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 WITH_DF_BLOCK_ALLOCATOR
45 #define GANG_ALLOCATION(flags) \
46 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
49 * Metaslab granularity, in bytes. This is roughly similar to what would be
50 * referred to as the "stripe size" in traditional RAID arrays. In normal
51 * operation, we will try to write this amount of data to each disk before
52 * moving on to the next top-level vdev.
54 static uint64_t metaslab_aliquot = 1024 * 1024;
57 * For testing, make some blocks above a certain size be gang blocks.
59 uint64_t metaslab_force_ganging = SPA_MAXBLOCKSIZE + 1;
62 * In pools where the log space map feature is not enabled we touch
63 * multiple metaslabs (and their respective space maps) with each
64 * transaction group. Thus, we benefit from having a small space map
65 * block size since it allows us to issue more I/O operations scattered
66 * around the disk. So a sane default for the space map block size
69 int zfs_metaslab_sm_blksz_no_log = (1 << 14);
72 * When the log space map feature is enabled, we accumulate a lot of
73 * changes per metaslab that are flushed once in a while so we benefit
74 * from a bigger block size like 128K for the metaslab space maps.
76 int zfs_metaslab_sm_blksz_with_log = (1 << 17);
79 * The in-core space map representation is more compact than its on-disk form.
80 * The zfs_condense_pct determines how much more compact the in-core
81 * space map representation must be before we compact it on-disk.
82 * Values should be greater than or equal to 100.
84 uint_t zfs_condense_pct = 200;
87 * Condensing a metaslab is not guaranteed to actually reduce the amount of
88 * space used on disk. In particular, a space map uses data in increments of
89 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
90 * same number of blocks after condensing. Since the goal of condensing is to
91 * reduce the number of IOPs required to read the space map, we only want to
92 * condense when we can be sure we will reduce the number of blocks used by the
93 * space map. Unfortunately, we cannot precisely compute whether or not this is
94 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
95 * we apply the following heuristic: do not condense a spacemap unless the
96 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
99 static const int zfs_metaslab_condense_block_threshold = 4;
102 * The zfs_mg_noalloc_threshold defines which metaslab groups should
103 * be eligible for allocation. The value is defined as a percentage of
104 * free space. Metaslab groups that have more free space than
105 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
106 * a metaslab group's free space is less than or equal to the
107 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
108 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
109 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
110 * groups are allowed to accept allocations. Gang blocks are always
111 * eligible to allocate on any metaslab group. The default value of 0 means
112 * no metaslab group will be excluded based on this criterion.
114 static uint_t zfs_mg_noalloc_threshold = 0;
117 * Metaslab groups are considered eligible for allocations if their
118 * fragmentation metric (measured as a percentage) is less than or
119 * equal to zfs_mg_fragmentation_threshold. If a metaslab group
120 * exceeds this threshold then it will be skipped unless all metaslab
121 * groups within the metaslab class have also crossed this threshold.
123 * This tunable was introduced to avoid edge cases where we continue
124 * allocating from very fragmented disks in our pool while other, less
125 * fragmented disks, exists. On the other hand, if all disks in the
126 * pool are uniformly approaching the threshold, the threshold can
127 * be a speed bump in performance, where we keep switching the disks
128 * that we allocate from (e.g. we allocate some segments from disk A
129 * making it bypassing the threshold while freeing segments from disk
130 * B getting its fragmentation below the threshold).
132 * Empirically, we've seen that our vdev selection for allocations is
133 * good enough that fragmentation increases uniformly across all vdevs
134 * the majority of the time. Thus we set the threshold percentage high
135 * enough to avoid hitting the speed bump on pools that are being pushed
138 static uint_t zfs_mg_fragmentation_threshold = 95;
141 * Allow metaslabs to keep their active state as long as their fragmentation
142 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
143 * active metaslab that exceeds this threshold will no longer keep its active
144 * status allowing better metaslabs to be selected.
146 static uint_t zfs_metaslab_fragmentation_threshold = 70;
149 * When set will load all metaslabs when pool is first opened.
151 int metaslab_debug_load = B_FALSE;
154 * When set will prevent metaslabs from being unloaded.
156 static int metaslab_debug_unload = B_FALSE;
159 * Minimum size which forces the dynamic allocator to change
160 * it's allocation strategy. Once the space map cannot satisfy
161 * an allocation of this size then it switches to using more
162 * aggressive strategy (i.e search by size rather than offset).
164 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
167 * The minimum free space, in percent, which must be available
168 * in a space map to continue allocations in a first-fit fashion.
169 * Once the space map's free space drops below this level we dynamically
170 * switch to using best-fit allocations.
172 uint_t metaslab_df_free_pct = 4;
175 * Maximum distance to search forward from the last offset. Without this
176 * limit, fragmented pools can see >100,000 iterations and
177 * metaslab_block_picker() becomes the performance limiting factor on
178 * high-performance storage.
180 * With the default setting of 16MB, we typically see less than 500
181 * iterations, even with very fragmented, ashift=9 pools. The maximum number
182 * of iterations possible is:
183 * metaslab_df_max_search / (2 * (1<<ashift))
184 * With the default setting of 16MB this is 16*1024 (with ashift=9) or
185 * 2048 (with ashift=12).
187 static uint_t metaslab_df_max_search = 16 * 1024 * 1024;
190 * Forces the metaslab_block_picker function to search for at least this many
191 * segments forwards until giving up on finding a segment that the allocation
194 static const uint32_t metaslab_min_search_count = 100;
197 * If we are not searching forward (due to metaslab_df_max_search,
198 * metaslab_df_free_pct, or metaslab_df_alloc_threshold), this tunable
199 * controls what segment is used. If it is set, we will use the largest free
200 * segment. If it is not set, we will use a segment of exactly the requested
203 static int metaslab_df_use_largest_segment = B_FALSE;
206 * Percentage of all cpus that can be used by the metaslab taskq.
208 int metaslab_load_pct = 50;
211 * These tunables control how long a metaslab will remain loaded after the
212 * last allocation from it. A metaslab can't be unloaded until at least
213 * metaslab_unload_delay TXG's and metaslab_unload_delay_ms milliseconds
214 * have elapsed. However, zfs_metaslab_mem_limit may cause it to be
215 * unloaded sooner. These settings are intended to be generous -- to keep
216 * metaslabs loaded for a long time, reducing the rate of metaslab loading.
218 static uint_t metaslab_unload_delay = 32;
219 static uint_t metaslab_unload_delay_ms = 10 * 60 * 1000; /* ten minutes */
222 * Max number of metaslabs per group to preload.
224 uint_t metaslab_preload_limit = 10;
227 * Enable/disable preloading of metaslab.
229 static int metaslab_preload_enabled = B_TRUE;
232 * Enable/disable fragmentation weighting on metaslabs.
234 static int metaslab_fragmentation_factor_enabled = B_TRUE;
237 * Enable/disable lba weighting (i.e. outer tracks are given preference).
239 static int metaslab_lba_weighting_enabled = B_TRUE;
242 * Enable/disable metaslab group biasing.
244 static int metaslab_bias_enabled = B_TRUE;
247 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
249 static const boolean_t zfs_remap_blkptr_enable = B_TRUE;
252 * Enable/disable segment-based metaslab selection.
254 static int zfs_metaslab_segment_weight_enabled = B_TRUE;
257 * When using segment-based metaslab selection, we will continue
258 * allocating from the active metaslab until we have exhausted
259 * zfs_metaslab_switch_threshold of its buckets.
261 static int zfs_metaslab_switch_threshold = 2;
264 * Internal switch to enable/disable the metaslab allocation tracing
267 static const boolean_t metaslab_trace_enabled = B_FALSE;
270 * Maximum entries that the metaslab allocation tracing facility will keep
271 * in a given list when running in non-debug mode. We limit the number
272 * of entries in non-debug mode to prevent us from using up too much memory.
273 * The limit should be sufficiently large that we don't expect any allocation
274 * to every exceed this value. In debug mode, the system will panic if this
275 * limit is ever reached allowing for further investigation.
277 static const uint64_t metaslab_trace_max_entries = 5000;
280 * Maximum number of metaslabs per group that can be disabled
283 static const int max_disabled_ms = 3;
286 * Time (in seconds) to respect ms_max_size when the metaslab is not loaded.
287 * To avoid 64-bit overflow, don't set above UINT32_MAX.
289 static uint64_t zfs_metaslab_max_size_cache_sec = 1 * 60 * 60; /* 1 hour */
292 * Maximum percentage of memory to use on storing loaded metaslabs. If loading
293 * a metaslab would take it over this percentage, the oldest selected metaslab
294 * is automatically unloaded.
296 static uint_t zfs_metaslab_mem_limit = 25;
299 * Force the per-metaslab range trees to use 64-bit integers to store
300 * segments. Used for debugging purposes.
302 static const boolean_t zfs_metaslab_force_large_segs = B_FALSE;
305 * By default we only store segments over a certain size in the size-sorted
306 * metaslab trees (ms_allocatable_by_size and
307 * ms_unflushed_frees_by_size). This dramatically reduces memory usage and
308 * improves load and unload times at the cost of causing us to use slightly
309 * larger segments than we would otherwise in some cases.
311 static const uint32_t metaslab_by_size_min_shift = 14;
314 * If not set, we will first try normal allocation. If that fails then
315 * we will do a gang allocation. If that fails then we will do a "try hard"
316 * gang allocation. If that fails then we will have a multi-layer gang
319 * If set, we will first try normal allocation. If that fails then
320 * we will do a "try hard" allocation. If that fails we will do a gang
321 * allocation. If that fails we will do a "try hard" gang allocation. If
322 * that fails then we will have a multi-layer gang block.
324 static int zfs_metaslab_try_hard_before_gang = B_FALSE;
327 * When not trying hard, we only consider the best zfs_metaslab_find_max_tries
328 * metaslabs. This improves performance, especially when there are many
329 * metaslabs per vdev and the allocation can't actually be satisfied (so we
330 * would otherwise iterate all the metaslabs). If there is a metaslab with a
331 * worse weight but it can actually satisfy the allocation, we won't find it
332 * until trying hard. This may happen if the worse metaslab is not loaded
333 * (and the true weight is better than we have calculated), or due to weight
334 * bucketization. E.g. we are looking for a 60K segment, and the best
335 * metaslabs all have free segments in the 32-63K bucket, but the best
336 * zfs_metaslab_find_max_tries metaslabs have ms_max_size <60KB, and a
337 * subsequent metaslab has ms_max_size >60KB (but fewer segments in this
338 * bucket, and therefore a lower weight).
340 static uint_t zfs_metaslab_find_max_tries = 100;
342 static uint64_t metaslab_weight(metaslab_t *, boolean_t);
343 static void metaslab_set_fragmentation(metaslab_t *, boolean_t);
344 static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t);
345 static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t);
347 static void metaslab_passivate(metaslab_t *msp, uint64_t weight);
348 static uint64_t metaslab_weight_from_range_tree(metaslab_t *msp);
349 static void metaslab_flush_update(metaslab_t *, dmu_tx_t *);
350 static unsigned int metaslab_idx_func(multilist_t *, void *);
351 static void metaslab_evict(metaslab_t *, uint64_t);
352 static void metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg);
353 kmem_cache_t *metaslab_alloc_trace_cache;
355 typedef struct metaslab_stats {
356 kstat_named_t metaslabstat_trace_over_limit;
357 kstat_named_t metaslabstat_reload_tree;
358 kstat_named_t metaslabstat_too_many_tries;
359 kstat_named_t metaslabstat_try_hard;
362 static metaslab_stats_t metaslab_stats = {
363 { "trace_over_limit", KSTAT_DATA_UINT64 },
364 { "reload_tree", KSTAT_DATA_UINT64 },
365 { "too_many_tries", KSTAT_DATA_UINT64 },
366 { "try_hard", KSTAT_DATA_UINT64 },
369 #define METASLABSTAT_BUMP(stat) \
370 atomic_inc_64(&metaslab_stats.stat.value.ui64);
373 static kstat_t *metaslab_ksp;
376 metaslab_stat_init(void)
378 ASSERT(metaslab_alloc_trace_cache == NULL);
379 metaslab_alloc_trace_cache = kmem_cache_create(
380 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
381 0, NULL, NULL, NULL, NULL, NULL, 0);
382 metaslab_ksp = kstat_create("zfs", 0, "metaslab_stats",
383 "misc", KSTAT_TYPE_NAMED, sizeof (metaslab_stats) /
384 sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
385 if (metaslab_ksp != NULL) {
386 metaslab_ksp->ks_data = &metaslab_stats;
387 kstat_install(metaslab_ksp);
392 metaslab_stat_fini(void)
394 if (metaslab_ksp != NULL) {
395 kstat_delete(metaslab_ksp);
399 kmem_cache_destroy(metaslab_alloc_trace_cache);
400 metaslab_alloc_trace_cache = NULL;
404 * ==========================================================================
406 * ==========================================================================
409 metaslab_class_create(spa_t *spa, const metaslab_ops_t *ops)
411 metaslab_class_t *mc;
413 mc = kmem_zalloc(offsetof(metaslab_class_t,
414 mc_allocator[spa->spa_alloc_count]), KM_SLEEP);
418 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
419 multilist_create(&mc->mc_metaslab_txg_list, sizeof (metaslab_t),
420 offsetof(metaslab_t, ms_class_txg_node), metaslab_idx_func);
421 for (int i = 0; i < spa->spa_alloc_count; i++) {
422 metaslab_class_allocator_t *mca = &mc->mc_allocator[i];
423 mca->mca_rotor = NULL;
424 zfs_refcount_create_tracked(&mca->mca_alloc_slots);
431 metaslab_class_destroy(metaslab_class_t *mc)
433 spa_t *spa = mc->mc_spa;
435 ASSERT(mc->mc_alloc == 0);
436 ASSERT(mc->mc_deferred == 0);
437 ASSERT(mc->mc_space == 0);
438 ASSERT(mc->mc_dspace == 0);
440 for (int i = 0; i < spa->spa_alloc_count; i++) {
441 metaslab_class_allocator_t *mca = &mc->mc_allocator[i];
442 ASSERT(mca->mca_rotor == NULL);
443 zfs_refcount_destroy(&mca->mca_alloc_slots);
445 mutex_destroy(&mc->mc_lock);
446 multilist_destroy(&mc->mc_metaslab_txg_list);
447 kmem_free(mc, offsetof(metaslab_class_t,
448 mc_allocator[spa->spa_alloc_count]));
452 metaslab_class_validate(metaslab_class_t *mc)
454 metaslab_group_t *mg;
458 * Must hold one of the spa_config locks.
460 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
461 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
463 if ((mg = mc->mc_allocator[0].mca_rotor) == NULL)
468 ASSERT(vd->vdev_mg != NULL);
469 ASSERT3P(vd->vdev_top, ==, vd);
470 ASSERT3P(mg->mg_class, ==, mc);
471 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
472 } while ((mg = mg->mg_next) != mc->mc_allocator[0].mca_rotor);
478 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
479 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
481 atomic_add_64(&mc->mc_alloc, alloc_delta);
482 atomic_add_64(&mc->mc_deferred, defer_delta);
483 atomic_add_64(&mc->mc_space, space_delta);
484 atomic_add_64(&mc->mc_dspace, dspace_delta);
488 metaslab_class_get_alloc(metaslab_class_t *mc)
490 return (mc->mc_alloc);
494 metaslab_class_get_deferred(metaslab_class_t *mc)
496 return (mc->mc_deferred);
500 metaslab_class_get_space(metaslab_class_t *mc)
502 return (mc->mc_space);
506 metaslab_class_get_dspace(metaslab_class_t *mc)
508 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
512 metaslab_class_histogram_verify(metaslab_class_t *mc)
514 spa_t *spa = mc->mc_spa;
515 vdev_t *rvd = spa->spa_root_vdev;
519 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
522 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
525 mutex_enter(&mc->mc_lock);
526 for (int c = 0; c < rvd->vdev_children; c++) {
527 vdev_t *tvd = rvd->vdev_child[c];
528 metaslab_group_t *mg = vdev_get_mg(tvd, mc);
531 * Skip any holes, uninitialized top-levels, or
532 * vdevs that are not in this metalab class.
534 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
535 mg->mg_class != mc) {
539 IMPLY(mg == mg->mg_vd->vdev_log_mg,
540 mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
542 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
543 mc_hist[i] += mg->mg_histogram[i];
546 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
547 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
550 mutex_exit(&mc->mc_lock);
551 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
555 * Calculate the metaslab class's fragmentation metric. The metric
556 * is weighted based on the space contribution of each metaslab group.
557 * The return value will be a number between 0 and 100 (inclusive), or
558 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
559 * zfs_frag_table for more information about the metric.
562 metaslab_class_fragmentation(metaslab_class_t *mc)
564 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
565 uint64_t fragmentation = 0;
567 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
569 for (int c = 0; c < rvd->vdev_children; c++) {
570 vdev_t *tvd = rvd->vdev_child[c];
571 metaslab_group_t *mg = tvd->vdev_mg;
574 * Skip any holes, uninitialized top-levels,
575 * or vdevs that are not in this metalab class.
577 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
578 mg->mg_class != mc) {
583 * If a metaslab group does not contain a fragmentation
584 * metric then just bail out.
586 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
587 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
588 return (ZFS_FRAG_INVALID);
592 * Determine how much this metaslab_group is contributing
593 * to the overall pool fragmentation metric.
595 fragmentation += mg->mg_fragmentation *
596 metaslab_group_get_space(mg);
598 fragmentation /= metaslab_class_get_space(mc);
600 ASSERT3U(fragmentation, <=, 100);
601 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
602 return (fragmentation);
606 * Calculate the amount of expandable space that is available in
607 * this metaslab class. If a device is expanded then its expandable
608 * space will be the amount of allocatable space that is currently not
609 * part of this metaslab class.
612 metaslab_class_expandable_space(metaslab_class_t *mc)
614 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
617 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
618 for (int c = 0; c < rvd->vdev_children; c++) {
619 vdev_t *tvd = rvd->vdev_child[c];
620 metaslab_group_t *mg = tvd->vdev_mg;
622 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
623 mg->mg_class != mc) {
628 * Calculate if we have enough space to add additional
629 * metaslabs. We report the expandable space in terms
630 * of the metaslab size since that's the unit of expansion.
632 space += P2ALIGN(tvd->vdev_max_asize - tvd->vdev_asize,
633 1ULL << tvd->vdev_ms_shift);
635 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
640 metaslab_class_evict_old(metaslab_class_t *mc, uint64_t txg)
642 multilist_t *ml = &mc->mc_metaslab_txg_list;
643 for (int i = 0; i < multilist_get_num_sublists(ml); i++) {
644 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
645 metaslab_t *msp = multilist_sublist_head(mls);
646 multilist_sublist_unlock(mls);
647 while (msp != NULL) {
648 mutex_enter(&msp->ms_lock);
651 * If the metaslab has been removed from the list
652 * (which could happen if we were at the memory limit
653 * and it was evicted during this loop), then we can't
654 * proceed and we should restart the sublist.
656 if (!multilist_link_active(&msp->ms_class_txg_node)) {
657 mutex_exit(&msp->ms_lock);
661 mls = multilist_sublist_lock(ml, i);
662 metaslab_t *next_msp = multilist_sublist_next(mls, msp);
663 multilist_sublist_unlock(mls);
665 msp->ms_selected_txg + metaslab_unload_delay &&
666 gethrtime() > msp->ms_selected_time +
667 (uint64_t)MSEC2NSEC(metaslab_unload_delay_ms)) {
668 metaslab_evict(msp, txg);
671 * Once we've hit a metaslab selected too
672 * recently to evict, we're done evicting for
675 mutex_exit(&msp->ms_lock);
678 mutex_exit(&msp->ms_lock);
685 metaslab_compare(const void *x1, const void *x2)
687 const metaslab_t *m1 = (const metaslab_t *)x1;
688 const metaslab_t *m2 = (const metaslab_t *)x2;
692 if (m1->ms_allocator != -1 && m1->ms_primary)
694 else if (m1->ms_allocator != -1 && !m1->ms_primary)
696 if (m2->ms_allocator != -1 && m2->ms_primary)
698 else if (m2->ms_allocator != -1 && !m2->ms_primary)
702 * Sort inactive metaslabs first, then primaries, then secondaries. When
703 * selecting a metaslab to allocate from, an allocator first tries its
704 * primary, then secondary active metaslab. If it doesn't have active
705 * metaslabs, or can't allocate from them, it searches for an inactive
706 * metaslab to activate. If it can't find a suitable one, it will steal
707 * a primary or secondary metaslab from another allocator.
714 int cmp = TREE_CMP(m2->ms_weight, m1->ms_weight);
718 IMPLY(TREE_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2);
720 return (TREE_CMP(m1->ms_start, m2->ms_start));
724 * ==========================================================================
726 * ==========================================================================
729 * Update the allocatable flag and the metaslab group's capacity.
730 * The allocatable flag is set to true if the capacity is below
731 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
732 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
733 * transitions from allocatable to non-allocatable or vice versa then the
734 * metaslab group's class is updated to reflect the transition.
737 metaslab_group_alloc_update(metaslab_group_t *mg)
739 vdev_t *vd = mg->mg_vd;
740 metaslab_class_t *mc = mg->mg_class;
741 vdev_stat_t *vs = &vd->vdev_stat;
742 boolean_t was_allocatable;
743 boolean_t was_initialized;
745 ASSERT(vd == vd->vdev_top);
746 ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==,
749 mutex_enter(&mg->mg_lock);
750 was_allocatable = mg->mg_allocatable;
751 was_initialized = mg->mg_initialized;
753 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
756 mutex_enter(&mc->mc_lock);
759 * If the metaslab group was just added then it won't
760 * have any space until we finish syncing out this txg.
761 * At that point we will consider it initialized and available
762 * for allocations. We also don't consider non-activated
763 * metaslab groups (e.g. vdevs that are in the middle of being removed)
764 * to be initialized, because they can't be used for allocation.
766 mg->mg_initialized = metaslab_group_initialized(mg);
767 if (!was_initialized && mg->mg_initialized) {
769 } else if (was_initialized && !mg->mg_initialized) {
770 ASSERT3U(mc->mc_groups, >, 0);
773 if (mg->mg_initialized)
774 mg->mg_no_free_space = B_FALSE;
777 * A metaslab group is considered allocatable if it has plenty
778 * of free space or is not heavily fragmented. We only take
779 * fragmentation into account if the metaslab group has a valid
780 * fragmentation metric (i.e. a value between 0 and 100).
782 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
783 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
784 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
785 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
788 * The mc_alloc_groups maintains a count of the number of
789 * groups in this metaslab class that are still above the
790 * zfs_mg_noalloc_threshold. This is used by the allocating
791 * threads to determine if they should avoid allocations to
792 * a given group. The allocator will avoid allocations to a group
793 * if that group has reached or is below the zfs_mg_noalloc_threshold
794 * and there are still other groups that are above the threshold.
795 * When a group transitions from allocatable to non-allocatable or
796 * vice versa we update the metaslab class to reflect that change.
797 * When the mc_alloc_groups value drops to 0 that means that all
798 * groups have reached the zfs_mg_noalloc_threshold making all groups
799 * eligible for allocations. This effectively means that all devices
800 * are balanced again.
802 if (was_allocatable && !mg->mg_allocatable)
803 mc->mc_alloc_groups--;
804 else if (!was_allocatable && mg->mg_allocatable)
805 mc->mc_alloc_groups++;
806 mutex_exit(&mc->mc_lock);
808 mutex_exit(&mg->mg_lock);
812 metaslab_sort_by_flushed(const void *va, const void *vb)
814 const metaslab_t *a = va;
815 const metaslab_t *b = vb;
817 int cmp = TREE_CMP(a->ms_unflushed_txg, b->ms_unflushed_txg);
821 uint64_t a_vdev_id = a->ms_group->mg_vd->vdev_id;
822 uint64_t b_vdev_id = b->ms_group->mg_vd->vdev_id;
823 cmp = TREE_CMP(a_vdev_id, b_vdev_id);
827 return (TREE_CMP(a->ms_id, b->ms_id));
831 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
833 metaslab_group_t *mg;
835 mg = kmem_zalloc(offsetof(metaslab_group_t,
836 mg_allocator[allocators]), KM_SLEEP);
837 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
838 mutex_init(&mg->mg_ms_disabled_lock, NULL, MUTEX_DEFAULT, NULL);
839 cv_init(&mg->mg_ms_disabled_cv, NULL, CV_DEFAULT, NULL);
840 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
841 sizeof (metaslab_t), offsetof(metaslab_t, ms_group_node));
844 mg->mg_activation_count = 0;
845 mg->mg_initialized = B_FALSE;
846 mg->mg_no_free_space = B_TRUE;
847 mg->mg_allocators = allocators;
849 for (int i = 0; i < allocators; i++) {
850 metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
851 zfs_refcount_create_tracked(&mga->mga_alloc_queue_depth);
854 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
855 maxclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT | TASKQ_DYNAMIC);
861 metaslab_group_destroy(metaslab_group_t *mg)
863 ASSERT(mg->mg_prev == NULL);
864 ASSERT(mg->mg_next == NULL);
866 * We may have gone below zero with the activation count
867 * either because we never activated in the first place or
868 * because we're done, and possibly removing the vdev.
870 ASSERT(mg->mg_activation_count <= 0);
872 taskq_destroy(mg->mg_taskq);
873 avl_destroy(&mg->mg_metaslab_tree);
874 mutex_destroy(&mg->mg_lock);
875 mutex_destroy(&mg->mg_ms_disabled_lock);
876 cv_destroy(&mg->mg_ms_disabled_cv);
878 for (int i = 0; i < mg->mg_allocators; i++) {
879 metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
880 zfs_refcount_destroy(&mga->mga_alloc_queue_depth);
882 kmem_free(mg, offsetof(metaslab_group_t,
883 mg_allocator[mg->mg_allocators]));
887 metaslab_group_activate(metaslab_group_t *mg)
889 metaslab_class_t *mc = mg->mg_class;
890 spa_t *spa = mc->mc_spa;
891 metaslab_group_t *mgprev, *mgnext;
893 ASSERT3U(spa_config_held(spa, SCL_ALLOC, RW_WRITER), !=, 0);
895 ASSERT(mg->mg_prev == NULL);
896 ASSERT(mg->mg_next == NULL);
897 ASSERT(mg->mg_activation_count <= 0);
899 if (++mg->mg_activation_count <= 0)
902 mg->mg_aliquot = metaslab_aliquot * MAX(1,
903 vdev_get_ndisks(mg->mg_vd) - vdev_get_nparity(mg->mg_vd));
904 metaslab_group_alloc_update(mg);
906 if ((mgprev = mc->mc_allocator[0].mca_rotor) == NULL) {
910 mgnext = mgprev->mg_next;
911 mg->mg_prev = mgprev;
912 mg->mg_next = mgnext;
913 mgprev->mg_next = mg;
914 mgnext->mg_prev = mg;
916 for (int i = 0; i < spa->spa_alloc_count; i++) {
917 mc->mc_allocator[i].mca_rotor = mg;
923 * Passivate a metaslab group and remove it from the allocation rotor.
924 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
925 * a metaslab group. This function will momentarily drop spa_config_locks
926 * that are lower than the SCL_ALLOC lock (see comment below).
929 metaslab_group_passivate(metaslab_group_t *mg)
931 metaslab_class_t *mc = mg->mg_class;
932 spa_t *spa = mc->mc_spa;
933 metaslab_group_t *mgprev, *mgnext;
934 int locks = spa_config_held(spa, SCL_ALL, RW_WRITER);
936 ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==,
937 (SCL_ALLOC | SCL_ZIO));
939 if (--mg->mg_activation_count != 0) {
940 for (int i = 0; i < spa->spa_alloc_count; i++)
941 ASSERT(mc->mc_allocator[i].mca_rotor != mg);
942 ASSERT(mg->mg_prev == NULL);
943 ASSERT(mg->mg_next == NULL);
944 ASSERT(mg->mg_activation_count < 0);
949 * The spa_config_lock is an array of rwlocks, ordered as
950 * follows (from highest to lowest):
951 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
952 * SCL_ZIO > SCL_FREE > SCL_VDEV
953 * (For more information about the spa_config_lock see spa_misc.c)
954 * The higher the lock, the broader its coverage. When we passivate
955 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
956 * config locks. However, the metaslab group's taskq might be trying
957 * to preload metaslabs so we must drop the SCL_ZIO lock and any
958 * lower locks to allow the I/O to complete. At a minimum,
959 * we continue to hold the SCL_ALLOC lock, which prevents any future
960 * allocations from taking place and any changes to the vdev tree.
962 spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa);
963 taskq_wait_outstanding(mg->mg_taskq, 0);
964 spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
965 metaslab_group_alloc_update(mg);
966 for (int i = 0; i < mg->mg_allocators; i++) {
967 metaslab_group_allocator_t *mga = &mg->mg_allocator[i];
968 metaslab_t *msp = mga->mga_primary;
970 mutex_enter(&msp->ms_lock);
971 metaslab_passivate(msp,
972 metaslab_weight_from_range_tree(msp));
973 mutex_exit(&msp->ms_lock);
975 msp = mga->mga_secondary;
977 mutex_enter(&msp->ms_lock);
978 metaslab_passivate(msp,
979 metaslab_weight_from_range_tree(msp));
980 mutex_exit(&msp->ms_lock);
984 mgprev = mg->mg_prev;
985 mgnext = mg->mg_next;
990 mgprev->mg_next = mgnext;
991 mgnext->mg_prev = mgprev;
993 for (int i = 0; i < spa->spa_alloc_count; i++) {
994 if (mc->mc_allocator[i].mca_rotor == mg)
995 mc->mc_allocator[i].mca_rotor = mgnext;
1003 metaslab_group_initialized(metaslab_group_t *mg)
1005 vdev_t *vd = mg->mg_vd;
1006 vdev_stat_t *vs = &vd->vdev_stat;
1008 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
1012 metaslab_group_get_space(metaslab_group_t *mg)
1015 * Note that the number of nodes in mg_metaslab_tree may be one less
1016 * than vdev_ms_count, due to the embedded log metaslab.
1018 mutex_enter(&mg->mg_lock);
1019 uint64_t ms_count = avl_numnodes(&mg->mg_metaslab_tree);
1020 mutex_exit(&mg->mg_lock);
1021 return ((1ULL << mg->mg_vd->vdev_ms_shift) * ms_count);
1025 metaslab_group_histogram_verify(metaslab_group_t *mg)
1028 avl_tree_t *t = &mg->mg_metaslab_tree;
1029 uint64_t ashift = mg->mg_vd->vdev_ashift;
1031 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
1034 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
1037 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
1038 SPACE_MAP_HISTOGRAM_SIZE + ashift);
1040 mutex_enter(&mg->mg_lock);
1041 for (metaslab_t *msp = avl_first(t);
1042 msp != NULL; msp = AVL_NEXT(t, msp)) {
1043 VERIFY3P(msp->ms_group, ==, mg);
1044 /* skip if not active */
1045 if (msp->ms_sm == NULL)
1048 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1049 mg_hist[i + ashift] +=
1050 msp->ms_sm->sm_phys->smp_histogram[i];
1054 for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
1055 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
1057 mutex_exit(&mg->mg_lock);
1059 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
1063 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
1065 metaslab_class_t *mc = mg->mg_class;
1066 uint64_t ashift = mg->mg_vd->vdev_ashift;
1068 ASSERT(MUTEX_HELD(&msp->ms_lock));
1069 if (msp->ms_sm == NULL)
1072 mutex_enter(&mg->mg_lock);
1073 mutex_enter(&mc->mc_lock);
1074 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1075 IMPLY(mg == mg->mg_vd->vdev_log_mg,
1076 mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
1077 mg->mg_histogram[i + ashift] +=
1078 msp->ms_sm->sm_phys->smp_histogram[i];
1079 mc->mc_histogram[i + ashift] +=
1080 msp->ms_sm->sm_phys->smp_histogram[i];
1082 mutex_exit(&mc->mc_lock);
1083 mutex_exit(&mg->mg_lock);
1087 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
1089 metaslab_class_t *mc = mg->mg_class;
1090 uint64_t ashift = mg->mg_vd->vdev_ashift;
1092 ASSERT(MUTEX_HELD(&msp->ms_lock));
1093 if (msp->ms_sm == NULL)
1096 mutex_enter(&mg->mg_lock);
1097 mutex_enter(&mc->mc_lock);
1098 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1099 ASSERT3U(mg->mg_histogram[i + ashift], >=,
1100 msp->ms_sm->sm_phys->smp_histogram[i]);
1101 ASSERT3U(mc->mc_histogram[i + ashift], >=,
1102 msp->ms_sm->sm_phys->smp_histogram[i]);
1103 IMPLY(mg == mg->mg_vd->vdev_log_mg,
1104 mc == spa_embedded_log_class(mg->mg_vd->vdev_spa));
1106 mg->mg_histogram[i + ashift] -=
1107 msp->ms_sm->sm_phys->smp_histogram[i];
1108 mc->mc_histogram[i + ashift] -=
1109 msp->ms_sm->sm_phys->smp_histogram[i];
1111 mutex_exit(&mc->mc_lock);
1112 mutex_exit(&mg->mg_lock);
1116 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
1118 ASSERT(msp->ms_group == NULL);
1119 mutex_enter(&mg->mg_lock);
1122 avl_add(&mg->mg_metaslab_tree, msp);
1123 mutex_exit(&mg->mg_lock);
1125 mutex_enter(&msp->ms_lock);
1126 metaslab_group_histogram_add(mg, msp);
1127 mutex_exit(&msp->ms_lock);
1131 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
1133 mutex_enter(&msp->ms_lock);
1134 metaslab_group_histogram_remove(mg, msp);
1135 mutex_exit(&msp->ms_lock);
1137 mutex_enter(&mg->mg_lock);
1138 ASSERT(msp->ms_group == mg);
1139 avl_remove(&mg->mg_metaslab_tree, msp);
1141 metaslab_class_t *mc = msp->ms_group->mg_class;
1142 multilist_sublist_t *mls =
1143 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
1144 if (multilist_link_active(&msp->ms_class_txg_node))
1145 multilist_sublist_remove(mls, msp);
1146 multilist_sublist_unlock(mls);
1148 msp->ms_group = NULL;
1149 mutex_exit(&mg->mg_lock);
1153 metaslab_group_sort_impl(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
1155 ASSERT(MUTEX_HELD(&msp->ms_lock));
1156 ASSERT(MUTEX_HELD(&mg->mg_lock));
1157 ASSERT(msp->ms_group == mg);
1159 avl_remove(&mg->mg_metaslab_tree, msp);
1160 msp->ms_weight = weight;
1161 avl_add(&mg->mg_metaslab_tree, msp);
1166 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
1169 * Although in principle the weight can be any value, in
1170 * practice we do not use values in the range [1, 511].
1172 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
1173 ASSERT(MUTEX_HELD(&msp->ms_lock));
1175 mutex_enter(&mg->mg_lock);
1176 metaslab_group_sort_impl(mg, msp, weight);
1177 mutex_exit(&mg->mg_lock);
1181 * Calculate the fragmentation for a given metaslab group. We can use
1182 * a simple average here since all metaslabs within the group must have
1183 * the same size. The return value will be a value between 0 and 100
1184 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
1185 * group have a fragmentation metric.
1188 metaslab_group_fragmentation(metaslab_group_t *mg)
1190 vdev_t *vd = mg->mg_vd;
1191 uint64_t fragmentation = 0;
1192 uint64_t valid_ms = 0;
1194 for (int m = 0; m < vd->vdev_ms_count; m++) {
1195 metaslab_t *msp = vd->vdev_ms[m];
1197 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
1199 if (msp->ms_group != mg)
1203 fragmentation += msp->ms_fragmentation;
1206 if (valid_ms <= mg->mg_vd->vdev_ms_count / 2)
1207 return (ZFS_FRAG_INVALID);
1209 fragmentation /= valid_ms;
1210 ASSERT3U(fragmentation, <=, 100);
1211 return (fragmentation);
1215 * Determine if a given metaslab group should skip allocations. A metaslab
1216 * group should avoid allocations if its free capacity is less than the
1217 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
1218 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
1219 * that can still handle allocations. If the allocation throttle is enabled
1220 * then we skip allocations to devices that have reached their maximum
1221 * allocation queue depth unless the selected metaslab group is the only
1222 * eligible group remaining.
1225 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
1226 int flags, uint64_t psize, int allocator, int d)
1228 spa_t *spa = mg->mg_vd->vdev_spa;
1229 metaslab_class_t *mc = mg->mg_class;
1232 * We can only consider skipping this metaslab group if it's
1233 * in the normal metaslab class and there are other metaslab
1234 * groups to select from. Otherwise, we always consider it eligible
1237 if ((mc != spa_normal_class(spa) &&
1238 mc != spa_special_class(spa) &&
1239 mc != spa_dedup_class(spa)) ||
1244 * If the metaslab group's mg_allocatable flag is set (see comments
1245 * in metaslab_group_alloc_update() for more information) and
1246 * the allocation throttle is disabled then allow allocations to this
1247 * device. However, if the allocation throttle is enabled then
1248 * check if we have reached our allocation limit (mga_alloc_queue_depth)
1249 * to determine if we should allow allocations to this metaslab group.
1250 * If all metaslab groups are no longer considered allocatable
1251 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1252 * gang block size then we allow allocations on this metaslab group
1253 * regardless of the mg_allocatable or throttle settings.
1255 if (mg->mg_allocatable) {
1256 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
1258 uint64_t qmax = mga->mga_cur_max_alloc_queue_depth;
1260 if (!mc->mc_alloc_throttle_enabled)
1264 * If this metaslab group does not have any free space, then
1265 * there is no point in looking further.
1267 if (mg->mg_no_free_space)
1271 * Some allocations (e.g., those coming from device removal
1272 * where the * allocations are not even counted in the
1273 * metaslab * allocation queues) are allowed to bypass
1276 if (flags & METASLAB_DONT_THROTTLE)
1280 * Relax allocation throttling for ditto blocks. Due to
1281 * random imbalances in allocation it tends to push copies
1282 * to one vdev, that looks a bit better at the moment.
1284 qmax = qmax * (4 + d) / 4;
1286 qdepth = zfs_refcount_count(&mga->mga_alloc_queue_depth);
1289 * If this metaslab group is below its qmax or it's
1290 * the only allocatable metasable group, then attempt
1291 * to allocate from it.
1293 if (qdepth < qmax || mc->mc_alloc_groups == 1)
1295 ASSERT3U(mc->mc_alloc_groups, >, 1);
1298 * Since this metaslab group is at or over its qmax, we
1299 * need to determine if there are metaslab groups after this
1300 * one that might be able to handle this allocation. This is
1301 * racy since we can't hold the locks for all metaslab
1302 * groups at the same time when we make this check.
1304 for (metaslab_group_t *mgp = mg->mg_next;
1305 mgp != rotor; mgp = mgp->mg_next) {
1306 metaslab_group_allocator_t *mgap =
1307 &mgp->mg_allocator[allocator];
1308 qmax = mgap->mga_cur_max_alloc_queue_depth;
1309 qmax = qmax * (4 + d) / 4;
1311 zfs_refcount_count(&mgap->mga_alloc_queue_depth);
1314 * If there is another metaslab group that
1315 * might be able to handle the allocation, then
1316 * we return false so that we skip this group.
1318 if (qdepth < qmax && !mgp->mg_no_free_space)
1323 * We didn't find another group to handle the allocation
1324 * so we can't skip this metaslab group even though
1325 * we are at or over our qmax.
1329 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
1336 * ==========================================================================
1337 * Range tree callbacks
1338 * ==========================================================================
1342 * Comparison function for the private size-ordered tree using 32-bit
1343 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1345 __attribute__((always_inline)) inline
1347 metaslab_rangesize32_compare(const void *x1, const void *x2)
1349 const range_seg32_t *r1 = x1;
1350 const range_seg32_t *r2 = x2;
1352 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1353 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1355 int cmp = TREE_CMP(rs_size1, rs_size2);
1357 return (cmp + !cmp * TREE_CMP(r1->rs_start, r2->rs_start));
1361 * Comparison function for the private size-ordered tree using 64-bit
1362 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1364 __attribute__((always_inline)) inline
1366 metaslab_rangesize64_compare(const void *x1, const void *x2)
1368 const range_seg64_t *r1 = x1;
1369 const range_seg64_t *r2 = x2;
1371 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1372 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1374 int cmp = TREE_CMP(rs_size1, rs_size2);
1376 return (cmp + !cmp * TREE_CMP(r1->rs_start, r2->rs_start));
1379 typedef struct metaslab_rt_arg {
1380 zfs_btree_t *mra_bt;
1381 uint32_t mra_floor_shift;
1382 } metaslab_rt_arg_t;
1386 metaslab_rt_arg_t *mra;
1390 metaslab_size_sorted_add(void *arg, uint64_t start, uint64_t size)
1392 struct mssa_arg *mssap = arg;
1393 range_tree_t *rt = mssap->rt;
1394 metaslab_rt_arg_t *mrap = mssap->mra;
1395 range_seg_max_t seg = {0};
1396 rs_set_start(&seg, rt, start);
1397 rs_set_end(&seg, rt, start + size);
1398 metaslab_rt_add(rt, &seg, mrap);
1402 metaslab_size_tree_full_load(range_tree_t *rt)
1404 metaslab_rt_arg_t *mrap = rt->rt_arg;
1405 METASLABSTAT_BUMP(metaslabstat_reload_tree);
1406 ASSERT0(zfs_btree_numnodes(mrap->mra_bt));
1407 mrap->mra_floor_shift = 0;
1408 struct mssa_arg arg = {0};
1411 range_tree_walk(rt, metaslab_size_sorted_add, &arg);
1415 ZFS_BTREE_FIND_IN_BUF_FUNC(metaslab_rt_find_rangesize32_in_buf,
1416 range_seg32_t, metaslab_rangesize32_compare)
1418 ZFS_BTREE_FIND_IN_BUF_FUNC(metaslab_rt_find_rangesize64_in_buf,
1419 range_seg64_t, metaslab_rangesize64_compare)
1422 * Create any block allocator specific components. The current allocators
1423 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1426 metaslab_rt_create(range_tree_t *rt, void *arg)
1428 metaslab_rt_arg_t *mrap = arg;
1429 zfs_btree_t *size_tree = mrap->mra_bt;
1432 int (*compare) (const void *, const void *);
1433 bt_find_in_buf_f bt_find;
1434 switch (rt->rt_type) {
1436 size = sizeof (range_seg32_t);
1437 compare = metaslab_rangesize32_compare;
1438 bt_find = metaslab_rt_find_rangesize32_in_buf;
1441 size = sizeof (range_seg64_t);
1442 compare = metaslab_rangesize64_compare;
1443 bt_find = metaslab_rt_find_rangesize64_in_buf;
1446 panic("Invalid range seg type %d", rt->rt_type);
1448 zfs_btree_create(size_tree, compare, bt_find, size);
1449 mrap->mra_floor_shift = metaslab_by_size_min_shift;
1453 metaslab_rt_destroy(range_tree_t *rt, void *arg)
1456 metaslab_rt_arg_t *mrap = arg;
1457 zfs_btree_t *size_tree = mrap->mra_bt;
1459 zfs_btree_destroy(size_tree);
1460 kmem_free(mrap, sizeof (*mrap));
1464 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1466 metaslab_rt_arg_t *mrap = arg;
1467 zfs_btree_t *size_tree = mrap->mra_bt;
1469 if (rs_get_end(rs, rt) - rs_get_start(rs, rt) <
1470 (1ULL << mrap->mra_floor_shift))
1473 zfs_btree_add(size_tree, rs);
1477 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1479 metaslab_rt_arg_t *mrap = arg;
1480 zfs_btree_t *size_tree = mrap->mra_bt;
1482 if (rs_get_end(rs, rt) - rs_get_start(rs, rt) < (1ULL <<
1483 mrap->mra_floor_shift))
1486 zfs_btree_remove(size_tree, rs);
1490 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1492 metaslab_rt_arg_t *mrap = arg;
1493 zfs_btree_t *size_tree = mrap->mra_bt;
1494 zfs_btree_clear(size_tree);
1495 zfs_btree_destroy(size_tree);
1497 metaslab_rt_create(rt, arg);
1500 static const range_tree_ops_t metaslab_rt_ops = {
1501 .rtop_create = metaslab_rt_create,
1502 .rtop_destroy = metaslab_rt_destroy,
1503 .rtop_add = metaslab_rt_add,
1504 .rtop_remove = metaslab_rt_remove,
1505 .rtop_vacate = metaslab_rt_vacate
1509 * ==========================================================================
1510 * Common allocator routines
1511 * ==========================================================================
1515 * Return the maximum contiguous segment within the metaslab.
1518 metaslab_largest_allocatable(metaslab_t *msp)
1520 zfs_btree_t *t = &msp->ms_allocatable_by_size;
1525 if (zfs_btree_numnodes(t) == 0)
1526 metaslab_size_tree_full_load(msp->ms_allocatable);
1528 rs = zfs_btree_last(t, NULL);
1532 return (rs_get_end(rs, msp->ms_allocatable) - rs_get_start(rs,
1533 msp->ms_allocatable));
1537 * Return the maximum contiguous segment within the unflushed frees of this
1541 metaslab_largest_unflushed_free(metaslab_t *msp)
1543 ASSERT(MUTEX_HELD(&msp->ms_lock));
1545 if (msp->ms_unflushed_frees == NULL)
1548 if (zfs_btree_numnodes(&msp->ms_unflushed_frees_by_size) == 0)
1549 metaslab_size_tree_full_load(msp->ms_unflushed_frees);
1550 range_seg_t *rs = zfs_btree_last(&msp->ms_unflushed_frees_by_size,
1556 * When a range is freed from the metaslab, that range is added to
1557 * both the unflushed frees and the deferred frees. While the block
1558 * will eventually be usable, if the metaslab were loaded the range
1559 * would not be added to the ms_allocatable tree until TXG_DEFER_SIZE
1560 * txgs had passed. As a result, when attempting to estimate an upper
1561 * bound for the largest currently-usable free segment in the
1562 * metaslab, we need to not consider any ranges currently in the defer
1563 * trees. This algorithm approximates the largest available chunk in
1564 * the largest range in the unflushed_frees tree by taking the first
1565 * chunk. While this may be a poor estimate, it should only remain so
1566 * briefly and should eventually self-correct as frees are no longer
1567 * deferred. Similar logic applies to the ms_freed tree. See
1568 * metaslab_load() for more details.
1570 * There are two primary sources of inaccuracy in this estimate. Both
1571 * are tolerated for performance reasons. The first source is that we
1572 * only check the largest segment for overlaps. Smaller segments may
1573 * have more favorable overlaps with the other trees, resulting in
1574 * larger usable chunks. Second, we only look at the first chunk in
1575 * the largest segment; there may be other usable chunks in the
1576 * largest segment, but we ignore them.
1578 uint64_t rstart = rs_get_start(rs, msp->ms_unflushed_frees);
1579 uint64_t rsize = rs_get_end(rs, msp->ms_unflushed_frees) - rstart;
1580 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1583 boolean_t found = range_tree_find_in(msp->ms_defer[t], rstart,
1584 rsize, &start, &size);
1586 if (rstart == start)
1588 rsize = start - rstart;
1594 boolean_t found = range_tree_find_in(msp->ms_freed, rstart,
1595 rsize, &start, &size);
1597 rsize = start - rstart;
1602 static range_seg_t *
1603 metaslab_block_find(zfs_btree_t *t, range_tree_t *rt, uint64_t start,
1604 uint64_t size, zfs_btree_index_t *where)
1607 range_seg_max_t rsearch;
1609 rs_set_start(&rsearch, rt, start);
1610 rs_set_end(&rsearch, rt, start + size);
1612 rs = zfs_btree_find(t, &rsearch, where);
1614 rs = zfs_btree_next(t, where, where);
1620 #if defined(WITH_DF_BLOCK_ALLOCATOR) || \
1621 defined(WITH_CF_BLOCK_ALLOCATOR)
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);
1657 #endif /* WITH_DF/CF_BLOCK_ALLOCATOR */
1659 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1661 * ==========================================================================
1662 * Dynamic Fit (df) block allocator
1664 * Search for a free chunk of at least this size, starting from the last
1665 * offset (for this alignment of block) looking for up to
1666 * metaslab_df_max_search bytes (16MB). If a large enough free chunk is not
1667 * found within 16MB, then return a free chunk of exactly the requested size (or
1670 * If it seems like searching from the last offset will be unproductive, skip
1671 * that and just return a free chunk of exactly the requested size (or larger).
1672 * This is based on metaslab_df_alloc_threshold and metaslab_df_free_pct. This
1673 * mechanism is probably not very useful and may be removed in the future.
1675 * The behavior when not searching can be changed to return the largest free
1676 * chunk, instead of a free chunk of exactly the requested size, by setting
1677 * metaslab_df_use_largest_segment.
1678 * ==========================================================================
1681 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1684 * Find the largest power of 2 block size that evenly divides the
1685 * requested size. This is used to try to allocate blocks with similar
1686 * alignment from the same area of the metaslab (i.e. same cursor
1687 * bucket) but it does not guarantee that other allocations sizes
1688 * may exist in the same region.
1690 uint64_t align = size & -size;
1691 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1692 range_tree_t *rt = msp->ms_allocatable;
1693 uint_t free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1696 ASSERT(MUTEX_HELD(&msp->ms_lock));
1699 * If we're running low on space, find a segment based on size,
1700 * rather than iterating based on offset.
1702 if (metaslab_largest_allocatable(msp) < metaslab_df_alloc_threshold ||
1703 free_pct < metaslab_df_free_pct) {
1706 offset = metaslab_block_picker(rt,
1707 cursor, size, metaslab_df_max_search);
1712 if (zfs_btree_numnodes(&msp->ms_allocatable_by_size) == 0)
1713 metaslab_size_tree_full_load(msp->ms_allocatable);
1715 if (metaslab_df_use_largest_segment) {
1716 /* use largest free segment */
1717 rs = zfs_btree_last(&msp->ms_allocatable_by_size, NULL);
1719 zfs_btree_index_t where;
1720 /* use segment of this size, or next largest */
1721 rs = metaslab_block_find(&msp->ms_allocatable_by_size,
1722 rt, msp->ms_start, size, &where);
1724 if (rs != NULL && rs_get_start(rs, rt) + size <= rs_get_end(rs,
1726 offset = rs_get_start(rs, rt);
1727 *cursor = offset + size;
1734 const metaslab_ops_t zfs_metaslab_ops = {
1737 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1739 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1741 * ==========================================================================
1742 * Cursor fit block allocator -
1743 * Select the largest region in the metaslab, set the cursor to the beginning
1744 * of the range and the cursor_end to the end of the range. As allocations
1745 * are made advance the cursor. Continue allocating from the cursor until
1746 * the range is exhausted and then find a new range.
1747 * ==========================================================================
1750 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1752 range_tree_t *rt = msp->ms_allocatable;
1753 zfs_btree_t *t = &msp->ms_allocatable_by_size;
1754 uint64_t *cursor = &msp->ms_lbas[0];
1755 uint64_t *cursor_end = &msp->ms_lbas[1];
1756 uint64_t offset = 0;
1758 ASSERT(MUTEX_HELD(&msp->ms_lock));
1760 ASSERT3U(*cursor_end, >=, *cursor);
1762 if ((*cursor + size) > *cursor_end) {
1765 if (zfs_btree_numnodes(t) == 0)
1766 metaslab_size_tree_full_load(msp->ms_allocatable);
1767 rs = zfs_btree_last(t, NULL);
1768 if (rs == NULL || (rs_get_end(rs, rt) - rs_get_start(rs, rt)) <
1772 *cursor = rs_get_start(rs, rt);
1773 *cursor_end = rs_get_end(rs, rt);
1782 const metaslab_ops_t zfs_metaslab_ops = {
1785 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1787 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1789 * ==========================================================================
1790 * New dynamic fit allocator -
1791 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1792 * contiguous blocks. If no region is found then just use the largest segment
1794 * ==========================================================================
1798 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1799 * to request from the allocator.
1801 uint64_t metaslab_ndf_clump_shift = 4;
1804 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1806 zfs_btree_t *t = &msp->ms_allocatable->rt_root;
1807 range_tree_t *rt = msp->ms_allocatable;
1808 zfs_btree_index_t where;
1810 range_seg_max_t rsearch;
1811 uint64_t hbit = highbit64(size);
1812 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1813 uint64_t max_size = metaslab_largest_allocatable(msp);
1815 ASSERT(MUTEX_HELD(&msp->ms_lock));
1817 if (max_size < size)
1820 rs_set_start(&rsearch, rt, *cursor);
1821 rs_set_end(&rsearch, rt, *cursor + size);
1823 rs = zfs_btree_find(t, &rsearch, &where);
1824 if (rs == NULL || (rs_get_end(rs, rt) - rs_get_start(rs, rt)) < size) {
1825 t = &msp->ms_allocatable_by_size;
1827 rs_set_start(&rsearch, rt, 0);
1828 rs_set_end(&rsearch, rt, MIN(max_size, 1ULL << (hbit +
1829 metaslab_ndf_clump_shift)));
1831 rs = zfs_btree_find(t, &rsearch, &where);
1833 rs = zfs_btree_next(t, &where, &where);
1837 if ((rs_get_end(rs, rt) - rs_get_start(rs, rt)) >= size) {
1838 *cursor = rs_get_start(rs, rt) + size;
1839 return (rs_get_start(rs, rt));
1844 const metaslab_ops_t zfs_metaslab_ops = {
1847 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1851 * ==========================================================================
1853 * ==========================================================================
1857 * Wait for any in-progress metaslab loads to complete.
1860 metaslab_load_wait(metaslab_t *msp)
1862 ASSERT(MUTEX_HELD(&msp->ms_lock));
1864 while (msp->ms_loading) {
1865 ASSERT(!msp->ms_loaded);
1866 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1871 * Wait for any in-progress flushing to complete.
1874 metaslab_flush_wait(metaslab_t *msp)
1876 ASSERT(MUTEX_HELD(&msp->ms_lock));
1878 while (msp->ms_flushing)
1879 cv_wait(&msp->ms_flush_cv, &msp->ms_lock);
1883 metaslab_idx_func(multilist_t *ml, void *arg)
1885 metaslab_t *msp = arg;
1888 * ms_id values are allocated sequentially, so full 64bit
1889 * division would be a waste of time, so limit it to 32 bits.
1891 return ((unsigned int)msp->ms_id % multilist_get_num_sublists(ml));
1895 metaslab_allocated_space(metaslab_t *msp)
1897 return (msp->ms_allocated_space);
1901 * Verify that the space accounting on disk matches the in-core range_trees.
1904 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
1906 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1907 uint64_t allocating = 0;
1908 uint64_t sm_free_space, msp_free_space;
1910 ASSERT(MUTEX_HELD(&msp->ms_lock));
1911 ASSERT(!msp->ms_condensing);
1913 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
1917 * We can only verify the metaslab space when we're called
1918 * from syncing context with a loaded metaslab that has an
1919 * allocated space map. Calling this in non-syncing context
1920 * does not provide a consistent view of the metaslab since
1921 * we're performing allocations in the future.
1923 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
1928 * Even though the smp_alloc field can get negative,
1929 * when it comes to a metaslab's space map, that should
1930 * never be the case.
1932 ASSERT3S(space_map_allocated(msp->ms_sm), >=, 0);
1934 ASSERT3U(space_map_allocated(msp->ms_sm), >=,
1935 range_tree_space(msp->ms_unflushed_frees));
1937 ASSERT3U(metaslab_allocated_space(msp), ==,
1938 space_map_allocated(msp->ms_sm) +
1939 range_tree_space(msp->ms_unflushed_allocs) -
1940 range_tree_space(msp->ms_unflushed_frees));
1942 sm_free_space = msp->ms_size - metaslab_allocated_space(msp);
1945 * Account for future allocations since we would have
1946 * already deducted that space from the ms_allocatable.
1948 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
1950 range_tree_space(msp->ms_allocating[(txg + t) & TXG_MASK]);
1952 ASSERT3U(allocating + msp->ms_allocated_this_txg, ==,
1953 msp->ms_allocating_total);
1955 ASSERT3U(msp->ms_deferspace, ==,
1956 range_tree_space(msp->ms_defer[0]) +
1957 range_tree_space(msp->ms_defer[1]));
1959 msp_free_space = range_tree_space(msp->ms_allocatable) + allocating +
1960 msp->ms_deferspace + range_tree_space(msp->ms_freed);
1962 VERIFY3U(sm_free_space, ==, msp_free_space);
1966 metaslab_aux_histograms_clear(metaslab_t *msp)
1969 * Auxiliary histograms are only cleared when resetting them,
1970 * which can only happen while the metaslab is loaded.
1972 ASSERT(msp->ms_loaded);
1974 memset(msp->ms_synchist, 0, sizeof (msp->ms_synchist));
1975 for (int t = 0; t < TXG_DEFER_SIZE; t++)
1976 memset(msp->ms_deferhist[t], 0, sizeof (msp->ms_deferhist[t]));
1980 metaslab_aux_histogram_add(uint64_t *histogram, uint64_t shift,
1984 * This is modeled after space_map_histogram_add(), so refer to that
1985 * function for implementation details. We want this to work like
1986 * the space map histogram, and not the range tree histogram, as we
1987 * are essentially constructing a delta that will be later subtracted
1988 * from the space map histogram.
1991 for (int i = shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
1992 ASSERT3U(i, >=, idx + shift);
1993 histogram[idx] += rt->rt_histogram[i] << (i - idx - shift);
1995 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
1996 ASSERT3U(idx + shift, ==, i);
1998 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
2004 * Called at every sync pass that the metaslab gets synced.
2006 * The reason is that we want our auxiliary histograms to be updated
2007 * wherever the metaslab's space map histogram is updated. This way
2008 * we stay consistent on which parts of the metaslab space map's
2009 * histogram are currently not available for allocations (e.g because
2010 * they are in the defer, freed, and freeing trees).
2013 metaslab_aux_histograms_update(metaslab_t *msp)
2015 space_map_t *sm = msp->ms_sm;
2019 * This is similar to the metaslab's space map histogram updates
2020 * that take place in metaslab_sync(). The only difference is that
2021 * we only care about segments that haven't made it into the
2022 * ms_allocatable tree yet.
2024 if (msp->ms_loaded) {
2025 metaslab_aux_histograms_clear(msp);
2027 metaslab_aux_histogram_add(msp->ms_synchist,
2028 sm->sm_shift, msp->ms_freed);
2030 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2031 metaslab_aux_histogram_add(msp->ms_deferhist[t],
2032 sm->sm_shift, msp->ms_defer[t]);
2036 metaslab_aux_histogram_add(msp->ms_synchist,
2037 sm->sm_shift, msp->ms_freeing);
2041 * Called every time we are done syncing (writing to) the metaslab,
2042 * i.e. at the end of each sync pass.
2043 * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist]
2046 metaslab_aux_histograms_update_done(metaslab_t *msp, boolean_t defer_allowed)
2048 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2049 space_map_t *sm = msp->ms_sm;
2053 * We came here from metaslab_init() when creating/opening a
2054 * pool, looking at a metaslab that hasn't had any allocations
2061 * This is similar to the actions that we take for the ms_freed
2062 * and ms_defer trees in metaslab_sync_done().
2064 uint64_t hist_index = spa_syncing_txg(spa) % TXG_DEFER_SIZE;
2065 if (defer_allowed) {
2066 memcpy(msp->ms_deferhist[hist_index], msp->ms_synchist,
2067 sizeof (msp->ms_synchist));
2069 memset(msp->ms_deferhist[hist_index], 0,
2070 sizeof (msp->ms_deferhist[hist_index]));
2072 memset(msp->ms_synchist, 0, sizeof (msp->ms_synchist));
2076 * Ensure that the metaslab's weight and fragmentation are consistent
2077 * with the contents of the histogram (either the range tree's histogram
2078 * or the space map's depending whether the metaslab is loaded).
2081 metaslab_verify_weight_and_frag(metaslab_t *msp)
2083 ASSERT(MUTEX_HELD(&msp->ms_lock));
2085 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
2089 * We can end up here from vdev_remove_complete(), in which case we
2090 * cannot do these assertions because we hold spa config locks and
2091 * thus we are not allowed to read from the DMU.
2093 * We check if the metaslab group has been removed and if that's
2094 * the case we return immediately as that would mean that we are
2095 * here from the aforementioned code path.
2097 if (msp->ms_group == NULL)
2101 * Devices being removed always return a weight of 0 and leave
2102 * fragmentation and ms_max_size as is - there is nothing for
2103 * us to verify here.
2105 vdev_t *vd = msp->ms_group->mg_vd;
2106 if (vd->vdev_removing)
2110 * If the metaslab is dirty it probably means that we've done
2111 * some allocations or frees that have changed our histograms
2112 * and thus the weight.
2114 for (int t = 0; t < TXG_SIZE; t++) {
2115 if (txg_list_member(&vd->vdev_ms_list, msp, t))
2120 * This verification checks that our in-memory state is consistent
2121 * with what's on disk. If the pool is read-only then there aren't
2122 * any changes and we just have the initially-loaded state.
2124 if (!spa_writeable(msp->ms_group->mg_vd->vdev_spa))
2127 /* some extra verification for in-core tree if you can */
2128 if (msp->ms_loaded) {
2129 range_tree_stat_verify(msp->ms_allocatable);
2130 VERIFY(space_map_histogram_verify(msp->ms_sm,
2131 msp->ms_allocatable));
2134 uint64_t weight = msp->ms_weight;
2135 uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2136 boolean_t space_based = WEIGHT_IS_SPACEBASED(msp->ms_weight);
2137 uint64_t frag = msp->ms_fragmentation;
2138 uint64_t max_segsize = msp->ms_max_size;
2141 msp->ms_fragmentation = 0;
2144 * This function is used for verification purposes and thus should
2145 * not introduce any side-effects/mutations on the system's state.
2147 * Regardless of whether metaslab_weight() thinks this metaslab
2148 * should be active or not, we want to ensure that the actual weight
2149 * (and therefore the value of ms_weight) would be the same if it
2150 * was to be recalculated at this point.
2152 * In addition we set the nodirty flag so metaslab_weight() does
2153 * not dirty the metaslab for future TXGs (e.g. when trying to
2154 * force condensing to upgrade the metaslab spacemaps).
2156 msp->ms_weight = metaslab_weight(msp, B_TRUE) | was_active;
2158 VERIFY3U(max_segsize, ==, msp->ms_max_size);
2161 * If the weight type changed then there is no point in doing
2162 * verification. Revert fields to their original values.
2164 if ((space_based && !WEIGHT_IS_SPACEBASED(msp->ms_weight)) ||
2165 (!space_based && WEIGHT_IS_SPACEBASED(msp->ms_weight))) {
2166 msp->ms_fragmentation = frag;
2167 msp->ms_weight = weight;
2171 VERIFY3U(msp->ms_fragmentation, ==, frag);
2172 VERIFY3U(msp->ms_weight, ==, weight);
2176 * If we're over the zfs_metaslab_mem_limit, select the loaded metaslab from
2177 * this class that was used longest ago, and attempt to unload it. We don't
2178 * want to spend too much time in this loop to prevent performance
2179 * degradation, and we expect that most of the time this operation will
2180 * succeed. Between that and the normal unloading processing during txg sync,
2181 * we expect this to keep the metaslab memory usage under control.
2184 metaslab_potentially_evict(metaslab_class_t *mc)
2187 uint64_t allmem = arc_all_memory();
2188 uint64_t inuse = spl_kmem_cache_inuse(zfs_btree_leaf_cache);
2189 uint64_t size = spl_kmem_cache_entry_size(zfs_btree_leaf_cache);
2191 for (; allmem * zfs_metaslab_mem_limit / 100 < inuse * size &&
2192 tries < multilist_get_num_sublists(&mc->mc_metaslab_txg_list) * 2;
2194 unsigned int idx = multilist_get_random_index(
2195 &mc->mc_metaslab_txg_list);
2196 multilist_sublist_t *mls =
2197 multilist_sublist_lock(&mc->mc_metaslab_txg_list, idx);
2198 metaslab_t *msp = multilist_sublist_head(mls);
2199 multilist_sublist_unlock(mls);
2200 while (msp != NULL && allmem * zfs_metaslab_mem_limit / 100 <
2202 VERIFY3P(mls, ==, multilist_sublist_lock(
2203 &mc->mc_metaslab_txg_list, idx));
2205 metaslab_idx_func(&mc->mc_metaslab_txg_list, msp));
2207 if (!multilist_link_active(&msp->ms_class_txg_node)) {
2208 multilist_sublist_unlock(mls);
2211 metaslab_t *next_msp = multilist_sublist_next(mls, msp);
2212 multilist_sublist_unlock(mls);
2214 * If the metaslab is currently loading there are two
2215 * cases. If it's the metaslab we're evicting, we
2216 * can't continue on or we'll panic when we attempt to
2217 * recursively lock the mutex. If it's another
2218 * metaslab that's loading, it can be safely skipped,
2219 * since we know it's very new and therefore not a
2220 * good eviction candidate. We check later once the
2221 * lock is held that the metaslab is fully loaded
2222 * before actually unloading it.
2224 if (msp->ms_loading) {
2227 spl_kmem_cache_inuse(zfs_btree_leaf_cache);
2231 * We can't unload metaslabs with no spacemap because
2232 * they're not ready to be unloaded yet. We can't
2233 * unload metaslabs with outstanding allocations
2234 * because doing so could cause the metaslab's weight
2235 * to decrease while it's unloaded, which violates an
2236 * invariant that we use to prevent unnecessary
2237 * loading. We also don't unload metaslabs that are
2238 * currently active because they are high-weight
2239 * metaslabs that are likely to be used in the near
2242 mutex_enter(&msp->ms_lock);
2243 if (msp->ms_allocator == -1 && msp->ms_sm != NULL &&
2244 msp->ms_allocating_total == 0) {
2245 metaslab_unload(msp);
2247 mutex_exit(&msp->ms_lock);
2249 inuse = spl_kmem_cache_inuse(zfs_btree_leaf_cache);
2253 (void) mc, (void) zfs_metaslab_mem_limit;
2258 metaslab_load_impl(metaslab_t *msp)
2262 ASSERT(MUTEX_HELD(&msp->ms_lock));
2263 ASSERT(msp->ms_loading);
2264 ASSERT(!msp->ms_condensing);
2267 * We temporarily drop the lock to unblock other operations while we
2268 * are reading the space map. Therefore, metaslab_sync() and
2269 * metaslab_sync_done() can run at the same time as we do.
2271 * If we are using the log space maps, metaslab_sync() can't write to
2272 * the metaslab's space map while we are loading as we only write to
2273 * it when we are flushing the metaslab, and that can't happen while
2274 * we are loading it.
2276 * If we are not using log space maps though, metaslab_sync() can
2277 * append to the space map while we are loading. Therefore we load
2278 * only entries that existed when we started the load. Additionally,
2279 * metaslab_sync_done() has to wait for the load to complete because
2280 * there are potential races like metaslab_load() loading parts of the
2281 * space map that are currently being appended by metaslab_sync(). If
2282 * we didn't, the ms_allocatable would have entries that
2283 * metaslab_sync_done() would try to re-add later.
2285 * That's why before dropping the lock we remember the synced length
2286 * of the metaslab and read up to that point of the space map,
2287 * ignoring entries appended by metaslab_sync() that happen after we
2290 uint64_t length = msp->ms_synced_length;
2291 mutex_exit(&msp->ms_lock);
2293 hrtime_t load_start = gethrtime();
2294 metaslab_rt_arg_t *mrap;
2295 if (msp->ms_allocatable->rt_arg == NULL) {
2296 mrap = kmem_zalloc(sizeof (*mrap), KM_SLEEP);
2298 mrap = msp->ms_allocatable->rt_arg;
2299 msp->ms_allocatable->rt_ops = NULL;
2300 msp->ms_allocatable->rt_arg = NULL;
2302 mrap->mra_bt = &msp->ms_allocatable_by_size;
2303 mrap->mra_floor_shift = metaslab_by_size_min_shift;
2305 if (msp->ms_sm != NULL) {
2306 error = space_map_load_length(msp->ms_sm, msp->ms_allocatable,
2309 /* Now, populate the size-sorted tree. */
2310 metaslab_rt_create(msp->ms_allocatable, mrap);
2311 msp->ms_allocatable->rt_ops = &metaslab_rt_ops;
2312 msp->ms_allocatable->rt_arg = mrap;
2314 struct mssa_arg arg = {0};
2315 arg.rt = msp->ms_allocatable;
2317 range_tree_walk(msp->ms_allocatable, metaslab_size_sorted_add,
2321 * Add the size-sorted tree first, since we don't need to load
2322 * the metaslab from the spacemap.
2324 metaslab_rt_create(msp->ms_allocatable, mrap);
2325 msp->ms_allocatable->rt_ops = &metaslab_rt_ops;
2326 msp->ms_allocatable->rt_arg = mrap;
2328 * The space map has not been allocated yet, so treat
2329 * all the space in the metaslab as free and add it to the
2330 * ms_allocatable tree.
2332 range_tree_add(msp->ms_allocatable,
2333 msp->ms_start, msp->ms_size);
2337 * If the ms_sm doesn't exist, this means that this
2338 * metaslab hasn't gone through metaslab_sync() and
2339 * thus has never been dirtied. So we shouldn't
2340 * expect any unflushed allocs or frees from previous
2343 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
2344 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
2349 * We need to grab the ms_sync_lock to prevent metaslab_sync() from
2350 * changing the ms_sm (or log_sm) and the metaslab's range trees
2351 * while we are about to use them and populate the ms_allocatable.
2352 * The ms_lock is insufficient for this because metaslab_sync() doesn't
2353 * hold the ms_lock while writing the ms_checkpointing tree to disk.
2355 mutex_enter(&msp->ms_sync_lock);
2356 mutex_enter(&msp->ms_lock);
2358 ASSERT(!msp->ms_condensing);
2359 ASSERT(!msp->ms_flushing);
2362 mutex_exit(&msp->ms_sync_lock);
2366 ASSERT3P(msp->ms_group, !=, NULL);
2367 msp->ms_loaded = B_TRUE;
2370 * Apply all the unflushed changes to ms_allocatable right
2371 * away so any manipulations we do below have a clear view
2372 * of what is allocated and what is free.
2374 range_tree_walk(msp->ms_unflushed_allocs,
2375 range_tree_remove, msp->ms_allocatable);
2376 range_tree_walk(msp->ms_unflushed_frees,
2377 range_tree_add, msp->ms_allocatable);
2379 ASSERT3P(msp->ms_group, !=, NULL);
2380 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2381 if (spa_syncing_log_sm(spa) != NULL) {
2382 ASSERT(spa_feature_is_enabled(spa,
2383 SPA_FEATURE_LOG_SPACEMAP));
2386 * If we use a log space map we add all the segments
2387 * that are in ms_unflushed_frees so they are available
2390 * ms_allocatable needs to contain all free segments
2391 * that are ready for allocations (thus not segments
2392 * from ms_freeing, ms_freed, and the ms_defer trees).
2393 * But if we grab the lock in this code path at a sync
2394 * pass later that 1, then it also contains the
2395 * segments of ms_freed (they were added to it earlier
2396 * in this path through ms_unflushed_frees). So we
2397 * need to remove all the segments that exist in
2398 * ms_freed from ms_allocatable as they will be added
2399 * later in metaslab_sync_done().
2401 * When there's no log space map, the ms_allocatable
2402 * correctly doesn't contain any segments that exist
2403 * in ms_freed [see ms_synced_length].
2405 range_tree_walk(msp->ms_freed,
2406 range_tree_remove, msp->ms_allocatable);
2410 * If we are not using the log space map, ms_allocatable
2411 * contains the segments that exist in the ms_defer trees
2412 * [see ms_synced_length]. Thus we need to remove them
2413 * from ms_allocatable as they will be added again in
2414 * metaslab_sync_done().
2416 * If we are using the log space map, ms_allocatable still
2417 * contains the segments that exist in the ms_defer trees.
2418 * Not because it read them through the ms_sm though. But
2419 * because these segments are part of ms_unflushed_frees
2420 * whose segments we add to ms_allocatable earlier in this
2423 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2424 range_tree_walk(msp->ms_defer[t],
2425 range_tree_remove, msp->ms_allocatable);
2429 * Call metaslab_recalculate_weight_and_sort() now that the
2430 * metaslab is loaded so we get the metaslab's real weight.
2432 * Unless this metaslab was created with older software and
2433 * has not yet been converted to use segment-based weight, we
2434 * expect the new weight to be better or equal to the weight
2435 * that the metaslab had while it was not loaded. This is
2436 * because the old weight does not take into account the
2437 * consolidation of adjacent segments between TXGs. [see
2438 * comment for ms_synchist and ms_deferhist[] for more info]
2440 uint64_t weight = msp->ms_weight;
2441 uint64_t max_size = msp->ms_max_size;
2442 metaslab_recalculate_weight_and_sort(msp);
2443 if (!WEIGHT_IS_SPACEBASED(weight))
2444 ASSERT3U(weight, <=, msp->ms_weight);
2445 msp->ms_max_size = metaslab_largest_allocatable(msp);
2446 ASSERT3U(max_size, <=, msp->ms_max_size);
2447 hrtime_t load_end = gethrtime();
2448 msp->ms_load_time = load_end;
2449 zfs_dbgmsg("metaslab_load: txg %llu, spa %s, vdev_id %llu, "
2450 "ms_id %llu, smp_length %llu, "
2451 "unflushed_allocs %llu, unflushed_frees %llu, "
2452 "freed %llu, defer %llu + %llu, unloaded time %llu ms, "
2453 "loading_time %lld ms, ms_max_size %llu, "
2454 "max size error %lld, "
2455 "old_weight %llx, new_weight %llx",
2456 (u_longlong_t)spa_syncing_txg(spa), spa_name(spa),
2457 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
2458 (u_longlong_t)msp->ms_id,
2459 (u_longlong_t)space_map_length(msp->ms_sm),
2460 (u_longlong_t)range_tree_space(msp->ms_unflushed_allocs),
2461 (u_longlong_t)range_tree_space(msp->ms_unflushed_frees),
2462 (u_longlong_t)range_tree_space(msp->ms_freed),
2463 (u_longlong_t)range_tree_space(msp->ms_defer[0]),
2464 (u_longlong_t)range_tree_space(msp->ms_defer[1]),
2465 (longlong_t)((load_start - msp->ms_unload_time) / 1000000),
2466 (longlong_t)((load_end - load_start) / 1000000),
2467 (u_longlong_t)msp->ms_max_size,
2468 (u_longlong_t)msp->ms_max_size - max_size,
2469 (u_longlong_t)weight, (u_longlong_t)msp->ms_weight);
2471 metaslab_verify_space(msp, spa_syncing_txg(spa));
2472 mutex_exit(&msp->ms_sync_lock);
2477 metaslab_load(metaslab_t *msp)
2479 ASSERT(MUTEX_HELD(&msp->ms_lock));
2482 * There may be another thread loading the same metaslab, if that's
2483 * the case just wait until the other thread is done and return.
2485 metaslab_load_wait(msp);
2488 VERIFY(!msp->ms_loading);
2489 ASSERT(!msp->ms_condensing);
2492 * We set the loading flag BEFORE potentially dropping the lock to
2493 * wait for an ongoing flush (see ms_flushing below). This way other
2494 * threads know that there is already a thread that is loading this
2497 msp->ms_loading = B_TRUE;
2500 * Wait for any in-progress flushing to finish as we drop the ms_lock
2501 * both here (during space_map_load()) and in metaslab_flush() (when
2502 * we flush our changes to the ms_sm).
2504 if (msp->ms_flushing)
2505 metaslab_flush_wait(msp);
2508 * In the possibility that we were waiting for the metaslab to be
2509 * flushed (where we temporarily dropped the ms_lock), ensure that
2510 * no one else loaded the metaslab somehow.
2512 ASSERT(!msp->ms_loaded);
2515 * If we're loading a metaslab in the normal class, consider evicting
2516 * another one to keep our memory usage under the limit defined by the
2517 * zfs_metaslab_mem_limit tunable.
2519 if (spa_normal_class(msp->ms_group->mg_class->mc_spa) ==
2520 msp->ms_group->mg_class) {
2521 metaslab_potentially_evict(msp->ms_group->mg_class);
2524 int error = metaslab_load_impl(msp);
2526 ASSERT(MUTEX_HELD(&msp->ms_lock));
2527 msp->ms_loading = B_FALSE;
2528 cv_broadcast(&msp->ms_load_cv);
2534 metaslab_unload(metaslab_t *msp)
2536 ASSERT(MUTEX_HELD(&msp->ms_lock));
2539 * This can happen if a metaslab is selected for eviction (in
2540 * metaslab_potentially_evict) and then unloaded during spa_sync (via
2541 * metaslab_class_evict_old).
2543 if (!msp->ms_loaded)
2546 range_tree_vacate(msp->ms_allocatable, NULL, NULL);
2547 msp->ms_loaded = B_FALSE;
2548 msp->ms_unload_time = gethrtime();
2550 msp->ms_activation_weight = 0;
2551 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
2553 if (msp->ms_group != NULL) {
2554 metaslab_class_t *mc = msp->ms_group->mg_class;
2555 multilist_sublist_t *mls =
2556 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
2557 if (multilist_link_active(&msp->ms_class_txg_node))
2558 multilist_sublist_remove(mls, msp);
2559 multilist_sublist_unlock(mls);
2561 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2562 zfs_dbgmsg("metaslab_unload: txg %llu, spa %s, vdev_id %llu, "
2563 "ms_id %llu, weight %llx, "
2564 "selected txg %llu (%llu ms ago), alloc_txg %llu, "
2565 "loaded %llu ms ago, max_size %llu",
2566 (u_longlong_t)spa_syncing_txg(spa), spa_name(spa),
2567 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
2568 (u_longlong_t)msp->ms_id,
2569 (u_longlong_t)msp->ms_weight,
2570 (u_longlong_t)msp->ms_selected_txg,
2571 (u_longlong_t)(msp->ms_unload_time -
2572 msp->ms_selected_time) / 1000 / 1000,
2573 (u_longlong_t)msp->ms_alloc_txg,
2574 (u_longlong_t)(msp->ms_unload_time -
2575 msp->ms_load_time) / 1000 / 1000,
2576 (u_longlong_t)msp->ms_max_size);
2580 * We explicitly recalculate the metaslab's weight based on its space
2581 * map (as it is now not loaded). We want unload metaslabs to always
2582 * have their weights calculated from the space map histograms, while
2583 * loaded ones have it calculated from their in-core range tree
2584 * [see metaslab_load()]. This way, the weight reflects the information
2585 * available in-core, whether it is loaded or not.
2587 * If ms_group == NULL means that we came here from metaslab_fini(),
2588 * at which point it doesn't make sense for us to do the recalculation
2591 if (msp->ms_group != NULL)
2592 metaslab_recalculate_weight_and_sort(msp);
2596 * We want to optimize the memory use of the per-metaslab range
2597 * trees. To do this, we store the segments in the range trees in
2598 * units of sectors, zero-indexing from the start of the metaslab. If
2599 * the vdev_ms_shift - the vdev_ashift is less than 32, we can store
2600 * the ranges using two uint32_ts, rather than two uint64_ts.
2603 metaslab_calculate_range_tree_type(vdev_t *vdev, metaslab_t *msp,
2604 uint64_t *start, uint64_t *shift)
2606 if (vdev->vdev_ms_shift - vdev->vdev_ashift < 32 &&
2607 !zfs_metaslab_force_large_segs) {
2608 *shift = vdev->vdev_ashift;
2609 *start = msp->ms_start;
2610 return (RANGE_SEG32);
2614 return (RANGE_SEG64);
2619 metaslab_set_selected_txg(metaslab_t *msp, uint64_t txg)
2621 ASSERT(MUTEX_HELD(&msp->ms_lock));
2622 metaslab_class_t *mc = msp->ms_group->mg_class;
2623 multilist_sublist_t *mls =
2624 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
2625 if (multilist_link_active(&msp->ms_class_txg_node))
2626 multilist_sublist_remove(mls, msp);
2627 msp->ms_selected_txg = txg;
2628 msp->ms_selected_time = gethrtime();
2629 multilist_sublist_insert_tail(mls, msp);
2630 multilist_sublist_unlock(mls);
2634 metaslab_space_update(vdev_t *vd, metaslab_class_t *mc, int64_t alloc_delta,
2635 int64_t defer_delta, int64_t space_delta)
2637 vdev_space_update(vd, alloc_delta, defer_delta, space_delta);
2639 ASSERT3P(vd->vdev_spa->spa_root_vdev, ==, vd->vdev_parent);
2640 ASSERT(vd->vdev_ms_count != 0);
2642 metaslab_class_space_update(mc, alloc_delta, defer_delta, space_delta,
2643 vdev_deflated_space(vd, space_delta));
2647 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object,
2648 uint64_t txg, metaslab_t **msp)
2650 vdev_t *vd = mg->mg_vd;
2651 spa_t *spa = vd->vdev_spa;
2652 objset_t *mos = spa->spa_meta_objset;
2656 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
2657 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
2658 mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL);
2659 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
2660 cv_init(&ms->ms_flush_cv, NULL, CV_DEFAULT, NULL);
2661 multilist_link_init(&ms->ms_class_txg_node);
2664 ms->ms_start = id << vd->vdev_ms_shift;
2665 ms->ms_size = 1ULL << vd->vdev_ms_shift;
2666 ms->ms_allocator = -1;
2667 ms->ms_new = B_TRUE;
2669 vdev_ops_t *ops = vd->vdev_ops;
2670 if (ops->vdev_op_metaslab_init != NULL)
2671 ops->vdev_op_metaslab_init(vd, &ms->ms_start, &ms->ms_size);
2674 * We only open space map objects that already exist. All others
2675 * will be opened when we finally allocate an object for it. For
2676 * readonly pools there is no need to open the space map object.
2679 * When called from vdev_expand(), we can't call into the DMU as
2680 * we are holding the spa_config_lock as a writer and we would
2681 * deadlock [see relevant comment in vdev_metaslab_init()]. in
2682 * that case, the object parameter is zero though, so we won't
2683 * call into the DMU.
2685 if (object != 0 && !(spa->spa_mode == SPA_MODE_READ &&
2686 !spa->spa_read_spacemaps)) {
2687 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
2688 ms->ms_size, vd->vdev_ashift);
2691 kmem_free(ms, sizeof (metaslab_t));
2695 ASSERT(ms->ms_sm != NULL);
2696 ms->ms_allocated_space = space_map_allocated(ms->ms_sm);
2699 uint64_t shift, start;
2700 range_seg_type_t type =
2701 metaslab_calculate_range_tree_type(vd, ms, &start, &shift);
2703 ms->ms_allocatable = range_tree_create(NULL, type, NULL, start, shift);
2704 for (int t = 0; t < TXG_SIZE; t++) {
2705 ms->ms_allocating[t] = range_tree_create(NULL, type,
2706 NULL, start, shift);
2708 ms->ms_freeing = range_tree_create(NULL, type, NULL, start, shift);
2709 ms->ms_freed = range_tree_create(NULL, type, NULL, start, shift);
2710 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2711 ms->ms_defer[t] = range_tree_create(NULL, type, NULL,
2714 ms->ms_checkpointing =
2715 range_tree_create(NULL, type, NULL, start, shift);
2716 ms->ms_unflushed_allocs =
2717 range_tree_create(NULL, type, NULL, start, shift);
2719 metaslab_rt_arg_t *mrap = kmem_zalloc(sizeof (*mrap), KM_SLEEP);
2720 mrap->mra_bt = &ms->ms_unflushed_frees_by_size;
2721 mrap->mra_floor_shift = metaslab_by_size_min_shift;
2722 ms->ms_unflushed_frees = range_tree_create(&metaslab_rt_ops,
2723 type, mrap, start, shift);
2725 ms->ms_trim = range_tree_create(NULL, type, NULL, start, shift);
2727 metaslab_group_add(mg, ms);
2728 metaslab_set_fragmentation(ms, B_FALSE);
2731 * If we're opening an existing pool (txg == 0) or creating
2732 * a new one (txg == TXG_INITIAL), all space is available now.
2733 * If we're adding space to an existing pool, the new space
2734 * does not become available until after this txg has synced.
2735 * The metaslab's weight will also be initialized when we sync
2736 * out this txg. This ensures that we don't attempt to allocate
2737 * from it before we have initialized it completely.
2739 if (txg <= TXG_INITIAL) {
2740 metaslab_sync_done(ms, 0);
2741 metaslab_space_update(vd, mg->mg_class,
2742 metaslab_allocated_space(ms), 0, 0);
2746 vdev_dirty(vd, 0, NULL, txg);
2747 vdev_dirty(vd, VDD_METASLAB, ms, txg);
2756 metaslab_fini_flush_data(metaslab_t *msp)
2758 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2760 if (metaslab_unflushed_txg(msp) == 0) {
2761 ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL),
2765 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
2767 mutex_enter(&spa->spa_flushed_ms_lock);
2768 avl_remove(&spa->spa_metaslabs_by_flushed, msp);
2769 mutex_exit(&spa->spa_flushed_ms_lock);
2771 spa_log_sm_decrement_mscount(spa, metaslab_unflushed_txg(msp));
2772 spa_log_summary_decrement_mscount(spa, metaslab_unflushed_txg(msp),
2773 metaslab_unflushed_dirty(msp));
2777 metaslab_unflushed_changes_memused(metaslab_t *ms)
2779 return ((range_tree_numsegs(ms->ms_unflushed_allocs) +
2780 range_tree_numsegs(ms->ms_unflushed_frees)) *
2781 ms->ms_unflushed_allocs->rt_root.bt_elem_size);
2785 metaslab_fini(metaslab_t *msp)
2787 metaslab_group_t *mg = msp->ms_group;
2788 vdev_t *vd = mg->mg_vd;
2789 spa_t *spa = vd->vdev_spa;
2791 metaslab_fini_flush_data(msp);
2793 metaslab_group_remove(mg, msp);
2795 mutex_enter(&msp->ms_lock);
2796 VERIFY(msp->ms_group == NULL);
2799 * If this metaslab hasn't been through metaslab_sync_done() yet its
2800 * space hasn't been accounted for in its vdev and doesn't need to be
2804 metaslab_space_update(vd, mg->mg_class,
2805 -metaslab_allocated_space(msp), 0, -msp->ms_size);
2808 space_map_close(msp->ms_sm);
2811 metaslab_unload(msp);
2813 range_tree_destroy(msp->ms_allocatable);
2814 range_tree_destroy(msp->ms_freeing);
2815 range_tree_destroy(msp->ms_freed);
2817 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
2818 metaslab_unflushed_changes_memused(msp));
2819 spa->spa_unflushed_stats.sus_memused -=
2820 metaslab_unflushed_changes_memused(msp);
2821 range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
2822 range_tree_destroy(msp->ms_unflushed_allocs);
2823 range_tree_destroy(msp->ms_checkpointing);
2824 range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
2825 range_tree_destroy(msp->ms_unflushed_frees);
2827 for (int t = 0; t < TXG_SIZE; t++) {
2828 range_tree_destroy(msp->ms_allocating[t]);
2830 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2831 range_tree_destroy(msp->ms_defer[t]);
2833 ASSERT0(msp->ms_deferspace);
2835 for (int t = 0; t < TXG_SIZE; t++)
2836 ASSERT(!txg_list_member(&vd->vdev_ms_list, msp, t));
2838 range_tree_vacate(msp->ms_trim, NULL, NULL);
2839 range_tree_destroy(msp->ms_trim);
2841 mutex_exit(&msp->ms_lock);
2842 cv_destroy(&msp->ms_load_cv);
2843 cv_destroy(&msp->ms_flush_cv);
2844 mutex_destroy(&msp->ms_lock);
2845 mutex_destroy(&msp->ms_sync_lock);
2846 ASSERT3U(msp->ms_allocator, ==, -1);
2848 kmem_free(msp, sizeof (metaslab_t));
2851 #define FRAGMENTATION_TABLE_SIZE 17
2854 * This table defines a segment size based fragmentation metric that will
2855 * allow each metaslab to derive its own fragmentation value. This is done
2856 * by calculating the space in each bucket of the spacemap histogram and
2857 * multiplying that by the fragmentation metric in this table. Doing
2858 * this for all buckets and dividing it by the total amount of free
2859 * space in this metaslab (i.e. the total free space in all buckets) gives
2860 * us the fragmentation metric. This means that a high fragmentation metric
2861 * equates to most of the free space being comprised of small segments.
2862 * Conversely, if the metric is low, then most of the free space is in
2863 * large segments. A 10% change in fragmentation equates to approximately
2864 * double the number of segments.
2866 * This table defines 0% fragmented space using 16MB segments. Testing has
2867 * shown that segments that are greater than or equal to 16MB do not suffer
2868 * from drastic performance problems. Using this value, we derive the rest
2869 * of the table. Since the fragmentation value is never stored on disk, it
2870 * is possible to change these calculations in the future.
2872 static const int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
2892 * Calculate the metaslab's fragmentation metric and set ms_fragmentation.
2893 * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not
2894 * been upgraded and does not support this metric. Otherwise, the return
2895 * value should be in the range [0, 100].
2898 metaslab_set_fragmentation(metaslab_t *msp, boolean_t nodirty)
2900 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2901 uint64_t fragmentation = 0;
2903 boolean_t feature_enabled = spa_feature_is_enabled(spa,
2904 SPA_FEATURE_SPACEMAP_HISTOGRAM);
2906 if (!feature_enabled) {
2907 msp->ms_fragmentation = ZFS_FRAG_INVALID;
2912 * A null space map means that the entire metaslab is free
2913 * and thus is not fragmented.
2915 if (msp->ms_sm == NULL) {
2916 msp->ms_fragmentation = 0;
2921 * If this metaslab's space map has not been upgraded, flag it
2922 * so that we upgrade next time we encounter it.
2924 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
2925 uint64_t txg = spa_syncing_txg(spa);
2926 vdev_t *vd = msp->ms_group->mg_vd;
2929 * If we've reached the final dirty txg, then we must
2930 * be shutting down the pool. We don't want to dirty
2931 * any data past this point so skip setting the condense
2932 * flag. We can retry this action the next time the pool
2933 * is imported. We also skip marking this metaslab for
2934 * condensing if the caller has explicitly set nodirty.
2937 spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) {
2938 msp->ms_condense_wanted = B_TRUE;
2939 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2940 zfs_dbgmsg("txg %llu, requesting force condense: "
2941 "ms_id %llu, vdev_id %llu", (u_longlong_t)txg,
2942 (u_longlong_t)msp->ms_id,
2943 (u_longlong_t)vd->vdev_id);
2945 msp->ms_fragmentation = ZFS_FRAG_INVALID;
2949 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
2951 uint8_t shift = msp->ms_sm->sm_shift;
2953 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
2954 FRAGMENTATION_TABLE_SIZE - 1);
2956 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
2959 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
2962 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
2963 fragmentation += space * zfs_frag_table[idx];
2967 fragmentation /= total;
2968 ASSERT3U(fragmentation, <=, 100);
2970 msp->ms_fragmentation = fragmentation;
2974 * Compute a weight -- a selection preference value -- for the given metaslab.
2975 * This is based on the amount of free space, the level of fragmentation,
2976 * the LBA range, and whether the metaslab is loaded.
2979 metaslab_space_weight(metaslab_t *msp)
2981 metaslab_group_t *mg = msp->ms_group;
2982 vdev_t *vd = mg->mg_vd;
2983 uint64_t weight, space;
2985 ASSERT(MUTEX_HELD(&msp->ms_lock));
2988 * The baseline weight is the metaslab's free space.
2990 space = msp->ms_size - metaslab_allocated_space(msp);
2992 if (metaslab_fragmentation_factor_enabled &&
2993 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
2995 * Use the fragmentation information to inversely scale
2996 * down the baseline weight. We need to ensure that we
2997 * don't exclude this metaslab completely when it's 100%
2998 * fragmented. To avoid this we reduce the fragmented value
3001 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
3004 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
3005 * this metaslab again. The fragmentation metric may have
3006 * decreased the space to something smaller than
3007 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
3008 * so that we can consume any remaining space.
3010 if (space > 0 && space < SPA_MINBLOCKSIZE)
3011 space = SPA_MINBLOCKSIZE;
3016 * Modern disks have uniform bit density and constant angular velocity.
3017 * Therefore, the outer recording zones are faster (higher bandwidth)
3018 * than the inner zones by the ratio of outer to inner track diameter,
3019 * which is typically around 2:1. We account for this by assigning
3020 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
3021 * In effect, this means that we'll select the metaslab with the most
3022 * free bandwidth rather than simply the one with the most free space.
3024 if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
3025 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
3026 ASSERT(weight >= space && weight <= 2 * space);
3030 * If this metaslab is one we're actively using, adjust its
3031 * weight to make it preferable to any inactive metaslab so
3032 * we'll polish it off. If the fragmentation on this metaslab
3033 * has exceed our threshold, then don't mark it active.
3035 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
3036 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
3037 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
3040 WEIGHT_SET_SPACEBASED(weight);
3045 * Return the weight of the specified metaslab, according to the segment-based
3046 * weighting algorithm. The metaslab must be loaded. This function can
3047 * be called within a sync pass since it relies only on the metaslab's
3048 * range tree which is always accurate when the metaslab is loaded.
3051 metaslab_weight_from_range_tree(metaslab_t *msp)
3053 uint64_t weight = 0;
3054 uint32_t segments = 0;
3056 ASSERT(msp->ms_loaded);
3058 for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
3060 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
3061 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
3064 segments += msp->ms_allocatable->rt_histogram[i];
3067 * The range tree provides more precision than the space map
3068 * and must be downgraded so that all values fit within the
3069 * space map's histogram. This allows us to compare loaded
3070 * vs. unloaded metaslabs to determine which metaslab is
3071 * considered "best".
3076 if (segments != 0) {
3077 WEIGHT_SET_COUNT(weight, segments);
3078 WEIGHT_SET_INDEX(weight, i);
3079 WEIGHT_SET_ACTIVE(weight, 0);
3087 * Calculate the weight based on the on-disk histogram. Should be applied
3088 * only to unloaded metaslabs (i.e no incoming allocations) in-order to
3089 * give results consistent with the on-disk state
3092 metaslab_weight_from_spacemap(metaslab_t *msp)
3094 space_map_t *sm = msp->ms_sm;
3095 ASSERT(!msp->ms_loaded);
3097 ASSERT3U(space_map_object(sm), !=, 0);
3098 ASSERT3U(sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
3101 * Create a joint histogram from all the segments that have made
3102 * it to the metaslab's space map histogram, that are not yet
3103 * available for allocation because they are still in the freeing
3104 * pipeline (e.g. freeing, freed, and defer trees). Then subtract
3105 * these segments from the space map's histogram to get a more
3108 uint64_t deferspace_histogram[SPACE_MAP_HISTOGRAM_SIZE] = {0};
3109 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
3110 deferspace_histogram[i] += msp->ms_synchist[i];
3111 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3112 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
3113 deferspace_histogram[i] += msp->ms_deferhist[t][i];
3117 uint64_t weight = 0;
3118 for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
3119 ASSERT3U(sm->sm_phys->smp_histogram[i], >=,
3120 deferspace_histogram[i]);
3122 sm->sm_phys->smp_histogram[i] - deferspace_histogram[i];
3124 WEIGHT_SET_COUNT(weight, count);
3125 WEIGHT_SET_INDEX(weight, i + sm->sm_shift);
3126 WEIGHT_SET_ACTIVE(weight, 0);
3134 * Compute a segment-based weight for the specified metaslab. The weight
3135 * is determined by highest bucket in the histogram. The information
3136 * for the highest bucket is encoded into the weight value.
3139 metaslab_segment_weight(metaslab_t *msp)
3141 metaslab_group_t *mg = msp->ms_group;
3142 uint64_t weight = 0;
3143 uint8_t shift = mg->mg_vd->vdev_ashift;
3145 ASSERT(MUTEX_HELD(&msp->ms_lock));
3148 * The metaslab is completely free.
3150 if (metaslab_allocated_space(msp) == 0) {
3151 int idx = highbit64(msp->ms_size) - 1;
3152 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
3154 if (idx < max_idx) {
3155 WEIGHT_SET_COUNT(weight, 1ULL);
3156 WEIGHT_SET_INDEX(weight, idx);
3158 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
3159 WEIGHT_SET_INDEX(weight, max_idx);
3161 WEIGHT_SET_ACTIVE(weight, 0);
3162 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
3166 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
3169 * If the metaslab is fully allocated then just make the weight 0.
3171 if (metaslab_allocated_space(msp) == msp->ms_size)
3174 * If the metaslab is already loaded, then use the range tree to
3175 * determine the weight. Otherwise, we rely on the space map information
3176 * to generate the weight.
3178 if (msp->ms_loaded) {
3179 weight = metaslab_weight_from_range_tree(msp);
3181 weight = metaslab_weight_from_spacemap(msp);
3185 * If the metaslab was active the last time we calculated its weight
3186 * then keep it active. We want to consume the entire region that
3187 * is associated with this weight.
3189 if (msp->ms_activation_weight != 0 && weight != 0)
3190 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
3195 * Determine if we should attempt to allocate from this metaslab. If the
3196 * metaslab is loaded, then we can determine if the desired allocation
3197 * can be satisfied by looking at the size of the maximum free segment
3198 * on that metaslab. Otherwise, we make our decision based on the metaslab's
3199 * weight. For segment-based weighting we can determine the maximum
3200 * allocation based on the index encoded in its value. For space-based
3201 * weights we rely on the entire weight (excluding the weight-type bit).
3204 metaslab_should_allocate(metaslab_t *msp, uint64_t asize, boolean_t try_hard)
3207 * If the metaslab is loaded, ms_max_size is definitive and we can use
3208 * the fast check. If it's not, the ms_max_size is a lower bound (once
3209 * set), and we should use the fast check as long as we're not in
3210 * try_hard and it's been less than zfs_metaslab_max_size_cache_sec
3211 * seconds since the metaslab was unloaded.
3213 if (msp->ms_loaded ||
3214 (msp->ms_max_size != 0 && !try_hard && gethrtime() <
3215 msp->ms_unload_time + SEC2NSEC(zfs_metaslab_max_size_cache_sec)))
3216 return (msp->ms_max_size >= asize);
3218 boolean_t should_allocate;
3219 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
3221 * The metaslab segment weight indicates segments in the
3222 * range [2^i, 2^(i+1)), where i is the index in the weight.
3223 * Since the asize might be in the middle of the range, we
3224 * should attempt the allocation if asize < 2^(i+1).
3226 should_allocate = (asize <
3227 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
3229 should_allocate = (asize <=
3230 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
3233 return (should_allocate);
3237 metaslab_weight(metaslab_t *msp, boolean_t nodirty)
3239 vdev_t *vd = msp->ms_group->mg_vd;
3240 spa_t *spa = vd->vdev_spa;
3243 ASSERT(MUTEX_HELD(&msp->ms_lock));
3245 metaslab_set_fragmentation(msp, nodirty);
3248 * Update the maximum size. If the metaslab is loaded, this will
3249 * ensure that we get an accurate maximum size if newly freed space
3250 * has been added back into the free tree. If the metaslab is
3251 * unloaded, we check if there's a larger free segment in the
3252 * unflushed frees. This is a lower bound on the largest allocatable
3253 * segment size. Coalescing of adjacent entries may reveal larger
3254 * allocatable segments, but we aren't aware of those until loading
3255 * the space map into a range tree.
3257 if (msp->ms_loaded) {
3258 msp->ms_max_size = metaslab_largest_allocatable(msp);
3260 msp->ms_max_size = MAX(msp->ms_max_size,
3261 metaslab_largest_unflushed_free(msp));
3265 * Segment-based weighting requires space map histogram support.
3267 if (zfs_metaslab_segment_weight_enabled &&
3268 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
3269 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
3270 sizeof (space_map_phys_t))) {
3271 weight = metaslab_segment_weight(msp);
3273 weight = metaslab_space_weight(msp);
3279 metaslab_recalculate_weight_and_sort(metaslab_t *msp)
3281 ASSERT(MUTEX_HELD(&msp->ms_lock));
3283 /* note: we preserve the mask (e.g. indication of primary, etc..) */
3284 uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
3285 metaslab_group_sort(msp->ms_group, msp,
3286 metaslab_weight(msp, B_FALSE) | was_active);
3290 metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
3291 int allocator, uint64_t activation_weight)
3293 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
3294 ASSERT(MUTEX_HELD(&msp->ms_lock));
3297 * If we're activating for the claim code, we don't want to actually
3298 * set the metaslab up for a specific allocator.
3300 if (activation_weight == METASLAB_WEIGHT_CLAIM) {
3301 ASSERT0(msp->ms_activation_weight);
3302 msp->ms_activation_weight = msp->ms_weight;
3303 metaslab_group_sort(mg, msp, msp->ms_weight |
3308 metaslab_t **mspp = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
3309 &mga->mga_primary : &mga->mga_secondary);
3311 mutex_enter(&mg->mg_lock);
3312 if (*mspp != NULL) {
3313 mutex_exit(&mg->mg_lock);
3318 ASSERT3S(msp->ms_allocator, ==, -1);
3319 msp->ms_allocator = allocator;
3320 msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY);
3322 ASSERT0(msp->ms_activation_weight);
3323 msp->ms_activation_weight = msp->ms_weight;
3324 metaslab_group_sort_impl(mg, msp,
3325 msp->ms_weight | activation_weight);
3326 mutex_exit(&mg->mg_lock);
3332 metaslab_activate(metaslab_t *msp, int allocator, uint64_t activation_weight)
3334 ASSERT(MUTEX_HELD(&msp->ms_lock));
3337 * The current metaslab is already activated for us so there
3338 * is nothing to do. Already activated though, doesn't mean
3339 * that this metaslab is activated for our allocator nor our
3340 * requested activation weight. The metaslab could have started
3341 * as an active one for our allocator but changed allocators
3342 * while we were waiting to grab its ms_lock or we stole it
3343 * [see find_valid_metaslab()]. This means that there is a
3344 * possibility of passivating a metaslab of another allocator
3345 * or from a different activation mask, from this thread.
3347 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
3348 ASSERT(msp->ms_loaded);
3352 int error = metaslab_load(msp);
3354 metaslab_group_sort(msp->ms_group, msp, 0);
3359 * When entering metaslab_load() we may have dropped the
3360 * ms_lock because we were loading this metaslab, or we
3361 * were waiting for another thread to load it for us. In
3362 * that scenario, we recheck the weight of the metaslab
3363 * to see if it was activated by another thread.
3365 * If the metaslab was activated for another allocator or
3366 * it was activated with a different activation weight (e.g.
3367 * we wanted to make it a primary but it was activated as
3368 * secondary) we return error (EBUSY).
3370 * If the metaslab was activated for the same allocator
3371 * and requested activation mask, skip activating it.
3373 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
3374 if (msp->ms_allocator != allocator)
3377 if ((msp->ms_weight & activation_weight) == 0)
3378 return (SET_ERROR(EBUSY));
3380 EQUIV((activation_weight == METASLAB_WEIGHT_PRIMARY),
3386 * If the metaslab has literally 0 space, it will have weight 0. In
3387 * that case, don't bother activating it. This can happen if the
3388 * metaslab had space during find_valid_metaslab, but another thread
3389 * loaded it and used all that space while we were waiting to grab the
3392 if (msp->ms_weight == 0) {
3393 ASSERT0(range_tree_space(msp->ms_allocatable));
3394 return (SET_ERROR(ENOSPC));
3397 if ((error = metaslab_activate_allocator(msp->ms_group, msp,
3398 allocator, activation_weight)) != 0) {
3402 ASSERT(msp->ms_loaded);
3403 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
3409 metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp,
3412 ASSERT(MUTEX_HELD(&msp->ms_lock));
3413 ASSERT(msp->ms_loaded);
3415 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
3416 metaslab_group_sort(mg, msp, weight);
3420 mutex_enter(&mg->mg_lock);
3421 ASSERT3P(msp->ms_group, ==, mg);
3422 ASSERT3S(0, <=, msp->ms_allocator);
3423 ASSERT3U(msp->ms_allocator, <, mg->mg_allocators);
3425 metaslab_group_allocator_t *mga = &mg->mg_allocator[msp->ms_allocator];
3426 if (msp->ms_primary) {
3427 ASSERT3P(mga->mga_primary, ==, msp);
3428 ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
3429 mga->mga_primary = NULL;
3431 ASSERT3P(mga->mga_secondary, ==, msp);
3432 ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
3433 mga->mga_secondary = NULL;
3435 msp->ms_allocator = -1;
3436 metaslab_group_sort_impl(mg, msp, weight);
3437 mutex_exit(&mg->mg_lock);
3441 metaslab_passivate(metaslab_t *msp, uint64_t weight)
3443 uint64_t size __maybe_unused = weight & ~METASLAB_WEIGHT_TYPE;
3446 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
3447 * this metaslab again. In that case, it had better be empty,
3448 * or we would be leaving space on the table.
3450 ASSERT(!WEIGHT_IS_SPACEBASED(msp->ms_weight) ||
3451 size >= SPA_MINBLOCKSIZE ||
3452 range_tree_space(msp->ms_allocatable) == 0);
3453 ASSERT0(weight & METASLAB_ACTIVE_MASK);
3455 ASSERT(msp->ms_activation_weight != 0);
3456 msp->ms_activation_weight = 0;
3457 metaslab_passivate_allocator(msp->ms_group, msp, weight);
3458 ASSERT0(msp->ms_weight & METASLAB_ACTIVE_MASK);
3462 * Segment-based metaslabs are activated once and remain active until
3463 * we either fail an allocation attempt (similar to space-based metaslabs)
3464 * or have exhausted the free space in zfs_metaslab_switch_threshold
3465 * buckets since the metaslab was activated. This function checks to see
3466 * if we've exhausted the zfs_metaslab_switch_threshold buckets in the
3467 * metaslab and passivates it proactively. This will allow us to select a
3468 * metaslab with a larger contiguous region, if any, remaining within this
3469 * metaslab group. If we're in sync pass > 1, then we continue using this
3470 * metaslab so that we don't dirty more block and cause more sync passes.
3473 metaslab_segment_may_passivate(metaslab_t *msp)
3475 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3477 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
3481 * Since we are in the middle of a sync pass, the most accurate
3482 * information that is accessible to us is the in-core range tree
3483 * histogram; calculate the new weight based on that information.
3485 uint64_t weight = metaslab_weight_from_range_tree(msp);
3486 int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
3487 int current_idx = WEIGHT_GET_INDEX(weight);
3489 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
3490 metaslab_passivate(msp, weight);
3494 metaslab_preload(void *arg)
3496 metaslab_t *msp = arg;
3497 metaslab_class_t *mc = msp->ms_group->mg_class;
3498 spa_t *spa = mc->mc_spa;
3499 fstrans_cookie_t cookie = spl_fstrans_mark();
3501 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
3503 mutex_enter(&msp->ms_lock);
3504 (void) metaslab_load(msp);
3505 metaslab_set_selected_txg(msp, spa_syncing_txg(spa));
3506 mutex_exit(&msp->ms_lock);
3507 spl_fstrans_unmark(cookie);
3511 metaslab_group_preload(metaslab_group_t *mg)
3513 spa_t *spa = mg->mg_vd->vdev_spa;
3515 avl_tree_t *t = &mg->mg_metaslab_tree;
3518 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
3519 taskq_wait_outstanding(mg->mg_taskq, 0);
3523 mutex_enter(&mg->mg_lock);
3526 * Load the next potential metaslabs
3528 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
3529 ASSERT3P(msp->ms_group, ==, mg);
3532 * We preload only the maximum number of metaslabs specified
3533 * by metaslab_preload_limit. If a metaslab is being forced
3534 * to condense then we preload it too. This will ensure
3535 * that force condensing happens in the next txg.
3537 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
3541 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
3542 msp, TQ_SLEEP) != TASKQID_INVALID);
3544 mutex_exit(&mg->mg_lock);
3548 * Determine if the space map's on-disk footprint is past our tolerance for
3549 * inefficiency. We would like to use the following criteria to make our
3552 * 1. Do not condense if the size of the space map object would dramatically
3553 * increase as a result of writing out the free space range tree.
3555 * 2. Condense if the on on-disk space map representation is at least
3556 * zfs_condense_pct/100 times the size of the optimal representation
3557 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, optimal = 1.1MB).
3559 * 3. Do not condense if the on-disk size of the space map does not actually
3562 * Unfortunately, we cannot compute the on-disk size of the space map in this
3563 * context because we cannot accurately compute the effects of compression, etc.
3564 * Instead, we apply the heuristic described in the block comment for
3565 * zfs_metaslab_condense_block_threshold - we only condense if the space used
3566 * is greater than a threshold number of blocks.
3569 metaslab_should_condense(metaslab_t *msp)
3571 space_map_t *sm = msp->ms_sm;
3572 vdev_t *vd = msp->ms_group->mg_vd;
3573 uint64_t vdev_blocksize = 1ULL << vd->vdev_ashift;
3575 ASSERT(MUTEX_HELD(&msp->ms_lock));
3576 ASSERT(msp->ms_loaded);
3578 ASSERT3U(spa_sync_pass(vd->vdev_spa), ==, 1);
3581 * We always condense metaslabs that are empty and metaslabs for
3582 * which a condense request has been made.
3584 if (range_tree_numsegs(msp->ms_allocatable) == 0 ||
3585 msp->ms_condense_wanted)
3588 uint64_t record_size = MAX(sm->sm_blksz, vdev_blocksize);
3589 uint64_t object_size = space_map_length(sm);
3590 uint64_t optimal_size = space_map_estimate_optimal_size(sm,
3591 msp->ms_allocatable, SM_NO_VDEVID);
3593 return (object_size >= (optimal_size * zfs_condense_pct / 100) &&
3594 object_size > zfs_metaslab_condense_block_threshold * record_size);
3598 * Condense the on-disk space map representation to its minimized form.
3599 * The minimized form consists of a small number of allocations followed
3600 * by the entries of the free range tree (ms_allocatable). The condensed
3601 * spacemap contains all the entries of previous TXGs (including those in
3602 * the pool-wide log spacemaps; thus this is effectively a superset of
3603 * metaslab_flush()), but this TXG's entries still need to be written.
3606 metaslab_condense(metaslab_t *msp, dmu_tx_t *tx)
3608 range_tree_t *condense_tree;
3609 space_map_t *sm = msp->ms_sm;
3610 uint64_t txg = dmu_tx_get_txg(tx);
3611 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3613 ASSERT(MUTEX_HELD(&msp->ms_lock));
3614 ASSERT(msp->ms_loaded);
3615 ASSERT(msp->ms_sm != NULL);
3618 * In order to condense the space map, we need to change it so it
3619 * only describes which segments are currently allocated and free.
3621 * All the current free space resides in the ms_allocatable, all
3622 * the ms_defer trees, and all the ms_allocating trees. We ignore
3623 * ms_freed because it is empty because we're in sync pass 1. We
3624 * ignore ms_freeing because these changes are not yet reflected
3625 * in the spacemap (they will be written later this txg).
3627 * So to truncate the space map to represent all the entries of
3628 * previous TXGs we do the following:
3630 * 1] We create a range tree (condense tree) that is 100% empty.
3631 * 2] We add to it all segments found in the ms_defer trees
3632 * as those segments are marked as free in the original space
3633 * map. We do the same with the ms_allocating trees for the same
3634 * reason. Adding these segments should be a relatively
3635 * inexpensive operation since we expect these trees to have a
3636 * small number of nodes.
3637 * 3] We vacate any unflushed allocs, since they are not frees we
3638 * need to add to the condense tree. Then we vacate any
3639 * unflushed frees as they should already be part of ms_allocatable.
3640 * 4] At this point, we would ideally like to add all segments
3641 * in the ms_allocatable tree from the condense tree. This way
3642 * we would write all the entries of the condense tree as the
3643 * condensed space map, which would only contain freed
3644 * segments with everything else assumed to be allocated.
3646 * Doing so can be prohibitively expensive as ms_allocatable can
3647 * be large, and therefore computationally expensive to add to
3648 * the condense_tree. Instead we first sync out an entry marking
3649 * everything as allocated, then the condense_tree and then the
3650 * ms_allocatable, in the condensed space map. While this is not
3651 * optimal, it is typically close to optimal and more importantly
3652 * much cheaper to compute.
3654 * 5] Finally, as both of the unflushed trees were written to our
3655 * new and condensed metaslab space map, we basically flushed
3656 * all the unflushed changes to disk, thus we call
3657 * metaslab_flush_update().
3659 ASSERT3U(spa_sync_pass(spa), ==, 1);
3660 ASSERT(range_tree_is_empty(msp->ms_freed)); /* since it is pass 1 */
3662 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %px, vdev id %llu, "
3663 "spa %s, smp size %llu, segments %llu, forcing condense=%s",
3664 (u_longlong_t)txg, (u_longlong_t)msp->ms_id, msp,
3665 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
3666 spa->spa_name, (u_longlong_t)space_map_length(msp->ms_sm),
3667 (u_longlong_t)range_tree_numsegs(msp->ms_allocatable),
3668 msp->ms_condense_wanted ? "TRUE" : "FALSE");
3670 msp->ms_condense_wanted = B_FALSE;
3672 range_seg_type_t type;
3673 uint64_t shift, start;
3674 type = metaslab_calculate_range_tree_type(msp->ms_group->mg_vd, msp,
3677 condense_tree = range_tree_create(NULL, type, NULL, start, shift);
3679 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3680 range_tree_walk(msp->ms_defer[t],
3681 range_tree_add, condense_tree);
3684 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
3685 range_tree_walk(msp->ms_allocating[(txg + t) & TXG_MASK],
3686 range_tree_add, condense_tree);
3689 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3690 metaslab_unflushed_changes_memused(msp));
3691 spa->spa_unflushed_stats.sus_memused -=
3692 metaslab_unflushed_changes_memused(msp);
3693 range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
3694 range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
3697 * We're about to drop the metaslab's lock thus allowing other
3698 * consumers to change it's content. Set the metaslab's ms_condensing
3699 * flag to ensure that allocations on this metaslab do not occur
3700 * while we're in the middle of committing it to disk. This is only
3701 * critical for ms_allocatable as all other range trees use per TXG
3702 * views of their content.
3704 msp->ms_condensing = B_TRUE;
3706 mutex_exit(&msp->ms_lock);
3707 uint64_t object = space_map_object(msp->ms_sm);
3708 space_map_truncate(sm,
3709 spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ?
3710 zfs_metaslab_sm_blksz_with_log : zfs_metaslab_sm_blksz_no_log, tx);
3713 * space_map_truncate() may have reallocated the spacemap object.
3714 * If so, update the vdev_ms_array.
3716 if (space_map_object(msp->ms_sm) != object) {
3717 object = space_map_object(msp->ms_sm);
3718 dmu_write(spa->spa_meta_objset,
3719 msp->ms_group->mg_vd->vdev_ms_array, sizeof (uint64_t) *
3720 msp->ms_id, sizeof (uint64_t), &object, tx);
3725 * When the log space map feature is enabled, each space map will
3726 * always have ALLOCS followed by FREES for each sync pass. This is
3727 * typically true even when the log space map feature is disabled,
3728 * except from the case where a metaslab goes through metaslab_sync()
3729 * and gets condensed. In that case the metaslab's space map will have
3730 * ALLOCS followed by FREES (due to condensing) followed by ALLOCS
3731 * followed by FREES (due to space_map_write() in metaslab_sync()) for
3734 range_tree_t *tmp_tree = range_tree_create(NULL, type, NULL, start,
3736 range_tree_add(tmp_tree, msp->ms_start, msp->ms_size);
3737 space_map_write(sm, tmp_tree, SM_ALLOC, SM_NO_VDEVID, tx);
3738 space_map_write(sm, msp->ms_allocatable, SM_FREE, SM_NO_VDEVID, tx);
3739 space_map_write(sm, condense_tree, SM_FREE, SM_NO_VDEVID, tx);
3741 range_tree_vacate(condense_tree, NULL, NULL);
3742 range_tree_destroy(condense_tree);
3743 range_tree_vacate(tmp_tree, NULL, NULL);
3744 range_tree_destroy(tmp_tree);
3745 mutex_enter(&msp->ms_lock);
3747 msp->ms_condensing = B_FALSE;
3748 metaslab_flush_update(msp, tx);
3752 metaslab_unflushed_add(metaslab_t *msp, dmu_tx_t *tx)
3754 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3755 ASSERT(spa_syncing_log_sm(spa) != NULL);
3756 ASSERT(msp->ms_sm != NULL);
3757 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3758 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3760 mutex_enter(&spa->spa_flushed_ms_lock);
3761 metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx);
3762 metaslab_set_unflushed_dirty(msp, B_TRUE);
3763 avl_add(&spa->spa_metaslabs_by_flushed, msp);
3764 mutex_exit(&spa->spa_flushed_ms_lock);
3766 spa_log_sm_increment_current_mscount(spa);
3767 spa_log_summary_add_flushed_metaslab(spa, B_TRUE);
3771 metaslab_unflushed_bump(metaslab_t *msp, dmu_tx_t *tx, boolean_t dirty)
3773 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3774 ASSERT(spa_syncing_log_sm(spa) != NULL);
3775 ASSERT(msp->ms_sm != NULL);
3776 ASSERT(metaslab_unflushed_txg(msp) != 0);
3777 ASSERT3P(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL), ==, msp);
3778 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
3779 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
3781 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(spa));
3783 /* update metaslab's position in our flushing tree */
3784 uint64_t ms_prev_flushed_txg = metaslab_unflushed_txg(msp);
3785 boolean_t ms_prev_flushed_dirty = metaslab_unflushed_dirty(msp);
3786 mutex_enter(&spa->spa_flushed_ms_lock);
3787 avl_remove(&spa->spa_metaslabs_by_flushed, msp);
3788 metaslab_set_unflushed_txg(msp, spa_syncing_txg(spa), tx);
3789 metaslab_set_unflushed_dirty(msp, dirty);
3790 avl_add(&spa->spa_metaslabs_by_flushed, msp);
3791 mutex_exit(&spa->spa_flushed_ms_lock);
3793 /* update metaslab counts of spa_log_sm_t nodes */
3794 spa_log_sm_decrement_mscount(spa, ms_prev_flushed_txg);
3795 spa_log_sm_increment_current_mscount(spa);
3797 /* update log space map summary */
3798 spa_log_summary_decrement_mscount(spa, ms_prev_flushed_txg,
3799 ms_prev_flushed_dirty);
3800 spa_log_summary_add_flushed_metaslab(spa, dirty);
3802 /* cleanup obsolete logs if any */
3803 spa_cleanup_old_sm_logs(spa, tx);
3807 * Called when the metaslab has been flushed (its own spacemap now reflects
3808 * all the contents of the pool-wide spacemap log). Updates the metaslab's
3809 * metadata and any pool-wide related log space map data (e.g. summary,
3810 * obsolete logs, etc..) to reflect that.
3813 metaslab_flush_update(metaslab_t *msp, dmu_tx_t *tx)
3815 metaslab_group_t *mg = msp->ms_group;
3816 spa_t *spa = mg->mg_vd->vdev_spa;
3818 ASSERT(MUTEX_HELD(&msp->ms_lock));
3820 ASSERT3U(spa_sync_pass(spa), ==, 1);
3823 * Just because a metaslab got flushed, that doesn't mean that
3824 * it will pass through metaslab_sync_done(). Thus, make sure to
3825 * update ms_synced_length here in case it doesn't.
3827 msp->ms_synced_length = space_map_length(msp->ms_sm);
3830 * We may end up here from metaslab_condense() without the
3831 * feature being active. In that case this is a no-op.
3833 if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP) ||
3834 metaslab_unflushed_txg(msp) == 0)
3837 metaslab_unflushed_bump(msp, tx, B_FALSE);
3841 metaslab_flush(metaslab_t *msp, dmu_tx_t *tx)
3843 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
3845 ASSERT(MUTEX_HELD(&msp->ms_lock));
3846 ASSERT3U(spa_sync_pass(spa), ==, 1);
3847 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
3849 ASSERT(msp->ms_sm != NULL);
3850 ASSERT(metaslab_unflushed_txg(msp) != 0);
3851 ASSERT(avl_find(&spa->spa_metaslabs_by_flushed, msp, NULL) != NULL);
3854 * There is nothing wrong with flushing the same metaslab twice, as
3855 * this codepath should work on that case. However, the current
3856 * flushing scheme makes sure to avoid this situation as we would be
3857 * making all these calls without having anything meaningful to write
3858 * to disk. We assert this behavior here.
3860 ASSERT3U(metaslab_unflushed_txg(msp), <, dmu_tx_get_txg(tx));
3863 * We can not flush while loading, because then we would
3864 * not load the ms_unflushed_{allocs,frees}.
3866 if (msp->ms_loading)
3869 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3870 metaslab_verify_weight_and_frag(msp);
3873 * Metaslab condensing is effectively flushing. Therefore if the
3874 * metaslab can be condensed we can just condense it instead of
3877 * Note that metaslab_condense() does call metaslab_flush_update()
3878 * so we can just return immediately after condensing. We also
3879 * don't need to care about setting ms_flushing or broadcasting
3880 * ms_flush_cv, even if we temporarily drop the ms_lock in
3881 * metaslab_condense(), as the metaslab is already loaded.
3883 if (msp->ms_loaded && metaslab_should_condense(msp)) {
3884 metaslab_group_t *mg = msp->ms_group;
3887 * For all histogram operations below refer to the
3888 * comments of metaslab_sync() where we follow a
3889 * similar procedure.
3891 metaslab_group_histogram_verify(mg);
3892 metaslab_class_histogram_verify(mg->mg_class);
3893 metaslab_group_histogram_remove(mg, msp);
3895 metaslab_condense(msp, tx);
3897 space_map_histogram_clear(msp->ms_sm);
3898 space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
3899 ASSERT(range_tree_is_empty(msp->ms_freed));
3900 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
3901 space_map_histogram_add(msp->ms_sm,
3902 msp->ms_defer[t], tx);
3904 metaslab_aux_histograms_update(msp);
3906 metaslab_group_histogram_add(mg, msp);
3907 metaslab_group_histogram_verify(mg);
3908 metaslab_class_histogram_verify(mg->mg_class);
3910 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3913 * Since we recreated the histogram (and potentially
3914 * the ms_sm too while condensing) ensure that the
3915 * weight is updated too because we are not guaranteed
3916 * that this metaslab is dirty and will go through
3917 * metaslab_sync_done().
3919 metaslab_recalculate_weight_and_sort(msp);
3923 msp->ms_flushing = B_TRUE;
3924 uint64_t sm_len_before = space_map_length(msp->ms_sm);
3926 mutex_exit(&msp->ms_lock);
3927 space_map_write(msp->ms_sm, msp->ms_unflushed_allocs, SM_ALLOC,
3929 space_map_write(msp->ms_sm, msp->ms_unflushed_frees, SM_FREE,
3931 mutex_enter(&msp->ms_lock);
3933 uint64_t sm_len_after = space_map_length(msp->ms_sm);
3934 if (zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) {
3935 zfs_dbgmsg("flushing: txg %llu, spa %s, vdev_id %llu, "
3936 "ms_id %llu, unflushed_allocs %llu, unflushed_frees %llu, "
3937 "appended %llu bytes", (u_longlong_t)dmu_tx_get_txg(tx),
3939 (u_longlong_t)msp->ms_group->mg_vd->vdev_id,
3940 (u_longlong_t)msp->ms_id,
3941 (u_longlong_t)range_tree_space(msp->ms_unflushed_allocs),
3942 (u_longlong_t)range_tree_space(msp->ms_unflushed_frees),
3943 (u_longlong_t)(sm_len_after - sm_len_before));
3946 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
3947 metaslab_unflushed_changes_memused(msp));
3948 spa->spa_unflushed_stats.sus_memused -=
3949 metaslab_unflushed_changes_memused(msp);
3950 range_tree_vacate(msp->ms_unflushed_allocs, NULL, NULL);
3951 range_tree_vacate(msp->ms_unflushed_frees, NULL, NULL);
3953 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3954 metaslab_verify_weight_and_frag(msp);
3956 metaslab_flush_update(msp, tx);
3958 metaslab_verify_space(msp, dmu_tx_get_txg(tx));
3959 metaslab_verify_weight_and_frag(msp);
3961 msp->ms_flushing = B_FALSE;
3962 cv_broadcast(&msp->ms_flush_cv);
3967 * Write a metaslab to disk in the context of the specified transaction group.
3970 metaslab_sync(metaslab_t *msp, uint64_t txg)
3972 metaslab_group_t *mg = msp->ms_group;
3973 vdev_t *vd = mg->mg_vd;
3974 spa_t *spa = vd->vdev_spa;
3975 objset_t *mos = spa_meta_objset(spa);
3976 range_tree_t *alloctree = msp->ms_allocating[txg & TXG_MASK];
3979 ASSERT(!vd->vdev_ishole);
3982 * This metaslab has just been added so there's no work to do now.
3985 ASSERT0(range_tree_space(alloctree));
3986 ASSERT0(range_tree_space(msp->ms_freeing));
3987 ASSERT0(range_tree_space(msp->ms_freed));
3988 ASSERT0(range_tree_space(msp->ms_checkpointing));
3989 ASSERT0(range_tree_space(msp->ms_trim));
3994 * Normally, we don't want to process a metaslab if there are no
3995 * allocations or frees to perform. However, if the metaslab is being
3996 * forced to condense, it's loaded and we're not beyond the final
3997 * dirty txg, we need to let it through. Not condensing beyond the
3998 * final dirty txg prevents an issue where metaslabs that need to be
3999 * condensed but were loaded for other reasons could cause a panic
4000 * here. By only checking the txg in that branch of the conditional,
4001 * we preserve the utility of the VERIFY statements in all other
4004 if (range_tree_is_empty(alloctree) &&
4005 range_tree_is_empty(msp->ms_freeing) &&
4006 range_tree_is_empty(msp->ms_checkpointing) &&
4007 !(msp->ms_loaded && msp->ms_condense_wanted &&
4008 txg <= spa_final_dirty_txg(spa)))
4012 VERIFY3U(txg, <=, spa_final_dirty_txg(spa));
4015 * The only state that can actually be changing concurrently
4016 * with metaslab_sync() is the metaslab's ms_allocatable. No
4017 * other thread can be modifying this txg's alloc, freeing,
4018 * freed, or space_map_phys_t. We drop ms_lock whenever we
4019 * could call into the DMU, because the DMU can call down to
4020 * us (e.g. via zio_free()) at any time.
4022 * The spa_vdev_remove_thread() can be reading metaslab state
4023 * concurrently, and it is locked out by the ms_sync_lock.
4024 * Note that the ms_lock is insufficient for this, because it
4025 * is dropped by space_map_write().
4027 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
4030 * Generate a log space map if one doesn't exist already.
4032 spa_generate_syncing_log_sm(spa, tx);
4034 if (msp->ms_sm == NULL) {
4035 uint64_t new_object = space_map_alloc(mos,
4036 spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP) ?
4037 zfs_metaslab_sm_blksz_with_log :
4038 zfs_metaslab_sm_blksz_no_log, tx);
4039 VERIFY3U(new_object, !=, 0);
4041 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
4042 msp->ms_id, sizeof (uint64_t), &new_object, tx);
4044 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
4045 msp->ms_start, msp->ms_size, vd->vdev_ashift));
4046 ASSERT(msp->ms_sm != NULL);
4048 ASSERT(range_tree_is_empty(msp->ms_unflushed_allocs));
4049 ASSERT(range_tree_is_empty(msp->ms_unflushed_frees));
4050 ASSERT0(metaslab_allocated_space(msp));
4053 if (!range_tree_is_empty(msp->ms_checkpointing) &&
4054 vd->vdev_checkpoint_sm == NULL) {
4055 ASSERT(spa_has_checkpoint(spa));
4057 uint64_t new_object = space_map_alloc(mos,
4058 zfs_vdev_standard_sm_blksz, tx);
4059 VERIFY3U(new_object, !=, 0);
4061 VERIFY0(space_map_open(&vd->vdev_checkpoint_sm,
4062 mos, new_object, 0, vd->vdev_asize, vd->vdev_ashift));
4063 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
4066 * We save the space map object as an entry in vdev_top_zap
4067 * so it can be retrieved when the pool is reopened after an
4068 * export or through zdb.
4070 VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset,
4071 vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM,
4072 sizeof (new_object), 1, &new_object, tx));
4075 mutex_enter(&msp->ms_sync_lock);
4076 mutex_enter(&msp->ms_lock);
4079 * Note: metaslab_condense() clears the space map's histogram.
4080 * Therefore we must verify and remove this histogram before
4083 metaslab_group_histogram_verify(mg);
4084 metaslab_class_histogram_verify(mg->mg_class);
4085 metaslab_group_histogram_remove(mg, msp);
4087 if (spa->spa_sync_pass == 1 && msp->ms_loaded &&
4088 metaslab_should_condense(msp))
4089 metaslab_condense(msp, tx);
4092 * We'll be going to disk to sync our space accounting, thus we
4093 * drop the ms_lock during that time so allocations coming from
4094 * open-context (ZIL) for future TXGs do not block.
4096 mutex_exit(&msp->ms_lock);
4097 space_map_t *log_sm = spa_syncing_log_sm(spa);
4098 if (log_sm != NULL) {
4099 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP));
4100 if (metaslab_unflushed_txg(msp) == 0)
4101 metaslab_unflushed_add(msp, tx);
4102 else if (!metaslab_unflushed_dirty(msp))
4103 metaslab_unflushed_bump(msp, tx, B_TRUE);
4105 space_map_write(log_sm, alloctree, SM_ALLOC,
4107 space_map_write(log_sm, msp->ms_freeing, SM_FREE,
4109 mutex_enter(&msp->ms_lock);
4111 ASSERT3U(spa->spa_unflushed_stats.sus_memused, >=,
4112 metaslab_unflushed_changes_memused(msp));
4113 spa->spa_unflushed_stats.sus_memused -=
4114 metaslab_unflushed_changes_memused(msp);
4115 range_tree_remove_xor_add(alloctree,
4116 msp->ms_unflushed_frees, msp->ms_unflushed_allocs);
4117 range_tree_remove_xor_add(msp->ms_freeing,
4118 msp->ms_unflushed_allocs, msp->ms_unflushed_frees);
4119 spa->spa_unflushed_stats.sus_memused +=
4120 metaslab_unflushed_changes_memused(msp);
4122 ASSERT(!spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP));
4124 space_map_write(msp->ms_sm, alloctree, SM_ALLOC,
4126 space_map_write(msp->ms_sm, msp->ms_freeing, SM_FREE,
4128 mutex_enter(&msp->ms_lock);
4131 msp->ms_allocated_space += range_tree_space(alloctree);
4132 ASSERT3U(msp->ms_allocated_space, >=,
4133 range_tree_space(msp->ms_freeing));
4134 msp->ms_allocated_space -= range_tree_space(msp->ms_freeing);
4136 if (!range_tree_is_empty(msp->ms_checkpointing)) {
4137 ASSERT(spa_has_checkpoint(spa));
4138 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
4141 * Since we are doing writes to disk and the ms_checkpointing
4142 * tree won't be changing during that time, we drop the
4143 * ms_lock while writing to the checkpoint space map, for the
4144 * same reason mentioned above.
4146 mutex_exit(&msp->ms_lock);
4147 space_map_write(vd->vdev_checkpoint_sm,
4148 msp->ms_checkpointing, SM_FREE, SM_NO_VDEVID, tx);
4149 mutex_enter(&msp->ms_lock);
4151 spa->spa_checkpoint_info.sci_dspace +=
4152 range_tree_space(msp->ms_checkpointing);
4153 vd->vdev_stat.vs_checkpoint_space +=
4154 range_tree_space(msp->ms_checkpointing);
4155 ASSERT3U(vd->vdev_stat.vs_checkpoint_space, ==,
4156 -space_map_allocated(vd->vdev_checkpoint_sm));
4158 range_tree_vacate(msp->ms_checkpointing, NULL, NULL);
4161 if (msp->ms_loaded) {
4163 * When the space map is loaded, we have an accurate
4164 * histogram in the range tree. This gives us an opportunity
4165 * to bring the space map's histogram up-to-date so we clear
4166 * it first before updating it.
4168 space_map_histogram_clear(msp->ms_sm);
4169 space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
4172 * Since we've cleared the histogram we need to add back
4173 * any free space that has already been processed, plus
4174 * any deferred space. This allows the on-disk histogram
4175 * to accurately reflect all free space even if some space
4176 * is not yet available for allocation (i.e. deferred).
4178 space_map_histogram_add(msp->ms_sm, msp->ms_freed, tx);
4181 * Add back any deferred free space that has not been
4182 * added back into the in-core free tree yet. This will
4183 * ensure that we don't end up with a space map histogram
4184 * that is completely empty unless the metaslab is fully
4187 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
4188 space_map_histogram_add(msp->ms_sm,
4189 msp->ms_defer[t], tx);
4194 * Always add the free space from this sync pass to the space
4195 * map histogram. We want to make sure that the on-disk histogram
4196 * accounts for all free space. If the space map is not loaded,
4197 * then we will lose some accuracy but will correct it the next
4198 * time we load the space map.
4200 space_map_histogram_add(msp->ms_sm, msp->ms_freeing, tx);
4201 metaslab_aux_histograms_update(msp);
4203 metaslab_group_histogram_add(mg, msp);
4204 metaslab_group_histogram_verify(mg);
4205 metaslab_class_histogram_verify(mg->mg_class);
4208 * For sync pass 1, we avoid traversing this txg's free range tree
4209 * and instead will just swap the pointers for freeing and freed.
4210 * We can safely do this since the freed_tree is guaranteed to be
4211 * empty on the initial pass.
4213 * Keep in mind that even if we are currently using a log spacemap
4214 * we want current frees to end up in the ms_allocatable (but not
4215 * get appended to the ms_sm) so their ranges can be reused as usual.
4217 if (spa_sync_pass(spa) == 1) {
4218 range_tree_swap(&msp->ms_freeing, &msp->ms_freed);
4219 ASSERT0(msp->ms_allocated_this_txg);
4221 range_tree_vacate(msp->ms_freeing,
4222 range_tree_add, msp->ms_freed);
4224 msp->ms_allocated_this_txg += range_tree_space(alloctree);
4225 range_tree_vacate(alloctree, NULL, NULL);
4227 ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
4228 ASSERT0(range_tree_space(msp->ms_allocating[TXG_CLEAN(txg)
4230 ASSERT0(range_tree_space(msp->ms_freeing));
4231 ASSERT0(range_tree_space(msp->ms_checkpointing));
4233 mutex_exit(&msp->ms_lock);
4236 * Verify that the space map object ID has been recorded in the
4240 VERIFY0(dmu_read(mos, vd->vdev_ms_array,
4241 msp->ms_id * sizeof (uint64_t), sizeof (uint64_t), &object, 0));
4242 VERIFY3U(object, ==, space_map_object(msp->ms_sm));
4244 mutex_exit(&msp->ms_sync_lock);
4249 metaslab_evict(metaslab_t *msp, uint64_t txg)
4251 if (!msp->ms_loaded || msp->ms_disabled != 0)
4254 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
4255 VERIFY0(range_tree_space(
4256 msp->ms_allocating[(txg + t) & TXG_MASK]));
4258 if (msp->ms_allocator != -1)
4259 metaslab_passivate(msp, msp->ms_weight & ~METASLAB_ACTIVE_MASK);
4261 if (!metaslab_debug_unload)
4262 metaslab_unload(msp);
4266 * Called after a transaction group has completely synced to mark
4267 * all of the metaslab's free space as usable.
4270 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
4272 metaslab_group_t *mg = msp->ms_group;
4273 vdev_t *vd = mg->mg_vd;
4274 spa_t *spa = vd->vdev_spa;
4275 range_tree_t **defer_tree;
4276 int64_t alloc_delta, defer_delta;
4277 boolean_t defer_allowed = B_TRUE;
4279 ASSERT(!vd->vdev_ishole);
4281 mutex_enter(&msp->ms_lock);
4284 /* this is a new metaslab, add its capacity to the vdev */
4285 metaslab_space_update(vd, mg->mg_class, 0, 0, msp->ms_size);
4287 /* there should be no allocations nor frees at this point */
4288 VERIFY0(msp->ms_allocated_this_txg);
4289 VERIFY0(range_tree_space(msp->ms_freed));
4292 ASSERT0(range_tree_space(msp->ms_freeing));
4293 ASSERT0(range_tree_space(msp->ms_checkpointing));
4295 defer_tree = &msp->ms_defer[txg % TXG_DEFER_SIZE];
4297 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
4298 metaslab_class_get_alloc(spa_normal_class(spa));
4299 if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) {
4300 defer_allowed = B_FALSE;
4304 alloc_delta = msp->ms_allocated_this_txg -
4305 range_tree_space(msp->ms_freed);
4307 if (defer_allowed) {
4308 defer_delta = range_tree_space(msp->ms_freed) -
4309 range_tree_space(*defer_tree);
4311 defer_delta -= range_tree_space(*defer_tree);
4313 metaslab_space_update(vd, mg->mg_class, alloc_delta + defer_delta,
4316 if (spa_syncing_log_sm(spa) == NULL) {
4318 * If there's a metaslab_load() in progress and we don't have
4319 * a log space map, it means that we probably wrote to the
4320 * metaslab's space map. If this is the case, we need to
4321 * make sure that we wait for the load to complete so that we
4322 * have a consistent view at the in-core side of the metaslab.
4324 metaslab_load_wait(msp);
4326 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
4330 * When auto-trimming is enabled, free ranges which are added to
4331 * ms_allocatable are also be added to ms_trim. The ms_trim tree is
4332 * periodically consumed by the vdev_autotrim_thread() which issues
4333 * trims for all ranges and then vacates the tree. The ms_trim tree
4334 * can be discarded at any time with the sole consequence of recent
4335 * frees not being trimmed.
4337 if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON) {
4338 range_tree_walk(*defer_tree, range_tree_add, msp->ms_trim);
4339 if (!defer_allowed) {
4340 range_tree_walk(msp->ms_freed, range_tree_add,
4344 range_tree_vacate(msp->ms_trim, NULL, NULL);
4348 * Move the frees from the defer_tree back to the free
4349 * range tree (if it's loaded). Swap the freed_tree and
4350 * the defer_tree -- this is safe to do because we've
4351 * just emptied out the defer_tree.
4353 range_tree_vacate(*defer_tree,
4354 msp->ms_loaded ? range_tree_add : NULL, msp->ms_allocatable);
4355 if (defer_allowed) {
4356 range_tree_swap(&msp->ms_freed, defer_tree);
4358 range_tree_vacate(msp->ms_freed,
4359 msp->ms_loaded ? range_tree_add : NULL,
4360 msp->ms_allocatable);
4363 msp->ms_synced_length = space_map_length(msp->ms_sm);
4365 msp->ms_deferspace += defer_delta;
4366 ASSERT3S(msp->ms_deferspace, >=, 0);
4367 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
4368 if (msp->ms_deferspace != 0) {
4370 * Keep syncing this metaslab until all deferred frees
4371 * are back in circulation.
4373 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
4375 metaslab_aux_histograms_update_done(msp, defer_allowed);
4378 msp->ms_new = B_FALSE;
4379 mutex_enter(&mg->mg_lock);
4381 mutex_exit(&mg->mg_lock);
4385 * Re-sort metaslab within its group now that we've adjusted
4386 * its allocatable space.
4388 metaslab_recalculate_weight_and_sort(msp);
4390 ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
4391 ASSERT0(range_tree_space(msp->ms_freeing));
4392 ASSERT0(range_tree_space(msp->ms_freed));
4393 ASSERT0(range_tree_space(msp->ms_checkpointing));
4394 msp->ms_allocating_total -= msp->ms_allocated_this_txg;
4395 msp->ms_allocated_this_txg = 0;
4396 mutex_exit(&msp->ms_lock);
4400 metaslab_sync_reassess(metaslab_group_t *mg)
4402 spa_t *spa = mg->mg_class->mc_spa;
4404 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4405 metaslab_group_alloc_update(mg);
4406 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
4409 * Preload the next potential metaslabs but only on active
4410 * metaslab groups. We can get into a state where the metaslab
4411 * is no longer active since we dirty metaslabs as we remove a
4412 * a device, thus potentially making the metaslab group eligible
4415 if (mg->mg_activation_count > 0) {
4416 metaslab_group_preload(mg);
4418 spa_config_exit(spa, SCL_ALLOC, FTAG);
4422 * When writing a ditto block (i.e. more than one DVA for a given BP) on
4423 * the same vdev as an existing DVA of this BP, then try to allocate it
4424 * on a different metaslab than existing DVAs (i.e. a unique metaslab).
4427 metaslab_is_unique(metaslab_t *msp, dva_t *dva)
4431 if (DVA_GET_ASIZE(dva) == 0)
4434 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
4437 dva_ms_id = DVA_GET_OFFSET(dva) >> msp->ms_group->mg_vd->vdev_ms_shift;
4439 return (msp->ms_id != dva_ms_id);
4443 * ==========================================================================
4444 * Metaslab allocation tracing facility
4445 * ==========================================================================
4449 * Add an allocation trace element to the allocation tracing list.
4452 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
4453 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset,
4456 metaslab_alloc_trace_t *mat;
4458 if (!metaslab_trace_enabled)
4462 * When the tracing list reaches its maximum we remove
4463 * the second element in the list before adding a new one.
4464 * By removing the second element we preserve the original
4465 * entry as a clue to what allocations steps have already been
4468 if (zal->zal_size == metaslab_trace_max_entries) {
4469 metaslab_alloc_trace_t *mat_next;
4471 panic("too many entries in allocation list");
4473 METASLABSTAT_BUMP(metaslabstat_trace_over_limit);
4475 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
4476 list_remove(&zal->zal_list, mat_next);
4477 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
4480 mat = kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
4481 list_link_init(&mat->mat_list_node);
4484 mat->mat_size = psize;
4485 mat->mat_dva_id = dva_id;
4486 mat->mat_offset = offset;
4487 mat->mat_weight = 0;
4488 mat->mat_allocator = allocator;
4491 mat->mat_weight = msp->ms_weight;
4494 * The list is part of the zio so locking is not required. Only
4495 * a single thread will perform allocations for a given zio.
4497 list_insert_tail(&zal->zal_list, mat);
4500 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
4504 metaslab_trace_init(zio_alloc_list_t *zal)
4506 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
4507 offsetof(metaslab_alloc_trace_t, mat_list_node));
4512 metaslab_trace_fini(zio_alloc_list_t *zal)
4514 metaslab_alloc_trace_t *mat;
4516 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
4517 kmem_cache_free(metaslab_alloc_trace_cache, mat);
4518 list_destroy(&zal->zal_list);
4523 * ==========================================================================
4524 * Metaslab block operations
4525 * ==========================================================================
4529 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, const void *tag,
4530 int flags, int allocator)
4532 if (!(flags & METASLAB_ASYNC_ALLOC) ||
4533 (flags & METASLAB_DONT_THROTTLE))
4536 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4537 if (!mg->mg_class->mc_alloc_throttle_enabled)
4540 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4541 (void) zfs_refcount_add(&mga->mga_alloc_queue_depth, tag);
4545 metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator)
4547 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4548 metaslab_class_allocator_t *mca =
4549 &mg->mg_class->mc_allocator[allocator];
4550 uint64_t max = mg->mg_max_alloc_queue_depth;
4551 uint64_t cur = mga->mga_cur_max_alloc_queue_depth;
4553 if (atomic_cas_64(&mga->mga_cur_max_alloc_queue_depth,
4554 cur, cur + 1) == cur) {
4555 atomic_inc_64(&mca->mca_alloc_max_slots);
4558 cur = mga->mga_cur_max_alloc_queue_depth;
4563 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, const void *tag,
4564 int flags, int allocator, boolean_t io_complete)
4566 if (!(flags & METASLAB_ASYNC_ALLOC) ||
4567 (flags & METASLAB_DONT_THROTTLE))
4570 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4571 if (!mg->mg_class->mc_alloc_throttle_enabled)
4574 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4575 (void) zfs_refcount_remove(&mga->mga_alloc_queue_depth, tag);
4577 metaslab_group_increment_qdepth(mg, allocator);
4581 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, const void *tag,
4585 const dva_t *dva = bp->blk_dva;
4586 int ndvas = BP_GET_NDVAS(bp);
4588 for (int d = 0; d < ndvas; d++) {
4589 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
4590 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
4591 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4592 VERIFY(zfs_refcount_not_held(&mga->mga_alloc_queue_depth, tag));
4598 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
4601 range_tree_t *rt = msp->ms_allocatable;
4602 metaslab_class_t *mc = msp->ms_group->mg_class;
4604 ASSERT(MUTEX_HELD(&msp->ms_lock));
4605 VERIFY(!msp->ms_condensing);
4606 VERIFY0(msp->ms_disabled);
4608 start = mc->mc_ops->msop_alloc(msp, size);
4609 if (start != -1ULL) {
4610 metaslab_group_t *mg = msp->ms_group;
4611 vdev_t *vd = mg->mg_vd;
4613 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
4614 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
4615 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
4616 range_tree_remove(rt, start, size);
4617 range_tree_clear(msp->ms_trim, start, size);
4619 if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
4620 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
4622 range_tree_add(msp->ms_allocating[txg & TXG_MASK], start, size);
4623 msp->ms_allocating_total += size;
4625 /* Track the last successful allocation */
4626 msp->ms_alloc_txg = txg;
4627 metaslab_verify_space(msp, txg);
4631 * Now that we've attempted the allocation we need to update the
4632 * metaslab's maximum block size since it may have changed.
4634 msp->ms_max_size = metaslab_largest_allocatable(msp);
4639 * Find the metaslab with the highest weight that is less than what we've
4640 * already tried. In the common case, this means that we will examine each
4641 * metaslab at most once. Note that concurrent callers could reorder metaslabs
4642 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
4643 * activated by another thread, and we fail to allocate from the metaslab we
4644 * have selected, we may not try the newly-activated metaslab, and instead
4645 * activate another metaslab. This is not optimal, but generally does not cause
4646 * any problems (a possible exception being if every metaslab is completely full
4647 * except for the newly-activated metaslab which we fail to examine).
4650 find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight,
4651 dva_t *dva, int d, boolean_t want_unique, uint64_t asize, int allocator,
4652 boolean_t try_hard, zio_alloc_list_t *zal, metaslab_t *search,
4653 boolean_t *was_active)
4656 avl_tree_t *t = &mg->mg_metaslab_tree;
4657 metaslab_t *msp = avl_find(t, search, &idx);
4659 msp = avl_nearest(t, idx, AVL_AFTER);
4662 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
4665 if (!try_hard && tries > zfs_metaslab_find_max_tries) {
4666 METASLABSTAT_BUMP(metaslabstat_too_many_tries);
4671 if (!metaslab_should_allocate(msp, asize, try_hard)) {
4672 metaslab_trace_add(zal, mg, msp, asize, d,
4673 TRACE_TOO_SMALL, allocator);
4678 * If the selected metaslab is condensing or disabled,
4681 if (msp->ms_condensing || msp->ms_disabled > 0)
4684 *was_active = msp->ms_allocator != -1;
4686 * If we're activating as primary, this is our first allocation
4687 * from this disk, so we don't need to check how close we are.
4688 * If the metaslab under consideration was already active,
4689 * we're getting desperate enough to steal another allocator's
4690 * metaslab, so we still don't care about distances.
4692 if (activation_weight == METASLAB_WEIGHT_PRIMARY || *was_active)
4695 for (i = 0; i < d; i++) {
4697 !metaslab_is_unique(msp, &dva[i]))
4698 break; /* try another metaslab */
4705 search->ms_weight = msp->ms_weight;
4706 search->ms_start = msp->ms_start + 1;
4707 search->ms_allocator = msp->ms_allocator;
4708 search->ms_primary = msp->ms_primary;
4714 metaslab_active_mask_verify(metaslab_t *msp)
4716 ASSERT(MUTEX_HELD(&msp->ms_lock));
4718 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
4721 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0)
4724 if (msp->ms_weight & METASLAB_WEIGHT_PRIMARY) {
4725 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
4726 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM);
4727 VERIFY3S(msp->ms_allocator, !=, -1);
4728 VERIFY(msp->ms_primary);
4732 if (msp->ms_weight & METASLAB_WEIGHT_SECONDARY) {
4733 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
4734 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_CLAIM);
4735 VERIFY3S(msp->ms_allocator, !=, -1);
4736 VERIFY(!msp->ms_primary);
4740 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
4741 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
4742 VERIFY0(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
4743 VERIFY3S(msp->ms_allocator, ==, -1);
4749 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
4750 uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d,
4751 int allocator, boolean_t try_hard)
4753 metaslab_t *msp = NULL;
4754 uint64_t offset = -1ULL;
4756 uint64_t activation_weight = METASLAB_WEIGHT_PRIMARY;
4757 for (int i = 0; i < d; i++) {
4758 if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
4759 DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
4760 activation_weight = METASLAB_WEIGHT_SECONDARY;
4761 } else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
4762 DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
4763 activation_weight = METASLAB_WEIGHT_CLAIM;
4769 * If we don't have enough metaslabs active to fill the entire array, we
4770 * just use the 0th slot.
4772 if (mg->mg_ms_ready < mg->mg_allocators * 3)
4774 metaslab_group_allocator_t *mga = &mg->mg_allocator[allocator];
4776 ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2);
4778 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
4779 search->ms_weight = UINT64_MAX;
4780 search->ms_start = 0;
4782 * At the end of the metaslab tree are the already-active metaslabs,
4783 * first the primaries, then the secondaries. When we resume searching
4784 * through the tree, we need to consider ms_allocator and ms_primary so
4785 * we start in the location right after where we left off, and don't
4786 * accidentally loop forever considering the same metaslabs.
4788 search->ms_allocator = -1;
4789 search->ms_primary = B_TRUE;
4791 boolean_t was_active = B_FALSE;
4793 mutex_enter(&mg->mg_lock);
4795 if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
4796 mga->mga_primary != NULL) {
4797 msp = mga->mga_primary;
4800 * Even though we don't hold the ms_lock for the
4801 * primary metaslab, those fields should not
4802 * change while we hold the mg_lock. Thus it is
4803 * safe to make assertions on them.
4805 ASSERT(msp->ms_primary);
4806 ASSERT3S(msp->ms_allocator, ==, allocator);
4807 ASSERT(msp->ms_loaded);
4809 was_active = B_TRUE;
4810 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
4811 } else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
4812 mga->mga_secondary != NULL) {
4813 msp = mga->mga_secondary;
4816 * See comment above about the similar assertions
4817 * for the primary metaslab.
4819 ASSERT(!msp->ms_primary);
4820 ASSERT3S(msp->ms_allocator, ==, allocator);
4821 ASSERT(msp->ms_loaded);
4823 was_active = B_TRUE;
4824 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
4826 msp = find_valid_metaslab(mg, activation_weight, dva, d,
4827 want_unique, asize, allocator, try_hard, zal,
4828 search, &was_active);
4831 mutex_exit(&mg->mg_lock);
4833 kmem_free(search, sizeof (*search));
4836 mutex_enter(&msp->ms_lock);
4838 metaslab_active_mask_verify(msp);
4841 * This code is disabled out because of issues with
4842 * tracepoints in non-gpl kernel modules.
4845 DTRACE_PROBE3(ms__activation__attempt,
4846 metaslab_t *, msp, uint64_t, activation_weight,
4847 boolean_t, was_active);
4851 * Ensure that the metaslab we have selected is still
4852 * capable of handling our request. It's possible that
4853 * another thread may have changed the weight while we
4854 * were blocked on the metaslab lock. We check the
4855 * active status first to see if we need to set_selected_txg
4858 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
4859 ASSERT3S(msp->ms_allocator, ==, -1);
4860 mutex_exit(&msp->ms_lock);
4865 * If the metaslab was activated for another allocator
4866 * while we were waiting in the ms_lock above, or it's
4867 * a primary and we're seeking a secondary (or vice versa),
4868 * we go back and select a new metaslab.
4870 if (!was_active && (msp->ms_weight & METASLAB_ACTIVE_MASK) &&
4871 (msp->ms_allocator != -1) &&
4872 (msp->ms_allocator != allocator || ((activation_weight ==
4873 METASLAB_WEIGHT_PRIMARY) != msp->ms_primary))) {
4874 ASSERT(msp->ms_loaded);
4875 ASSERT((msp->ms_weight & METASLAB_WEIGHT_CLAIM) ||
4876 msp->ms_allocator != -1);
4877 mutex_exit(&msp->ms_lock);
4882 * This metaslab was used for claiming regions allocated
4883 * by the ZIL during pool import. Once these regions are
4884 * claimed we don't need to keep the CLAIM bit set
4885 * anymore. Passivate this metaslab to zero its activation
4888 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM &&
4889 activation_weight != METASLAB_WEIGHT_CLAIM) {
4890 ASSERT(msp->ms_loaded);
4891 ASSERT3S(msp->ms_allocator, ==, -1);
4892 metaslab_passivate(msp, msp->ms_weight &
4893 ~METASLAB_WEIGHT_CLAIM);
4894 mutex_exit(&msp->ms_lock);
4898 metaslab_set_selected_txg(msp, txg);
4900 int activation_error =
4901 metaslab_activate(msp, allocator, activation_weight);
4902 metaslab_active_mask_verify(msp);
4905 * If the metaslab was activated by another thread for
4906 * another allocator or activation_weight (EBUSY), or it
4907 * failed because another metaslab was assigned as primary
4908 * for this allocator (EEXIST) we continue using this
4909 * metaslab for our allocation, rather than going on to a
4910 * worse metaslab (we waited for that metaslab to be loaded
4913 * If the activation failed due to an I/O error or ENOSPC we
4914 * skip to the next metaslab.
4916 boolean_t activated;
4917 if (activation_error == 0) {
4919 } else if (activation_error == EBUSY ||
4920 activation_error == EEXIST) {
4921 activated = B_FALSE;
4923 mutex_exit(&msp->ms_lock);
4926 ASSERT(msp->ms_loaded);
4929 * Now that we have the lock, recheck to see if we should
4930 * continue to use this metaslab for this allocation. The
4931 * the metaslab is now loaded so metaslab_should_allocate()
4932 * can accurately determine if the allocation attempt should
4935 if (!metaslab_should_allocate(msp, asize, try_hard)) {
4936 /* Passivate this metaslab and select a new one. */
4937 metaslab_trace_add(zal, mg, msp, asize, d,
4938 TRACE_TOO_SMALL, allocator);
4943 * If this metaslab is currently condensing then pick again
4944 * as we can't manipulate this metaslab until it's committed
4945 * to disk. If this metaslab is being initialized, we shouldn't
4946 * allocate from it since the allocated region might be
4947 * overwritten after allocation.
4949 if (msp->ms_condensing) {
4950 metaslab_trace_add(zal, mg, msp, asize, d,
4951 TRACE_CONDENSING, allocator);
4953 metaslab_passivate(msp, msp->ms_weight &
4954 ~METASLAB_ACTIVE_MASK);
4956 mutex_exit(&msp->ms_lock);
4958 } else if (msp->ms_disabled > 0) {
4959 metaslab_trace_add(zal, mg, msp, asize, d,
4960 TRACE_DISABLED, allocator);
4962 metaslab_passivate(msp, msp->ms_weight &
4963 ~METASLAB_ACTIVE_MASK);
4965 mutex_exit(&msp->ms_lock);
4969 offset = metaslab_block_alloc(msp, asize, txg);
4970 metaslab_trace_add(zal, mg, msp, asize, d, offset, allocator);
4972 if (offset != -1ULL) {
4973 /* Proactively passivate the metaslab, if needed */
4975 metaslab_segment_may_passivate(msp);
4979 ASSERT(msp->ms_loaded);
4982 * This code is disabled out because of issues with
4983 * tracepoints in non-gpl kernel modules.
4986 DTRACE_PROBE2(ms__alloc__failure, metaslab_t *, msp,
4991 * We were unable to allocate from this metaslab so determine
4992 * a new weight for this metaslab. Now that we have loaded
4993 * the metaslab we can provide a better hint to the metaslab
4996 * For space-based metaslabs, we use the maximum block size.
4997 * This information is only available when the metaslab
4998 * is loaded and is more accurate than the generic free
4999 * space weight that was calculated by metaslab_weight().
5000 * This information allows us to quickly compare the maximum
5001 * available allocation in the metaslab to the allocation
5002 * size being requested.
5004 * For segment-based metaslabs, determine the new weight
5005 * based on the highest bucket in the range tree. We
5006 * explicitly use the loaded segment weight (i.e. the range
5007 * tree histogram) since it contains the space that is
5008 * currently available for allocation and is accurate
5009 * even within a sync pass.
5012 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
5013 weight = metaslab_largest_allocatable(msp);
5014 WEIGHT_SET_SPACEBASED(weight);
5016 weight = metaslab_weight_from_range_tree(msp);
5020 metaslab_passivate(msp, weight);
5023 * For the case where we use the metaslab that is
5024 * active for another allocator we want to make
5025 * sure that we retain the activation mask.
5027 * Note that we could attempt to use something like
5028 * metaslab_recalculate_weight_and_sort() that
5029 * retains the activation mask here. That function
5030 * uses metaslab_weight() to set the weight though
5031 * which is not as accurate as the calculations
5034 weight |= msp->ms_weight & METASLAB_ACTIVE_MASK;
5035 metaslab_group_sort(mg, msp, weight);
5037 metaslab_active_mask_verify(msp);
5040 * We have just failed an allocation attempt, check
5041 * that metaslab_should_allocate() agrees. Otherwise,
5042 * we may end up in an infinite loop retrying the same
5045 ASSERT(!metaslab_should_allocate(msp, asize, try_hard));
5047 mutex_exit(&msp->ms_lock);
5049 mutex_exit(&msp->ms_lock);
5050 kmem_free(search, sizeof (*search));
5055 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
5056 uint64_t asize, uint64_t txg, boolean_t want_unique, dva_t *dva, int d,
5057 int allocator, boolean_t try_hard)
5060 ASSERT(mg->mg_initialized);
5062 offset = metaslab_group_alloc_normal(mg, zal, asize, txg, want_unique,
5063 dva, d, allocator, try_hard);
5065 mutex_enter(&mg->mg_lock);
5066 if (offset == -1ULL) {
5067 mg->mg_failed_allocations++;
5068 metaslab_trace_add(zal, mg, NULL, asize, d,
5069 TRACE_GROUP_FAILURE, allocator);
5070 if (asize == SPA_GANGBLOCKSIZE) {
5072 * This metaslab group was unable to allocate
5073 * the minimum gang block size so it must be out of
5074 * space. We must notify the allocation throttle
5075 * to start skipping allocation attempts to this
5076 * metaslab group until more space becomes available.
5077 * Note: this failure cannot be caused by the
5078 * allocation throttle since the allocation throttle
5079 * is only responsible for skipping devices and
5080 * not failing block allocations.
5082 mg->mg_no_free_space = B_TRUE;
5085 mg->mg_allocations++;
5086 mutex_exit(&mg->mg_lock);
5091 * Allocate a block for the specified i/o.
5094 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
5095 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
5096 zio_alloc_list_t *zal, int allocator)
5098 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
5099 metaslab_group_t *mg, *fast_mg, *rotor;
5101 boolean_t try_hard = B_FALSE;
5103 ASSERT(!DVA_IS_VALID(&dva[d]));
5106 * For testing, make some blocks above a certain size be gang blocks.
5107 * This will result in more split blocks when using device removal,
5108 * and a large number of split blocks coupled with ztest-induced
5109 * damage can result in extremely long reconstruction times. This
5110 * will also test spilling from special to normal.
5112 if (psize >= metaslab_force_ganging && (random_in_range(100) < 3)) {
5113 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG,
5115 return (SET_ERROR(ENOSPC));
5119 * Start at the rotor and loop through all mgs until we find something.
5120 * Note that there's no locking on mca_rotor or mca_aliquot because
5121 * nothing actually breaks if we miss a few updates -- we just won't
5122 * allocate quite as evenly. It all balances out over time.
5124 * If we are doing ditto or log blocks, try to spread them across
5125 * consecutive vdevs. If we're forced to reuse a vdev before we've
5126 * allocated all of our ditto blocks, then try and spread them out on
5127 * that vdev as much as possible. If it turns out to not be possible,
5128 * gradually lower our standards until anything becomes acceptable.
5129 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
5130 * gives us hope of containing our fault domains to something we're
5131 * able to reason about. Otherwise, any two top-level vdev failures
5132 * will guarantee the loss of data. With consecutive allocation,
5133 * only two adjacent top-level vdev failures will result in data loss.
5135 * If we are doing gang blocks (hintdva is non-NULL), try to keep
5136 * ourselves on the same vdev as our gang block header. That
5137 * way, we can hope for locality in vdev_cache, plus it makes our
5138 * fault domains something tractable.
5141 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
5144 * It's possible the vdev we're using as the hint no
5145 * longer exists or its mg has been closed (e.g. by
5146 * device removal). Consult the rotor when
5149 if (vd != NULL && vd->vdev_mg != NULL) {
5150 mg = vdev_get_mg(vd, mc);
5152 if (flags & METASLAB_HINTBP_AVOID)
5155 mg = mca->mca_rotor;
5157 } else if (d != 0) {
5158 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
5159 mg = vd->vdev_mg->mg_next;
5160 } else if (flags & METASLAB_FASTWRITE) {
5161 mg = fast_mg = mca->mca_rotor;
5164 if (fast_mg->mg_vd->vdev_pending_fastwrite <
5165 mg->mg_vd->vdev_pending_fastwrite)
5167 } while ((fast_mg = fast_mg->mg_next) != mca->mca_rotor);
5170 ASSERT(mca->mca_rotor != NULL);
5171 mg = mca->mca_rotor;
5175 * If the hint put us into the wrong metaslab class, or into a
5176 * metaslab group that has been passivated, just follow the rotor.
5178 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
5179 mg = mca->mca_rotor;
5184 boolean_t allocatable;
5186 ASSERT(mg->mg_activation_count == 1);
5190 * Don't allocate from faulted devices.
5193 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
5194 allocatable = vdev_allocatable(vd);
5195 spa_config_exit(spa, SCL_ZIO, FTAG);
5197 allocatable = vdev_allocatable(vd);
5201 * Determine if the selected metaslab group is eligible
5202 * for allocations. If we're ganging then don't allow
5203 * this metaslab group to skip allocations since that would
5204 * inadvertently return ENOSPC and suspend the pool
5205 * even though space is still available.
5207 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
5208 allocatable = metaslab_group_allocatable(mg, rotor,
5209 flags, psize, allocator, d);
5213 metaslab_trace_add(zal, mg, NULL, psize, d,
5214 TRACE_NOT_ALLOCATABLE, allocator);
5218 ASSERT(mg->mg_initialized);
5221 * Avoid writing single-copy data to an unhealthy,
5222 * non-redundant vdev, unless we've already tried all
5225 if (vd->vdev_state < VDEV_STATE_HEALTHY &&
5226 d == 0 && !try_hard && vd->vdev_children == 0) {
5227 metaslab_trace_add(zal, mg, NULL, psize, d,
5228 TRACE_VDEV_ERROR, allocator);
5232 ASSERT(mg->mg_class == mc);
5234 uint64_t asize = vdev_psize_to_asize(vd, psize);
5235 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
5238 * If we don't need to try hard, then require that the
5239 * block be on a different metaslab from any other DVAs
5240 * in this BP (unique=true). If we are trying hard, then
5241 * allow any metaslab to be used (unique=false).
5243 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
5244 !try_hard, dva, d, allocator, try_hard);
5246 if (offset != -1ULL) {
5248 * If we've just selected this metaslab group,
5249 * figure out whether the corresponding vdev is
5250 * over- or under-used relative to the pool,
5251 * and set an allocation bias to even it out.
5253 * Bias is also used to compensate for unequally
5254 * sized vdevs so that space is allocated fairly.
5256 if (mca->mca_aliquot == 0 && metaslab_bias_enabled) {
5257 vdev_stat_t *vs = &vd->vdev_stat;
5258 int64_t vs_free = vs->vs_space - vs->vs_alloc;
5259 int64_t mc_free = mc->mc_space - mc->mc_alloc;
5263 * Calculate how much more or less we should
5264 * try to allocate from this device during
5265 * this iteration around the rotor.
5267 * This basically introduces a zero-centered
5268 * bias towards the devices with the most
5269 * free space, while compensating for vdev
5273 * vdev V1 = 16M/128M
5274 * vdev V2 = 16M/128M
5275 * ratio(V1) = 100% ratio(V2) = 100%
5277 * vdev V1 = 16M/128M
5278 * vdev V2 = 64M/128M
5279 * ratio(V1) = 127% ratio(V2) = 72%
5281 * vdev V1 = 16M/128M
5282 * vdev V2 = 64M/512M
5283 * ratio(V1) = 40% ratio(V2) = 160%
5285 ratio = (vs_free * mc->mc_alloc_groups * 100) /
5287 mg->mg_bias = ((ratio - 100) *
5288 (int64_t)mg->mg_aliquot) / 100;
5289 } else if (!metaslab_bias_enabled) {
5293 if ((flags & METASLAB_FASTWRITE) ||
5294 atomic_add_64_nv(&mca->mca_aliquot, asize) >=
5295 mg->mg_aliquot + mg->mg_bias) {
5296 mca->mca_rotor = mg->mg_next;
5297 mca->mca_aliquot = 0;
5300 DVA_SET_VDEV(&dva[d], vd->vdev_id);
5301 DVA_SET_OFFSET(&dva[d], offset);
5302 DVA_SET_GANG(&dva[d],
5303 ((flags & METASLAB_GANG_HEADER) ? 1 : 0));
5304 DVA_SET_ASIZE(&dva[d], asize);
5306 if (flags & METASLAB_FASTWRITE) {
5307 atomic_add_64(&vd->vdev_pending_fastwrite,
5314 mca->mca_rotor = mg->mg_next;
5315 mca->mca_aliquot = 0;
5316 } while ((mg = mg->mg_next) != rotor);
5319 * If we haven't tried hard, perhaps do so now.
5321 if (!try_hard && (zfs_metaslab_try_hard_before_gang ||
5322 GANG_ALLOCATION(flags) || (flags & METASLAB_ZIL) != 0 ||
5323 psize <= 1 << spa->spa_min_ashift)) {
5324 METASLABSTAT_BUMP(metaslabstat_try_hard);
5329 memset(&dva[d], 0, sizeof (dva_t));
5331 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC, allocator);
5332 return (SET_ERROR(ENOSPC));
5336 metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize,
5337 boolean_t checkpoint)
5340 spa_t *spa = vd->vdev_spa;
5342 ASSERT(vdev_is_concrete(vd));
5343 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5344 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
5346 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5348 VERIFY(!msp->ms_condensing);
5349 VERIFY3U(offset, >=, msp->ms_start);
5350 VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size);
5351 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5352 VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift));
5354 metaslab_check_free_impl(vd, offset, asize);
5356 mutex_enter(&msp->ms_lock);
5357 if (range_tree_is_empty(msp->ms_freeing) &&
5358 range_tree_is_empty(msp->ms_checkpointing)) {
5359 vdev_dirty(vd, VDD_METASLAB, msp, spa_syncing_txg(spa));
5363 ASSERT(spa_has_checkpoint(spa));
5364 range_tree_add(msp->ms_checkpointing, offset, asize);
5366 range_tree_add(msp->ms_freeing, offset, asize);
5368 mutex_exit(&msp->ms_lock);
5372 metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5373 uint64_t size, void *arg)
5375 (void) inner_offset;
5376 boolean_t *checkpoint = arg;
5378 ASSERT3P(checkpoint, !=, NULL);
5380 if (vd->vdev_ops->vdev_op_remap != NULL)
5381 vdev_indirect_mark_obsolete(vd, offset, size);
5383 metaslab_free_impl(vd, offset, size, *checkpoint);
5387 metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
5388 boolean_t checkpoint)
5390 spa_t *spa = vd->vdev_spa;
5392 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5394 if (spa_syncing_txg(spa) > spa_freeze_txg(spa))
5397 if (spa->spa_vdev_removal != NULL &&
5398 spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id &&
5399 vdev_is_concrete(vd)) {
5401 * Note: we check if the vdev is concrete because when
5402 * we complete the removal, we first change the vdev to be
5403 * an indirect vdev (in open context), and then (in syncing
5404 * context) clear spa_vdev_removal.
5406 free_from_removing_vdev(vd, offset, size);
5407 } else if (vd->vdev_ops->vdev_op_remap != NULL) {
5408 vdev_indirect_mark_obsolete(vd, offset, size);
5409 vd->vdev_ops->vdev_op_remap(vd, offset, size,
5410 metaslab_free_impl_cb, &checkpoint);
5412 metaslab_free_concrete(vd, offset, size, checkpoint);
5416 typedef struct remap_blkptr_cb_arg {
5418 spa_remap_cb_t rbca_cb;
5419 vdev_t *rbca_remap_vd;
5420 uint64_t rbca_remap_offset;
5422 } remap_blkptr_cb_arg_t;
5425 remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5426 uint64_t size, void *arg)
5428 remap_blkptr_cb_arg_t *rbca = arg;
5429 blkptr_t *bp = rbca->rbca_bp;
5431 /* We can not remap split blocks. */
5432 if (size != DVA_GET_ASIZE(&bp->blk_dva[0]))
5434 ASSERT0(inner_offset);
5436 if (rbca->rbca_cb != NULL) {
5438 * At this point we know that we are not handling split
5439 * blocks and we invoke the callback on the previous
5440 * vdev which must be indirect.
5442 ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops);
5444 rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id,
5445 rbca->rbca_remap_offset, size, rbca->rbca_cb_arg);
5447 /* set up remap_blkptr_cb_arg for the next call */
5448 rbca->rbca_remap_vd = vd;
5449 rbca->rbca_remap_offset = offset;
5453 * The phys birth time is that of dva[0]. This ensures that we know
5454 * when each dva was written, so that resilver can determine which
5455 * blocks need to be scrubbed (i.e. those written during the time
5456 * the vdev was offline). It also ensures that the key used in
5457 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
5458 * we didn't change the phys_birth, a lookup in the ARC for a
5459 * remapped BP could find the data that was previously stored at
5460 * this vdev + offset.
5462 vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa,
5463 DVA_GET_VDEV(&bp->blk_dva[0]));
5464 vdev_indirect_births_t *vib = oldvd->vdev_indirect_births;
5465 bp->blk_phys_birth = vdev_indirect_births_physbirth(vib,
5466 DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0]));
5468 DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id);
5469 DVA_SET_OFFSET(&bp->blk_dva[0], offset);
5473 * If the block pointer contains any indirect DVAs, modify them to refer to
5474 * concrete DVAs. Note that this will sometimes not be possible, leaving
5475 * the indirect DVA in place. This happens if the indirect DVA spans multiple
5476 * segments in the mapping (i.e. it is a "split block").
5478 * If the BP was remapped, calls the callback on the original dva (note the
5479 * callback can be called multiple times if the original indirect DVA refers
5480 * to another indirect DVA, etc).
5482 * Returns TRUE if the BP was remapped.
5485 spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg)
5487 remap_blkptr_cb_arg_t rbca;
5489 if (!zfs_remap_blkptr_enable)
5492 if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS))
5496 * Dedup BP's can not be remapped, because ddt_phys_select() depends
5497 * on DVA[0] being the same in the BP as in the DDT (dedup table).
5499 if (BP_GET_DEDUP(bp))
5503 * Gang blocks can not be remapped, because
5504 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
5505 * the BP used to read the gang block header (GBH) being the same
5506 * as the DVA[0] that we allocated for the GBH.
5512 * Embedded BP's have no DVA to remap.
5514 if (BP_GET_NDVAS(bp) < 1)
5518 * Note: we only remap dva[0]. If we remapped other dvas, we
5519 * would no longer know what their phys birth txg is.
5521 dva_t *dva = &bp->blk_dva[0];
5523 uint64_t offset = DVA_GET_OFFSET(dva);
5524 uint64_t size = DVA_GET_ASIZE(dva);
5525 vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
5527 if (vd->vdev_ops->vdev_op_remap == NULL)
5531 rbca.rbca_cb = callback;
5532 rbca.rbca_remap_vd = vd;
5533 rbca.rbca_remap_offset = offset;
5534 rbca.rbca_cb_arg = arg;
5537 * remap_blkptr_cb() will be called in order for each level of
5538 * indirection, until a concrete vdev is reached or a split block is
5539 * encountered. old_vd and old_offset are updated within the callback
5540 * as we go from the one indirect vdev to the next one (either concrete
5541 * or indirect again) in that order.
5543 vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca);
5545 /* Check if the DVA wasn't remapped because it is a split block */
5546 if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id)
5553 * Undo the allocation of a DVA which happened in the given transaction group.
5556 metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
5560 uint64_t vdev = DVA_GET_VDEV(dva);
5561 uint64_t offset = DVA_GET_OFFSET(dva);
5562 uint64_t size = DVA_GET_ASIZE(dva);
5564 ASSERT(DVA_IS_VALID(dva));
5565 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5567 if (txg > spa_freeze_txg(spa))
5570 if ((vd = vdev_lookup_top(spa, vdev)) == NULL || !DVA_IS_VALID(dva) ||
5571 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
5572 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
5573 (u_longlong_t)vdev, (u_longlong_t)offset,
5574 (u_longlong_t)size);
5578 ASSERT(!vd->vdev_removing);
5579 ASSERT(vdev_is_concrete(vd));
5580 ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
5581 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
5583 if (DVA_GET_GANG(dva))
5584 size = vdev_gang_header_asize(vd);
5586 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5588 mutex_enter(&msp->ms_lock);
5589 range_tree_remove(msp->ms_allocating[txg & TXG_MASK],
5591 msp->ms_allocating_total -= size;
5593 VERIFY(!msp->ms_condensing);
5594 VERIFY3U(offset, >=, msp->ms_start);
5595 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
5596 VERIFY3U(range_tree_space(msp->ms_allocatable) + size, <=,
5598 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5599 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
5600 range_tree_add(msp->ms_allocatable, offset, size);
5601 mutex_exit(&msp->ms_lock);
5605 * Free the block represented by the given DVA.
5608 metaslab_free_dva(spa_t *spa, const dva_t *dva, boolean_t checkpoint)
5610 uint64_t vdev = DVA_GET_VDEV(dva);
5611 uint64_t offset = DVA_GET_OFFSET(dva);
5612 uint64_t size = DVA_GET_ASIZE(dva);
5613 vdev_t *vd = vdev_lookup_top(spa, vdev);
5615 ASSERT(DVA_IS_VALID(dva));
5616 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
5618 if (DVA_GET_GANG(dva)) {
5619 size = vdev_gang_header_asize(vd);
5622 metaslab_free_impl(vd, offset, size, checkpoint);
5626 * Reserve some allocation slots. The reservation system must be called
5627 * before we call into the allocator. If there aren't any available slots
5628 * then the I/O will be throttled until an I/O completes and its slots are
5629 * freed up. The function returns true if it was successful in placing
5633 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
5634 zio_t *zio, int flags)
5636 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
5637 uint64_t max = mca->mca_alloc_max_slots;
5639 ASSERT(mc->mc_alloc_throttle_enabled);
5640 if (GANG_ALLOCATION(flags) || (flags & METASLAB_MUST_RESERVE) ||
5641 zfs_refcount_count(&mca->mca_alloc_slots) + slots <= max) {
5643 * The potential race between _count() and _add() is covered
5644 * by the allocator lock in most cases, or irrelevant due to
5645 * GANG_ALLOCATION() or METASLAB_MUST_RESERVE set in others.
5646 * But even if we assume some other non-existing scenario, the
5647 * worst that can happen is few more I/Os get to allocation
5648 * earlier, that is not a problem.
5650 * We reserve the slots individually so that we can unreserve
5651 * them individually when an I/O completes.
5653 zfs_refcount_add_few(&mca->mca_alloc_slots, slots, zio);
5654 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
5661 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots,
5662 int allocator, zio_t *zio)
5664 metaslab_class_allocator_t *mca = &mc->mc_allocator[allocator];
5666 ASSERT(mc->mc_alloc_throttle_enabled);
5667 zfs_refcount_remove_few(&mca->mca_alloc_slots, slots, zio);
5671 metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
5675 spa_t *spa = vd->vdev_spa;
5678 if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count)
5679 return (SET_ERROR(ENXIO));
5681 ASSERT3P(vd->vdev_ms, !=, NULL);
5682 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
5684 mutex_enter(&msp->ms_lock);
5686 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded) {
5687 error = metaslab_activate(msp, 0, METASLAB_WEIGHT_CLAIM);
5688 if (error == EBUSY) {
5689 ASSERT(msp->ms_loaded);
5690 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
5696 !range_tree_contains(msp->ms_allocatable, offset, size))
5697 error = SET_ERROR(ENOENT);
5699 if (error || txg == 0) { /* txg == 0 indicates dry run */
5700 mutex_exit(&msp->ms_lock);
5704 VERIFY(!msp->ms_condensing);
5705 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
5706 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
5707 VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=,
5709 range_tree_remove(msp->ms_allocatable, offset, size);
5710 range_tree_clear(msp->ms_trim, offset, size);
5712 if (spa_writeable(spa)) { /* don't dirty if we're zdb(8) */
5713 metaslab_class_t *mc = msp->ms_group->mg_class;
5714 multilist_sublist_t *mls =
5715 multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
5716 if (!multilist_link_active(&msp->ms_class_txg_node)) {
5717 msp->ms_selected_txg = txg;
5718 multilist_sublist_insert_head(mls, msp);
5720 multilist_sublist_unlock(mls);
5722 if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
5723 vdev_dirty(vd, VDD_METASLAB, msp, txg);
5724 range_tree_add(msp->ms_allocating[txg & TXG_MASK],
5726 msp->ms_allocating_total += size;
5729 mutex_exit(&msp->ms_lock);
5734 typedef struct metaslab_claim_cb_arg_t {
5737 } metaslab_claim_cb_arg_t;
5740 metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
5741 uint64_t size, void *arg)
5743 (void) inner_offset;
5744 metaslab_claim_cb_arg_t *mcca_arg = arg;
5746 if (mcca_arg->mcca_error == 0) {
5747 mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset,
5748 size, mcca_arg->mcca_txg);
5753 metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
5755 if (vd->vdev_ops->vdev_op_remap != NULL) {
5756 metaslab_claim_cb_arg_t arg;
5759 * Only zdb(8) can claim on indirect vdevs. This is used
5760 * to detect leaks of mapped space (that are not accounted
5761 * for in the obsolete counts, spacemap, or bpobj).
5763 ASSERT(!spa_writeable(vd->vdev_spa));
5767 vd->vdev_ops->vdev_op_remap(vd, offset, size,
5768 metaslab_claim_impl_cb, &arg);
5770 if (arg.mcca_error == 0) {
5771 arg.mcca_error = metaslab_claim_concrete(vd,
5774 return (arg.mcca_error);
5776 return (metaslab_claim_concrete(vd, offset, size, txg));
5781 * Intent log support: upon opening the pool after a crash, notify the SPA
5782 * of blocks that the intent log has allocated for immediate write, but
5783 * which are still considered free by the SPA because the last transaction
5784 * group didn't commit yet.
5787 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
5789 uint64_t vdev = DVA_GET_VDEV(dva);
5790 uint64_t offset = DVA_GET_OFFSET(dva);
5791 uint64_t size = DVA_GET_ASIZE(dva);
5794 if ((vd = vdev_lookup_top(spa, vdev)) == NULL) {
5795 return (SET_ERROR(ENXIO));
5798 ASSERT(DVA_IS_VALID(dva));
5800 if (DVA_GET_GANG(dva))
5801 size = vdev_gang_header_asize(vd);
5803 return (metaslab_claim_impl(vd, offset, size, txg));
5807 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
5808 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
5809 zio_alloc_list_t *zal, zio_t *zio, int allocator)
5811 dva_t *dva = bp->blk_dva;
5812 dva_t *hintdva = (hintbp != NULL) ? hintbp->blk_dva : NULL;
5815 ASSERT(bp->blk_birth == 0);
5816 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
5818 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
5820 if (mc->mc_allocator[allocator].mca_rotor == NULL) {
5821 /* no vdevs in this class */
5822 spa_config_exit(spa, SCL_ALLOC, FTAG);
5823 return (SET_ERROR(ENOSPC));
5826 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
5827 ASSERT(BP_GET_NDVAS(bp) == 0);
5828 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
5829 ASSERT3P(zal, !=, NULL);
5831 for (int d = 0; d < ndvas; d++) {
5832 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
5833 txg, flags, zal, allocator);
5835 for (d--; d >= 0; d--) {
5836 metaslab_unalloc_dva(spa, &dva[d], txg);
5837 metaslab_group_alloc_decrement(spa,
5838 DVA_GET_VDEV(&dva[d]), zio, flags,
5839 allocator, B_FALSE);
5840 memset(&dva[d], 0, sizeof (dva_t));
5842 spa_config_exit(spa, SCL_ALLOC, FTAG);
5846 * Update the metaslab group's queue depth
5847 * based on the newly allocated dva.
5849 metaslab_group_alloc_increment(spa,
5850 DVA_GET_VDEV(&dva[d]), zio, flags, allocator);
5854 ASSERT(BP_GET_NDVAS(bp) == ndvas);
5856 spa_config_exit(spa, SCL_ALLOC, FTAG);
5858 BP_SET_BIRTH(bp, txg, 0);
5864 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
5866 const dva_t *dva = bp->blk_dva;
5867 int ndvas = BP_GET_NDVAS(bp);
5869 ASSERT(!BP_IS_HOLE(bp));
5870 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
5873 * If we have a checkpoint for the pool we need to make sure that
5874 * the blocks that we free that are part of the checkpoint won't be
5875 * reused until the checkpoint is discarded or we revert to it.
5877 * The checkpoint flag is passed down the metaslab_free code path
5878 * and is set whenever we want to add a block to the checkpoint's
5879 * accounting. That is, we "checkpoint" blocks that existed at the
5880 * time the checkpoint was created and are therefore referenced by
5881 * the checkpointed uberblock.
5883 * Note that, we don't checkpoint any blocks if the current
5884 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
5885 * normally as they will be referenced by the checkpointed uberblock.
5887 boolean_t checkpoint = B_FALSE;
5888 if (bp->blk_birth <= spa->spa_checkpoint_txg &&
5889 spa_syncing_txg(spa) > spa->spa_checkpoint_txg) {
5891 * At this point, if the block is part of the checkpoint
5892 * there is no way it was created in the current txg.
5895 ASSERT3U(spa_syncing_txg(spa), ==, txg);
5896 checkpoint = B_TRUE;
5899 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
5901 for (int d = 0; d < ndvas; d++) {
5903 metaslab_unalloc_dva(spa, &dva[d], txg);
5905 ASSERT3U(txg, ==, spa_syncing_txg(spa));
5906 metaslab_free_dva(spa, &dva[d], checkpoint);
5910 spa_config_exit(spa, SCL_FREE, FTAG);
5914 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
5916 const dva_t *dva = bp->blk_dva;
5917 int ndvas = BP_GET_NDVAS(bp);
5920 ASSERT(!BP_IS_HOLE(bp));
5924 * First do a dry run to make sure all DVAs are claimable,
5925 * so we don't have to unwind from partial failures below.
5927 if ((error = metaslab_claim(spa, bp, 0)) != 0)
5931 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
5933 for (int d = 0; d < ndvas; d++) {
5934 error = metaslab_claim_dva(spa, &dva[d], txg);
5939 spa_config_exit(spa, SCL_ALLOC, FTAG);
5941 ASSERT(error == 0 || txg == 0);
5947 metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp)
5949 const dva_t *dva = bp->blk_dva;
5950 int ndvas = BP_GET_NDVAS(bp);
5951 uint64_t psize = BP_GET_PSIZE(bp);
5955 ASSERT(!BP_IS_HOLE(bp));
5956 ASSERT(!BP_IS_EMBEDDED(bp));
5959 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
5961 for (d = 0; d < ndvas; d++) {
5962 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
5964 atomic_add_64(&vd->vdev_pending_fastwrite, psize);
5967 spa_config_exit(spa, SCL_VDEV, FTAG);
5971 metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp)
5973 const dva_t *dva = bp->blk_dva;
5974 int ndvas = BP_GET_NDVAS(bp);
5975 uint64_t psize = BP_GET_PSIZE(bp);
5979 ASSERT(!BP_IS_HOLE(bp));
5980 ASSERT(!BP_IS_EMBEDDED(bp));
5983 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
5985 for (d = 0; d < ndvas; d++) {
5986 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
5988 ASSERT3U(vd->vdev_pending_fastwrite, >=, psize);
5989 atomic_sub_64(&vd->vdev_pending_fastwrite, psize);
5992 spa_config_exit(spa, SCL_VDEV, FTAG);
5996 metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset,
5997 uint64_t size, void *arg)
5999 (void) inner, (void) arg;
6001 if (vd->vdev_ops == &vdev_indirect_ops)
6004 metaslab_check_free_impl(vd, offset, size);
6008 metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size)
6011 spa_t *spa __maybe_unused = vd->vdev_spa;
6013 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
6016 if (vd->vdev_ops->vdev_op_remap != NULL) {
6017 vd->vdev_ops->vdev_op_remap(vd, offset, size,
6018 metaslab_check_free_impl_cb, NULL);
6022 ASSERT(vdev_is_concrete(vd));
6023 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
6024 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
6026 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
6028 mutex_enter(&msp->ms_lock);
6029 if (msp->ms_loaded) {
6030 range_tree_verify_not_present(msp->ms_allocatable,
6035 * Check all segments that currently exist in the freeing pipeline.
6037 * It would intuitively make sense to also check the current allocating
6038 * tree since metaslab_unalloc_dva() exists for extents that are
6039 * allocated and freed in the same sync pass within the same txg.
6040 * Unfortunately there are places (e.g. the ZIL) where we allocate a
6041 * segment but then we free part of it within the same txg
6042 * [see zil_sync()]. Thus, we don't call range_tree_verify() in the
6043 * current allocating tree.
6045 range_tree_verify_not_present(msp->ms_freeing, offset, size);
6046 range_tree_verify_not_present(msp->ms_checkpointing, offset, size);
6047 range_tree_verify_not_present(msp->ms_freed, offset, size);
6048 for (int j = 0; j < TXG_DEFER_SIZE; j++)
6049 range_tree_verify_not_present(msp->ms_defer[j], offset, size);
6050 range_tree_verify_not_present(msp->ms_trim, offset, size);
6051 mutex_exit(&msp->ms_lock);
6055 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
6057 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
6060 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
6061 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
6062 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
6063 vdev_t *vd = vdev_lookup_top(spa, vdev);
6064 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
6065 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
6067 if (DVA_GET_GANG(&bp->blk_dva[i]))
6068 size = vdev_gang_header_asize(vd);
6070 ASSERT3P(vd, !=, NULL);
6072 metaslab_check_free_impl(vd, offset, size);
6074 spa_config_exit(spa, SCL_VDEV, FTAG);
6078 metaslab_group_disable_wait(metaslab_group_t *mg)
6080 ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
6081 while (mg->mg_disabled_updating) {
6082 cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
6087 metaslab_group_disabled_increment(metaslab_group_t *mg)
6089 ASSERT(MUTEX_HELD(&mg->mg_ms_disabled_lock));
6090 ASSERT(mg->mg_disabled_updating);
6092 while (mg->mg_ms_disabled >= max_disabled_ms) {
6093 cv_wait(&mg->mg_ms_disabled_cv, &mg->mg_ms_disabled_lock);
6095 mg->mg_ms_disabled++;
6096 ASSERT3U(mg->mg_ms_disabled, <=, max_disabled_ms);
6100 * Mark the metaslab as disabled to prevent any allocations on this metaslab.
6101 * We must also track how many metaslabs are currently disabled within a
6102 * metaslab group and limit them to prevent allocation failures from
6103 * occurring because all metaslabs are disabled.
6106 metaslab_disable(metaslab_t *msp)
6108 ASSERT(!MUTEX_HELD(&msp->ms_lock));
6109 metaslab_group_t *mg = msp->ms_group;
6111 mutex_enter(&mg->mg_ms_disabled_lock);
6114 * To keep an accurate count of how many threads have disabled
6115 * a specific metaslab group, we only allow one thread to mark
6116 * the metaslab group at a time. This ensures that the value of
6117 * ms_disabled will be accurate when we decide to mark a metaslab
6118 * group as disabled. To do this we force all other threads
6119 * to wait till the metaslab's mg_disabled_updating flag is no
6122 metaslab_group_disable_wait(mg);
6123 mg->mg_disabled_updating = B_TRUE;
6124 if (msp->ms_disabled == 0) {
6125 metaslab_group_disabled_increment(mg);
6127 mutex_enter(&msp->ms_lock);
6129 mutex_exit(&msp->ms_lock);
6131 mg->mg_disabled_updating = B_FALSE;
6132 cv_broadcast(&mg->mg_ms_disabled_cv);
6133 mutex_exit(&mg->mg_ms_disabled_lock);
6137 metaslab_enable(metaslab_t *msp, boolean_t sync, boolean_t unload)
6139 metaslab_group_t *mg = msp->ms_group;
6140 spa_t *spa = mg->mg_vd->vdev_spa;
6143 * Wait for the outstanding IO to be synced to prevent newly
6144 * allocated blocks from being overwritten. This used by
6145 * initialize and TRIM which are modifying unallocated space.
6148 txg_wait_synced(spa_get_dsl(spa), 0);
6150 mutex_enter(&mg->mg_ms_disabled_lock);
6151 mutex_enter(&msp->ms_lock);
6152 if (--msp->ms_disabled == 0) {
6153 mg->mg_ms_disabled--;
6154 cv_broadcast(&mg->mg_ms_disabled_cv);
6156 metaslab_unload(msp);
6158 mutex_exit(&msp->ms_lock);
6159 mutex_exit(&mg->mg_ms_disabled_lock);
6163 metaslab_set_unflushed_dirty(metaslab_t *ms, boolean_t dirty)
6165 ms->ms_unflushed_dirty = dirty;
6169 metaslab_update_ondisk_flush_data(metaslab_t *ms, dmu_tx_t *tx)
6171 vdev_t *vd = ms->ms_group->mg_vd;
6172 spa_t *spa = vd->vdev_spa;
6173 objset_t *mos = spa_meta_objset(spa);
6175 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
6177 metaslab_unflushed_phys_t entry = {
6178 .msp_unflushed_txg = metaslab_unflushed_txg(ms),
6180 uint64_t entry_size = sizeof (entry);
6181 uint64_t entry_offset = ms->ms_id * entry_size;
6183 uint64_t object = 0;
6184 int err = zap_lookup(mos, vd->vdev_top_zap,
6185 VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1,
6187 if (err == ENOENT) {
6188 object = dmu_object_alloc(mos, DMU_OTN_UINT64_METADATA,
6189 SPA_OLD_MAXBLOCKSIZE, DMU_OT_NONE, 0, tx);
6190 VERIFY0(zap_add(mos, vd->vdev_top_zap,
6191 VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1,
6197 dmu_write(spa_meta_objset(spa), object, entry_offset, entry_size,
6202 metaslab_set_unflushed_txg(metaslab_t *ms, uint64_t txg, dmu_tx_t *tx)
6204 ms->ms_unflushed_txg = txg;
6205 metaslab_update_ondisk_flush_data(ms, tx);
6209 metaslab_unflushed_dirty(metaslab_t *ms)
6211 return (ms->ms_unflushed_dirty);
6215 metaslab_unflushed_txg(metaslab_t *ms)
6217 return (ms->ms_unflushed_txg);
6220 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, aliquot, U64, ZMOD_RW,
6221 "Allocation granularity (a.k.a. stripe size)");
6223 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, debug_load, INT, ZMOD_RW,
6224 "Load all metaslabs when pool is first opened");
6226 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, debug_unload, INT, ZMOD_RW,
6227 "Prevent metaslabs from being unloaded");
6229 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, preload_enabled, INT, ZMOD_RW,
6230 "Preload potential metaslabs during reassessment");
6232 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, unload_delay, UINT, ZMOD_RW,
6233 "Delay in txgs after metaslab was last used before unloading");
6235 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, unload_delay_ms, UINT, ZMOD_RW,
6236 "Delay in milliseconds after metaslab was last used before unloading");
6239 ZFS_MODULE_PARAM(zfs_mg, zfs_mg_, noalloc_threshold, UINT, ZMOD_RW,
6240 "Percentage of metaslab group size that should be free to make it "
6241 "eligible for allocation");
6243 ZFS_MODULE_PARAM(zfs_mg, zfs_mg_, fragmentation_threshold, UINT, ZMOD_RW,
6244 "Percentage of metaslab group size that should be considered eligible "
6245 "for allocations unless all metaslab groups within the metaslab class "
6246 "have also crossed this threshold");
6248 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, fragmentation_factor_enabled, INT,
6250 "Use the fragmentation metric to prefer less fragmented metaslabs");
6253 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, fragmentation_threshold, UINT,
6254 ZMOD_RW, "Fragmentation for metaslab to allow allocation");
6256 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, lba_weighting_enabled, INT, ZMOD_RW,
6257 "Prefer metaslabs with lower LBAs");
6259 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, bias_enabled, INT, ZMOD_RW,
6260 "Enable metaslab group biasing");
6262 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, segment_weight_enabled, INT,
6263 ZMOD_RW, "Enable segment-based metaslab selection");
6265 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, switch_threshold, INT, ZMOD_RW,
6266 "Segment-based metaslab selection maximum buckets before switching");
6268 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, force_ganging, U64, ZMOD_RW,
6269 "Blocks larger than this size are forced to be gang blocks");
6271 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, df_max_search, UINT, ZMOD_RW,
6272 "Max distance (bytes) to search forward before using size tree");
6274 ZFS_MODULE_PARAM(zfs_metaslab, metaslab_, df_use_largest_segment, INT, ZMOD_RW,
6275 "When looking in size tree, use largest segment instead of exact fit");
6277 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, max_size_cache_sec, U64,
6278 ZMOD_RW, "How long to trust the cached max chunk size of a metaslab");
6280 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, mem_limit, UINT, ZMOD_RW,
6281 "Percentage of memory that can be used to store metaslab range trees");
6283 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, try_hard_before_gang, INT,
6284 ZMOD_RW, "Try hard to allocate before ganging");
6286 ZFS_MODULE_PARAM(zfs_metaslab, zfs_metaslab_, find_max_tries, UINT, ZMOD_RW,
6287 "Normally only consider this many of the best metaslabs in each vdev");