582 uint64_t rsize = rs_get_end(rs, msp->ms_unflushed_frees) - rstart;
1583 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1586 boolean_t found = range_tree_find_in(msp->ms_defer[t], rstart,
1587 rsize, &start, &size);
1589 if (rstart == start)
1591 rsize = start - rstart;
1597 boolean_t found = range_tree_find_in(msp->ms_freed, rstart,
1598 rsize, &start, &size);
1600 rsize = start - rstart;
1605 static range_seg_t *
1606 metaslab_block_find(zfs_btree_t *t, range_tree_t *rt, uint64_t start,
1607 uint64_t size, zfs_btree_index_t *where)
1610 range_seg_max_t rsearch;
1612 rs_set_start(&rsearch, rt, start);
1613 rs_set_end(&rsearch, rt, start + size);
1615 rs = zfs_btree_find(t, &rsearch, where);
1617 rs = zfs_btree_next(t, where, where);
1624 * This is a helper function that can be used by the allocator to find a
1625 * suitable block to allocate. This will search the specified B-tree looking
1626 * for a block that matches the specified criteria.
1629 metaslab_block_picker(range_tree_t *rt, uint64_t *cursor, uint64_t size,
1630 uint64_t max_search)
1633 *cursor = rt->rt_start;
1634 zfs_btree_t *bt = &rt->rt_root;
1635 zfs_btree_index_t where;
1636 range_seg_t *rs = metaslab_block_find(bt, rt, *cursor, size, &where);
1637 uint64_t first_found;
1638 int count_searched = 0;
1641 first_found = rs_get_start(rs, rt);
1643 while (rs != NULL && (rs_get_start(rs, rt) - first_found <=
1644 max_search || count_searched < metaslab_min_search_count)) {
1645 uint64_t offset = rs_get_start(rs, rt);
1646 if (offset + size <= rs_get_end(rs, rt)) {
1647 *cursor = offset + size;
1650 rs = zfs_btree_next(bt, &where, &where);
1658 static uint64_t metaslab_df_alloc(metaslab_t *msp, uint64_t size);
1659 static uint64_t metaslab_cf_alloc(metaslab_t *msp, uint64_t size);
1660 static uint64_t metaslab_ndf_alloc(metaslab_t *msp, uint64_t size);
1661 metaslab_ops_t *metaslab_allocator(spa_t *spa);
1663 static metaslab_ops_t metaslab_allocators[] = {
1664 { "dynamic", metaslab_df_alloc },
1665 { "cursor", metaslab_cf_alloc },
1666 { "new-dynamic", metaslab_ndf_alloc },
1670 spa_find_allocator_byname(const char *val)
1672 int a = ARRAY_SIZE(metaslab_allocators) - 1;
1673 if (strcmp("new-dynamic", val) == 0)
1674 return (-1); /* remove when ndf is working */
1675 for (; a >= 0; a--) {
1676 if (strcmp(val, metaslab_allocators[a].msop_name) == 0)
1683 spa_set_allocator(spa_t *spa, const char *allocator)
1685 int a = spa_find_allocator_byname(allocator);
1687 spa->spa_active_allocator = a;
1688 zfs_dbgmsg("spa allocator: %s\n", metaslab_allocators[a].msop_name);
1692 spa_get_allocator(spa_t *spa)
1694 return (spa->spa_active_allocator);
1697 #if defined(_KERNEL)
1699 param_set_active_allocator_common(const char *val)
1704 return (SET_ERROR(EINVAL));
1706 if ((p = strchr(val, '\n')) != NULL)
1709 int a = spa_find_allocator_byname(val);
1711 return (SET_ERROR(EINVAL));
1713 zfs_active_allocator = metaslab_allocators[a].msop_name;
1719 metaslab_allocator(spa_t *spa)
1721 int allocator = spa_get_allocator(spa);
1722 return (&metaslab_allocators[allocator]);
1726 * ==========================================================================
1727 * Dynamic Fit (df) block allocator
1729 * Search for a free chunk of at least this size, starting from the last
1730 * offset (for this alignment of block) looking for up to
1731 * metaslab_df_max_search bytes (16MB). If a large enough free chunk is not
1732 * found within 16MB, then return a free chunk of exactly the requested size (or
1735 * If it seems like searching from the last offset will be unproductive, skip
1736 * that and just return a free chunk of exactly the requested size (or larger).
1737 * This is based on metaslab_df_alloc_threshold and metaslab_df_free_pct. This
1738 * mechanism is probably not very useful and may be removed in the future.
1740 * The behavior when not searching can be changed to return the largest free
1741 * chunk, instead of a free chunk of exactly the requested size, by setting
1742 * metaslab_df_use_largest_segment.
1743 * ==========================================================================
1746 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1749 * Find the largest power of 2 block size that evenly divides the
1750 * requested size. This is used to try to allocate blocks with similar
1751 * alignment from the same area of the metaslab (i.e. same cursor
1752 * bucket) but it does not guarantee that other allocations sizes
1753 * may exist in the same region.
1755 uint64_t align = size & -size;
1756 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1757 range_tree_t *rt = msp->ms_allocatable;
1758 uint_t free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1761 ASSERT(MUTEX_HELD(&msp->ms_lock));
1764 * If we're running low on space, find a segment based on size,
1765 * rather than iterating based on offset.
1767 if (metaslab_largest_allocatable(msp) < metaslab_df_alloc_threshold ||
1768 free_pct < metaslab_df_free_pct) {
1771 offset = metaslab_block_picker(rt,
1772 cursor, size, metaslab_df_max_search);
1777 if (zfs_btree_numnodes(&msp->ms_allocatable_by_size) == 0)
1778 metaslab_size_tree_full_load(msp->ms_allocatable);
1780 if (metaslab_df_use_largest_segment) {
1781 /* use largest free segment */
1782 rs = zfs_btree_last(&msp->ms_allocatable_by_size, NULL);
1784 zfs_btree_index_t where;
1785 /* use segment of this size, or next largest */
1786 rs = metaslab_block_find(&msp->ms_allocatable_by_size,
1787 rt, msp->ms_start, size, &where);
1789 if (rs != NULL && rs_get_start(rs, rt) + size <= rs_get_end(rs,
1791 offset = rs_get_start(rs, rt);
1792 *cursor = offset + size;
1800 * ==========================================================================
1801 * Cursor fit block allocator -
1802 * Select the largest region in the metaslab, set the cursor to the beginning
1803 * of the range and the cursor_end to the end of the range. As allocations
1804 * are made advance the cursor. Continue allocating from the cursor until
1805 * the range is exhausted and then find a new range.
1806 * ==========================================================================
1809 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1811 range_tree_t *rt = msp->ms_allocatable;
1812 zfs_btree_t *t = &msp->ms_allocatable_by_size;
1813 uint64_t *cursor = &msp->ms_lbas[0];
1814 uint64_t *cursor_end = &msp->ms_lbas[1];
1815 uint64_t offset = 0;
1817 ASSERT(MUTEX_HELD(&msp->ms_lock));
1819 ASSERT3U(*cursor_end, >=, *cursor);
1821 if ((*cursor + size) > *cursor_end) {
1824 if (zfs_btree_numnodes(t) == 0)
1825 metaslab_size_tree_full_load(msp->ms_allocatable);
1826 rs = zfs_btree_last(t, NULL);
1827 if (rs == NULL || (rs_get_end(rs, rt) - rs_get_start(rs, rt)) <
1831 *cursor = rs_get_start(rs, rt);
1832 *cursor_end = rs_get_end(rs, rt);
1842 * ==========================================================================
1843 * New dynamic fit allocator -
1844 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1845 * contiguous blocks. If no region is found then just use the largest segment
1847 * ==========================================================================
1851 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1852 * to request from the allocator.
1854 uint64_t metaslab_ndf_clump_shift = 4;
1857 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1859 zfs_btree_t *t = &msp->ms_allocatable->rt_root;
1860 range_tree_t *rt = msp->ms_allocatable;
1861 zfs_btree_index_t where;
1863 range_seg_max_t rsearch;
1864 uint64_t hbit = highbit64(size);
1865 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1866 uint64_t max_size = metaslab_largest_allocatable(msp);
1868 ASSERT(MUTEX_HELD(&msp->ms_lock));
1870 if (max_size < size)
1873 rs_set_start(&rsearch, rt, *cursor);
1874 rs_set_end(&rsearch, rt, *cursor + size);
1876 rs = zfs_btree_find(t, &rsearch, &where);
1877 if (rs == NULL || (rs_get_end(rs, rt) - rs_get_start(rs, rt)) < size) {
1878 t = &msp->ms_allocatable_by_size;
1880 rs_set_start(&rsearch, rt, 0);
1881 rs_set_end(&rsearch, rt, MIN(max_size, 1ULL << (hbit +
1882 metaslab_ndf_clump_shift)));
1884 rs = zfs_btree_find(t, &rsearch, &where);
1886 rs = zfs_btree_next(t, &where, &where);
1890 if ((rs_get_end(rs, rt) - rs_get_start(rs, rt)) >= size) {
1891 *cursor = rs_get_start(rs, rt) + size;
1892 return (rs_get_start(rs, rt));
1898 * ==========================================================================
1900 * ==========================================================================
1904 * Wait for any in-progress metaslab loads to complete.
1907 metaslab_load_wait(metaslab_t *msp)
1909 ASSERT(MUTEX_HELD(&msp->ms_lock));
1911 while (msp->ms_loading) {
1912 ASSERT(!msp->ms_loaded);
1913 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1918 * Wait for any in-progress flushing to complete.
1921 metaslab_flush_wait(metaslab_t *msp)
1923 ASSERT(MUTEX_HELD(&msp->ms_lock));
1925 while (msp->ms_flushing)
1926 cv_wait(&msp->ms_flush_cv, &msp->ms_lock);
1930 metaslab_idx_func(multilist_t *ml, void *arg)
1932 metaslab_t *msp = arg;
1935 * ms_id values are allocated sequentially, so full 64bit
1936 * division would be a waste of time, so limit it to 32 bits.
1938 return ((unsigned int)msp->ms_id % multilist_get_num_sublists(ml));
1942 metaslab_allocated_space(metaslab_t *msp)
1944 return (msp->ms_allocated_space);
1948 * Verify that the space accounting on disk matches the in-core range_trees.
1951 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
1953 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1954 uint64_t allocating = 0;
1955 uint64_t sm_free_space, msp_free_space;
1957 ASSERT(MUTEX_HELD(&msp->ms_lock));
1958 ASSERT(!msp->ms_condensing);
1960 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
1964 * We can only verify the metaslab space when we're called
1965 * from syncing context with a loaded metaslab that has an
1966 * allocated space map. Calling this in non-syncing context
1967 * does not provide a consistent view of the metaslab since
1968 * we're performing allocations in the future.
1970 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
1975 * Even though the smp_alloc field can get negative,
1976 * when it comes to a metaslab's space map, that should
1977 * never be the case.
1979 ASSERT3S(space_map_allocated(msp->ms_sm), >=, 0);
1981 ASSERT3U(space_map_allocated(msp->ms_sm), >=,
1982 range_tree_space(msp->ms_unflushed_frees));
1984 ASSERT3U(metaslab_allocated_space(msp), ==,
1985 space_map_allocated(msp->ms_sm) +
1986 range_tree_space(msp->ms_unflushed_allocs) -
1987 range_tree_space(msp->ms_unflushed_frees));
1989 sm_free_space = msp->ms_size - metaslab_allocated_space(msp);
1992 * Account for future allocations since we would have
1993 * already deducted that space from the ms_allocatable.
1995 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
1997 range_tree_space(msp->ms_allocating[(txg + t) & TXG_MASK]);
1999 ASSERT3U(allocating + msp->ms_allocated_this_txg, ==,
2000 msp->ms_allocating_total);
2002 ASSERT3U(msp->ms_deferspace, ==,
2003 range_tree_space(msp->ms_defer[0]) +
2004 range_tree_space(msp->ms_defer[1]));
2006 msp_free_space = range_tree_space(msp->ms_allocatable) + allocating +
2007 msp->ms_deferspace + range_tree_space(msp->ms_freed);
2009 VERIFY3U(sm_free_space, ==, msp_free_space);
2013 metaslab_aux_histograms_clear(metaslab_t *msp)
2016 * Auxiliary histograms are only cleared when resetting them,
2017 * which can only happen while the metaslab is loaded.
2019 ASSERT(msp->ms_loaded);
2021 memset(msp->ms_synchist, 0, sizeof (msp->ms_synchist));
2022 for (int t = 0; t < TXG_DEFER_SIZE; t++)
2023 memset(msp->ms_deferhist[t], 0, sizeof (msp->ms_deferhist[t]));
2027 metaslab_aux_histogram_add(uint64_t *histogram, uint64_t shift,
2031 * This is modeled after space_map_histogram_add(), so refer to that
2032 * function for implementation details. We want this to work like
2033 * the space map histogram, and not the range tree histogram, as we
2034 * are essentially constructing a delta that will be later subtracted
2035 * from the space map histogram.
2038 for (int i = shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
2039 ASSERT3U(i, >=, idx + shift);
2040 histogram[idx] += rt->rt_histogram[i] << (i - idx - shift);
2042 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
2043 ASSERT3U(idx + shift, ==, i);
2045 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
2051 * Called at every sync pass that the metaslab gets synced.
2053 * The reason is that we want our auxiliary histograms to be updated
2054 * wherever the metaslab's space map histogram is updated. This way
2055 * we stay consistent on which parts of the metaslab space map's
2056 * histogram are currently not available for allocations (e.g because
2057 * they are in the defer, freed, and freeing trees).
2060 metaslab_aux_histograms_update(metaslab_t *msp)
2062 space_map_t *sm = msp->ms_sm;
2066 * This is similar to the metaslab's space map histogram updates
2067 * that take place in metaslab_sync(). The only difference is that
2068 * we only care about segments that haven't made it into the
2069 * ms_allocatable tree yet.
2071 if (msp->ms_loaded) {
2072 metaslab_aux_histograms_clear(msp);
2074 metaslab_aux_histogram_add(msp->ms_synchist,
2075 sm->sm_shift, msp->ms_freed);
2077 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2078 metaslab_aux_histogram_add(msp->ms_deferhist[t],
2079 sm->sm_shift, msp->ms_defer[t]);
2083 metaslab_aux_histogram_add(msp->ms_synchist,
2084 sm->sm_shift, msp->ms_freeing);
2088 * Called every time we are done syncing (writing to) the metaslab,
2089 * i.e. at the end of each sync pass.
2090 * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist]
2093 metaslab_aux_histograms_update_done(metaslab_t *msp, boolean_t defer_allowed)
2095 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2096 space_map_t *sm = msp->ms_sm;
2100 * We came here from metaslab_init() when creating/opening a
2101 * pool, looking at a metaslab that hasn't had any allocations
2108 * This is similar to the actions that we take for the ms_freed
2109 * and ms_defer trees in metaslab_sync_done().
2111 uint64_t hist_index = spa_syncing_txg(spa) % TXG_DEFER_SIZE;
2112 if (defer_allowed) {
2113 memcpy(msp->ms_deferhist[hist_index], msp->ms_synchist,
2114 sizeof (msp->ms_synchist));
2116 memset(msp->ms_deferhist[hist_index], 0,
2117 sizeof (msp->ms_deferhist[hist_index]));
2119 memset(msp->ms_synchist, 0, sizeof (msp->ms_synchist));
2123 * Ensure that the metaslab's weight and fragmentation are consistent
2124 * with the contents of the histogram (either the range tree's histogram
2125 * or the space map's depending whether the metaslab is loaded).
2128 metaslab_verify_weight_and_frag(metaslab_t *msp)
2130 ASSERT(MUTEX_HELD(&msp->ms_lock));
2132 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
2136 * We can end up here from vdev_remove_complete(), in which case we
2137 * cannot do these assertions because we hold spa config locks and
2138 * thus we are not allowed to read from the DMU.
2140 * We check if the metaslab group has been removed and if that's
2141 * the case we return immediately as that would mean that we are
2142 * here from the aforementioned code path.
2144 if (msp->ms_group == NULL)
2148 * Devices being removed always return a weight of 0 and leave
2149 * fragmentation and ms_max_size as is - there is nothing for
2150 * us to verify here.
2152 vdev_t *vd = msp->ms_group->mg_vd;
2153 if (vd->vdev_removing)
2157 * If the metaslab is dirty it probably means that we've done
2158 * some allocations or frees that have changed our histograms
2159 * and thus the weight.
2161 for (int t = 0; t < TXG_SIZE; t++) {
2162 if (txg_list_member(&vd->vdev_ms_list, msp, t))
2167 * This verification checks that our in-memory state is consistent
2168 * with what's on disk. If the pool is read-only then there aren't
2169 * any changes and we just have the initially-loaded state.
2171 if (!spa_writeable(msp->ms_group->mg_vd->vdev_spa))
2174 /* some extra verification for in-core tree if you can */
2175 if (msp->ms_loaded) {
2176 range_tree_stat_verify(msp->ms_allocatable);
2177 VERIFY(space_map_histogram_verify(msp->ms_sm,
2178 msp->ms_allocatable));
2181 uint64_t weight = msp->ms_weight;
2182 uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2183 boolean_t space_based = WEIGHT_IS_SPACEBASED(msp->ms_weight);
2184 uint64_t frag = msp->ms_fragmentation;
2185 uint64_t max_segsize = msp->ms_max_size;
2188 msp->ms_fragmentation = 0;
2191 * This function is used for verification purposes and thus should
2192 * not introduce any side-effects/mutations on the system's state.
2194 * Regardless of whether metaslab_weight() thinks this metaslab
2195 * should be active or not, we want to ensure that the actual weight
2196 * (and therefore the value of ms_weight) would be the same if it
2197 * was to be recalculated at this point.
2199 * In addition we set the nodirty flag so metaslab_weight() does
2200 * not dirty the metaslab for future TXGs (e.g. when trying to
2201 * force condensing to upgrade the metaslab spacemaps).
2203 msp->ms_weight = metaslab_weight(msp, B_TRUE) | was_active;
2205 VERIFY3U(max_segsize, ==, msp->ms_max_size);
2208 * If the weight type changed then there is no point in doing
2209 * verification. Revert fields to their original values.
2211 if ((space_based && !WEIGHT_IS_SPACEBASED(msp->ms_weight)) ||
2212 (!space_based && WEIGHT_IS_SPACEBASED(msp->ms_weight))) {
2213 msp->ms_fragmentation = frag;
2214 msp->ms_weight = weight;
2218 VERIFY3U(msp->ms_fragmentation, ==, frag);
2219 VERIFY3U(msp->ms_weight, ==, weight);
2223 * If we're over the zfs_metaslab_mem_limit, select the loaded metaslab from
2224 * this class that was used longest ago, and attempt to unload it. We don't
2225 * want to spend too much time in this loop to prevent performance
2226 * degradation, and we expect that most of the time this operation will
2227 * succeed. Between that and the normal unloading processing during txg sync,
2228 * we expect this to keep the metaslab memory usage under control.
2231 metaslab_potentially_evict(metaslab_class_t *mc)
2234 uint64_t allmem = arc_all_memory();
2235 uint64_t inuse = spl_kmem_cache_inuse(zfs_btree_leaf_cache);
2236 uint64_t size = spl_kmem_cache_entry_size(zfs_btree_leaf_cache);
2238 for (; allmem * zfs_metaslab_mem_limit / 100 < inuse * size &&
2239 tries < multilist_get_num_sublists(&mc->mc_metaslab_txg_list) * 2;
2241 unsigned int idx = multilist_get_random_index(
2242 &mc->mc_metaslab_txg_list);
2243 multilist_sublist_t *mls =
2244 multilist_sublist_lock(&mc->mc_metaslab_txg_list, idx);
2245 metaslab_t *msp = multilist_sublist_head(mls);
2246 multilist_sublist_unlock(mls);
2247 while (msp != NULL && allmem * zfs_metaslab_mem_limit / 100 <
2249 VERIFY3P(mls, ==, multilist_sublist_lock(
2250 &mc->mc_metaslab_txg_list, idx));
2252 metaslab_idx_func(&mc->mc_metaslab_txg_list, msp));
2254 if (!multilist_link_active(&msp->ms_class_txg_node)) {
2255 multilist_sublist_unlock(mls);
2258 metaslab_t *next_msp = multilist_sublist_next(mls, msp);
2259 multilist_sublist_unlock(mls);
2261 * If the metaslab is currently loading there are two
2262 * cases. If it's the metaslab we're evicting, we
2263 * can't continue on or we'll panic when we attempt to
2264 * recursively lock the mutex. If it's another
2265 * metaslab that's loading, it can be safely skipped,
2266 * since we know it's very new and therefore not a
2267 * good eviction candidate. We check later once the
2268 * lock is held that the metaslab is fully loaded
2269 * before actually unloading it.
2271 if (msp->ms_loading) {
2274 spl_kmem_cache_inuse(zfs_btree_leaf_cache);
2278 * We can't unload metaslabs with no spacemap because
2279 * they're not ready to be unloaded yet. We can't
2280 * unload metaslabs with outstanding allocations
2281 * because doing so could cause the metaslab's weight
2282 * to decrease while it's unloaded, which violates an
2283 * invariant that we use to prevent unnecessary
2284 * loading. We also don't unload metaslabs that are
2285 * currently active because they are high-weight
2286 * metaslabs that are likely to be used in the near
2289 mutex_enter(&msp->ms_lock);
2290 if (msp->ms_allocator == -1 && msp->ms_sm != NULL &&
2291 msp->ms_allocating_total == 0) {
2292 metaslab_unload(msp);
2294 mutex_exit(&msp->ms_lock);
2296 inuse = spl_kmem_cache_inuse(zfs_btree_leaf_cache);
2300 (void) mc, (void) zfs_metaslab_mem_limit;
2305 metaslab_load_impl(metaslab_t *msp)
2309 ASSERT(MUTEX_HELD(&msp->ms_lock));
2310 ASSERT(msp->ms_loading);
2311 ASSERT(!msp->ms_condensing);
2314 * We temporarily drop the lock to unblock other operations while we
2315 * are reading the space map. Therefore, metaslab_sync() and
2316 * metaslab_sync_done() can run at the same time as we do.
2318 * If we are using the log space maps, metaslab_sync() can't write to
2319 * the metaslab's space map while we are loading as we only write to
2320 * it when we are flushing the metaslab, and that can't happen while
2321 * we are loading it.
2323 * If we are not using log space maps though, metaslab_sync() can
2324 * append to the space map while we are loading. Therefore we load
2325 * only entries that existed when we started the load. Additionally,
2326 * metaslab_sync_done() has to wait for the load to complete because
2327 * there are potential races like metaslab_load() loading parts of the
2328 * space map that are currently being appended by metaslab_sync(). If
2329 * we didn't, the ms_allocatable would have entries that
2330 * metaslab_sync_done() would try to re-add later.
2332 * That's why before dropping the lock we remember the synced length
2333 * of the metaslab and read up to that point of the space map,
2334 * ignoring entries appended by metaslab_sync() that happen after we
2337 uint64_t length = msp->ms_synced_length;
2338 mutex_exit(&msp->ms_lock);
2340 hrtime_t load_start = gethrtime();
2341 metaslab_rt_arg_t *mrap;
2342 if (msp->ms_allocatable->rt_arg == NULL) {
2343 mrap = kmem_zalloc(sizeof (*mrap), KM_SLEEP);
2345 mrap = msp->ms_allocatable->rt_arg;
2346 msp->ms_allocatable->rt_ops = NULL;
2347 msp->ms_allocatable->rt_arg = NULL;
2349 mrap->mra_bt = &msp->ms_allocatable_by_size;
2350 mrap->mra_floor_shift = metaslab_by_size_min_shift;
2352 if (msp->ms_sm != NULL) {
2353 error = space_map_load_length(msp->ms_sm, msp->ms_allocatable,
2356 /* Now, populate the size-sorted tree. */
2357 metaslab_rt_create(msp->ms_allocatable, mrap);
2358 msp->ms_allocatable->rt_ops = &metaslab_rt_ops;
2359 msp->ms_allocatable->rt_arg = mrap;
2361 struct mssa_arg arg = {0};
2362 arg.rt = msp->ms_allocatable;
2364 range_tree_walk(msp->ms_allocatable, metaslab_size_sorted_add,
2368 * Add the size-sorted tree first, since we don't need to load
2369 * the metaslab from the spacemap.
2371 metaslab_rt_create(msp->ms_allocatable, mrap);
2372 msp->ms_allocatable->rt_ops = &metaslab_rt_ops;
2373 msp->ms_allocatable->rt_arg = mrap;
2375 * The space map has not been allocated yet, so treat
2376 * all the space in the metaslab as free and add it to the
2377 * ms_allocatable tree.
2379 range_tree_add(msp->ms_allocatable,
2380 msp->ms_start, msp->ms_size);
2384 * If the ms_sm doesn't exist, this means that this
2385 * metaslab hasn't gone through metaslab_sync() and
2386 * thus has never been dirtied. So we shouldn't
2387 * expect any unflushed allocs or frees from previous
2390 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
2391 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
2396 * We need to grab the ms_sync_lock to prevent metaslab_sync() from
2397 * changing the ms_sm (or log_sm) and the metaslab's range trees
2398 * while we are about to use them and populate the ms_allocatable.
2399 * The ms_lock is insufficient for this because metaslab_sync() doesn't
2400 * hold the ms_lock while writing the ms_checkpointing tree to disk.
2402 mutex_enter(&msp->ms_sync_lock);
2403 mutex_enter(&msp->ms_lock);
2405 ASSERT(!msp->ms_condensing);
2406 ASSERT(!msp->ms_flushing);
2409 mutex_exit(&msp->ms_sync_lock);
2413 ASSERT3P(msp->ms_group, !=, NULL);
2414 msp->ms_loaded = B_TRUE;
2417 * Apply all the unflushed changes to ms_allocatable right
2418 * away so any manipulations we do below have a clear view
2419 * of what is allocated and what is free.
2421 range_tree_walk(msp->ms_unflushed_allocs,
2422 range_tree_remove, msp->ms_allocatable);
2423 range_tree_walk(msp->ms_unflushed_frees,
2424 range_tree_add, msp->ms_allocatable);
2426 ASSERT3P(msp->ms_group, !=, NULL);
2427 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2428 if (spa_syncing_log_sm(spa) != NULL) {
2429 ASSERT(spa_feature_is_enabled(spa,
2430 SPA_FEATURE_LOG_SPACEMAP));
2433 * If we use a log space map we add all the segments
2434 * that are in ms_unflushed_frees so they are available
2437 * ms_allocatable needs to contain all free segments
2438 * that are ready for allocations (thus not segments
2439 * from ms_freeing, ms_freed, and the ms_defer trees).
2440 * But if we grab the lock in this code path at a sync
2441 * pass later that 1, then it also contains the
2442 * segments of ms_freed (they were added to it earlier
2443 * in this path through ms_unflushed_frees). So we
2444 * need to remove all the segments that exist in
2445 * ms_freed from ms_allocatable as they will be added
2446 * later in metaslab_sync_done().
2448 * When there's no log space map, the ms_allocatable
2449 * correctly doesn't contain any segments that exist
2450 * in ms_freed [see ms_synced_length].
2452 range_tree_walk(msp->ms_freed,
2453 range_tree_remove, msp->ms_allocatable);
2457 * If we are not using the log space map, ms_allocatable
2458 * contains the segments that exist in the ms_defer trees
2459 * [see ms_synced_length]. Thus we need to remove them
2460 * from ms_allocatable as they will be added again in
2461 * metaslab_sync_done().
2463 * If we are using the log space map, ms_allocatable still
2464 * contains the segments that exist in the ms_defer trees.
2465 * Not because it read them through the ms_sm though. But
2466 * because these segments are part of ms_unflushed_frees
2467 * whose segments we add to ms_allocatable earlier in this
2470 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2471 range_tree_walk(msp->ms_defer[t],
2472 range_tree_remove, msp->ms_allocatable);
2476 * Call metaslab_recalculate_weight_and_sort() now that the
2477 * metaslab is loaded so we get the metaslab's real weight.
2479 * Unless this metaslab was created with older software and
2480 * has not yet been converted to use segment-based weight, we
2481 * expect the new weight to be better or equal to the weight
2482 * that the metaslab had while it was not loaded. This is
2483 * because the old weight does not take into account the
2484 * consolidation of adjacent segments between TXGs. [see
2485 * comment for ms_synchist and ms_deferhist[] for more info]
2487 uint64_t weight = msp->ms_weight;
2488 uint64_t max_size = msp->ms_max_size;
2489 metaslab_recalculate_weight_and_sort(msp);
2490 if (!WEIGHT_IS_SPACEBASED(weight))
2491 ASSERT3U(weight, <=, msp->ms_weight);
2492 msp->ms_max_size = metaslab_largest_allocatable(msp);
2493 ASSERT3U(max_size, <=, msp->ms_max_size);
2494 hrtime_t load_end = gethrtime();
2495 msp->ms_load_time = load_end;
2496 zfs_dbgmsg("metaslab_load: txg %llu, spa %s, vdev_id %llu, "
2497 "ms_id %llu, smp_length %llu, "
2498 "unflushed_allocs %llu, unflushed_frees %llu, "
2499 "freed %llu, defer %llu + %llu, unloaded time %llu ms, "
2500 "loading_time %lld ms, ms_max_size %llu, "
2501 "max size error %lld, "
2502 "old_weight %llx, new_weight %llx",
2503 (u_longlong_t)spa_syncing_txg(spa), spa_name(spa),
2504 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
2505 (u_longlong_t)msp->ms_id,
2506 (u_longlong_t)space_map_length(msp->ms_sm),
2507 (u_longlong_t)range_tree_space(msp->ms_unflushed_allocs),
2508 (u_longlong_t)range_tree_space(msp->ms_unflushed_frees),
2509 (u_longlong_t)range_tree_space(msp->ms_freed),
2510 (u_longlong_t)range_tree_space(msp->ms_defer[0]),
2511 (u_longlong_t)range_tree_space(msp->ms_defer[1]),
2512 (longlong_t)((load_start - msp->ms_unload_time) / 1000000),
2513 (longlong_t)((load_end - load_start) / 1000000),
2514 (u_longlong_t)msp->ms_max_size,
2515 (u_longlong_t)msp->ms_max_size - max_size,
2516 (u_longlong_t)weight, (u_longlong_t)msp->ms_weight);
2518 metaslab_verify_space(msp, spa_syncing_txg(spa));
2519 mutex_exit(&msp->ms_sync_lock);
2524 metaslab_load(metaslab_t *msp)
2526 ASSERT(MUTEX_HELD(&msp->ms_lock));
2529 * There may be another thread loading the same metaslab, if that's
2530 * the case just wait until the other thread is done and return.
2532 metaslab_load_wait(msp);
2535 VERIFY(!msp->ms_loading);
2536 ASSERT(!msp->ms_condensing);
2539 * We set the loading flag BEFORE potentially dropping the lock to
2540 * wait for an ongoing flush (see ms_flushing below). This way other
2541 * threads know that there is already a thread that is loading this
2544 msp->ms_loading = B_TRUE;
2547 * Wait for any in-progress flushing to finish as we drop the ms_lock
2548 * both here (during space_map_load()) and in metaslab_flush() (when
2549 * we flush our changes to the ms_sm).
2551 if (msp->ms_flushing)
2552 metaslab_flush_wait(msp);
2555 * In the possibility that we were waiting for the metaslab to be
2556 * flushed (where we temporarily dropped the ms_lock), ensure that
2557 * no one else loaded the metaslab somehow.
2559 ASSERT(!msp->ms_loaded);
2562 * If we're loading a metaslab in the normal class, consider evicting
2563 * another one to keep our memory usage under the limit defined by the
2564 * zfs_metaslab_mem_limit tunable.
2566 if (spa_normal_class(msp->ms_group->mg_class->mc_spa) ==
2567 msp->ms_group->mg_class) {
2568 metaslab_potentially_evict(msp->ms_group->mg_class);
2571 int error = metaslab_load_impl(msp);
2573 ASSERT(MUTEX_HELD(&msp->ms_lock));
2574 msp->ms_loading = B_FALSE;
2575 cv_broadcast(&msp->ms_load_cv);
2581 metaslab_unload(metaslab_t *msp)
2583 ASSERT(MUTEX_HELD(&msp->ms_lock));
2586 * This can happen if a metaslab is selected for eviction (in
2587 * metaslab_potentially_evict) and then unloaded during spa_sync (via
2588 * metaslab_class_evict_old).
2590 if (!msp->ms_loaded)
2593 range_tree_vacate(msp->ms_allocatable, NULL, NULL);
2594 msp->ms_loaded = B_FALSE;
2595 msp->ms_unload_time = gethrtime();
2597 msp->ms_activation_weight = 0;
2598 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
2600 if (msp->ms_group != NULL) {
2601 metaslab_class_t *mc = msp->ms_group->mg_class;
2602 multilist_sublist_t *mls =
2603 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
2604 if (multilist_link_active(&msp->ms_class_txg_node))
2605 multilist_sublist_remove(mls, msp);
2606 multilist_sublist_unlock(mls);
2608 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2609 zfs_dbgmsg("metaslab_unload: txg %llu, spa %s, vdev_id %llu, "
2610 "ms_id %llu, weight %llx, "
2611 "selected txg %llu (%llu ms ago), alloc_txg %llu, "
2612 "loaded %llu ms ago, max_size %llu",
2613 (u_longlong_t)spa_syncing_txg(spa), spa_name(spa),
2614 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
2615 (u_longlong_t)msp->ms_id,
2616 (u_longlong_t)msp->ms_weight,
2617 (u_longlong_t)msp->ms_selected_txg,
2618 (u_longlong_t)(msp->ms_unload_time -
2619 msp->ms_selected_time) / 1000 / 1000,
2620 (u_longlong_t)msp->ms_alloc_txg,
2621 (u_longlong_t)(msp->ms_unload_time -
2622 msp->ms_load_time) / 1000 / 1000,
2623 (u_longlong_t)msp->ms_max_size);
2627 * We explicitly recalculate the metaslab's weight based on its space
2628 * map (as it is now not loaded). We want unload metaslabs to always
2629 * have their weights calculated from the space map histograms, while
2630 * loaded ones have it calculated from their in-core range tree
2631 * [see metaslab_load()]. This way, the weight reflects the information
2632 * available in-core, whether it is loaded or not.
2634 * If ms_group == NULL means that we came here from metaslab_fini(),
2635 * at which point it doesn't make sense for us to do the recalculation
2638 if (msp->ms_group != NULL)
2639 metaslab_recalculate_weight_and_sort(msp);
2643 * We want to optimize the memory use of the per-metaslab range
2644 * trees. To do this, we store the segments in the range trees in
2645 * units of sectors, zero-indexing from the start of the metaslab. If
2646 * the vdev_ms_shift - the vdev_ashift is less than 32, we can store
2647 * the ranges using two uint32_ts, rather than two uint64_ts.
2650 metaslab_calculate_range_tree_type(vdev_t *vdev, metaslab_t *msp,
2651 uint64_t *start, uint64_t *shift)
2653 if (vdev->vdev_ms_shift - vdev->vdev_ashift < 32 &&
2654 !zfs_metaslab_force_large_segs) {
2655 *shift = vdev->vdev_ashift;
2656 *start = msp->ms_start;
2657 return (RANGE_SEG32);
2661 return (RANGE_SEG64);
2666 metaslab_set_selected_txg(metaslab_t *msp, uint64_t txg)
2668 ASSERT(MUTEX_HELD(&msp->ms_lock));
2669 metaslab_class_t *mc = msp->ms_group->mg_class;
2670 multilist_sublist_t *mls =
2671 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
2672 if (multilist_link_active(&msp->ms_class_txg_node))
2673 multilist_sublist_remove(mls, msp);
2674 msp->ms_selected_txg = txg;
2675 msp->ms_selected_time = gethrtime();
2676 multilist_sublist_insert_tail(mls, msp);
2677 multilist_sublist_unlock(mls);
2681 metaslab_space_update(vdev_t *vd, metaslab_class_t *mc, int64_t alloc_delta,
2682 int64_t defer_delta, int64_t space_delta)
2684 vdev_space_update(vd, alloc_delta, defer_delta, space_delta);
2686 ASSERT3P(vd->vdev_spa->spa_root_vdev, ==, vd->vdev_parent);
2687 ASSERT(vd->vdev_ms_count != 0);
2689 metaslab_class_space_update(mc, alloc_delta, defer_delta, space_delta,
2690 vdev_deflated_space(vd, space_delta));
2694 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object,
2695 uint64_t txg, metaslab_t **msp)
2697 vdev_t *vd = mg->mg_vd;
2698 spa_t *spa = vd->vdev_spa;
2699 objset_t *mos = spa->spa_meta_objset;
2703 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
2704 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
2705 mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL);
2706 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
2707 cv_init(&ms->ms_flush_cv, NULL, CV_DEFAULT, NULL);
2708 multilist_link_init(&ms->ms_class_txg_node);
2711 ms->ms_start = id << vd->vdev_ms_shift;
2712 ms->ms_size = 1ULL << vd->vdev_ms_shift;
2713 ms->ms_allocator = -1;
2714 ms->ms_new = B_TRUE;
2716 vdev_ops_t *ops = vd->vdev_ops;
2717 if (ops->vdev_op_metaslab_init != NULL)
2718 ops->vdev_op_metaslab_init(vd, &ms->ms_start, &ms->ms_size);
2721 * We only open space map objects that already exist. All others
2722 * will be opened when we finally allocate an object for it. For
2723 * readonly pools there is no need to open the space map object.
2726 * When called from vdev_expand(), we can't call into the DMU as
2727 * we are holding the spa_config_lock as a writer and we would
2728 * deadlock [see relevant comment in vdev_metaslab_init()]. in
2729 * that case, the object parameter is zero though, so we won't
2730 * call into the DMU.
2732 if (object != 0 && !(spa->spa_mode == SPA_MODE_READ &&
2733 !spa->spa_read_spacemaps)) {
2734 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
2735 ms->ms_size, vd->vdev_ashift);
2738 kmem_free(ms, sizeof (metaslab_t));
2742 ASSERT(ms->ms_sm != NULL);
2743 ms->ms_allocated_space = space_map_allocated(ms->ms_sm);
2746 uint64_t shift, start;
2747 range_seg_type_t type =
2748 metaslab_calculate_range_tree_type(vd, ms, &start, &shift);
2750 ms->ms_allocatable = range_tree_create(NULL, type, NULL, start, shift);
2751 for (int t = 0; t < TXG_SIZE; t++) {
2752 ms->ms_allocating[t] = range_tree_create(NULL, type,
2753 NULL, start, shift);
2755 ms->ms_freeing = range_tree_create(NULL, type, NULL, start, shift);
2756 ms->ms_freed = range_tree_create(NULL, type, NULL, start, shift);
2757 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2758 ms->ms_defer[t] = range_tree_create(NULL, type, NULL,
2761 ms->ms_checkpointing =
2762 range_tree_create(NULL, type, NULL, start, shift);
2763 ms->ms_unflushed_allocs =
2764 range_tree_create(NULL, type, NULL, start, shift);
2766 metaslab_rt_arg_t *mrap = kmem_zalloc(sizeof (*mrap), KM_SLEEP);
2767 mrap->mra_bt = &ms->ms_unflushed_frees_by_size;
2768 mrap->mra_floor_shift = metaslab_by_size_min_shift;
2769 ms->ms_unflushed_frees = range_tree_create(&metaslab_rt_ops,
2770 type, mrap, start, shift);
2772 ms->ms_trim = range_tree_create(NULL, type, NULL, start, shift);
2774 metaslab_group_add(mg, ms);
2775 metaslab_set_fragmentation(ms, B_FALSE);
2778 * If we're opening an existing pool (txg == 0) or creating
2779 * a new one (txg == TXG_INITIAL), all space is available now.
2780 * If we're adding space to an existing pool, the new space
2781 * does not become available until after this txg has synced.
2782 * The metaslab's weight will also be initialized when we sync
2783 * out this txg. This ensures that we don't attempt to allocate
2784 * from it before we have initialized it completely.
2786 if (txg <= TXG_INITIAL) {
2787 metaslab_sync_done(ms, 0);
2788 metaslab_space_update(vd, mg->mg_class,
2789 metaslab_allocated_space(ms), 0, 0);
2793 vdev_dirty(vd, 0, NULL, txg);
2794 vdev_dirty(vd, VDD_METASLAB, ms, txg);
2803 metaslab_fini_flush_data(metaslab_t *msp)
2805 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2807 if (metaslab_unflushed_txg(msp) == 0) {
2808 ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL),
2812 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
2814 mutex_enter(&spa->spa_flushed_ms_lock);
2815 avl_remove(&spa->spa_metaslabs_by_flushed, msp);
2816 mutex_exit(&spa->spa_flushed_ms_lock);
2818 spa_log_sm_decrement_mscount(spa, metaslab_unflushed_txg(msp));
2819 spa_log_summary_decrement_mscount(spa, metaslab_unflushed_txg(msp),
2820 metaslab_unflushed_dirty(msp));
2824 metaslab_unflushed_changes_memused(metaslab_t *ms)
2826 return ((range_tree_numsegs(ms->ms_unflushed_allocs) +
2827 range_tree_numsegs(ms->ms_unflushed_frees)) *
2828 ms->ms_unflushed_allocs->rt_root.bt_elem_size);
2832 metaslab_fini(metaslab_t *msp)
2834 metaslab_group_t *mg = msp->ms_group;
2835 vdev_t *vd = mg->mg_vd;
2836 spa_t *spa = vd->vdev_spa;
2838 metaslab_fini_flush_data(msp);
2840 metaslab_group_remove(mg, msp);
2842 mutex_enter(&msp->ms_lock);
2843 VERIFY(msp->ms_group == NULL);
2846 * If this metaslab hasn't been through metaslab_sync_done() yet its
2847 * space hasn't been accounted for in its vdev and doesn't need to be
2851 metaslab_space_update(vd, mg->mg_class,
2852 -metaslab_allocated_space(msp), 0, -msp->ms_size);
2855 space_map_close(msp->ms_sm);
2858 metaslab_unload(msp);
2860 range_tree_destroy(msp->ms_allocatable);
2861 range_tree_destroy(msp->ms_freeing);
2862 range_tree_destroy(msp->ms_freed);
2864 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
2865 metaslab_unflushed_changes_memused(msp));
2866 spa->spa_unflushed_stats.sus_memused -=
2867 metaslab_unflushed_changes_memused(msp);
2868 range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
2869 range_tree_destroy(msp->ms_unflushed_allocs);
2870 range_tree_destroy(msp->ms_checkpointing);
2871 range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
2872 range_tree_destroy(msp->ms_unflushed_frees);
2874 for (int t = 0; t < TXG_SIZE; t++) {
2875 range_tree_destroy(msp->ms_allocating[t]);
2877 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2878 range_tree_destroy(msp->ms_defer[t]);
2880 ASSERT0(msp->ms_deferspace);
2882 for (int t = 0; t < TXG_SIZE; t++)
2883 ASSERT(!txg_list_member(&vd->vdev_ms_list, msp, t));
2885 range_tree_vacate(msp->ms_trim, NULL, NULL);
2886 range_tree_destroy(msp->ms_trim);
2888 mutex_exit(&msp->ms_lock);
2889 cv_destroy(&msp->ms_load_cv);
2890 cv_destroy(&msp->ms_flush_cv);
2891 mutex_destroy(&msp->ms_lock);
2892 mutex_destroy(&msp->ms_sync_lock);
2893 ASSERT3U(msp->ms_allocator, ==, -1);
2895 kmem_free(msp, sizeof (metaslab_t));
2898 #define FRAGMENTATION_TABLE_SIZE 17
2901 * This table defines a segment size based fragmentation metric that will
2902 * allow each metaslab to derive its own fragmentation value. This is done
2903 * by calculating the space in each bucket of the spacemap histogram and
2904 * multiplying that by the fragmentation metric in this table. Doing
2905 * this for all buckets and dividing it by the total amount of free
2906 * space in this metaslab (i.e. the total free space in all buckets) gives
2907 * us the fragmentation metric. This means that a high fragmentation metric
2908 * equates to most of the free space being comprised of small segments.
2909 * Conversely, if the metric is low, then most of the free space is in
2910 * large segments. A 10% change in fragmentation equates to approximately
2911 * double the number of segments.
2913 * This table defines 0% fragmented space using 16MB segments. Testing has
2914 * shown that segments that are greater than or equal to 16MB do not suffer
2915 * from drastic performance problems. Using this value, we derive the rest
2916 * of the table. Since the fragmentation value is never stored on disk, it
2917 * is possible to change these calculations in the future.
2919 static const int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
2939 * Calculate the metaslab's fragmentation metric and set ms_fragmentation.
2940 * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not
2941 * been upgraded and does not support this metric. Otherwise, the return
2942 * value should be in the range [0, 100].
2945 metaslab_set_fragmentation(metaslab_t *msp, boolean_t nodirty)
2947 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2948 uint64_t fragmentation = 0;
2950 boolean_t feature_enabled = spa_feature_is_enabled(spa,
2951 SPA_FEATURE_SPACEMAP_HISTOGRAM);
2953 if (!feature_enabled) {
2954 msp->ms_fragmentation = ZFS_FRAG_INVALID;
2959 * A null space map means that the entire metaslab is free
2960 * and thus is not fragmented.
2962 if (msp->ms_sm == NULL) {
2963 msp->ms_fragmentation = 0;
2968 * If this metaslab's space map has not been upgraded, flag it
2969 * so that we upgrade next time we encounter it.
2971 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
2972 uint64_t txg = spa_syncing_txg(spa);
2973 vdev_t *vd = msp->ms_group->mg_vd;
2976 * If we've reached the final dirty txg, then we must
2977 * be shutting down the pool. We don't want to dirty
2978 * any data past this point so skip setting the condense
2979 * flag. We can retry this action the next time the pool
2980 * is imported. We also skip marking this metaslab for
2981 * condensing if the caller has explicitly set nodirty.
2984 spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) {
2985 msp->ms_condense_wanted = B_TRUE;
2986 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2987 zfs_dbgmsg("txg %llu, requesting force condense: "
2988 "ms_id %llu, vdev_id %llu", (u_longlong_t)txg,
2989 (u_longlong_t)msp->ms_id,
2990 (u_longlong_t)vd->vdev_id);
2992 msp->ms_fragmentation = ZFS_FRAG_INVALID;
2996 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
2998 uint8_t shift = msp->ms_sm->sm_shift;
3000 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
3001 FRAGMENTATION_TABLE_SIZE - 1);
3003 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
3006 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
3009 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
3010 fragmentation += space * zfs_frag_table[idx];
3014 fragmentation /= total;
3015 ASSERT3U(fragmentation, <=, 100);
3017 msp->ms_fragmentation = fragmentation;
3021 * Compute a weight -- a selection preference value -- for the given metaslab.
3022 * This is based on the amount of free space, the level of fragmentation,
3023 * the LBA range, and whether the metaslab is loaded.
3026 metaslab_space_weight(metaslab_t *msp)
3028 metaslab_group_t *mg = msp->ms_group;
3029 vdev_t *vd = mg->mg_vd;
3030 uint64_t weight, space;
3032 ASSERT(MUTEX_HELD(&msp->ms_lock));
3035 * The baseline weight is the metaslab's free space.
3037 space = msp->ms_size - metaslab_allocated_space(msp);
3039 if (metaslab_fragmentation_factor_enabled &&
3040 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
3042 * Use the fragmentation information to inversely scale
3043 * down the baseline weight. We need to ensure that we
3044 * don't exclude this metaslab completely when it's 100%
3045 * fragmented. To avoid this we reduce the fragmented value
3048 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
3051 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
3052 * this metaslab again. The fragmentation metric may have
3053 * decreased the space to something smaller than
3054 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
3055 * so that we can consume any remaining space.
3057 if (space > 0 && space < SPA_MINBLOCKSIZE)
3058 space = SPA_MINBLOCKSIZE;
3063 * Modern disks have uniform bit density and constant angular velocity.
3064 * Therefore, the outer recording zones are faster (higher bandwidth)
3065 * than the inner zones by the ratio of outer to inner track diameter,
3066 * which is typically around 2:1. We account for this by assigning
3067 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
3068 * In effect, this means that we'll select the metaslab with the most
3069 * free bandwidth rather than simply the one with the most free space.
3071 if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
3072 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
3073 ASSERT(weight >= space && weight <= 2 * space);
3077 * If this metaslab is one we're actively using, adjust its
3078 * weight to make it preferable to any inactive metaslab so
3079 * we'll polish it off. If the fragmentation on this metaslab
3080 * has exceed our threshold, then don't mark it active.
3082 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
3083 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
3084 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
3087 WEIGHT_SET_SPACEBASED(weight);
3092 * Return the weight of the specified metaslab, according to the segment-based
3093 * weighting algorithm. The metaslab must be loaded. This function can
3094 * be called within a sync pass since it relies only on the metaslab's
3095 * range tree which is always accurate when the metaslab is loaded.
3098 metaslab_weight_from_range_tree(metaslab_t *msp)
3100 uint64_t weight = 0;
3101 uint32_t segments = 0;
3103 ASSERT(msp->ms_loaded);
3105 for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
3107 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
3108 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
3111 segments += msp->ms_allocatable->rt_histogram[i];
3114 * The range tree provides more precision than the space map
3115 * and must be downgraded so that all values fit within the
3116 * space map's histogram. This allows us to compare loaded
3117 * vs. unloaded metaslabs to determine which metaslab is
3118 * considered "best".
3123 if (segments != 0) {
3124 WEIGHT_SET_COUNT(weight, segments);
3125 WEIGHT_SET_INDEX(weight, i);
3126 WEIGHT_SET_ACTIVE(weight, 0);
3134 * Calculate the weight based on the on-disk histogram. Should be applied
3135 * only to unloaded metaslabs (i.e no incoming allocations) in-order to
3136 * give results consistent with the on-disk state
3139 metaslab_weight_from_spacemap(metaslab_t *msp)
3141 space_map_t *sm = msp->ms_sm;
3142 ASSERT(!msp->ms_loaded);
3144 ASSERT3U(space_map_object(sm), !=, 0);
3145 ASSERT3U(sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
3148 * Create a joint histogram from all the segments that have made
3149 * it to the metaslab's space map histogram, that are not yet
3150 * available for allocation because they are still in the freeing
3151 * pipeline (e.g. freeing, freed, and defer trees). Then subtract
3152 * these segments from the space map's histogram to get a more
3155 uint64_t deferspace_histogram[SPACE_MAP_HISTOGRAM_SIZE] = {0};
3156 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
3157 deferspace_histogram[i] += msp->ms_synchist[i];
3158 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3159 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
3160 deferspace_histogram[i] += msp->ms_deferhist[t][i];
3164 uint64_t weight = 0;
3165 for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
3166 ASSERT3U(sm->sm_phys->smp_histogram[i], >=,
3167 deferspace_histogram[i]);
3169 sm->sm_phys->smp_histogram[i] - deferspace_histogram[i];
3171 WEIGHT_SET_COUNT(weight, count);
3172 WEIGHT_SET_INDEX(weight, i + sm->sm_shift);
3173 WEIGHT_SET_ACTIVE(weight, 0);
3181 * Compute a segment-based weight for the specified metaslab. The weight
3182 * is determined by highest bucket in the histogram. The information
3183 * for the highest bucket is encoded into the weight value.
3186 metaslab_segment_weight(metaslab_t *msp)
3188 metaslab_group_t *mg = msp->ms_group;
3189 uint64_t weight = 0;
3190 uint8_t shift = mg->mg_vd->vdev_ashift;
3192 ASSERT(MUTEX_HELD(&msp->ms_lock));
3195 * The metaslab is completely free.
3197 if (metaslab_allocated_space(msp) == 0) {
3198 int idx = highbit64(msp->ms_size) - 1;
3199 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
3201 if (idx < max_idx) {
3202 WEIGHT_SET_COUNT(weight, 1ULL);
3203 WEIGHT_SET_INDEX(weight, idx);
3205 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
3206 WEIGHT_SET_INDEX(weight, max_idx);
3208 WEIGHT_SET_ACTIVE(weight, 0);
3209 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
3213 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
3216 * If the metaslab is fully allocated then just make the weight 0.
3218 if (metaslab_allocated_space(msp) == msp->ms_size)
3221 * If the metaslab is already loaded, then use the range tree to
3222 * determine the weight. Otherwise, we rely on the space map information
3223 * to generate the weight.
3225 if (msp->ms_loaded) {
3226 weight = metaslab_weight_from_range_tree(msp);
3228 weight = metaslab_weight_from_spacemap(msp);
3232 * If the metaslab was active the last time we calculated its weight
3233 * then keep it active. We want to consume the entire region that
3234 * is associated with this weight.
3236 if (msp->ms_activation_weight != 0 && weight != 0)
3237 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
3242 * Determine if we should attempt to allocate from this metaslab. If the
3243 * metaslab is loaded, then we can determine if the desired allocation
3244 * can be satisfied by looking at the size of the maximum free segment
3245 * on that metaslab. Otherwise, we make our decision based on the metaslab's
3246 * weight. For segment-based weighting we can determine the maximum
3247 * allocation based on the index encoded in its value. For space-based
3248 * weights we rely on the entire weight (excluding the weight-type bit).
3251 metaslab_should_allocate(metaslab_t *msp, uint64_t asize, boolean_t try_hard)
3254 * If the metaslab is loaded, ms_max_size is definitive and we can use
3255 * the fast check. If it's not, the ms_max_size is a lower bound (once
3256 * set), and we should use the fast check as long as we're not in
3257 * try_hard and it's been less than zfs_metaslab_max_size_cache_sec
3258 * seconds since the metaslab was unloaded.
3260 if (msp->ms_loaded ||
3261 (msp->ms_max_size != 0 && !try_hard && gethrtime() <
3262 msp->ms_unload_time + SEC2NSEC(zfs_metaslab_max_size_cache_sec)))
3263 return (msp->ms_max_size >= asize);
3265 boolean_t should_allocate;
3266 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
3268 * The metaslab segment weight indicates segments in the
3269 * range [2^i, 2^(i+1)), where i is the index in the weight.
3270 * Since the asize might be in the middle of the range, we
3271 * should attempt the allocation if asize < 2^(i+1).
3273 should_allocate = (asize <
3274 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
3276 should_allocate = (asize <=
3277 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
3280 return (should_allocate);
3284 metaslab_weight(metaslab_t *msp, boolean_t nodirty)
3286 vdev_t *vd = msp->ms_group->mg_vd;
3287 spa_t *spa = vd->vdev_spa;
3290 ASSERT(MUTEX_HELD(&msp->ms_lock));
3292 metaslab_set_fragmentation(msp, nodirty);
3295 * Update the maximum size. If the metaslab is loaded, this will
3296 * ensure that we get an accurate maximum size if newly freed space
3297 * has been added back into the free tree. If the metaslab is
3298 * unloaded, we check if there's a larger free segment in the
3299 * unflushed frees. This is a lower bound on the largest allocatable
3300 * segment size. Coalescing of adjacent entries may reveal larger
3301 * allocatable segments, but we aren't aware of those until loading
3302 * the space map into a range tree.
3304 if (msp->ms_loaded) {
3305 msp->ms_max_size = metaslab_largest_allocatable(msp);
3307 msp->ms_max_size = MAX(msp->ms_max_size,
3308 metaslab_largest_unflushed_free(msp));
3312 * Segment-based weighting requires space map histogram support.
3314 if (zfs_metaslab_segment_weight_enabled &&
3315 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
3316 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
3317 sizeof (space_map_phys_t))) {
3318 weight = metaslab_segment_weight(msp);
3320 weight = metaslab_space_weight(msp);
3326 metaslab_recalculate_weight_and_sort(metaslab_t *msp)
3328 ASSERT(MUTEX_HELD(&msp->ms_lock));
3330 /* note: we preserve the mask (e.g. indication of primary, etc..) */
3331 uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
3332 metaslab_group_sort(msp->ms_group, msp,
3333 metaslab_weight(msp, B_FALSE) | was_active);
3337 metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
3338 int allocator, uint64_t activation_weight)
3340 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
3341 ASSERT(MUTEX_HELD(&msp->ms_lock));
3344 * If we're activating for the claim code, we don't want to actually
3345 * set the metaslab up for a specific allocator.
3347 if (activation_weight == METASLAB_WEIGHT_CLAIM) {
3348 ASSERT0(msp->ms_activation_weight);
3349 msp->ms_activation_weight = msp->ms_weight;
3350 metaslab_group_sort(mg, msp, msp->ms_weight |
3355 metaslab_t **mspp = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
3356 &mga->mga_primary : &mga->mga_secondary);
3358 mutex_enter(&mg->mg_lock);
3359 if (*mspp != NULL) {
3360 mutex_exit(&mg->mg_lock);
3365 ASSERT3S(msp->ms_allocator, ==, -1);
3366 msp->ms_allocator = allocator;
3367 msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY);
3369 ASSERT0(msp->ms_activation_weight);
3370 msp->ms_activation_weight = msp->ms_weight;
3371 metaslab_group_sort_impl(mg, msp,
3372 msp->ms_weight | activation_weight);
3373 mutex_exit(&mg->mg_lock);
3379 metaslab_activate(metaslab_t *msp, int allocator, uint64_t activation_weight)
3381 ASSERT(MUTEX_HELD(&msp->ms_lock));
3384 * The current metaslab is already activated for us so there
3385 * is nothing to do. Already activated though, doesn't mean
3386 * that this metaslab is activated for our allocator nor our
3387 * requested activation weight. The metaslab could have started
3388 * as an active one for our allocator but changed allocators
3389 * while we were waiting to grab its ms_lock or we stole it
3390 * [see find_valid_metaslab()]. This means that there is a
3391 * possibility of passivating a metaslab of another allocator
3392 * or from a different activation mask, from this thread.
3394 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
3395 ASSERT(msp->ms_loaded);
3399 int error = metaslab_load(msp);
3401 metaslab_group_sort(msp->ms_group, msp, 0);
3406 * When entering metaslab_load() we may have dropped the
3407 * ms_lock because we were loading this metaslab, or we
3408 * were waiting for another thread to load it for us. In
3409 * that scenario, we recheck the weight of the metaslab
3410 * to see if it was activated by another thread.
3412 * If the metaslab was activated for another allocator or
3413 * it was activated with a different activation weight (e.g.
3414 * we wanted to make it a primary but it was activated as
3415 * secondary) we return error (EBUSY).
3417 * If the metaslab was activated for the same allocator
3418 * and requested activation mask, skip activating it.
3420 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
3421 if (msp->ms_allocator != allocator)
3424 if ((msp->ms_weight & activation_weight) == 0)
3425 return (SET_ERROR(EBUSY));
3427 EQUIV((activation_weight == METASLAB_WEIGHT_PRIMARY),
3433 * If the metaslab has literally 0 space, it will have weight 0. In
3434 * that case, don't bother activating it. This can happen if the
3435 * metaslab had space during find_valid_metaslab, but another thread
3436 * loaded it and used all that space while we were waiting to grab the
3439 if (msp->ms_weight == 0) {
3440 ASSERT0(range_tree_space(msp->ms_allocatable));
3441 return (SET_ERROR(ENOSPC));
3444 if ((error = metaslab_activate_allocator(msp->ms_group, msp,
3445 allocator, activation_weight)) != 0) {
3449 ASSERT(msp->ms_loaded);
3450 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
3456 metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp,
3459 ASSERT(MUTEX_HELD(&msp->ms_lock));
3460 ASSERT(msp->ms_loaded);
3462 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
3463 metaslab_group_sort(mg, msp, weight);
3467 mutex_enter(&mg->mg_lock);
3468 ASSERT3P(msp->ms_group, ==, mg);
3469 ASSERT3S(0, <=, msp->ms_allocator);
3470 ASSERT3U(msp->ms_allocator, <, mg->mg_allocators);
3472 metaslab_group_allocator_t *mga = &mg->mg_allocator[msp->ms_allocator];
3473 if (msp->ms_primary) {
3474 ASSERT3P(mga->mga_primary, ==, msp);
3475 ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
3476 mga->mga_primary = NULL;
3478 ASSERT3P(mga->mga_secondary, ==, msp);
3479 ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
3480 mga->mga_secondary = NULL;
3482 msp->ms_allocator = -1;
3483 metaslab_group_sort_impl(mg, msp, weight);
3484 mutex_exit(&mg->mg_lock);
3488 metaslab_passivate(metaslab_t *msp, uint64_t weight)
3490 uint64_t size __maybe_unused = weight & ~METASLAB_WEIGHT_TYPE;
3493 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
3494 * this metaslab again. In that case, it had better be empty,
3495 * or we would be leaving space on the table.
3497 ASSERT(!WEIGHT_IS_SPACEBASED(msp->ms_weight) ||
3498 size >= SPA_MINBLOCKSIZE ||
3499 range_tree_space(msp->ms_allocatable) == 0);
3500 ASSERT0(weight & METASLAB_ACTIVE_MASK);
3502 ASSERT(msp->ms_activation_weight != 0);
3503 msp->ms_activation_weight = 0;
3504 metaslab_passivate_allocator(msp->ms_group, msp, weight);
3505 ASSERT0(msp->ms_weight & METASLAB_ACTIVE_MASK);
3509 * Segment-based metaslabs are activated once and remain active until
3510 * we either fail an allocation attempt (similar to space-based metaslabs)
3511 * or have exhausted the free space in zfs_metaslab_switch_threshold
3512 * buckets since the metaslab was activated. This function checks to see
3513 * if we've exhausted the zfs_metaslab_switch_threshold buckets in the
3514 * metaslab and passivates it proactively. This will allow us to select a
3515 * metaslab with a larger contiguous region, if any, remaining within this
3516 * metaslab group. If we're in sync pass > 1, then we continue using this
3517 * metaslab so that we don't dirty more block and cause more sync passes.
3520 metaslab_segment_may_passivate(metaslab_t *msp)
3522 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3524 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
3528 * Since we are in the middle of a sync pass, the most accurate
3529 * information that is accessible to us is the in-core range tree
3530 * histogram; calculate the new weight based on that information.
3532 uint64_t weight = metaslab_weight_from_range_tree(msp);
3533 int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
3534 int current_idx = WEIGHT_GET_INDEX(weight);
3536 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
3537 metaslab_passivate(msp, weight);
3541 metaslab_preload(void *arg)
3543 metaslab_t *msp = arg;
3544 metaslab_class_t *mc = msp->ms_group->mg_class;
3545 spa_t *spa = mc->mc_spa;
3546 fstrans_cookie_t cookie = spl_fstrans_mark();
3548 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
3550 mutex_enter(&msp->ms_lock);
3551 (void) metaslab_load(msp);
3552 metaslab_set_selected_txg(msp, spa_syncing_txg(spa));
3553 mutex_exit(&msp->ms_lock);
3554 spl_fstrans_unmark(cookie);
3558 metaslab_group_preload(metaslab_group_t *mg)
3560 spa_t *spa = mg->mg_vd->vdev_spa;
3562 avl_tree_t *t = &mg->mg_metaslab_tree;
3565 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
3566 taskq_wait_outstanding(mg->mg_taskq, 0);
3570 mutex_enter(&mg->mg_lock);
3573 * Load the next potential metaslabs
3575 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
3576 ASSERT3P(msp->ms_group, ==, mg);
3579 * We preload only the maximum number of metaslabs specified
3580 * by metaslab_preload_limit. If a metaslab is being forced
3581 * to condense then we preload it too. This will ensure
3582 * that force condensing happens in the next txg.
3584 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
3588 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
3589 msp, TQ_SLEEP) != TASKQID_INVALID);
3591 mutex_exit(&mg->mg_lock);
3595 * Determine if the space map's on-disk footprint is past our tolerance for
3596 * inefficiency. We would like to use the following criteria to make our
3599 * 1. Do not condense if the size of the space map object would dramatically
3600 * increase as a result of writing out the free space range tree.
3602 * 2. Condense if the on on-disk space map representation is at least
3603 * zfs_condense_pct/100 times the size of the optimal representation
3604 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, optimal = 1.1MB).
3606 * 3. Do not condense if the on-disk size of the space map does not actually
3609 * Unfortunately, we cannot compute the on-disk size of the space map in this
3610 * context because we cannot accurately compute the effects of compression, etc.
3611 * Instead, we apply the heuristic described in the block comment for
3612 * zfs_metaslab_condense_block_threshold - we only condense if the space used
3613 * is greater than a threshold number of blocks.
3616 metaslab_should_condense(metaslab_t *msp)
3618 space_map_t *sm = msp->ms_sm;
3619 vdev_t *vd = msp->ms_group->mg_vd;
3620 uint64_t vdev_blocksize = 1ULL << vd->vdev_ashift;
3622 ASSERT(MUTEX_HELD(&msp->ms_lock));
3623 ASSERT(msp->ms_loaded);
3625 ASSERT3U(spa_sync_pass(vd->vdev_spa), ==, 1);
3628 * We always condense metaslabs that are empty and metaslabs for
3629 * which a condense request has been made.
3631 if (range_tree_numsegs(msp->ms_allocatable) == 0 ||
3632 msp->ms_condense_wanted)
3635 uint64_t record_size = MAX(sm->sm_blksz, vdev_blocksize);
3636 uint64_t object_size = space_map_length(sm);
3637 uint64_t optimal_size = space_map_estimate_optimal_size(sm,
3638 msp->ms_allocatable, SM_NO_VDEVID);
3640 return (object_size >= (optimal_size * zfs_condense_pct / 100) &&
3641 object_size > zfs_metaslab_condense_block_threshold * record_size);
3645 * Condense the on-disk space map representation to its minimized form.
3646 * The minimized form consists of a small number of allocations followed
3647 * by the entries of the free range tree (ms_allocatable). The condensed
3648 * spacemap contains all the entries of previous TXGs (including those in
3649 * the pool-wide log spacemaps; thus this is effectively a superset of
3650 * metaslab_flush()), but this TXG's entries still need to be written.
3653 metaslab_condense(metaslab_t *msp, dmu_tx_t *tx)
3655 range_tree_t *condense_tree;
3656 space_map_t *sm = msp->ms_sm;
3657 uint64_t txg = dmu_tx_get_txg(tx);
3658 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3660 ASSERT(MUTEX_HELD(&msp->ms_lock));
3661 ASSERT(msp->ms_loaded);
3662 ASSERT(msp->ms_sm != NULL);
3665 * In order to condense the space map, we need to change it so it
3666 * only describes which segments are currently allocated and free.
3668 * All the current free space resides in the ms_allocatable, all
3669 * the ms_defer trees, and all the ms_allocating trees. We ignore
3670 * ms_freed because it is empty because we're in sync pass 1. We
3671 * ignore ms_freeing because these changes are not yet reflected
3672 * in the spacemap (they will be written later this txg).
3674 * So to truncate the space map to represent all the entries of
3675 * previous TXGs we do the following:
3677 * 1] We create a range tree (condense tree) that is 100% empty.
3678 * 2] We add to it all segments found in the ms_defer trees
3679 * as those segments are marked as free in the original space
3680 * map. We do the same with the ms_allocating trees for the same
3681 * reason. Adding these segments should be a relatively
3682 * inexpensive operation since we expect these trees to have a
3683 * small number of nodes.
3684 * 3] We vacate any unflushed allocs, since they are not frees we
3685 * need to add to the condense tree. Then we vacate any
3686 * unflushed frees as they should already be part of ms_allocatable.
3687 * 4] At this point, we would ideally like to add all segments
3688 * in the ms_allocatable tree from the condense tree. This way
3689 * we would write all the entries of the condense tree as the
3690 * condensed space map, which would only contain freed
3691 * segments with everything else assumed to be allocated.
3693 * Doing so can be prohibitively expensive as ms_allocatable can
3694 * be large, and therefore computationally expensive to add to
3695 * the condense_tree. Instead we first sync out an entry marking
3696 * everything as allocated, then the condense_tree and then the
3697 * ms_allocatable, in the condensed space map. While this is not
3698 * optimal, it is typically close to optimal and more importantly
3699 * much cheaper to compute.
3701 * 5] Finally, as both of the unflushed trees were written to our
3702 * new and condensed metaslab space map, we basically flushed
3703 * all the unflushed changes to disk, thus we call
3704 * metaslab_flush_update().
3706 ASSERT3U(spa_sync_pass(spa), ==, 1);
3707 ASSERT(range_tree_is_empty(msp->ms_freed)); /* since it is pass 1 */
3709 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %px, vdev id %llu, "
3710 "spa %s, smp size %llu, segments %llu, forcing condense=%s",
3711 (u_longlong_t)txg, (u_longlong_t)msp->ms_id, msp,
3712 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
3713 spa->spa_name, (u_longlong_t)space_map_length(msp->ms_sm),
3714 (u_longlong_t)range_tree_numsegs(msp->ms_allocatable),
3715 msp->ms_condense_wanted ? "TRUE" : "FALSE");
3717 msp->ms_condense_wanted = B_FALSE;
3719 range_seg_type_t type;
3720 uint64_t shift, start;
3721 type = metaslab_calculate_range_tree_type(msp->ms_group->mg_vd, msp,
3724 condense_tree = range_tree_create(NULL, type, NULL, start, shift);
3726 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3727 range_tree_walk(msp->ms_defer[t],
3728 range_tree_add, condense_tree);
3731 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
3732 range_tree_walk(msp->ms_allocating[(txg + t) & TXG_MASK],
3733 range_tree_add, condense_tree);
3736 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3737 metaslab_unflushed_changes_memused(msp));
3738 spa->spa_unflushed_stats.sus_memused -=
3739 metaslab_unflushed_changes_memused(msp);
3740 range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
3741 range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
3744 * We're about to drop the metaslab's lock thus allowing other
3745 * consumers to change it's content. Set the metaslab's ms_condensing
3746 * flag to ensure that allocations on this metaslab do not occur
3747 * while we're in the middle of committing it to disk. This is only
3748 * critical for ms_allocatable as all other range trees use per TXG
3749 * views of their content.
3751 msp->ms_condensing = B_TRUE;
3753 mutex_exit(&msp->ms_lock);
3754 uint64_t object = space_map_object(msp->ms_sm);
3755 space_map_truncate(sm,
3756 spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ?
3757 zfs_metaslab_sm_blksz_with_log : zfs_metaslab_sm_blksz_no_log, tx);
3760 * space_map_truncate() may have reallocated the spacemap object.
3761 * If so, update the vdev_ms_array.
3763 if (space_map_object(msp->ms_sm) != object) {
3764 object = space_map_object(msp->ms_sm);
3765 dmu_write(spa->spa_meta_objset,
3766 msp->ms_group->mg_vd->vdev_ms_array, sizeof (uint64_t) *
3767 msp->ms_id, sizeof (uint64_t), &object, tx);
3772 * When the log space map feature is enabled, each space map will
3773 * always have ALLOCS followed by FREES for each sync pass. This is
3774 * typically true even when the log space map feature is disabled,
3775 * except from the case where a metaslab goes through metaslab_sync()
3776 * and gets condensed. In that case the metaslab's space map will have
3777 * ALLOCS followed by FREES (due to condensing) followed by ALLOCS
3778 * followed by FREES (due to space_map_write() in metaslab_sync()) for
3781 range_tree_t *tmp_tree = range_tree_create(NULL, type, NULL, start,
3783 range_tree_add(tmp_tree, msp->ms_start, msp->ms_size);
3784 space_map_write(sm, tmp_tree, SM_ALLOC, SM_NO_VDEVID, tx);
3785 space_map_write(sm, msp->ms_allocatable, SM_FREE, SM_NO_VDEVID, tx);
3786 space_map_write(sm, condense_tree, SM_FREE, SM_NO_VDEVID, tx);
3788 range_tree_vacate(condense_tree, NULL, NULL);
3789 range_tree_destroy(condense_tree);
3790 range_tree_vacate(tmp_tree, NULL, NULL);
3791 range_tree_destroy(tmp_tree);
3792 mutex_enter(&msp->ms_lock);
3794 msp->ms_condensing = B_FALSE;
3795 metaslab_flush_update(msp, tx);
3799 metaslab_unflushed_add(metaslab_t *msp, dmu_tx_t *tx)
3801 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3802 ASSERT(spa_syncing_log_sm(spa) != NULL);
3803 ASSERT(msp->ms_sm != NULL);
3804 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3805 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3807 mutex_enter(&spa->spa_flushed_ms_lock);
3808 metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx);
3809 metaslab_set_unflushed_dirty(msp, B_TRUE);
3810 avl_add(&spa->spa_metaslabs_by_flushed, msp);
3811 mutex_exit(&spa->spa_flushed_ms_lock);
3813 spa_log_sm_increment_current_mscount(spa);
3814 spa_log_summary_add_flushed_metaslab(spa, B_TRUE);
3818 metaslab_unflushed_bump(metaslab_t *msp, dmu_tx_t *tx, boolean_t dirty)
3820 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3821 ASSERT(spa_syncing_log_sm(spa) != NULL);
3822 ASSERT(msp->ms_sm != NULL);
3823 ASSERT(metaslab_unflushed_txg(msp) != 0);
3824 ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL), ==, msp);
3825 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3826 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3828 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(spa));
3830 /* update metaslab's position in our flushing tree */
3831 uint64_t ms_prev_flushed_txg = metaslab_unflushed_txg(msp);
3832 boolean_t ms_prev_flushed_dirty = metaslab_unflushed_dirty(msp);
3833 mutex_enter(&spa->spa_flushed_ms_lock);
3834 avl_remove(&spa->spa_metaslabs_by_flushed, msp);
3835 metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx);
3836 metaslab_set_unflushed_dirty(msp, dirty);
3837 avl_add(&spa->spa_metaslabs_by_flushed, msp);
3838 mutex_exit(&spa->spa_flushed_ms_lock);
3840 /* update metaslab counts of spa_log_sm_t nodes */
3841 spa_log_sm_decrement_mscount(spa, ms_prev_flushed_txg);
3842 spa_log_sm_increment_current_mscount(spa);
3844 /* update log space map summary */
3845 spa_log_summary_decrement_mscount(spa, ms_prev_flushed_txg,
3846 ms_prev_flushed_dirty);
3847 spa_log_summary_add_flushed_metaslab(spa, dirty);
3849 /* cleanup obsolete logs if any */
3850 spa_cleanup_old_sm_logs(spa, tx);
3854 * Called when the metaslab has been flushed (its own spacemap now reflects
3855 * all the contents of the pool-wide spacemap log). Updates the metaslab's
3856 * metadata and any pool-wide related log space map data (e.g. summary,
3857 * obsolete logs, etc..) to reflect that.
3860 metaslab_flush_update(metaslab_t *msp, dmu_tx_t *tx)
3862 metaslab_group_t *mg = msp->ms_group;
3863 spa_t *spa = mg->mg_vd->vdev_spa;
3865 ASSERT(MUTEX_HELD(&msp->ms_lock));
3867 ASSERT3U(spa_sync_pass(spa), ==, 1);
3870 * Just because a metaslab got flushed, that doesn't mean that
3871 * it will pass through metaslab_sync_done(). Thus, make sure to
3872 * update ms_synced_length here in case it doesn't.
3874 msp->ms_synced_length = space_map_length(msp->ms_sm);
3877 * We may end up here from metaslab_condense() without the
3878 * feature being active. In that case this is a no-op.
3880 if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP) ||
3881 metaslab_unflushed_txg(msp) == 0)
3884 metaslab_unflushed_bump(msp, tx, B_FALSE);
3888 metaslab_flush(metaslab_t *msp, dmu_tx_t *tx)
3890 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3892 ASSERT(MUTEX_HELD(&msp->ms_lock));
3893 ASSERT3U(spa_sync_pass(spa), ==, 1);
3894 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
3896 ASSERT(msp->ms_sm != NULL);
3897 ASSERT(metaslab_unflushed_txg(msp) != 0);
3898 ASSERT(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL) != NULL);
3901 * There is nothing wrong with flushing the same metaslab twice, as
3902 * this codepath should work on that case. However, the current
3903 * flushing scheme makes sure to avoid this situation as we would be
3904 * making all these calls without having anything meaningful to write
3905 * to disk. We assert this behavior here.
3907 ASSERT3U(metaslab_unflushed_txg(msp), <, dmu_tx_get_txg(tx));
3910 * We can not flush while loading, because then we would
3911 * not load the ms_unflushed_{allocs,frees}.
3913 if (msp->ms_loading)
3916 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3917 metaslab_verify_weight_and_frag(msp);
3920 * Metaslab condensing is effectively flushing. Therefore if the
3921 * metaslab can be condensed we can just condense it instead of
3924 * Note that metaslab_condense() does call metaslab_flush_update()
3925 * so we can just return immediately after condensing. We also
3926 * don't need to care about setting ms_flushing or broadcasting
3927 * ms_flush_cv, even if we temporarily drop the ms_lock in
3928 * metaslab_condense(), as the metaslab is already loaded.
3930 if (msp->ms_loaded && metaslab_should_condense(msp)) {
3931 metaslab_group_t *mg = msp->ms_group;
3934 * For all histogram operations below refer to the
3935 * comments of metaslab_sync() where we follow a
3936 * similar procedure.
3938 metaslab_group_histogram_verify(mg);
3939 metaslab_class_histogram_verify(mg->mg_class);
3940 metaslab_group_histogram_remove(mg, msp);
3942 metaslab_condense(msp, tx);
3944 space_map_histogram_clear(msp->ms_sm);
3945 space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
3946 ASSERT(range_tree_is_empty(msp->ms_freed));
3947 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3948 space_map_histogram_add(msp->ms_sm,
3949 msp->ms_defer[t], tx);
3951 metaslab_aux_histograms_update(msp);
3953 metaslab_group_histogram_add(mg, msp);
3954 metaslab_group_histogram_verify(mg);
3955 metaslab_class_histogram_verify(mg->mg_class);
3957 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3960 * Since we recreated the histogram (and potentially
3961 * the ms_sm too while condensing) ensure that the
3962 * weight is updated too because we are not guaranteed
3963 * that this metaslab is dirty and will go through
3964 * metaslab_sync_done().
3966 metaslab_recalculate_weight_and_sort(msp);
3970 msp->ms_flushing = B_TRUE;
3971 uint64_t sm_len_before = space_map_length(msp->ms_sm);
3973 mutex_exit(&msp->ms_lock);
3974 space_map_write(msp->ms_sm, msp->ms_unflushed_allocs, SM_ALLOC,
3976 space_map_write(msp->ms_sm, msp->ms_unflushed_frees, SM_FREE,
3978 mutex_enter(&msp->ms_lock);
3980 uint64_t sm_len_after = space_map_length(msp->ms_sm);
3981 if (zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) {
3982 zfs_dbgmsg("flushing: txg %llu, spa %s, vdev_id %llu, "
3983 "ms_id %llu, unflushed_allocs %llu, unflushed_frees %llu, "
3984 "appended %llu bytes", (u_longlong_t)dmu_tx_get_txg(tx),
3986 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
3987 (u_longlong_t)msp->ms_id,
3988 (u_longlong_t)range_tree_space(msp->ms_unflushed_allocs),
3989 (u_longlong_t)range_tree_space(msp->ms_unflushed_frees),
3990 (u_longlong_t)(sm_len_after - sm_len_before));
3993 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3994 metaslab_unflushed_changes_memused(msp));
3995 spa->spa_unflushed_stats.sus_memused -=
3996 metaslab_unflushed_changes_memused(msp);
3997 range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
3998 range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
4000 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
4001 metaslab_verify_weight_and_frag(msp);
4003 metaslab_flush_update(msp, tx);
4005 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
4006 metaslab_verify_weight_and_frag(msp);
4008 msp->ms_flushing = B_FALSE;
4009 cv_broadcast(&msp->ms_flush_cv);
4014 * Write a metaslab to disk in the context of the specified transaction group.
4017 metaslab_sync(metaslab_t *msp, uint64_t txg)
4019 metaslab_group_t *mg = msp->ms_group;
4020 vdev_t *vd = mg->mg_vd;
4021 spa_t *spa = vd->vdev_spa;
4022 objset_t *mos = spa_meta_objset(spa);
4023 range_tree_t *alloctree = msp->ms_allocating[txg & TXG_MASK];
4026 ASSERT(!vd->vdev_ishole);
4029 * This metaslab has just been added so there's no work to do now.
4032 ASSERT0(range_tree_space(alloctree));
4033 ASSERT0(range_tree_space(msp->ms_freeing));
4034 ASSERT0(range_tree_space(msp->ms_freed));
4035 ASSERT0(range_tree_space(msp->ms_checkpointing));
4036 ASSERT0(range_tree_space(msp->ms_trim));
4041 * Normally, we don't want to process a metaslab if there are no
4042 * allocations or frees to perform. However, if the metaslab is being
4043 * forced to condense, it's loaded and we're not beyond the final
4044 * dirty txg, we need to let it through. Not condensing beyond the
4045 * final dirty txg prevents an issue where metaslabs that need to be
4046 * condensed but were loaded for other reasons could cause a panic
4047 * here. By only checking the txg in that branch of the conditional,
4048 * we preserve the utility of the VERIFY statements in all other
4051 if (range_tree_is_empty(alloctree) &&
4052 range_tree_is_empty(msp->ms_freeing) &&
4053 range_tree_is_empty(msp->ms_checkpointing) &&
4054 !(msp->ms_loaded && msp->ms_condense_wanted &&
4055 txg <= spa_final_dirty_txg(spa)))
4059 VERIFY3U(txg, <=, spa_final_dirty_txg(spa));
4062 * The only state that can actually be changing concurrently
4063 * with metaslab_sync() is the metaslab's ms_allocatable. No
4064 * other thread can be modifying this txg's alloc, freeing,
4065 * freed, or space_map_phys_t. We drop ms_lock whenever we
4066 * could call into the DMU, because the DMU can call down to
4067 * us (e.g. via zio_free()) at any time.
4069 * The spa_vdev_remove_thread() can be reading metaslab state
4070 * concurrently, and it is locked out by the ms_sync_lock.
4071 * Note that the ms_lock is insufficient for this, because it
4072 * is dropped by space_map_write().
4074 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
4077 * Generate a log space map if one doesn't exist already.
4079 spa_generate_syncing_log_sm(spa, tx);
4081 if (msp->ms_sm == NULL) {
4082 uint64_t new_object = space_map_alloc(mos,
4083 spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ?
4084 zfs_metaslab_sm_blksz_with_log :
4085 zfs_metaslab_sm_blksz_no_log, tx);
4086 VERIFY3U(new_object, !=, 0);
4088 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
4089 msp->ms_id, sizeof (uint64_t), &new_object, tx);
4091 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
4092 msp->ms_start, msp->ms_size, vd->vdev_ashift));
4093 ASSERT(msp->ms_sm != NULL);
4095 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
4096 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
4097 ASSERT0(metaslab_allocated_space(msp));
4100 if (!range_tree_is_empty(msp->ms_checkpointing) &&
4101 vd->vdev_checkpoint_sm == NULL) {
4102 ASSERT(spa_has_checkpoint(spa));
4104 uint64_t new_object = space_map_alloc(mos,
4105 zfs_vdev_standard_sm_blksz, tx);
4106 VERIFY3U(new_object, !=, 0);
4108 VERIFY0(space_map_open(&vd->vdev_checkpoint_sm,
4109 mos, new_object, 0, vd->vdev_asize, vd->vdev_ashift));
4110 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
4113 * We save the space map object as an entry in vdev_top_zap
4114 * so it can be retrieved when the pool is reopened after an
4115 * export or through zdb.
4117 VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset,
4118 vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM,
4119 sizeof (new_object), 1, &new_object, tx));
4122 mutex_enter(&msp->ms_sync_lock);
4123 mutex_enter(&msp->ms_lock);
4126 * Note: metaslab_condense() clears the space map's histogram.
4127 * Therefore we must verify and remove this histogram before
4130 metaslab_group_histogram_verify(mg);
4131 metaslab_class_histogram_verify(mg->mg_class);
4132 metaslab_group_histogram_remove(mg, msp);
4134 if (spa->spa_sync_pass == 1 && msp->ms_loaded &&
4135 metaslab_should_condense(msp))
4136 metaslab_condense(msp, tx);
4139 * We'll be going to disk to sync our space accounting, thus we
4140 * drop the ms_lock during that time so allocations coming from
4141 * open-context (ZIL) for future TXGs do not block.
4143 mutex_exit(&msp->ms_lock);
4144 space_map_t *log_sm = spa_syncing_log_sm(spa);
4145 if (log_sm != NULL) {
4146 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP));
4147 if (metaslab_unflushed_txg(msp) == 0)
4148 metaslab_unflushed_add(msp, tx);
4149 else if (!metaslab_unflushed_dirty(msp))
4150 metaslab_unflushed_bump(msp, tx, B_TRUE);
4152 space_map_write(log_sm, alloctree, SM_ALLOC,
4154 space_map_write(log_sm, msp->ms_freeing, SM_FREE,
4156 mutex_enter(&msp->ms_lock);
4158 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
4159 metaslab_unflushed_changes_memused(msp));
4160 spa->spa_unflushed_stats.sus_memused -=
4161 metaslab_unflushed_changes_memused(msp);
4162 range_tree_remove_xor_add(alloctree,
4163 msp->ms_unflushed_frees, msp->ms_unflushed_allocs);
4164 range_tree_remove_xor_add(msp->ms_freeing,
4165 msp->ms_unflushed_allocs, msp->ms_unflushed_frees);
4166 spa->spa_unflushed_stats.sus_memused +=
4167 metaslab_unflushed_changes_memused(msp);
4169 ASSERT(!spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP));
4171 space_map_write(msp->ms_sm, alloctree, SM_ALLOC,
4173 space_map_write(msp->ms_sm, msp->ms_freeing, SM_FREE,
4175 mutex_enter(&msp->ms_lock);
4178 msp->ms_allocated_space += range_tree_space(alloctree);
4179 ASSERT3U(msp->ms_allocated_space, >=,
4180 range_tree_space(msp->ms_freeing));
4181 msp->ms_allocated_space -= range_tree_space(msp->ms_freeing);
4183 if (!range_tree_is_empty(msp->ms_checkpointing)) {
4184 ASSERT(spa_has_checkpoint(spa));
4185 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
4188 * Since we are doing writes to disk and the ms_checkpointing
4189 * tree won't be changing during that time, we drop the
4190 * ms_lock while writing to the checkpoint space map, for the
4191 * same reason mentioned above.
4193 mutex_exit(&msp->ms_lock);
4194 space_map_write(vd->vdev_checkpoint_sm,
4195 msp->ms_checkpointing, SM_FREE, SM_NO_VDEVID, tx);
4196 mutex_enter(&msp->ms_lock);
4198 spa->spa_checkpoint_info.sci_dspace +=
4199 range_tree_space(msp->ms_checkpointing);
4200 vd->vdev_stat.vs_checkpoint_space +=
4201 range_tree_space(msp->ms_checkpointing);
4202 ASSERT3U(vd->vdev_stat.vs_checkpoint_space, ==,
4203 -space_map_allocated(vd->vdev_checkpoint_sm));
4205 range_tree_vacate(msp->ms_checkpointing, NULL, NULL);
4208 if (msp->ms_loaded) {
4210 * When the space map is loaded, we have an accurate
4211 * histogram in the range tree. This gives us an opportunity
4212 * to bring the space map's histogram up-to-date so we clear
4213 * it first before updating it.
4215 space_map_histogram_clear(msp->ms_sm);
4216 space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
4219 * Since we've cleared the histogram we need to add back
4220 * any free space that has already been processed, plus
4221 * any deferred space. This allows the on-disk histogram
4222 * to accurately reflect all free space even if some space
4223 * is not yet available for allocation (i.e. deferred).
4225 space_map_histogram_add(msp->ms_sm, msp->ms_freed, tx);
4228 * Add back any deferred free space that has not been
4229 * added back into the in-core free tree yet. This will
4230 * ensure that we don't end up with a space map histogram
4231 * that is completely empty unless the metaslab is fully
4234 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
4235 space_map_histogram_add(msp->ms_sm,
4236 msp->ms_defer[t], tx);
4241 * Always add the free space from this sync pass to the space
4242 * map histogram. We want to make sure that the on-disk histogram
4243 * accounts for all free space. If the space map is not loaded,
4244 * then we will lose some accuracy but will correct it the next
4245 * time we load the space map.
4247 space_map_histogram_add(msp->ms_sm, msp->ms_freeing, tx);
4248 metaslab_aux_histograms_update(msp);
4250 metaslab_group_histogram_add(mg, msp);
4251 metaslab_group_histogram_verify(mg);
4252 metaslab_class_histogram_verify(mg->mg_class);
4255 * For sync pass 1, we avoid traversing this txg's free range tree
4256 * and instead will just swap the pointers for freeing and freed.
4257 * We can safely do this since the freed_tree is guaranteed to be
4258 * empty on the initial pass.
4260 * Keep in mind that even if we are currently using a log spacemap
4261 * we want current frees to end up in the ms_allocatable (but not
4262 * get appended to the ms_sm) so their ranges can be reused as usual.
4264 if (spa_sync_pass(spa) == 1) {
4265 range_tree_swap(&msp->ms_freeing, &msp->ms_freed);
4266 ASSERT0(msp->ms_allocated_this_txg);
4268 range_tree_vacate(msp->ms_freeing,
4269 range_tree_add, msp->ms_freed);
4271 msp->ms_allocated_this_txg += range_tree_space(alloctree);
4272 range_tree_vacate(alloctree, NULL, NULL);
4274 ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
4275 ASSERT0(range_tree_space(msp->ms_allocating[TXG_CLEAN(txg)
4277 ASSERT0(range_tree_space(msp->ms_freeing));
4278 ASSERT0(range_tree_space(msp->ms_checkpointing));
4280 mutex_exit(&msp->ms_lock);
4283 * Verify that the space map object ID has been recorded in the
4287 VERIFY0(dmu_read(mos, vd->vdev_ms_array,
4288 msp->ms_id * sizeof (uint64_t), sizeof (uint64_t), &object, 0));
4289 VERIFY3U(object, ==, space_map_object(msp->ms_sm));
4291 mutex_exit(&msp->ms_sync_lock);
4296 metaslab_evict(metaslab_t *msp, uint64_t txg)
4298 if (!msp->ms_loaded || msp->ms_disabled != 0)
4301 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
4302 VERIFY0(range_tree_space(
4303 msp->ms_allocating[(txg + t) & TXG_MASK]));
4305 if (msp->ms_allocator != -1)
4306 metaslab_passivate(msp, msp->ms_weight & ~METASLAB_ACTIVE_MASK);
4308 if (!metaslab_debug_unload)
4309 metaslab_unload(msp);
4313 * Called after a transaction group has completely synced to mark
4314 * all of the metaslab's free space as usable.
4317 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
4319 metaslab_group_t *mg = msp->ms_group;
4320 vdev_t *vd = mg->mg_vd;
4321 spa_t *spa = vd->vdev_spa;
4322 range_tree_t **defer_tree;
4323 int64_t alloc_delta, defer_delta;
4324 boolean_t defer_allowed = B_TRUE;
4326 ASSERT(!vd->vdev_ishole);
4328 mutex_enter(&msp->ms_lock);
4331 /* this is a new metaslab, add its capacity to the vdev */
4332 metaslab_space_update(vd, mg->mg_class, 0, 0, msp->ms_size);
4334 /* there should be no allocations nor frees at this point */
4335 VERIFY0(msp->ms_allocated_this_txg);
4336 VERIFY0(range_tree_space(msp->ms_freed));
4339 ASSERT0(range_tree_space(msp->ms_freeing));
4340 ASSERT0(range_tree_space(msp->ms_checkpointing));
4342 defer_tree = &msp->ms_defer[txg % TXG_DEFER_SIZE];
4344 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
4345 metaslab_class_get_alloc(spa_normal_class(spa));
4346 if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) {
4347 defer_allowed = B_FALSE;
4351 alloc_delta = msp->ms_allocated_this_txg -
4352 range_tree_space(msp->ms_freed);
4354 if (defer_allowed) {
4355 defer_delta = range_tree_space(msp->ms_freed) -
4356 range_tree_space(*defer_tree);
4358 defer_delta -= range_tree_space(*defer_tree);
4360 metaslab_space_update(vd, mg->mg_class, alloc_delta + defer_delta,
4363 if (spa_syncing_log_sm(spa) == NULL) {
4365 * If there's a metaslab_load() in progress and we don't have
4366 * a log space map, it means that we probably wrote to the
4367 * metaslab's space map. If this is the case, we need to
4368 * make sure that we wait for the load to complete so that we
4369 * have a consistent view at the in-core side of the metaslab.
4371 metaslab_load_wait(msp);
4373 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
4377 * When auto-trimming is enabled, free ranges which are added to
4378 * ms_allocatable are also be added to ms_trim. The ms_trim tree is
4379 * periodically consumed by the vdev_autotrim_thread() which issues
4380 * trims for all ranges and then vacates the tree. The ms_trim tree
4381 * can be discarded at any time with the sole consequence of recent
4382 * frees not being trimmed.
4384 if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON) {
4385 range_tree_walk(*defer_tree, range_tree_add, msp->ms_trim);
4386 if (!defer_allowed) {
4387 range_tree_walk(msp->ms_freed, range_tree_add,
4391 range_tree_vacate(msp->ms_trim, NULL, NULL);
4395 * Move the frees from the defer_tree back to the free
4396 * range tree (if it's loaded). Swap the freed_tree and
4397 * the defer_tree -- this is safe to do because we've
4398 * just emptied out the defer_tree.
4400 range_tree_vacate(*defer_tree,
4401 msp->ms_loaded ? range_tree_add : NULL, msp->ms_allocatable);
4402 if (defer_allowed) {
4403 range_tree_swap(&msp->ms_freed, defer_tree);
4405 range_tree_vacate(msp->ms_freed,
4406 msp->ms_loaded ? range_tree_add : NULL,
4407 msp->ms_allocatable);
4410 msp->ms_synced_length = space_map_length(msp->ms_sm);
4412 msp->ms_deferspace += defer_delta;
4413 ASSERT3S(msp->ms_deferspace, >=, 0);
4414 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
4415 if (msp->ms_deferspace != 0) {
4417 * Keep syncing this metaslab until all deferred frees
4418 * are back in circulation.
4420 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
4422 metaslab_aux_histograms_update_done(msp, defer_allowed);
4425 msp->ms_new = B_FALSE;
4426 mutex_enter(&mg->mg_lock);
4428 mutex_exit(&mg->mg_lock);
4432 * Re-sort metaslab within its group now that we've adjusted
4433 * its allocatable space.
4435 metaslab_recalculate_weight_and_sort(msp);
4437 ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
4438 ASSERT0(range_tree_space(msp->ms_freeing));
4439 ASSERT0(range_tree_space(msp->ms_freed));
4440 ASSERT0(range_tree_space(msp->ms_checkpointing));
4441 msp->ms_allocating_total -= msp->ms_allocated_this_txg;
4442 msp->ms_allocated_this_txg = 0;
4443 mutex_exit(&msp->ms_lock);
4447 metaslab_sync_reassess(metaslab_group_t *mg)
4449 spa_t *spa = mg->mg_class->mc_spa;
4451 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4452 metaslab_group_alloc_update(mg);
4453 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
4456 * Preload the next potential metaslabs but only on active
4457 * metaslab groups. We can get into a state where the metaslab
4458 * is no longer active since we dirty metaslabs as we remove a
4459 * a device, thus potentially making the metaslab group eligible
4462 if (mg->mg_activation_count > 0) {
4463 metaslab_group_preload(mg);
4465 spa_config_exit(spa, SCL_ALLOC, FTAG);
4469 * When writing a ditto block (i.e. more than one DVA for a given BP) on
4470 * the same vdev as an existing DVA of this BP, then try to allocate it
4471 * on a different metaslab than existing DVAs (i.e. a unique metaslab).
4474 metaslab_is_unique(metaslab_t *msp, dva_t *dva)
4478 if (DVA_GET_ASIZE(dva) == 0)
4481 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
4484 dva_ms_id = DVA_GET_OFFSET(dva) >> msp->ms_group->mg_vd->vdev_ms_shift;
4486 return (msp->ms_id != dva_ms_id);
4490 * ==========================================================================
4491 * Metaslab allocation tracing facility
4492 * ==========================================================================
4496 * Add an allocation trace element to the allocation tracing list.
4499 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
4500 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset,
4503 metaslab_alloc_trace_t *mat;
4505 if (!metaslab_trace_enabled)
4509 * When the tracing list reaches its maximum we remove
4510 * the second element in the list before adding a new one.
4511 * By removing the second element we preserve the original
4512 * entry as a clue to what allocations steps have already been
4515 if (zal->zal_size == metaslab_trace_max_entries) {
4516 metaslab_alloc_trace_t *mat_next;
4518 panic("too many entries in allocation list");
4520 METASLABSTAT_BUMP(metaslabstat_trace_over_limit);
4522 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
4523 list_remove(&zal->zal_list, mat_next);
4524 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
4527 mat = kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
4528 list_link_init(&mat->mat_list_node);
4531 mat->mat_size = psize;
4532 mat->mat_dva_id = dva_id;
4533 mat->mat_offset = offset;
4534 mat->mat_weight = 0;
4535 mat->mat_allocator = allocator;
4538 mat->mat_weight = msp->ms_weight;
4541 * The list is part of the zio so locking is not required. Only
4542 * a single thread will perform allocations for a given zio.
4544 list_insert_tail(&zal->zal_list, mat);
4547 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
4551 metaslab_trace_init(zio_alloc_list_t *zal)
4553 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
4554 offsetof(metaslab_alloc_trace_t, mat_list_node));
4559 metaslab_trace_fini(zio_alloc_list_t *zal)
4561 metaslab_alloc_trace_t *mat;
4563 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
4564 kmem_cache_free(metaslab_alloc_trace_cache, mat);
4565 list_destroy(&zal->zal_list);
4570 * ==========================================================================
4571 * Metaslab block operations
4572 * ==========================================================================
4576 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, const void *tag,
4577 int flags, int allocator)
4579 if (!(flags & METASLAB_ASYNC_ALLOC) ||
4580 (flags & METASLAB_DONT_THROTTLE))
4583 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4584 if (!mg->mg_class->mc_alloc_throttle_enabled)
4587 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4588 (void) zfs_refcount_add(&mga->mga_alloc_queue_depth, tag);
4592 metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator)
4594 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4595 metaslab_class_allocator_t *mca =
4596 &mg->mg_class->mc_allocator[allocator];
4597 uint64_t max = mg->mg_max_alloc_queue_depth;
4598 uint64_t cur = mga->mga_cur_max_alloc_queue_depth;
4600 if (atomic_cas_64(&mga->mga_cur_max_alloc_queue_depth,
4601 cur, cur + 1) == cur) {
4602 atomic_inc_64(&mca->mca_alloc_max_slots);
4605 cur = mga->mga_cur_max_alloc_queue_depth;
4610 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, const void *tag,
4611 int flags, int allocator, boolean_t io_complete)
4613 if (!(flags & METASLAB_ASYNC_ALLOC) ||
4614 (flags & METASLAB_DONT_THROTTLE))
4617 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4618 if (!mg->mg_class->mc_alloc_throttle_enabled)
4621 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4622 (void) zfs_refcount_remove(&mga->mga_alloc_queue_depth, tag);
4624 metaslab_group_increment_qdepth(mg, allocator);
4628 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, const void *tag,
4632 const dva_t *dva = bp->blk_dva;
4633 int ndvas = BP_GET_NDVAS(bp);
4635 for (int d = 0; d < ndvas; d++) {
4636 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
4637 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4638 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4639 VERIFY(zfs_refcount_not_held(&mga->mga_alloc_queue_depth, tag));
4645 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
4648 range_tree_t *rt = msp->ms_allocatable;
4649 metaslab_class_t *mc = msp->ms_group->mg_class;
4651 ASSERT(MUTEX_HELD(&msp->ms_lock));
4652 VERIFY(!msp->ms_condensing);
4653 VERIFY0(msp->ms_disabled);
4655 start = mc->mc_ops->msop_alloc(msp, size);
4656 if (start != -1ULL) {
4657 metaslab_group_t *mg = msp->ms_group;
4658 vdev_t *vd = mg->mg_vd;
4660 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
4661 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
4662 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
4663 range_tree_remove(rt, start, size);
4664 range_tree_clear(msp->ms_trim, start, size);
4666 if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
4667 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
4669 range_tree_add(msp->ms_allocating[txg & TXG_MASK], start, size);
4670 msp->ms_allocating_total += size;
4672 /* Track the last successful allocation */
4673 msp->ms_alloc_txg = txg;
4674 metaslab_verify_space(msp, txg);
4678 * Now that we've attempted the allocation we need to update the
4679 * metaslab's maximum block size since it may have changed.
4681 msp->ms_max_size = metaslab_largest_allocatable(msp);
4686 * Find the metaslab with the highest weight that is less than what we've
4687 * already tried. In the common case, this means that we will examine each
4688 * metaslab at most once. Note that concurrent callers could reorder metaslabs
4689 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
4690 * activated by another thread, and we fail to allocate from the metaslab we
4691 * have selected, we may not try the newly-activated metaslab, and instead
4692 * activate another metaslab. This is not optimal, but generally does not cause
4693 * any problems (a possible exception being if every metaslab is completely full
4694 * except for the newly-activated metaslab which we fail to examine).
4697 find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight,
4698 dva_t *dva, int d, boolean_t want_unique, uint64_t asize, int allocator,
4699 boolean_t try_hard, zio_alloc_list_t *zal, metaslab_t *search,
4700 boolean_t *was_active)
4703 avl_tree_t *t = &mg->mg_metaslab_tree;
4704 metaslab_t *msp = avl_find(t, search, &idx);
4706 msp = avl_nearest(t, idx, AVL_AFTER);
4709 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
4712 if (!try_hard && tries > zfs_metaslab_find_max_tries) {
4713 METASLABSTAT_BUMP(metaslabstat_too_many_tries);
4718 if (!metaslab_should_allocate(msp, asize, try_hard)) {
4719 metaslab_trace_add(zal, mg, msp, asize, d,
4720 TRACE_TOO_SMALL, allocator);
4725 * If the selected metaslab is condensing or disabled,
4728 if (msp->ms_condensing || msp->ms_disabled > 0)
4731 *was_active = msp->ms_allocator != -1;
4733 * If we're activating as primary, this is our first allocation
4734 * from this disk, so we don't need to check how close we are.
4735 * If the metaslab under consideration was already active,
4736 * we're getting desperate enough to steal another allocator's
4737 * metaslab, so we still don't care about distances.
4739 if (activation_weight == METASLAB_WEIGHT_PRIMARY || *was_active)
4742 for (i = 0; i < d; i++) {
4744 !metaslab_is_unique(msp, &dva[i]))
4745 break; /* try another metaslab */
4752 search->ms_weight = msp->ms_weight;
4753 search->ms_start = msp->ms_start + 1;
4754 search->ms_allocator = msp->ms_allocator;
4755 search->ms_primary = msp->ms_primary;
4761 metaslab_active_mask_verify(metaslab_t *msp)
4763 ASSERT(MUTEX_HELD(&msp->ms_lock));
4765 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
4768 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0)
4771 if (msp->ms_weight & METASLAB_WEIGHT_PRIMARY) {
4772 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
4773 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM);
4774 VERIFY3S(msp->ms_allocator, !=, -1);
4775 VERIFY(msp->ms_primary);
4779 if (msp->ms_weight & METASLAB_WEIGHT_SECONDARY) {
4780 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
4781 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM);
4782 VERIFY3S(msp->ms_allocator, !=, -1);
4783 VERIFY(!msp->ms_primary);
4787 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
4788 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
4789 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
4790 VERIFY3S(msp->ms_allocator, ==, -1);
4796 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
4797 uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d,
4798 int allocator, boolean_t try_hard)
4800 metaslab_t *msp = NULL;
4801 uint64_t offset = -1ULL;
4803 uint64_t activation_weight = METASLAB_WEIGHT_PRIMARY;
4804 for (int i = 0; i < d; i++) {
4805 if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
4806 DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
4807 activation_weight = METASLAB_WEIGHT_SECONDARY;
4808 } else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
4809 DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
4810 activation_weight = METASLAB_WEIGHT_CLAIM;
4816 * If we don't have enough metaslabs active to fill the entire array, we
4817 * just use the 0th slot.
4819 if (mg->mg_ms_ready < mg->mg_allocators * 3)
4821 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4823 ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2);
4825 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
4826 search->ms_weight = UINT64_MAX;
4827 search->ms_start = 0;
4829 * At the end of the metaslab tree are the already-active metaslabs,
4830 * first the primaries, then the secondaries. When we resume searching
4831 * through the tree, we need to consider ms_allocator and ms_primary so
4832 * we start in the location right after where we left off, and don't
4833 * accidentally loop forever considering the same metaslabs.
4835 search->ms_allocator = -1;
4836 search->ms_primary = B_TRUE;
4838 boolean_t was_active = B_FALSE;
4840 mutex_enter(&mg->mg_lock);
4842 if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
4843 mga->mga_primary != NULL) {
4844 msp = mga->mga_primary;
4847 * Even though we don't hold the ms_lock for the
4848 * primary metaslab, those fields should not
4849 * change while we hold the mg_lock. Thus it is
4850 * safe to make assertions on them.
4852 ASSERT(msp->ms_primary);
4853 ASSERT3S(msp->ms_allocator, ==, allocator);
4854 ASSERT(msp->ms_loaded);
4856 was_active = B_TRUE;
4857 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
4858 } else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
4859 mga->mga_secondary != NULL) {
4860 msp = mga->mga_secondary;
4863 * See comment above about the similar assertions
4864 * for the primary metaslab.
4866 ASSERT(!msp->ms_primary);
4867 ASSERT3S(msp->ms_allocator, ==, allocator);
4868 ASSERT(msp->ms_loaded);
4870 was_active = B_TRUE;
4871 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
4873 msp = find_valid_metaslab(mg, activation_weight, dva, d,
4874 want_unique, asize, allocator, try_hard, zal,
4875 search, &was_active);
4878 mutex_exit(&mg->mg_lock);
4880 kmem_free(search, sizeof (*search));
4883 mutex_enter(&msp->ms_lock);
4885 metaslab_active_mask_verify(msp);
4888 * This code is disabled out because of issues with
4889 * tracepoints in non-gpl kernel modules.
4892 DTRACE_PROBE3(ms__activation__attempt,
4893 metaslab_t *, msp, uint64_t, activation_weight,
4894 boolean_t, was_active);
4898 * Ensure that the metaslab we have selected is still
4899 * capable of handling our request. It's possible that
4900 * another thread may have changed the weight while we
4901 * were blocked on the metaslab lock. We check the
4902 * active status first to see if we need to set_selected_txg
4905 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
4906 ASSERT3S(msp->ms_allocator, ==, -1);
4907 mutex_exit(&msp->ms_lock);
4912 * If the metaslab was activated for another allocator
4913 * while we were waiting in the ms_lock above, or it's
4914 * a primary and we're seeking a secondary (or vice versa),
4915 * we go back and select a new metaslab.
4917 if (!was_active && (msp->ms_weight & METASLAB_ACTIVE_MASK) &&
4918 (msp->ms_allocator != -1) &&
4919 (msp->ms_allocator != allocator || ((activation_weight ==
4920 METASLAB_WEIGHT_PRIMARY) != msp->ms_primary))) {
4921 ASSERT(msp->ms_loaded);
4922 ASSERT((msp->ms_weight & METASLAB_WEIGHT_CLAIM) ||
4923 msp->ms_allocator != -1);
4924 mutex_exit(&msp->ms_lock);
4929 * This metaslab was used for claiming regions allocated
4930 * by the ZIL during pool import. Once these regions are
4931 * claimed we don't need to keep the CLAIM bit set
4932 * anymore. Passivate this metaslab to zero its activation
4935 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM &&
4936 activation_weight != METASLAB_WEIGHT_CLAIM) {
4937 ASSERT(msp->ms_loaded);
4938 ASSERT3S(msp->ms_allocator, ==, -1);
4939 metaslab_passivate(msp, msp->ms_weight &
4940 ~METASLAB_WEIGHT_CLAIM);
4941 mutex_exit(&msp->ms_lock);
4945 metaslab_set_selected_txg(msp, txg);
4947 int activation_error =
4948 metaslab_activate(msp, allocator, activation_weight);
4949 metaslab_active_mask_verify(msp);
4952 * If the metaslab was activated by another thread for
4953 * another allocator or activation_weight (EBUSY), or it
4954 * failed because another metaslab was assigned as primary
4955 * for this allocator (EEXIST) we continue using this
4956 * metaslab for our allocation, rather than going on to a
4957 * worse metaslab (we waited for that metaslab to be loaded
4960 * If the activation failed due to an I/O error or ENOSPC we
4961 * skip to the next metaslab.
4963 boolean_t activated;
4964 if (activation_error == 0) {
4966 } else if (activation_error == EBUSY ||
4967 activation_error == EEXIST) {
4968 activated = B_FALSE;
4970 mutex_exit(&msp->ms_lock);
4973 ASSERT(msp->ms_loaded);
4976 * Now that we have the lock, recheck to see if we should
4977 * continue to use this metaslab for this allocation. The
4978 * the metaslab is now loaded so metaslab_should_allocate()
4979 * can accurately determine if the allocation attempt should
4982 if (!metaslab_should_allocate(msp, asize, try_hard)) {
4983 /* Passivate this metaslab and select a new one. */
4984 metaslab_trace_add(zal, mg, msp, asize, d,
4985 TRACE_TOO_SMALL, allocator);
4990 * If this metaslab is currently condensing then pick again
4991 * as we can't manipulate this metaslab until it's committed
4992 * to disk. If this metaslab is being initialized, we shouldn't
4993 * allocate from it since the allocated region might be
4994 * overwritten after allocation.
4996 if (msp->ms_condensing) {
4997 metaslab_trace_add(zal, mg, msp, asize, d,
4998 TRACE_CONDENSING, allocator);
5000 metaslab_passivate(msp, msp->ms_weight &
5001 ~METASLAB_ACTIVE_MASK);
5003 mutex_exit(&msp->ms_lock);
5005 } else if (msp->ms_disabled > 0) {
5006 metaslab_trace_add(zal, mg, msp, asize, d,
5007 TRACE_DISABLED, allocator);
5009 metaslab_passivate(msp, msp->ms_weight &
5010 ~METASLAB_ACTIVE_MASK);
5012 mutex_exit(&msp->ms_lock);
5016 offset = metaslab_block_alloc(msp, asize, txg);
5017 metaslab_trace_add(zal, mg, msp, asize, d, offset, allocator);
5019 if (offset != -1ULL) {
5020 /* Proactively passivate the metaslab, if needed */
5022 metaslab_segment_may_passivate(msp);
5026 ASSERT(msp->ms_loaded);
5029 * This code is disabled out because of issues with
5030 * tracepoints in non-gpl kernel modules.
5033 DTRACE_PROBE2(ms__alloc__failure, metaslab_t *, msp,
5038 * We were unable to allocate from this metaslab so determine
5039 * a new weight for this metaslab. Now that we have loaded
5040 * the metaslab we can provide a better hint to the metaslab
5043 * For space-based metaslabs, we use the maximum block size.
5044 * This information is only available when the metaslab
5045 * is loaded and is more accurate than the generic free
5046 * space weight that was calculated by metaslab_weight().
5047 * This information allows us to quickly compare the maximum
5048 * available allocation in the metaslab to the allocation
5049 * size being requested.
5051 * For segment-based metaslabs, determine the new weight
5052 * based on the highest bucket in the range tree. We
5053 * explicitly use the loaded segment weight (i.e. the range
5054 * tree histogram) since it contains the space that is
5055 * currently available for allocation and is accurate
5056 * even within a sync pass.
5059 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
5060 weight = metaslab_largest_allocatable(msp);
5061 WEIGHT_SET_SPACEBASED(weight);
5063 weight = metaslab_weight_from_range_tree(msp);
5067 metaslab_passivate(msp, weight);
5070 * For the case where we use the metaslab that is
5071 * active for another allocator we want to make
5072 * sure that we retain the activation mask.
5074 * Note that we could attempt to use something like
5075 * metaslab_recalculate_weight_and_sort() that
5076 * retains the activation mask here. That function
5077 * uses metaslab_weight() to set the weight though
5078 * which is not as accurate as the calculations
5081 weight |= msp->ms_weight & METASLAB_ACTIVE_MASK;
5082 metaslab_group_sort(mg, msp, weight);
5084 metaslab_active_mask_verify(msp);
5087 * We have just failed an allocation attempt, check
5088 * that metaslab_should_allocate() agrees. Otherwise,
5089 * we may end up in an infinite loop retrying the same
5092 ASSERT(!metaslab_should_allocate(msp, asize, try_hard));
5094 mutex_exit(&msp->ms_lock);
5096 mutex_exit(&msp->ms_lock);
5097 kmem_free(search, sizeof (*search));
5102 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
5103 uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d,
5104 int allocator, boolean_t try_hard)
5107 ASSERT(mg->mg_initialized);
5109 offset = metaslab_group_alloc_normal(mg, zal, asize, txg, want_unique,
5110 dva, d, allocator, try_hard);
5112 mutex_enter(&mg->mg_lock);
5113 if (offset == -1ULL) {
5114 mg->mg_failed_allocations++;
5115 metaslab_trace_add(zal, mg, NULL, asize, d,
5116 TRACE_GROUP_FAILURE, allocator);
5117 if (asize == SPA_GANGBLOCKSIZE) {
5119 * This metaslab group was unable to allocate
5120 * the minimum gang block size so it must be out of
5121 * space. We must notify the allocation throttle
5122 * to start skipping allocation attempts to this
5123 * metaslab group until more space becomes available.
5124 * Note: this failure cannot be caused by the
5125 * allocation throttle since the allocation throttle
5126 * is only responsible for skipping devices and
5127 * not failing block allocations.
5129 mg->mg_no_free_space = B_TRUE;
5132 mg->mg_allocations++;
5133 mutex_exit(&mg->mg_lock);
5138 * Allocate a block for the specified i/o.
5141 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
5142 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
5143 zio_alloc_list_t *zal, int allocator)
5145 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
5146 metaslab_group_t *mg, *rotor;
5148 boolean_t try_hard = B_FALSE;
5150 ASSERT(!DVA_IS_VALID(&dva[d]));
5153 * For testing, make some blocks above a certain size be gang blocks.
5154 * This will result in more split blocks when using device removal,
5155 * and a large number of split blocks coupled with ztest-induced
5156 * damage can result in extremely long reconstruction times. This
5157 * will also test spilling from special to normal.
5159 if (psize >= metaslab_force_ganging &&
5160 metaslab_force_ganging_pct > 0 &&
5161 (random_in_range(100) < MIN(metaslab_force_ganging_pct, 100))) {
5162 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG,
5164 return (SET_ERROR(ENOSPC));
5168 * Start at the rotor and loop through all mgs until we find something.
5169 * Note that there's no locking on mca_rotor or mca_aliquot because
5170 * nothing actually breaks if we miss a few updates -- we just won't
5171 * allocate quite as evenly. It all balances out over time.
5173 * If we are doing ditto or log blocks, try to spread them across
5174 * consecutive vdevs. If we're forced to reuse a vdev before we've
5175 * allocated all of our ditto blocks, then try and spread them out on
5176 * that vdev as much as possible. If it turns out to not be possible,
5177 * gradually lower our standards until anything becomes acceptable.
5178 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
5179 * gives us hope of containing our fault domains to something we're
5180 * able to reason about. Otherwise, any two top-level vdev failures
5181 * will guarantee the loss of data. With consecutive allocation,
5182 * only two adjacent top-level vdev failures will result in data loss.
5184 * If we are doing gang blocks (hintdva is non-NULL), try to keep
5185 * ourselves on the same vdev as our gang block header. That
5186 * way, we can hope for locality in vdev_cache, plus it makes our
5187 * fault domains something tractable.
5190 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
5193 * It's possible the vdev we're using as the hint no
5194 * longer exists or its mg has been closed (e.g. by
5195 * device removal). Consult the rotor when
5198 if (vd != NULL && vd->vdev_mg != NULL) {
5199 mg = vdev_get_mg(vd, mc);
5201 if (flags & METASLAB_HINTBP_AVOID)
5204 mg = mca->mca_rotor;
5206 } else if (d != 0) {
5207 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
5208 mg = vd->vdev_mg->mg_next;
5210 ASSERT(mca->mca_rotor != NULL);
5211 mg = mca->mca_rotor;
5215 * If the hint put us into the wrong metaslab class, or into a
5216 * metaslab group that has been passivated, just follow the rotor.
5218 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
5219 mg = mca->mca_rotor;
5224 boolean_t allocatable;
5226 ASSERT(mg->mg_activation_count == 1);
5230 * Don't allocate from faulted devices.
5233 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
5234 allocatable = vdev_allocatable(vd);
5235 spa_config_exit(spa, SCL_ZIO, FTAG);
5237 allocatable = vdev_allocatable(vd);
5241 * Determine if the selected metaslab group is eligible
5242 * for allocations. If we're ganging then don't allow
5243 * this metaslab group to skip allocations since that would
5244 * inadvertently return ENOSPC and suspend the pool
5245 * even though space is still available.
5247 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
5248 allocatable = metaslab_group_allocatable(mg, rotor,
5249 flags, psize, allocator, d);
5253 metaslab_trace_add(zal, mg, NULL, psize, d,
5254 TRACE_NOT_ALLOCATABLE, allocator);
5258 ASSERT(mg->mg_initialized);
5261 * Avoid writing single-copy data to an unhealthy,
5262 * non-redundant vdev, unless we've already tried all
5265 if (vd->vdev_state < VDEV_STATE_HEALTHY &&
5266 d == 0 && !try_hard && vd->vdev_children == 0) {
5267 metaslab_trace_add(zal, mg, NULL, psize, d,
5268 TRACE_VDEV_ERROR, allocator);
5272 ASSERT(mg->mg_class == mc);
5274 uint64_t asize = vdev_psize_to_asize(vd, psize);
5275 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
5278 * If we don't need to try hard, then require that the
5279 * block be on a different metaslab from any other DVAs
5280 * in this BP (unique=true). If we are trying hard, then
5281 * allow any metaslab to be used (unique=false).
5283 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
5284 !try_hard, dva, d, allocator, try_hard);
5286 if (offset != -1ULL) {
5288 * If we've just selected this metaslab group,
5289 * figure out whether the corresponding vdev is
5290 * over- or under-used relative to the pool,
5291 * and set an allocation bias to even it out.
5293 * Bias is also used to compensate for unequally
5294 * sized vdevs so that space is allocated fairly.
5296 if (mca->mca_aliquot == 0 && metaslab_bias_enabled) {
5297 vdev_stat_t *vs = &vd->vdev_stat;
5298 int64_t vs_free = vs->vs_space - vs->vs_alloc;
5299 int64_t mc_free = mc->mc_space - mc->mc_alloc;
5303 * Calculate how much more or less we should
5304 * try to allocate from this device during
5305 * this iteration around the rotor.
5307 * This basically introduces a zero-centered
5308 * bias towards the devices with the most
5309 * free space, while compensating for vdev
5313 * vdev V1 = 16M/128M
5314 * vdev V2 = 16M/128M
5315 * ratio(V1) = 100% ratio(V2) = 100%
5317 * vdev V1 = 16M/128M
5318 * vdev V2 = 64M/128M
5319 * ratio(V1) = 127% ratio(V2) = 72%
5321 * vdev V1 = 16M/128M
5322 * vdev V2 = 64M/512M
5323 * ratio(V1) = 40% ratio(V2) = 160%
5325 ratio = (vs_free * mc->mc_alloc_groups * 100) /
5327 mg->mg_bias = ((ratio - 100) *
5328 (int64_t)mg->mg_aliquot) / 100;
5329 } else if (!metaslab_bias_enabled) {
5333 if ((flags & METASLAB_ZIL) ||
5334 atomic_add_64_nv(&mca->mca_aliquot, asize) >=
5335 mg->mg_aliquot + mg->mg_bias) {
5336 mca->mca_rotor = mg->mg_next;
5337 mca->mca_aliquot = 0;
5340 DVA_SET_VDEV(&dva[d], vd->vdev_id);
5341 DVA_SET_OFFSET(&dva[d], offset);
5342 DVA_SET_GANG(&dva[d],
5343 ((flags & METASLAB_GANG_HEADER) ? 1 : 0));
5344 DVA_SET_ASIZE(&dva[d], asize);
5349 mca->mca_rotor = mg->mg_next;
5350 mca->mca_aliquot = 0;
5351 } while ((mg = mg->mg_next) != rotor);
5354 * If we haven't tried hard, perhaps do so now.
5356 if (!try_hard && (zfs_metaslab_try_hard_before_gang ||
5357 GANG_ALLOCATION(flags) || (flags & METASLAB_ZIL) != 0 ||
5358 psize <= 1 << spa->spa_min_ashift)) {
5359 METASLABSTAT_BUMP(metaslabstat_try_hard);
5364 memset(&dva[d], 0, sizeof (dva_t));
5366 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC, allocator);
5367 return (SET_ERROR(ENOSPC));
5371 metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize,
5372 boolean_t checkpoint)
5375 spa_t *spa = vd->vdev_spa;
5377 ASSERT(vdev_is_concrete(vd));
5378 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5379 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
5381 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5383 VERIFY(!msp->ms_condensing);
5384 VERIFY3U(offset, >=, msp->ms_start);
5385 VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size);
5386 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5387 VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift));
5389 metaslab_check_free_impl(vd, offset, asize);
5391 mutex_enter(&msp->ms_lock);
5392 if (range_tree_is_empty(msp->ms_freeing) &&
5393 range_tree_is_empty(msp->ms_checkpointing)) {
5394 vdev_dirty(vd, VDD_METASLAB, msp, spa_syncing_txg(spa));
5398 ASSERT(spa_has_checkpoint(spa));
5399 range_tree_add(msp->ms_checkpointing, offset, asize);
5401 range_tree_add(msp->ms_freeing, offset, asize);
5403 mutex_exit(&msp->ms_lock);
5407 metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5408 uint64_t size, void *arg)
5410 (void) inner_offset;
5411 boolean_t *checkpoint = arg;
5413 ASSERT3P(checkpoint, !=, NULL);
5415 if (vd->vdev_ops->vdev_op_remap != NULL)
5416 vdev_indirect_mark_obsolete(vd, offset, size);
5418 metaslab_free_impl(vd, offset, size, *checkpoint);
5422 metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
5423 boolean_t checkpoint)
5425 spa_t *spa = vd->vdev_spa;
5427 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5429 if (spa_syncing_txg(spa) > spa_freeze_txg(spa))
5432 if (spa->spa_vdev_removal != NULL &&
5433 spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id &&
5434 vdev_is_concrete(vd)) {
5436 * Note: we check if the vdev is concrete because when
5437 * we complete the removal, we first change the vdev to be
5438 * an indirect vdev (in open context), and then (in syncing
5439 * context) clear spa_vdev_removal.
5441 free_from_removing_vdev(vd, offset, size);
5442 } else if (vd->vdev_ops->vdev_op_remap != NULL) {
5443 vdev_indirect_mark_obsolete(vd, offset, size);
5444 vd->vdev_ops->vdev_op_remap(vd, offset, size,
5445 metaslab_free_impl_cb, &checkpoint);
5447 metaslab_free_concrete(vd, offset, size, checkpoint);
5451 typedef struct remap_blkptr_cb_arg {
5453 spa_remap_cb_t rbca_cb;
5454 vdev_t *rbca_remap_vd;
5455 uint64_t rbca_remap_offset;
5457 } remap_blkptr_cb_arg_t;
5460 remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5461 uint64_t size, void *arg)
5463 remap_blkptr_cb_arg_t *rbca = arg;
5464 blkptr_t *bp = rbca->rbca_bp;
5466 /* We can not remap split blocks. */
5467 if (size != DVA_GET_ASIZE(&bp->blk_dva[0]))
5469 ASSERT0(inner_offset);
5471 if (rbca->rbca_cb != NULL) {
5473 * At this point we know that we are not handling split
5474 * blocks and we invoke the callback on the previous
5475 * vdev which must be indirect.
5477 ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops);
5479 rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id,
5480 rbca->rbca_remap_offset, size, rbca->rbca_cb_arg);
5482 /* set up remap_blkptr_cb_arg for the next call */
5483 rbca->rbca_remap_vd = vd;
5484 rbca->rbca_remap_offset = offset;
5488 * The phys birth time is that of dva[0]. This ensures that we know
5489 * when each dva was written, so that resilver can determine which
5490 * blocks need to be scrubbed (i.e. those written during the time
5491 * the vdev was offline). It also ensures that the key used in
5492 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
5493 * we didn't change the phys_birth, a lookup in the ARC for a
5494 * remapped BP could find the data that was previously stored at
5495 * this vdev + offset.
5497 vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa,
5498 DVA_GET_VDEV(&bp->blk_dva[0]));
5499 vdev_indirect_births_t *vib = oldvd->vdev_indirect_births;
5500 bp->blk_phys_birth = vdev_indirect_births_physbirth(vib,
5501 DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0]));
5503 DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id);
5504 DVA_SET_OFFSET(&bp->blk_dva[0], offset);
5508 * If the block pointer contains any indirect DVAs, modify them to refer to
5509 * concrete DVAs. Note that this will sometimes not be possible, leaving
5510 * the indirect DVA in place. This happens if the indirect DVA spans multiple
5511 * segments in the mapping (i.e. it is a "split block").
5513 * If the BP was remapped, calls the callback on the original dva (note the
5514 * callback can be called multiple times if the original indirect DVA refers
5515 * to another indirect DVA, etc).
5517 * Returns TRUE if the BP was remapped.
5520 spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg)
5522 remap_blkptr_cb_arg_t rbca;
5524 if (!zfs_remap_blkptr_enable)
5527 if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS))
5531 * Dedup BP's can not be remapped, because ddt_phys_select() depends
5532 * on DVA[0] being the same in the BP as in the DDT (dedup table).
5534 if (BP_GET_DEDUP(bp))
5538 * Gang blocks can not be remapped, because
5539 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
5540 * the BP used to read the gang block header (GBH) being the same
5541 * as the DVA[0] that we allocated for the GBH.
5547 * Embedded BP's have no DVA to remap.
5549 if (BP_GET_NDVAS(bp) < 1)
5553 * Note: we only remap dva[0]. If we remapped other dvas, we
5554 * would no longer know what their phys birth txg is.
5556 dva_t *dva = &bp->blk_dva[0];
5558 uint64_t offset = DVA_GET_OFFSET(dva);
5559 uint64_t size = DVA_GET_ASIZE(dva);
5560 vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
5562 if (vd->vdev_ops->vdev_op_remap == NULL)
5566 rbca.rbca_cb = callback;
5567 rbca.rbca_remap_vd = vd;
5568 rbca.rbca_remap_offset = offset;
5569 rbca.rbca_cb_arg = arg;
5572 * remap_blkptr_cb() will be called in order for each level of
5573 * indirection, until a concrete vdev is reached or a split block is
5574 * encountered. old_vd and old_offset are updated within the callback
5575 * as we go from the one indirect vdev to the next one (either concrete
5576 * or indirect again) in that order.
5578 vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca);
5580 /* Check if the DVA wasn't remapped because it is a split block */
5581 if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id)
5588 * Undo the allocation of a DVA which happened in the given transaction group.
5591 metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
5595 uint64_t vdev = DVA_GET_VDEV(dva);
5596 uint64_t offset = DVA_GET_OFFSET(dva);
5597 uint64_t size = DVA_GET_ASIZE(dva);
5599 ASSERT(DVA_IS_VALID(dva));
5600 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5602 if (txg > spa_freeze_txg(spa))
5605 if ((vd = vdev_lookup_top(spa, vdev)) == NULL || !DVA_IS_VALID(dva) ||
5606 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
5607 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
5608 (u_longlong_t)vdev, (u_longlong_t)offset,
5609 (u_longlong_t)size);
5613 ASSERT(!vd->vdev_removing);
5614 ASSERT(vdev_is_concrete(vd));
5615 ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
5616 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
5618 if (DVA_GET_GANG(dva))
5619 size = vdev_gang_header_asize(vd);
5621 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5623 mutex_enter(&msp->ms_lock);
5624 range_tree_remove(msp->ms_allocating[txg & TXG_MASK],
5626 msp->ms_allocating_total -= size;
5628 VERIFY(!msp->ms_condensing);
5629 VERIFY3U(offset, >=, msp->ms_start);
5630 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
5631 VERIFY3U(range_tree_space(msp->ms_allocatable) + size, <=,
5633 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5634 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
5635 range_tree_add(msp->ms_allocatable, offset, size);
5636 mutex_exit(&msp->ms_lock);
5640 * Free the block represented by the given DVA.
5643 metaslab_free_dva(spa_t *spa, const dva_t *dva, boolean_t checkpoint)
5645 uint64_t vdev = DVA_GET_VDEV(dva);
5646 uint64_t offset = DVA_GET_OFFSET(dva);
5647 uint64_t size = DVA_GET_ASIZE(dva);
5648 vdev_t *vd = vdev_lookup_top(spa, vdev);
5650 ASSERT(DVA_IS_VALID(dva));
5651 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5653 if (DVA_GET_GANG(dva)) {
5654 size = vdev_gang_header_asize(vd);
5657 metaslab_free_impl(vd, offset, size, checkpoint);
5661 * Reserve some allocation slots. The reservation system must be called
5662 * before we call into the allocator. If there aren't any available slots
5663 * then the I/O will be throttled until an I/O completes and its slots are
5664 * freed up. The function returns true if it was successful in placing
5668 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
5669 zio_t *zio, int flags)
5671 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
5672 uint64_t max = mca->mca_alloc_max_slots;
5674 ASSERT(mc->mc_alloc_throttle_enabled);
5675 if (GANG_ALLOCATION(flags) || (flags & METASLAB_MUST_RESERVE) ||
5676 zfs_refcount_count(&mca->mca_alloc_slots) + slots <= max) {
5678 * The potential race between _count() and _add() is covered
5679 * by the allocator lock in most cases, or irrelevant due to
5680 * GANG_ALLOCATION() or METASLAB_MUST_RESERVE set in others.
5681 * But even if we assume some other non-existing scenario, the
5682 * worst that can happen is few more I/Os get to allocation
5683 * earlier, that is not a problem.
5685 * We reserve the slots individually so that we can unreserve
5686 * them individually when an I/O completes.
5688 zfs_refcount_add_few(&mca->mca_alloc_slots, slots, zio);
5689 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
5696 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots,
5697 int allocator, zio_t *zio)
5699 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
5701 ASSERT(mc->mc_alloc_throttle_enabled);
5702 zfs_refcount_remove_few(&mca->mca_alloc_slots, slots, zio);
5706 metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
5710 spa_t *spa = vd->vdev_spa;
5713 if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count)
5714 return (SET_ERROR(ENXIO));
5716 ASSERT3P(vd->vdev_ms, !=, NULL);
5717 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5719 mutex_enter(&msp->ms_lock);
5721 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded) {
5722 error = metaslab_activate(msp, 0, METASLAB_WEIGHT_CLAIM);
5723 if (error == EBUSY) {
5724 ASSERT(msp->ms_loaded);
5725 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
5731 !range_tree_contains(msp->ms_allocatable, offset, size))
5732 error = SET_ERROR(ENOENT);
5734 if (error || txg == 0) { /* txg == 0 indicates dry run */
5735 mutex_exit(&msp->ms_lock);
5739 VERIFY(!msp->ms_condensing);
5740 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5741 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
5742 VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=,
5744 range_tree_remove(msp->ms_allocatable, offset, size);
5745 range_tree_clear(msp->ms_trim, offset, size);
5747 if (spa_writeable(spa)) { /* don't dirty if we're zdb(8) */
5748 metaslab_class_t *mc = msp->ms_group->mg_class;
5749 multilist_sublist_t *mls =
5750 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
5751 if (!multilist_link_active(&msp->ms_class_txg_node)) {
5752 msp->ms_selected_txg = txg;
5753 multilist_sublist_insert_head(mls, msp);
5755 multilist_sublist_unlock(mls);
5757 if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
5758 vdev_dirty(vd, VDD_METASLAB, msp, txg);
5759 range_tree_add(msp->ms_allocating[txg & TXG_MASK],
5761 msp->ms_allocating_total += size;
5764 mutex_exit(&msp->ms_lock);
5769 typedef struct metaslab_claim_cb_arg_t {
5772 } metaslab_claim_cb_arg_t;
5775 metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5776 uint64_t size, void *arg)
5778 (void) inner_offset;
5779 metaslab_claim_cb_arg_t *mcca_arg = arg;
5781 if (mcca_arg->mcca_error == 0) {
5782 mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset,
5783 size, mcca_arg->mcca_txg);
5788 metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
5790 if (vd->vdev_ops->vdev_op_remap != NULL) {
5791 metaslab_claim_cb_arg_t arg;
5794 * Only zdb(8) can claim on indirect vdevs. This is used
5795 * to detect leaks of mapped space (that are not accounted
5796 * for in the obsolete counts, spacemap, or bpobj).
5798 ASSERT(!spa_writeable(vd->vdev_spa));
5802 vd->vdev_ops->vdev_op_remap(vd, offset, size,
5803 metaslab_claim_impl_cb, &arg);
5805 if (arg.mcca_error == 0) {
5806 arg.mcca_error = metaslab_claim_concrete(vd,
5809 return (arg.mcca_error);
5811 return (metaslab_claim_concrete(vd, offset, size, txg));
5816 * Intent log support: upon opening the pool after a crash, notify the SPA
5817 * of blocks that the intent log has allocated for immediate write, but
5818 * which are still considered free by the SPA because the last transaction
5819 * group didn't commit yet.
5822 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
5824 uint64_t vdev = DVA_GET_VDEV(dva);
5825 uint64_t offset = DVA_GET_OFFSET(dva);
5826 uint64_t size = DVA_GET_ASIZE(dva);
5829 if ((vd = vdev_lookup_top(spa, vdev)) == NULL) {
5830 return (SET_ERROR(ENXIO));
5833 ASSERT(DVA_IS_VALID(dva));
5835 if (DVA_GET_GANG(dva))
5836 size = vdev_gang_header_asize(vd);
5838 return (metaslab_claim_impl(vd, offset, size, txg));
5842 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
5843 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
5844 zio_alloc_list_t *zal, zio_t *zio, int allocator)
5846 dva_t *dva = bp->blk_dva;
5847 dva_t *hintdva = (hintbp != NULL) ? hintbp->blk_dva : NULL;
5850 ASSERT(bp->blk_birth == 0);
5851 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
5853 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
5855 if (mc->mc_allocator[allocator].mca_rotor == NULL) {
5856 /* no vdevs in this class */
5857 spa_config_exit(spa, SCL_ALLOC, FTAG);
5858 return (SET_ERROR(ENOSPC));
5861 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
5862 ASSERT(BP_GET_NDVAS(bp) == 0);
5863 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
5864 ASSERT3P(zal, !=, NULL);
5866 for (int d = 0; d < ndvas; d++) {
5867 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
5868 txg, flags, zal, allocator);
5870 for (d--; d >= 0; d--) {
5871 metaslab_unalloc_dva(spa, &dva[d], txg);
5872 metaslab_group_alloc_decrement(spa,
5873 DVA_GET_VDEV(&dva[d]), zio, flags,
5874 allocator, B_FALSE);
5875 memset(&dva[d], 0, sizeof (dva_t));
5877 spa_config_exit(spa, SCL_ALLOC, FTAG);
5881 * Update the metaslab group's queue depth
5882 * based on the newly allocated dva.
5884 metaslab_group_alloc_increment(spa,
5885 DVA_GET_VDEV(&dva[d]), zio, flags, allocator);
5889 ASSERT(BP_GET_NDVAS(bp) == ndvas);
5891 spa_config_exit(spa, SCL_ALLOC, FTAG);
5893 BP_SET_BIRTH(bp, txg, 0);
5899 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
5901 const dva_t *dva = bp->blk_dva;
5902 int ndvas = BP_GET_NDVAS(bp);
5904 ASSERT(!BP_IS_HOLE(bp));
5905 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
5908 * If we have a checkpoint for the pool we need to make sure that
5909 * the blocks that we free that are part of the checkpoint won't be
5910 * reused until the checkpoint is discarded or we revert to it.
5912 * The checkpoint flag is passed down the metaslab_free code path
5913 * and is set whenever we want to add a block to the checkpoint's
5914 * accounting. That is, we "checkpoint" blocks that existed at the
5915 * time the checkpoint was created and are therefore referenced by
5916 * the checkpointed uberblock.
5918 * Note that, we don't checkpoint any blocks if the current
5919 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
5920 * normally as they will be referenced by the checkpointed uberblock.
5922 boolean_t checkpoint = B_FALSE;
5923 if (bp->blk_birth <= spa->spa_checkpoint_txg &&
5924 spa_syncing_txg(spa) > spa->spa_checkpoint_txg) {
5926 * At this point, if the block is part of the checkpoint
5927 * there is no way it was created in the current txg.
5930 ASSERT3U(spa_syncing_txg(spa), ==, txg);
5931 checkpoint = B_TRUE;
5934 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
5936 for (int d = 0; d < ndvas; d++) {
5938 metaslab_unalloc_dva(spa, &dva[d], txg);
5940 ASSERT3U(txg, ==, spa_syncing_txg(spa));
5941 metaslab_free_dva(spa, &dva[d], checkpoint);
5945 spa_config_exit(spa, SCL_FREE, FTAG);
5949 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
5951 const dva_t *dva = bp->blk_dva;
5952 int ndvas = BP_GET_NDVAS(bp);
5955 ASSERT(!BP_IS_HOLE(bp));
5959 * First do a dry run to make sure all DVAs are claimable,
5960 * so we don't have to unwind from partial failures below.
5962 if ((error = metaslab_claim(spa, bp, 0)) != 0)
5966 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
5968 for (int d = 0; d < ndvas; d++) {
5969 error = metaslab_claim_dva(spa, &dva[d], txg);
5974 spa_config_exit(spa, SCL_ALLOC, FTAG);
5976 ASSERT(error == 0 || txg == 0);
5982 metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset,
5983 uint64_t size, void *arg)
5985 (void) inner, (void) arg;
5987 if (vd->vdev_ops == &vdev_indirect_ops)
5990 metaslab_check_free_impl(vd, offset, size);
5994 metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size)
5997 spa_t *spa __maybe_unused = vd->vdev_spa;
5999 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
6002 if (vd->vdev_ops->vdev_op_remap != NULL) {
6003 vd->vdev_ops->vdev_op_remap(vd, offset, size,
6004 metaslab_check_free_impl_cb, NULL);
6008 ASSERT(vdev_is_concrete(vd));
6009 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
6010 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
6012 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
6014 mutex_enter(&msp->ms_lock);
6015 if (msp->ms_loaded) {
6016 range_tree_verify_not_present(msp->ms_allocatable,
6021 * Check all segments that currently exist in the freeing pipeline.
6023 * It would intuitively make sense to also check the current allocating
6024 * tree since metaslab_unalloc_dva() exists for extents that are
6025 * allocated and freed in the same sync pass within the same txg.
6026 * Unfortunately there are places (e.g. the ZIL) where we allocate a
6027 * segment but then we free part of it within the same txg
6028 * [see zil_sync()]. Thus, we don't call range_tree_verify() in the
6029 * current allocating tree.
6031 range_tree_verify_not_present(msp->ms_freeing, offset, size);
6032 range_tree_verify_not_present(msp->ms_checkpointing, offset, size);
6033 range_tree_verify_not_present(msp->ms_freed, offset, size);
6034 for (int j = 0; j < TXG_DEFER_SIZE; j++)
6035 range_tree_verify_not_present(msp->ms_defer[j], offset, size);
6036 range_tree_verify_not_present(msp->ms_trim, offset, size);
6037 mutex_exit(&msp->ms_lock);
6041 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
6043 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
6046 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
6047 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
6048 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
6049 vdev_t *vd = vdev_lookup_top(spa, vdev);
6050 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
6051 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
6053 if (DVA_GET_GANG(&bp->blk_dva[i]))
6054 size = vdev_gang_header_asize(vd);
6056 ASSERT3P(vd, !=, NULL);
6058 metaslab_check_free_impl(vd, offset, size);
6060 spa_config_exit(spa, SCL_VDEV, FTAG);
6064 metaslab_group_disable_wait(metaslab_group_t *mg)
6066 ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
6067 while (mg->mg_disabled_updating) {
6068 cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
6073 metaslab_group_disabled_increment(metaslab_group_t *mg)
6075 ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
6076 ASSERT(mg->mg_disabled_updating);
6078 while (mg->mg_ms_disabled >= max_disabled_ms) {
6079 cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
6081 mg->mg_ms_disabled++;
6082 ASSERT3U(mg->mg_ms_disabled, <=, max_disabled_ms);
6086 * Mark the metaslab as disabled to prevent any allocations on this metaslab.
6087 * We must also track how many metaslabs are currently disabled within a
6088 * metaslab group and limit them to prevent allocation failures from
6089 * occurring because all metaslabs are disabled.
6092 metaslab_disable(metaslab_t *msp)
6094 ASSERT(!MUTEX_HELD(&msp->ms_lock));
6095 metaslab_group_t *mg = msp->ms_group;
6097 mutex_enter(&mg->mg_ms_disabled_lock);
6100 * To keep an accurate count of how many threads have disabled
6101 * a specific metaslab group, we only allow one thread to mark
6102 * the metaslab group at a time. This ensures that the value of
6103 * ms_disabled will be accurate when we decide to mark a metaslab
6104 * group as disabled. To do this we force all other threads
6105 * to wait till the metaslab's mg_disabled_updating flag is no
6108 metaslab_group_disable_wait(mg);
6109 mg->mg_disabled_updating = B_TRUE;
6110 if (msp->ms_disabled == 0) {
6111 metaslab_group_disabled_increment(mg);
6113 mutex_enter(&msp->ms_lock);
6115 mutex_exit(&msp->ms_lock);
6117 mg->mg_disabled_updating = B_FALSE;
6118 cv_broadcast(&mg->mg_ms_disabled_cv);
6119 mutex_exit(&mg->mg_ms_disabled_lock);
6123 metaslab_enable(metaslab_t *msp, boolean_t sync, boolean_t unload)
6125 metaslab_group_t *mg = msp->ms_group;
6126 spa_t *spa = mg->mg_vd->vdev_spa;
6129 * Wait for the outstanding IO to be synced to prevent newly
6130 * allocated blocks from being overwritten. This used by
6131 * initialize and TRIM which are modifying unallocated space.
6134 txg_wait_synced(spa_get_dsl(spa), 0);
6136 mutex_enter(&mg->mg_ms_disabled_lock);
6137 mutex_enter(&msp->ms_lock);
6138 if (--msp->ms_disabled == 0) {
6139 mg->mg_ms_disabled--;
6140 cv_broadcast(&mg->mg_ms_disabled_cv);
6142 metaslab_unload(msp);
6144 mutex_exit(&msp->ms_lock);
6145 mutex_exit(&mg->mg_ms_disabled_lock);
6149 metaslab_set_unflushed_dirty(metaslab_t *ms, boolean_t dirty)
6151 ms->ms_unflushed_dirty = dirty;
6155 metaslab_update_ondisk_flush_data(metaslab_t *ms, dmu_tx_t *tx)
6157 vdev_t *vd = ms->ms_group->mg_vd;
6158 spa_t *spa = vd->vdev_spa;
6159 objset_t *mos = spa_meta_objset(spa);
6161 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
6163 metaslab_unflushed_phys_t entry = {
6164 .msp_unflushed_txg = metaslab_unflushed_txg(ms),
6166 uint64_t entry_size = sizeof (entry);
6167 uint64_t entry_offset = ms->ms_id * entry_size;
6169 uint64_t object = 0;
6170 int err = zap_lookup(mos, vd->vdev_top_zap,
6171 VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1,
6173 if (err == ENOENT) {
6174 object = dmu_object_alloc(mos, DMU_OTN_UINT64_METADATA,
6175 SPA_OLD_MAXBLOCKSIZE, DMU_OT_NONE, 0, tx);
6176 VERIFY0(zap_add(mos, vd->vdev_top_zap,
6177 VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1,
6183 dmu_write(spa_meta_objset(spa), object, entry_offset, entry_size,
6188 metaslab_set_unflushed_txg(metaslab_t *ms, uint64_t txg, dmu_tx_t *tx)
6190 ms->ms_unflushed_txg = txg;
6191 metaslab_update_ondisk_flush_data(ms, tx);
6195 metaslab_unflushed_dirty(metaslab_t *ms)
6197 return (ms->ms_unflushed_dirty);
6201 metaslab_unflushed_txg(metaslab_t *ms)
6203 return (ms->ms_unflushed_txg);
6206 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, aliquot, U64, ZMOD_RW,
6207 "Allocation granularity (a.k.a. stripe size)");
6209 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, debug_load, INT, ZMOD_RW,
6210 "Load all metaslabs when pool is first opened");
6212 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, debug_unload, INT, ZMOD_RW,
6213 "Prevent metaslabs from being unloaded");
6215 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, preload_enabled, INT, ZMOD_RW,
6216 "Preload potential metaslabs during reassessment");
6218 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, unload_delay, UINT, ZMOD_RW,
6219 "Delay in txgs after metaslab was last used before unloading");
6221 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, unload_delay_ms, UINT, ZMOD_RW,
6222 "Delay in milliseconds after metaslab was last used before unloading");
6225 ZFS_MODULE_PARAM(zfs_mg, zfs_mg_, noalloc_threshold, UINT, ZMOD_RW,
6226 "Percentage of metaslab group size that should be free to make it "
6227 "eligible for allocation");
6229 ZFS_MODULE_PARAM(zfs_mg, zfs_mg_, fragmentation_threshold, UINT, ZMOD_RW,
6230 "Percentage of metaslab group size that should be considered eligible "
6231 "for allocations unless all metaslab groups within the metaslab class "
6232 "have also crossed this threshold");
6234 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, fragmentation_factor_enabled, INT,
6236 "Use the fragmentation metric to prefer less fragmented metaslabs");
6239 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, fragmentation_threshold, UINT,
6240 ZMOD_RW, "Fragmentation for metaslab to allow allocation");
6242 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, lba_weighting_enabled, INT, ZMOD_RW,
6243 "Prefer metaslabs with lower LBAs");
6245 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, bias_enabled, INT, ZMOD_RW,
6246 "Enable metaslab group biasing");
6248 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, segment_weight_enabled, INT,
6249 ZMOD_RW, "Enable segment-based metaslab selection");
6251 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, switch_threshold, INT, ZMOD_RW,
6252 "Segment-based metaslab selection maximum buckets before switching");
6254 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, force_ganging, U64, ZMOD_RW,
6255 "Blocks larger than this size are sometimes forced to be gang blocks");
6257 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, force_ganging_pct, UINT, ZMOD_RW,
6258 "Percentage of large blocks that will be forced to be gang blocks");
6260 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, df_max_search, UINT, ZMOD_RW,
6261 "Max distance (bytes) to search forward before using size tree");
6263 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, df_use_largest_segment, INT, ZMOD_RW,
6264 "When looking in size tree, use largest segment instead of exact fit");
6266 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, max_size_cache_sec, U64,
6267 ZMOD_RW, "How long to trust the cached max chunk size of a metaslab");
6269 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, mem_limit, UINT, ZMOD_RW,
6270 "Percentage of memory that can be used to store metaslab range trees");
6272 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, try_hard_before_gang, INT,
6273 ZMOD_RW, "Try hard to allocate before ganging");
6275 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, find_max_tries, UINT, ZMOD_RW,
6276 "Normally only consider this many of the best metaslabs in each vdev");
6279 ZFS_MODULE_PARAM_CALL(zfs, zfs_, active_allocator,
6280 param_set_active_allocator, param_get_charp, ZMOD_RW,
6281 "SPA active allocator");