4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
27 #include <sys/zfs_context.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/space_map.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/vdev_impl.h>
34 #include <sys/spa_impl.h>
35 #include <sys/zfeature.h>
37 #define WITH_DF_BLOCK_ALLOCATOR
39 #define GANG_ALLOCATION(flags) \
40 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
43 * Metaslab granularity, in bytes. This is roughly similar to what would be
44 * referred to as the "stripe size" in traditional RAID arrays. In normal
45 * operation, we will try to write this amount of data to a top-level vdev
46 * before moving on to the next one.
48 unsigned long metaslab_aliquot = 512 << 10;
50 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
53 * The in-core space map representation is more compact than its on-disk form.
54 * The zfs_condense_pct determines how much more compact the in-core
55 * space map representation must be before we compact it on-disk.
56 * Values should be greater than or equal to 100.
58 int zfs_condense_pct = 200;
61 * Condensing a metaslab is not guaranteed to actually reduce the amount of
62 * space used on disk. In particular, a space map uses data in increments of
63 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
64 * same number of blocks after condensing. Since the goal of condensing is to
65 * reduce the number of IOPs required to read the space map, we only want to
66 * condense when we can be sure we will reduce the number of blocks used by the
67 * space map. Unfortunately, we cannot precisely compute whether or not this is
68 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
69 * we apply the following heuristic: do not condense a spacemap unless the
70 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
73 int zfs_metaslab_condense_block_threshold = 4;
76 * The zfs_mg_noalloc_threshold defines which metaslab groups should
77 * be eligible for allocation. The value is defined as a percentage of
78 * free space. Metaslab groups that have more free space than
79 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
80 * a metaslab group's free space is less than or equal to the
81 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
82 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
83 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
84 * groups are allowed to accept allocations. Gang blocks are always
85 * eligible to allocate on any metaslab group. The default value of 0 means
86 * no metaslab group will be excluded based on this criterion.
88 int zfs_mg_noalloc_threshold = 0;
91 * Metaslab groups are considered eligible for allocations if their
92 * fragmenation metric (measured as a percentage) is less than or equal to
93 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
94 * then it will be skipped unless all metaslab groups within the metaslab
95 * class have also crossed this threshold.
97 int zfs_mg_fragmentation_threshold = 85;
100 * Allow metaslabs to keep their active state as long as their fragmentation
101 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
102 * active metaslab that exceeds this threshold will no longer keep its active
103 * status allowing better metaslabs to be selected.
105 int zfs_metaslab_fragmentation_threshold = 70;
108 * When set will load all metaslabs when pool is first opened.
110 int metaslab_debug_load = 0;
113 * When set will prevent metaslabs from being unloaded.
115 int metaslab_debug_unload = 0;
118 * Minimum size which forces the dynamic allocator to change
119 * it's allocation strategy. Once the space map cannot satisfy
120 * an allocation of this size then it switches to using more
121 * aggressive strategy (i.e search by size rather than offset).
123 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
126 * The minimum free space, in percent, which must be available
127 * in a space map to continue allocations in a first-fit fashion.
128 * Once the space map's free space drops below this level we dynamically
129 * switch to using best-fit allocations.
131 int metaslab_df_free_pct = 4;
134 * Percentage of all cpus that can be used by the metaslab taskq.
136 int metaslab_load_pct = 50;
139 * Determines how many txgs a metaslab may remain loaded without having any
140 * allocations from it. As long as a metaslab continues to be used we will
143 int metaslab_unload_delay = TXG_SIZE * 2;
146 * Max number of metaslabs per group to preload.
148 int metaslab_preload_limit = SPA_DVAS_PER_BP;
151 * Enable/disable preloading of metaslab.
153 int metaslab_preload_enabled = B_TRUE;
156 * Enable/disable fragmentation weighting on metaslabs.
158 int metaslab_fragmentation_factor_enabled = B_TRUE;
161 * Enable/disable lba weighting (i.e. outer tracks are given preference).
163 int metaslab_lba_weighting_enabled = B_TRUE;
166 * Enable/disable metaslab group biasing.
168 int metaslab_bias_enabled = B_TRUE;
172 * Enable/disable segment-based metaslab selection.
174 int zfs_metaslab_segment_weight_enabled = B_TRUE;
177 * When using segment-based metaslab selection, we will continue
178 * allocating from the active metaslab until we have exhausted
179 * zfs_metaslab_switch_threshold of its buckets.
181 int zfs_metaslab_switch_threshold = 2;
184 * Internal switch to enable/disable the metaslab allocation tracing
187 #ifdef _METASLAB_TRACING
188 boolean_t metaslab_trace_enabled = B_TRUE;
192 * Maximum entries that the metaslab allocation tracing facility will keep
193 * in a given list when running in non-debug mode. We limit the number
194 * of entries in non-debug mode to prevent us from using up too much memory.
195 * The limit should be sufficiently large that we don't expect any allocation
196 * to every exceed this value. In debug mode, the system will panic if this
197 * limit is ever reached allowing for further investigation.
199 #ifdef _METASLAB_TRACING
200 uint64_t metaslab_trace_max_entries = 5000;
203 static uint64_t metaslab_weight(metaslab_t *);
204 static void metaslab_set_fragmentation(metaslab_t *);
206 #ifdef _METASLAB_TRACING
207 kmem_cache_t *metaslab_alloc_trace_cache;
211 * ==========================================================================
213 * ==========================================================================
216 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
218 metaslab_class_t *mc;
220 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
225 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
226 refcount_create_tracked(&mc->mc_alloc_slots);
232 metaslab_class_destroy(metaslab_class_t *mc)
234 ASSERT(mc->mc_rotor == NULL);
235 ASSERT(mc->mc_alloc == 0);
236 ASSERT(mc->mc_deferred == 0);
237 ASSERT(mc->mc_space == 0);
238 ASSERT(mc->mc_dspace == 0);
240 refcount_destroy(&mc->mc_alloc_slots);
241 mutex_destroy(&mc->mc_lock);
242 kmem_free(mc, sizeof (metaslab_class_t));
246 metaslab_class_validate(metaslab_class_t *mc)
248 metaslab_group_t *mg;
252 * Must hold one of the spa_config locks.
254 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
255 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
257 if ((mg = mc->mc_rotor) == NULL)
262 ASSERT(vd->vdev_mg != NULL);
263 ASSERT3P(vd->vdev_top, ==, vd);
264 ASSERT3P(mg->mg_class, ==, mc);
265 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
266 } while ((mg = mg->mg_next) != mc->mc_rotor);
272 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
273 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
275 atomic_add_64(&mc->mc_alloc, alloc_delta);
276 atomic_add_64(&mc->mc_deferred, defer_delta);
277 atomic_add_64(&mc->mc_space, space_delta);
278 atomic_add_64(&mc->mc_dspace, dspace_delta);
282 metaslab_class_get_alloc(metaslab_class_t *mc)
284 return (mc->mc_alloc);
288 metaslab_class_get_deferred(metaslab_class_t *mc)
290 return (mc->mc_deferred);
294 metaslab_class_get_space(metaslab_class_t *mc)
296 return (mc->mc_space);
300 metaslab_class_get_dspace(metaslab_class_t *mc)
302 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
306 metaslab_class_histogram_verify(metaslab_class_t *mc)
308 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
312 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
315 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
318 for (c = 0; c < rvd->vdev_children; c++) {
319 vdev_t *tvd = rvd->vdev_child[c];
320 metaslab_group_t *mg = tvd->vdev_mg;
323 * Skip any holes, uninitialized top-levels, or
324 * vdevs that are not in this metalab class.
326 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
327 mg->mg_class != mc) {
331 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
332 mc_hist[i] += mg->mg_histogram[i];
335 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
336 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
338 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
342 * Calculate the metaslab class's fragmentation metric. The metric
343 * is weighted based on the space contribution of each metaslab group.
344 * The return value will be a number between 0 and 100 (inclusive), or
345 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
346 * zfs_frag_table for more information about the metric.
349 metaslab_class_fragmentation(metaslab_class_t *mc)
351 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
352 uint64_t fragmentation = 0;
355 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
357 for (c = 0; c < rvd->vdev_children; c++) {
358 vdev_t *tvd = rvd->vdev_child[c];
359 metaslab_group_t *mg = tvd->vdev_mg;
362 * Skip any holes, uninitialized top-levels, or
363 * vdevs that are not in this metalab class.
365 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
366 mg->mg_class != mc) {
371 * If a metaslab group does not contain a fragmentation
372 * metric then just bail out.
374 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
375 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
376 return (ZFS_FRAG_INVALID);
380 * Determine how much this metaslab_group is contributing
381 * to the overall pool fragmentation metric.
383 fragmentation += mg->mg_fragmentation *
384 metaslab_group_get_space(mg);
386 fragmentation /= metaslab_class_get_space(mc);
388 ASSERT3U(fragmentation, <=, 100);
389 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
390 return (fragmentation);
394 * Calculate the amount of expandable space that is available in
395 * this metaslab class. If a device is expanded then its expandable
396 * space will be the amount of allocatable space that is currently not
397 * part of this metaslab class.
400 metaslab_class_expandable_space(metaslab_class_t *mc)
402 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
406 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
407 for (c = 0; c < rvd->vdev_children; c++) {
408 vdev_t *tvd = rvd->vdev_child[c];
409 metaslab_group_t *mg = tvd->vdev_mg;
411 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
412 mg->mg_class != mc) {
417 * Calculate if we have enough space to add additional
418 * metaslabs. We report the expandable space in terms
419 * of the metaslab size since that's the unit of expansion.
421 space += P2ALIGN(tvd->vdev_max_asize - tvd->vdev_asize,
422 1ULL << tvd->vdev_ms_shift);
424 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
429 metaslab_compare(const void *x1, const void *x2)
431 const metaslab_t *m1 = (const metaslab_t *)x1;
432 const metaslab_t *m2 = (const metaslab_t *)x2;
434 int cmp = AVL_CMP(m2->ms_weight, m1->ms_weight);
438 IMPLY(AVL_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2);
440 return (AVL_CMP(m1->ms_start, m2->ms_start));
444 * Verify that the space accounting on disk matches the in-core range_trees.
447 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
449 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
450 uint64_t allocated = 0;
451 uint64_t sm_free_space, msp_free_space;
454 ASSERT(MUTEX_HELD(&msp->ms_lock));
456 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
460 * We can only verify the metaslab space when we're called
461 * from syncing context with a loaded metaslab that has an allocated
462 * space map. Calling this in non-syncing context does not
463 * provide a consistent view of the metaslab since we're performing
464 * allocations in the future.
466 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
470 sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) -
471 space_map_alloc_delta(msp->ms_sm);
474 * Account for future allocations since we would have already
475 * deducted that space from the ms_freetree.
477 for (t = 0; t < TXG_CONCURRENT_STATES; t++) {
479 range_tree_space(msp->ms_alloctree[(txg + t) & TXG_MASK]);
482 msp_free_space = range_tree_space(msp->ms_tree) + allocated +
483 msp->ms_deferspace + range_tree_space(msp->ms_freedtree);
485 VERIFY3U(sm_free_space, ==, msp_free_space);
489 * ==========================================================================
491 * ==========================================================================
494 * Update the allocatable flag and the metaslab group's capacity.
495 * The allocatable flag is set to true if the capacity is below
496 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
497 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
498 * transitions from allocatable to non-allocatable or vice versa then the
499 * metaslab group's class is updated to reflect the transition.
502 metaslab_group_alloc_update(metaslab_group_t *mg)
504 vdev_t *vd = mg->mg_vd;
505 metaslab_class_t *mc = mg->mg_class;
506 vdev_stat_t *vs = &vd->vdev_stat;
507 boolean_t was_allocatable;
508 boolean_t was_initialized;
510 ASSERT(vd == vd->vdev_top);
512 mutex_enter(&mg->mg_lock);
513 was_allocatable = mg->mg_allocatable;
514 was_initialized = mg->mg_initialized;
516 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
519 mutex_enter(&mc->mc_lock);
522 * If the metaslab group was just added then it won't
523 * have any space until we finish syncing out this txg.
524 * At that point we will consider it initialized and available
525 * for allocations. We also don't consider non-activated
526 * metaslab groups (e.g. vdevs that are in the middle of being removed)
527 * to be initialized, because they can't be used for allocation.
529 mg->mg_initialized = metaslab_group_initialized(mg);
530 if (!was_initialized && mg->mg_initialized) {
532 } else if (was_initialized && !mg->mg_initialized) {
533 ASSERT3U(mc->mc_groups, >, 0);
536 if (mg->mg_initialized)
537 mg->mg_no_free_space = B_FALSE;
540 * A metaslab group is considered allocatable if it has plenty
541 * of free space or is not heavily fragmented. We only take
542 * fragmentation into account if the metaslab group has a valid
543 * fragmentation metric (i.e. a value between 0 and 100).
545 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
546 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
547 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
548 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
551 * The mc_alloc_groups maintains a count of the number of
552 * groups in this metaslab class that are still above the
553 * zfs_mg_noalloc_threshold. This is used by the allocating
554 * threads to determine if they should avoid allocations to
555 * a given group. The allocator will avoid allocations to a group
556 * if that group has reached or is below the zfs_mg_noalloc_threshold
557 * and there are still other groups that are above the threshold.
558 * When a group transitions from allocatable to non-allocatable or
559 * vice versa we update the metaslab class to reflect that change.
560 * When the mc_alloc_groups value drops to 0 that means that all
561 * groups have reached the zfs_mg_noalloc_threshold making all groups
562 * eligible for allocations. This effectively means that all devices
563 * are balanced again.
565 if (was_allocatable && !mg->mg_allocatable)
566 mc->mc_alloc_groups--;
567 else if (!was_allocatable && mg->mg_allocatable)
568 mc->mc_alloc_groups++;
569 mutex_exit(&mc->mc_lock);
571 mutex_exit(&mg->mg_lock);
575 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
577 metaslab_group_t *mg;
579 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
580 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
581 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
582 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
585 mg->mg_activation_count = 0;
586 mg->mg_initialized = B_FALSE;
587 mg->mg_no_free_space = B_TRUE;
588 refcount_create_tracked(&mg->mg_alloc_queue_depth);
590 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
591 maxclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT | TASKQ_DYNAMIC);
597 metaslab_group_destroy(metaslab_group_t *mg)
599 ASSERT(mg->mg_prev == NULL);
600 ASSERT(mg->mg_next == NULL);
602 * We may have gone below zero with the activation count
603 * either because we never activated in the first place or
604 * because we're done, and possibly removing the vdev.
606 ASSERT(mg->mg_activation_count <= 0);
608 taskq_destroy(mg->mg_taskq);
609 avl_destroy(&mg->mg_metaslab_tree);
610 mutex_destroy(&mg->mg_lock);
611 refcount_destroy(&mg->mg_alloc_queue_depth);
612 kmem_free(mg, sizeof (metaslab_group_t));
616 metaslab_group_activate(metaslab_group_t *mg)
618 metaslab_class_t *mc = mg->mg_class;
619 metaslab_group_t *mgprev, *mgnext;
621 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
623 ASSERT(mc->mc_rotor != mg);
624 ASSERT(mg->mg_prev == NULL);
625 ASSERT(mg->mg_next == NULL);
626 ASSERT(mg->mg_activation_count <= 0);
628 if (++mg->mg_activation_count <= 0)
631 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
632 metaslab_group_alloc_update(mg);
634 if ((mgprev = mc->mc_rotor) == NULL) {
638 mgnext = mgprev->mg_next;
639 mg->mg_prev = mgprev;
640 mg->mg_next = mgnext;
641 mgprev->mg_next = mg;
642 mgnext->mg_prev = mg;
648 metaslab_group_passivate(metaslab_group_t *mg)
650 metaslab_class_t *mc = mg->mg_class;
651 metaslab_group_t *mgprev, *mgnext;
653 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
655 if (--mg->mg_activation_count != 0) {
656 ASSERT(mc->mc_rotor != mg);
657 ASSERT(mg->mg_prev == NULL);
658 ASSERT(mg->mg_next == NULL);
659 ASSERT(mg->mg_activation_count < 0);
663 taskq_wait_outstanding(mg->mg_taskq, 0);
664 metaslab_group_alloc_update(mg);
666 mgprev = mg->mg_prev;
667 mgnext = mg->mg_next;
672 mc->mc_rotor = mgnext;
673 mgprev->mg_next = mgnext;
674 mgnext->mg_prev = mgprev;
682 metaslab_group_initialized(metaslab_group_t *mg)
684 vdev_t *vd = mg->mg_vd;
685 vdev_stat_t *vs = &vd->vdev_stat;
687 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
691 metaslab_group_get_space(metaslab_group_t *mg)
693 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
697 metaslab_group_histogram_verify(metaslab_group_t *mg)
700 vdev_t *vd = mg->mg_vd;
701 uint64_t ashift = vd->vdev_ashift;
704 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
707 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
710 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
711 SPACE_MAP_HISTOGRAM_SIZE + ashift);
713 for (m = 0; m < vd->vdev_ms_count; m++) {
714 metaslab_t *msp = vd->vdev_ms[m];
716 if (msp->ms_sm == NULL)
719 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
720 mg_hist[i + ashift] +=
721 msp->ms_sm->sm_phys->smp_histogram[i];
724 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
725 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
727 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
731 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
733 metaslab_class_t *mc = mg->mg_class;
734 uint64_t ashift = mg->mg_vd->vdev_ashift;
737 ASSERT(MUTEX_HELD(&msp->ms_lock));
738 if (msp->ms_sm == NULL)
741 mutex_enter(&mg->mg_lock);
742 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
743 mg->mg_histogram[i + ashift] +=
744 msp->ms_sm->sm_phys->smp_histogram[i];
745 mc->mc_histogram[i + ashift] +=
746 msp->ms_sm->sm_phys->smp_histogram[i];
748 mutex_exit(&mg->mg_lock);
752 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
754 metaslab_class_t *mc = mg->mg_class;
755 uint64_t ashift = mg->mg_vd->vdev_ashift;
758 ASSERT(MUTEX_HELD(&msp->ms_lock));
759 if (msp->ms_sm == NULL)
762 mutex_enter(&mg->mg_lock);
763 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
764 ASSERT3U(mg->mg_histogram[i + ashift], >=,
765 msp->ms_sm->sm_phys->smp_histogram[i]);
766 ASSERT3U(mc->mc_histogram[i + ashift], >=,
767 msp->ms_sm->sm_phys->smp_histogram[i]);
769 mg->mg_histogram[i + ashift] -=
770 msp->ms_sm->sm_phys->smp_histogram[i];
771 mc->mc_histogram[i + ashift] -=
772 msp->ms_sm->sm_phys->smp_histogram[i];
774 mutex_exit(&mg->mg_lock);
778 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
780 ASSERT(msp->ms_group == NULL);
781 mutex_enter(&mg->mg_lock);
784 avl_add(&mg->mg_metaslab_tree, msp);
785 mutex_exit(&mg->mg_lock);
787 mutex_enter(&msp->ms_lock);
788 metaslab_group_histogram_add(mg, msp);
789 mutex_exit(&msp->ms_lock);
793 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
795 mutex_enter(&msp->ms_lock);
796 metaslab_group_histogram_remove(mg, msp);
797 mutex_exit(&msp->ms_lock);
799 mutex_enter(&mg->mg_lock);
800 ASSERT(msp->ms_group == mg);
801 avl_remove(&mg->mg_metaslab_tree, msp);
802 msp->ms_group = NULL;
803 mutex_exit(&mg->mg_lock);
807 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
810 * Although in principle the weight can be any value, in
811 * practice we do not use values in the range [1, 511].
813 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
814 ASSERT(MUTEX_HELD(&msp->ms_lock));
816 mutex_enter(&mg->mg_lock);
817 ASSERT(msp->ms_group == mg);
818 avl_remove(&mg->mg_metaslab_tree, msp);
819 msp->ms_weight = weight;
820 avl_add(&mg->mg_metaslab_tree, msp);
821 mutex_exit(&mg->mg_lock);
825 * Calculate the fragmentation for a given metaslab group. We can use
826 * a simple average here since all metaslabs within the group must have
827 * the same size. The return value will be a value between 0 and 100
828 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
829 * group have a fragmentation metric.
832 metaslab_group_fragmentation(metaslab_group_t *mg)
834 vdev_t *vd = mg->mg_vd;
835 uint64_t fragmentation = 0;
836 uint64_t valid_ms = 0;
839 for (m = 0; m < vd->vdev_ms_count; m++) {
840 metaslab_t *msp = vd->vdev_ms[m];
842 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
846 fragmentation += msp->ms_fragmentation;
849 if (valid_ms <= vd->vdev_ms_count / 2)
850 return (ZFS_FRAG_INVALID);
852 fragmentation /= valid_ms;
853 ASSERT3U(fragmentation, <=, 100);
854 return (fragmentation);
858 * Determine if a given metaslab group should skip allocations. A metaslab
859 * group should avoid allocations if its free capacity is less than the
860 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
861 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
862 * that can still handle allocations. If the allocation throttle is enabled
863 * then we skip allocations to devices that have reached their maximum
864 * allocation queue depth unless the selected metaslab group is the only
865 * eligible group remaining.
868 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
871 spa_t *spa = mg->mg_vd->vdev_spa;
872 metaslab_class_t *mc = mg->mg_class;
875 * We can only consider skipping this metaslab group if it's
876 * in the normal metaslab class and there are other metaslab
877 * groups to select from. Otherwise, we always consider it eligible
880 if (mc != spa_normal_class(spa) || mc->mc_groups <= 1)
884 * If the metaslab group's mg_allocatable flag is set (see comments
885 * in metaslab_group_alloc_update() for more information) and
886 * the allocation throttle is disabled then allow allocations to this
887 * device. However, if the allocation throttle is enabled then
888 * check if we have reached our allocation limit (mg_alloc_queue_depth)
889 * to determine if we should allow allocations to this metaslab group.
890 * If all metaslab groups are no longer considered allocatable
891 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
892 * gang block size then we allow allocations on this metaslab group
893 * regardless of the mg_allocatable or throttle settings.
895 if (mg->mg_allocatable) {
896 metaslab_group_t *mgp;
898 uint64_t qmax = mg->mg_max_alloc_queue_depth;
900 if (!mc->mc_alloc_throttle_enabled)
904 * If this metaslab group does not have any free space, then
905 * there is no point in looking further.
907 if (mg->mg_no_free_space)
910 qdepth = refcount_count(&mg->mg_alloc_queue_depth);
913 * If this metaslab group is below its qmax or it's
914 * the only allocatable metasable group, then attempt
915 * to allocate from it.
917 if (qdepth < qmax || mc->mc_alloc_groups == 1)
919 ASSERT3U(mc->mc_alloc_groups, >, 1);
922 * Since this metaslab group is at or over its qmax, we
923 * need to determine if there are metaslab groups after this
924 * one that might be able to handle this allocation. This is
925 * racy since we can't hold the locks for all metaslab
926 * groups at the same time when we make this check.
928 for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
929 qmax = mgp->mg_max_alloc_queue_depth;
931 qdepth = refcount_count(&mgp->mg_alloc_queue_depth);
934 * If there is another metaslab group that
935 * might be able to handle the allocation, then
936 * we return false so that we skip this group.
938 if (qdepth < qmax && !mgp->mg_no_free_space)
943 * We didn't find another group to handle the allocation
944 * so we can't skip this metaslab group even though
945 * we are at or over our qmax.
949 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
956 * ==========================================================================
957 * Range tree callbacks
958 * ==========================================================================
962 * Comparison function for the private size-ordered tree. Tree is sorted
963 * by size, larger sizes at the end of the tree.
966 metaslab_rangesize_compare(const void *x1, const void *x2)
968 const range_seg_t *r1 = x1;
969 const range_seg_t *r2 = x2;
970 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
971 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
973 int cmp = AVL_CMP(rs_size1, rs_size2);
977 return (AVL_CMP(r1->rs_start, r2->rs_start));
981 * Create any block allocator specific components. The current allocators
982 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
985 metaslab_rt_create(range_tree_t *rt, void *arg)
987 metaslab_t *msp = arg;
989 ASSERT3P(rt->rt_arg, ==, msp);
990 ASSERT(msp->ms_tree == NULL);
992 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
993 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
997 * Destroy the block allocator specific components.
1000 metaslab_rt_destroy(range_tree_t *rt, void *arg)
1002 metaslab_t *msp = arg;
1004 ASSERT3P(rt->rt_arg, ==, msp);
1005 ASSERT3P(msp->ms_tree, ==, rt);
1006 ASSERT0(avl_numnodes(&msp->ms_size_tree));
1008 avl_destroy(&msp->ms_size_tree);
1012 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1014 metaslab_t *msp = arg;
1016 ASSERT3P(rt->rt_arg, ==, msp);
1017 ASSERT3P(msp->ms_tree, ==, rt);
1018 VERIFY(!msp->ms_condensing);
1019 avl_add(&msp->ms_size_tree, rs);
1023 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1025 metaslab_t *msp = arg;
1027 ASSERT3P(rt->rt_arg, ==, msp);
1028 ASSERT3P(msp->ms_tree, ==, rt);
1029 VERIFY(!msp->ms_condensing);
1030 avl_remove(&msp->ms_size_tree, rs);
1034 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1036 metaslab_t *msp = arg;
1038 ASSERT3P(rt->rt_arg, ==, msp);
1039 ASSERT3P(msp->ms_tree, ==, rt);
1042 * Normally one would walk the tree freeing nodes along the way.
1043 * Since the nodes are shared with the range trees we can avoid
1044 * walking all nodes and just reinitialize the avl tree. The nodes
1045 * will be freed by the range tree, so we don't want to free them here.
1047 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
1048 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1051 static range_tree_ops_t metaslab_rt_ops = {
1053 metaslab_rt_destroy,
1060 * ==========================================================================
1061 * Common allocator routines
1062 * ==========================================================================
1066 * Return the maximum contiguous segment within the metaslab.
1069 metaslab_block_maxsize(metaslab_t *msp)
1071 avl_tree_t *t = &msp->ms_size_tree;
1074 if (t == NULL || (rs = avl_last(t)) == NULL)
1077 return (rs->rs_end - rs->rs_start);
1080 static range_seg_t *
1081 metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
1083 range_seg_t *rs, rsearch;
1086 rsearch.rs_start = start;
1087 rsearch.rs_end = start + size;
1089 rs = avl_find(t, &rsearch, &where);
1091 rs = avl_nearest(t, where, AVL_AFTER);
1097 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
1098 defined(WITH_DF_BLOCK_ALLOCATOR) || \
1099 defined(WITH_CF_BLOCK_ALLOCATOR)
1101 * This is a helper function that can be used by the allocator to find
1102 * a suitable block to allocate. This will search the specified AVL
1103 * tree looking for a block that matches the specified criteria.
1106 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1109 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1111 while (rs != NULL) {
1112 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1114 if (offset + size <= rs->rs_end) {
1115 *cursor = offset + size;
1118 rs = AVL_NEXT(t, rs);
1122 * If we know we've searched the whole map (*cursor == 0), give up.
1123 * Otherwise, reset the cursor to the beginning and try again.
1129 return (metaslab_block_picker(t, cursor, size, align));
1131 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
1133 #if defined(WITH_FF_BLOCK_ALLOCATOR)
1135 * ==========================================================================
1136 * The first-fit block allocator
1137 * ==========================================================================
1140 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1143 * Find the largest power of 2 block size that evenly divides the
1144 * requested size. This is used to try to allocate blocks with similar
1145 * alignment from the same area of the metaslab (i.e. same cursor
1146 * bucket) but it does not guarantee that other allocations sizes
1147 * may exist in the same region.
1149 uint64_t align = size & -size;
1150 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1151 avl_tree_t *t = &msp->ms_tree->rt_root;
1153 return (metaslab_block_picker(t, cursor, size, align));
1156 static metaslab_ops_t metaslab_ff_ops = {
1160 metaslab_ops_t *zfs_metaslab_ops = &metaslab_ff_ops;
1161 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1163 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1165 * ==========================================================================
1166 * Dynamic block allocator -
1167 * Uses the first fit allocation scheme until space get low and then
1168 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1169 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1170 * ==========================================================================
1173 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1176 * Find the largest power of 2 block size that evenly divides the
1177 * requested size. This is used to try to allocate blocks with similar
1178 * alignment from the same area of the metaslab (i.e. same cursor
1179 * bucket) but it does not guarantee that other allocations sizes
1180 * may exist in the same region.
1182 uint64_t align = size & -size;
1183 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1184 range_tree_t *rt = msp->ms_tree;
1185 avl_tree_t *t = &rt->rt_root;
1186 uint64_t max_size = metaslab_block_maxsize(msp);
1187 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1189 ASSERT(MUTEX_HELD(&msp->ms_lock));
1190 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1192 if (max_size < size)
1196 * If we're running low on space switch to using the size
1197 * sorted AVL tree (best-fit).
1199 if (max_size < metaslab_df_alloc_threshold ||
1200 free_pct < metaslab_df_free_pct) {
1201 t = &msp->ms_size_tree;
1205 return (metaslab_block_picker(t, cursor, size, 1ULL));
1208 static metaslab_ops_t metaslab_df_ops = {
1212 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1213 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1215 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1217 * ==========================================================================
1218 * Cursor fit block allocator -
1219 * Select the largest region in the metaslab, set the cursor to the beginning
1220 * of the range and the cursor_end to the end of the range. As allocations
1221 * are made advance the cursor. Continue allocating from the cursor until
1222 * the range is exhausted and then find a new range.
1223 * ==========================================================================
1226 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1228 range_tree_t *rt = msp->ms_tree;
1229 avl_tree_t *t = &msp->ms_size_tree;
1230 uint64_t *cursor = &msp->ms_lbas[0];
1231 uint64_t *cursor_end = &msp->ms_lbas[1];
1232 uint64_t offset = 0;
1234 ASSERT(MUTEX_HELD(&msp->ms_lock));
1235 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1237 ASSERT3U(*cursor_end, >=, *cursor);
1239 if ((*cursor + size) > *cursor_end) {
1242 rs = avl_last(&msp->ms_size_tree);
1243 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1246 *cursor = rs->rs_start;
1247 *cursor_end = rs->rs_end;
1256 static metaslab_ops_t metaslab_cf_ops = {
1260 metaslab_ops_t *zfs_metaslab_ops = &metaslab_cf_ops;
1261 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1263 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1265 * ==========================================================================
1266 * New dynamic fit allocator -
1267 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1268 * contiguous blocks. If no region is found then just use the largest segment
1270 * ==========================================================================
1274 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1275 * to request from the allocator.
1277 uint64_t metaslab_ndf_clump_shift = 4;
1280 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1282 avl_tree_t *t = &msp->ms_tree->rt_root;
1284 range_seg_t *rs, rsearch;
1285 uint64_t hbit = highbit64(size);
1286 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1287 uint64_t max_size = metaslab_block_maxsize(msp);
1289 ASSERT(MUTEX_HELD(&msp->ms_lock));
1290 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1292 if (max_size < size)
1295 rsearch.rs_start = *cursor;
1296 rsearch.rs_end = *cursor + size;
1298 rs = avl_find(t, &rsearch, &where);
1299 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1300 t = &msp->ms_size_tree;
1302 rsearch.rs_start = 0;
1303 rsearch.rs_end = MIN(max_size,
1304 1ULL << (hbit + metaslab_ndf_clump_shift));
1305 rs = avl_find(t, &rsearch, &where);
1307 rs = avl_nearest(t, where, AVL_AFTER);
1311 if ((rs->rs_end - rs->rs_start) >= size) {
1312 *cursor = rs->rs_start + size;
1313 return (rs->rs_start);
1318 static metaslab_ops_t metaslab_ndf_ops = {
1322 metaslab_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops;
1323 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1327 * ==========================================================================
1329 * ==========================================================================
1333 * Wait for any in-progress metaslab loads to complete.
1336 metaslab_load_wait(metaslab_t *msp)
1338 ASSERT(MUTEX_HELD(&msp->ms_lock));
1340 while (msp->ms_loading) {
1341 ASSERT(!msp->ms_loaded);
1342 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1347 metaslab_load(metaslab_t *msp)
1351 boolean_t success = B_FALSE;
1353 ASSERT(MUTEX_HELD(&msp->ms_lock));
1354 ASSERT(!msp->ms_loaded);
1355 ASSERT(!msp->ms_loading);
1357 msp->ms_loading = B_TRUE;
1360 * If the space map has not been allocated yet, then treat
1361 * all the space in the metaslab as free and add it to the
1364 if (msp->ms_sm != NULL)
1365 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1367 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1369 success = (error == 0);
1370 msp->ms_loading = B_FALSE;
1373 ASSERT3P(msp->ms_group, !=, NULL);
1374 msp->ms_loaded = B_TRUE;
1376 for (t = 0; t < TXG_DEFER_SIZE; t++) {
1377 range_tree_walk(msp->ms_defertree[t],
1378 range_tree_remove, msp->ms_tree);
1380 msp->ms_max_size = metaslab_block_maxsize(msp);
1382 cv_broadcast(&msp->ms_load_cv);
1387 metaslab_unload(metaslab_t *msp)
1389 ASSERT(MUTEX_HELD(&msp->ms_lock));
1390 range_tree_vacate(msp->ms_tree, NULL, NULL);
1391 msp->ms_loaded = B_FALSE;
1392 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1393 msp->ms_max_size = 0;
1397 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1400 vdev_t *vd = mg->mg_vd;
1401 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1405 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1406 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1407 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1409 ms->ms_start = id << vd->vdev_ms_shift;
1410 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1413 * We only open space map objects that already exist. All others
1414 * will be opened when we finally allocate an object for it.
1417 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1418 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1421 kmem_free(ms, sizeof (metaslab_t));
1425 ASSERT(ms->ms_sm != NULL);
1429 * We create the main range tree here, but we don't create the
1430 * other range trees until metaslab_sync_done(). This serves
1431 * two purposes: it allows metaslab_sync_done() to detect the
1432 * addition of new space; and for debugging, it ensures that we'd
1433 * data fault on any attempt to use this metaslab before it's ready.
1435 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1436 metaslab_group_add(mg, ms);
1438 metaslab_set_fragmentation(ms);
1441 * If we're opening an existing pool (txg == 0) or creating
1442 * a new one (txg == TXG_INITIAL), all space is available now.
1443 * If we're adding space to an existing pool, the new space
1444 * does not become available until after this txg has synced.
1445 * The metaslab's weight will also be initialized when we sync
1446 * out this txg. This ensures that we don't attempt to allocate
1447 * from it before we have initialized it completely.
1449 if (txg <= TXG_INITIAL)
1450 metaslab_sync_done(ms, 0);
1453 * If metaslab_debug_load is set and we're initializing a metaslab
1454 * that has an allocated space map object then load the its space
1455 * map so that can verify frees.
1457 if (metaslab_debug_load && ms->ms_sm != NULL) {
1458 mutex_enter(&ms->ms_lock);
1459 VERIFY0(metaslab_load(ms));
1460 mutex_exit(&ms->ms_lock);
1464 vdev_dirty(vd, 0, NULL, txg);
1465 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1474 metaslab_fini(metaslab_t *msp)
1478 metaslab_group_t *mg = msp->ms_group;
1480 metaslab_group_remove(mg, msp);
1482 mutex_enter(&msp->ms_lock);
1483 VERIFY(msp->ms_group == NULL);
1484 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1486 space_map_close(msp->ms_sm);
1488 metaslab_unload(msp);
1489 range_tree_destroy(msp->ms_tree);
1490 range_tree_destroy(msp->ms_freeingtree);
1491 range_tree_destroy(msp->ms_freedtree);
1493 for (t = 0; t < TXG_SIZE; t++) {
1494 range_tree_destroy(msp->ms_alloctree[t]);
1497 for (t = 0; t < TXG_DEFER_SIZE; t++) {
1498 range_tree_destroy(msp->ms_defertree[t]);
1501 ASSERT0(msp->ms_deferspace);
1503 mutex_exit(&msp->ms_lock);
1504 cv_destroy(&msp->ms_load_cv);
1505 mutex_destroy(&msp->ms_lock);
1507 kmem_free(msp, sizeof (metaslab_t));
1510 #define FRAGMENTATION_TABLE_SIZE 17
1513 * This table defines a segment size based fragmentation metric that will
1514 * allow each metaslab to derive its own fragmentation value. This is done
1515 * by calculating the space in each bucket of the spacemap histogram and
1516 * multiplying that by the fragmetation metric in this table. Doing
1517 * this for all buckets and dividing it by the total amount of free
1518 * space in this metaslab (i.e. the total free space in all buckets) gives
1519 * us the fragmentation metric. This means that a high fragmentation metric
1520 * equates to most of the free space being comprised of small segments.
1521 * Conversely, if the metric is low, then most of the free space is in
1522 * large segments. A 10% change in fragmentation equates to approximately
1523 * double the number of segments.
1525 * This table defines 0% fragmented space using 16MB segments. Testing has
1526 * shown that segments that are greater than or equal to 16MB do not suffer
1527 * from drastic performance problems. Using this value, we derive the rest
1528 * of the table. Since the fragmentation value is never stored on disk, it
1529 * is possible to change these calculations in the future.
1531 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1551 * Calclate the metaslab's fragmentation metric. A return value
1552 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1553 * not support this metric. Otherwise, the return value should be in the
1557 metaslab_set_fragmentation(metaslab_t *msp)
1559 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1560 uint64_t fragmentation = 0;
1562 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1563 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1566 if (!feature_enabled) {
1567 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1572 * A null space map means that the entire metaslab is free
1573 * and thus is not fragmented.
1575 if (msp->ms_sm == NULL) {
1576 msp->ms_fragmentation = 0;
1581 * If this metaslab's space map has not been upgraded, flag it
1582 * so that we upgrade next time we encounter it.
1584 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1585 vdev_t *vd = msp->ms_group->mg_vd;
1587 if (spa_writeable(vd->vdev_spa)) {
1588 uint64_t txg = spa_syncing_txg(spa);
1590 msp->ms_condense_wanted = B_TRUE;
1591 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1592 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1593 "msp %p, vd %p", txg, msp, vd);
1595 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1599 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1601 uint8_t shift = msp->ms_sm->sm_shift;
1603 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1604 FRAGMENTATION_TABLE_SIZE - 1);
1606 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1609 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1612 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1613 fragmentation += space * zfs_frag_table[idx];
1617 fragmentation /= total;
1618 ASSERT3U(fragmentation, <=, 100);
1620 msp->ms_fragmentation = fragmentation;
1624 * Compute a weight -- a selection preference value -- for the given metaslab.
1625 * This is based on the amount of free space, the level of fragmentation,
1626 * the LBA range, and whether the metaslab is loaded.
1629 metaslab_space_weight(metaslab_t *msp)
1631 metaslab_group_t *mg = msp->ms_group;
1632 vdev_t *vd = mg->mg_vd;
1633 uint64_t weight, space;
1635 ASSERT(MUTEX_HELD(&msp->ms_lock));
1636 ASSERT(!vd->vdev_removing);
1639 * The baseline weight is the metaslab's free space.
1641 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1643 if (metaslab_fragmentation_factor_enabled &&
1644 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1646 * Use the fragmentation information to inversely scale
1647 * down the baseline weight. We need to ensure that we
1648 * don't exclude this metaslab completely when it's 100%
1649 * fragmented. To avoid this we reduce the fragmented value
1652 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1655 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1656 * this metaslab again. The fragmentation metric may have
1657 * decreased the space to something smaller than
1658 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1659 * so that we can consume any remaining space.
1661 if (space > 0 && space < SPA_MINBLOCKSIZE)
1662 space = SPA_MINBLOCKSIZE;
1667 * Modern disks have uniform bit density and constant angular velocity.
1668 * Therefore, the outer recording zones are faster (higher bandwidth)
1669 * than the inner zones by the ratio of outer to inner track diameter,
1670 * which is typically around 2:1. We account for this by assigning
1671 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1672 * In effect, this means that we'll select the metaslab with the most
1673 * free bandwidth rather than simply the one with the most free space.
1675 if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
1676 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1677 ASSERT(weight >= space && weight <= 2 * space);
1681 * If this metaslab is one we're actively using, adjust its
1682 * weight to make it preferable to any inactive metaslab so
1683 * we'll polish it off. If the fragmentation on this metaslab
1684 * has exceed our threshold, then don't mark it active.
1686 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1687 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1688 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1691 WEIGHT_SET_SPACEBASED(weight);
1696 * Return the weight of the specified metaslab, according to the segment-based
1697 * weighting algorithm. The metaslab must be loaded. This function can
1698 * be called within a sync pass since it relies only on the metaslab's
1699 * range tree which is always accurate when the metaslab is loaded.
1702 metaslab_weight_from_range_tree(metaslab_t *msp)
1704 uint64_t weight = 0;
1705 uint32_t segments = 0;
1708 ASSERT(msp->ms_loaded);
1710 for (i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT; i--) {
1711 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
1712 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1715 segments += msp->ms_tree->rt_histogram[i];
1718 * The range tree provides more precision than the space map
1719 * and must be downgraded so that all values fit within the
1720 * space map's histogram. This allows us to compare loaded
1721 * vs. unloaded metaslabs to determine which metaslab is
1722 * considered "best".
1727 if (segments != 0) {
1728 WEIGHT_SET_COUNT(weight, segments);
1729 WEIGHT_SET_INDEX(weight, i);
1730 WEIGHT_SET_ACTIVE(weight, 0);
1738 * Calculate the weight based on the on-disk histogram. This should only
1739 * be called after a sync pass has completely finished since the on-disk
1740 * information is updated in metaslab_sync().
1743 metaslab_weight_from_spacemap(metaslab_t *msp)
1745 uint64_t weight = 0;
1748 for (i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
1749 if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) {
1750 WEIGHT_SET_COUNT(weight,
1751 msp->ms_sm->sm_phys->smp_histogram[i]);
1752 WEIGHT_SET_INDEX(weight, i +
1753 msp->ms_sm->sm_shift);
1754 WEIGHT_SET_ACTIVE(weight, 0);
1762 * Compute a segment-based weight for the specified metaslab. The weight
1763 * is determined by highest bucket in the histogram. The information
1764 * for the highest bucket is encoded into the weight value.
1767 metaslab_segment_weight(metaslab_t *msp)
1769 metaslab_group_t *mg = msp->ms_group;
1770 uint64_t weight = 0;
1771 uint8_t shift = mg->mg_vd->vdev_ashift;
1773 ASSERT(MUTEX_HELD(&msp->ms_lock));
1776 * The metaslab is completely free.
1778 if (space_map_allocated(msp->ms_sm) == 0) {
1779 int idx = highbit64(msp->ms_size) - 1;
1780 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1782 if (idx < max_idx) {
1783 WEIGHT_SET_COUNT(weight, 1ULL);
1784 WEIGHT_SET_INDEX(weight, idx);
1786 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
1787 WEIGHT_SET_INDEX(weight, max_idx);
1789 WEIGHT_SET_ACTIVE(weight, 0);
1790 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
1795 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
1798 * If the metaslab is fully allocated then just make the weight 0.
1800 if (space_map_allocated(msp->ms_sm) == msp->ms_size)
1803 * If the metaslab is already loaded, then use the range tree to
1804 * determine the weight. Otherwise, we rely on the space map information
1805 * to generate the weight.
1807 if (msp->ms_loaded) {
1808 weight = metaslab_weight_from_range_tree(msp);
1810 weight = metaslab_weight_from_spacemap(msp);
1814 * If the metaslab was active the last time we calculated its weight
1815 * then keep it active. We want to consume the entire region that
1816 * is associated with this weight.
1818 if (msp->ms_activation_weight != 0 && weight != 0)
1819 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
1824 * Determine if we should attempt to allocate from this metaslab. If the
1825 * metaslab has a maximum size then we can quickly determine if the desired
1826 * allocation size can be satisfied. Otherwise, if we're using segment-based
1827 * weighting then we can determine the maximum allocation that this metaslab
1828 * can accommodate based on the index encoded in the weight. If we're using
1829 * space-based weights then rely on the entire weight (excluding the weight
1833 metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
1835 boolean_t should_allocate;
1837 if (msp->ms_max_size != 0)
1838 return (msp->ms_max_size >= asize);
1840 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
1842 * The metaslab segment weight indicates segments in the
1843 * range [2^i, 2^(i+1)), where i is the index in the weight.
1844 * Since the asize might be in the middle of the range, we
1845 * should attempt the allocation if asize < 2^(i+1).
1847 should_allocate = (asize <
1848 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1850 should_allocate = (asize <=
1851 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
1853 return (should_allocate);
1856 metaslab_weight(metaslab_t *msp)
1858 vdev_t *vd = msp->ms_group->mg_vd;
1859 spa_t *spa = vd->vdev_spa;
1862 ASSERT(MUTEX_HELD(&msp->ms_lock));
1865 * This vdev is in the process of being removed so there is nothing
1866 * for us to do here.
1868 if (vd->vdev_removing) {
1869 ASSERT0(space_map_allocated(msp->ms_sm));
1870 ASSERT0(vd->vdev_ms_shift);
1874 metaslab_set_fragmentation(msp);
1877 * Update the maximum size if the metaslab is loaded. This will
1878 * ensure that we get an accurate maximum size if newly freed space
1879 * has been added back into the free tree.
1882 msp->ms_max_size = metaslab_block_maxsize(msp);
1885 * Segment-based weighting requires space map histogram support.
1887 if (zfs_metaslab_segment_weight_enabled &&
1888 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
1889 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
1890 sizeof (space_map_phys_t))) {
1891 weight = metaslab_segment_weight(msp);
1893 weight = metaslab_space_weight(msp);
1899 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1901 ASSERT(MUTEX_HELD(&msp->ms_lock));
1903 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1904 metaslab_load_wait(msp);
1905 if (!msp->ms_loaded) {
1906 int error = metaslab_load(msp);
1908 metaslab_group_sort(msp->ms_group, msp, 0);
1913 msp->ms_activation_weight = msp->ms_weight;
1914 metaslab_group_sort(msp->ms_group, msp,
1915 msp->ms_weight | activation_weight);
1917 ASSERT(msp->ms_loaded);
1918 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1924 metaslab_passivate(metaslab_t *msp, uint64_t weight)
1926 ASSERTV(uint64_t size = weight & ~METASLAB_WEIGHT_TYPE);
1929 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1930 * this metaslab again. In that case, it had better be empty,
1931 * or we would be leaving space on the table.
1933 ASSERT(size >= SPA_MINBLOCKSIZE ||
1934 range_tree_space(msp->ms_tree) == 0);
1935 ASSERT0(weight & METASLAB_ACTIVE_MASK);
1937 msp->ms_activation_weight = 0;
1938 metaslab_group_sort(msp->ms_group, msp, weight);
1939 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1943 * Segment-based metaslabs are activated once and remain active until
1944 * we either fail an allocation attempt (similar to space-based metaslabs)
1945 * or have exhausted the free space in zfs_metaslab_switch_threshold
1946 * buckets since the metaslab was activated. This function checks to see
1947 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
1948 * metaslab and passivates it proactively. This will allow us to select a
1949 * metaslab with a larger contiguous region, if any, remaining within this
1950 * metaslab group. If we're in sync pass > 1, then we continue using this
1951 * metaslab so that we don't dirty more block and cause more sync passes.
1954 metaslab_segment_may_passivate(metaslab_t *msp)
1956 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1958 int activation_idx, current_idx;
1960 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
1964 * Since we are in the middle of a sync pass, the most accurate
1965 * information that is accessible to us is the in-core range tree
1966 * histogram; calculate the new weight based on that information.
1968 weight = metaslab_weight_from_range_tree(msp);
1969 activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
1970 current_idx = WEIGHT_GET_INDEX(weight);
1972 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
1973 metaslab_passivate(msp, weight);
1977 metaslab_preload(void *arg)
1979 metaslab_t *msp = arg;
1980 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1981 fstrans_cookie_t cookie = spl_fstrans_mark();
1983 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1985 mutex_enter(&msp->ms_lock);
1986 metaslab_load_wait(msp);
1987 if (!msp->ms_loaded)
1988 (void) metaslab_load(msp);
1989 msp->ms_selected_txg = spa_syncing_txg(spa);
1990 mutex_exit(&msp->ms_lock);
1991 spl_fstrans_unmark(cookie);
1995 metaslab_group_preload(metaslab_group_t *mg)
1997 spa_t *spa = mg->mg_vd->vdev_spa;
1999 avl_tree_t *t = &mg->mg_metaslab_tree;
2002 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
2003 taskq_wait_outstanding(mg->mg_taskq, 0);
2007 mutex_enter(&mg->mg_lock);
2009 * Load the next potential metaslabs
2011 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2013 * We preload only the maximum number of metaslabs specified
2014 * by metaslab_preload_limit. If a metaslab is being forced
2015 * to condense then we preload it too. This will ensure
2016 * that force condensing happens in the next txg.
2018 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2022 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2023 msp, TQ_SLEEP) != TASKQID_INVALID);
2025 mutex_exit(&mg->mg_lock);
2029 * Determine if the space map's on-disk footprint is past our tolerance
2030 * for inefficiency. We would like to use the following criteria to make
2033 * 1. The size of the space map object should not dramatically increase as a
2034 * result of writing out the free space range tree.
2036 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2037 * times the size than the free space range tree representation
2038 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
2040 * 3. The on-disk size of the space map should actually decrease.
2042 * Checking the first condition is tricky since we don't want to walk
2043 * the entire AVL tree calculating the estimated on-disk size. Instead we
2044 * use the size-ordered range tree in the metaslab and calculate the
2045 * size required to write out the largest segment in our free tree. If the
2046 * size required to represent that segment on disk is larger than the space
2047 * map object then we avoid condensing this map.
2049 * To determine the second criterion we use a best-case estimate and assume
2050 * each segment can be represented on-disk as a single 64-bit entry. We refer
2051 * to this best-case estimate as the space map's minimal form.
2053 * Unfortunately, we cannot compute the on-disk size of the space map in this
2054 * context because we cannot accurately compute the effects of compression, etc.
2055 * Instead, we apply the heuristic described in the block comment for
2056 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2057 * is greater than a threshold number of blocks.
2060 metaslab_should_condense(metaslab_t *msp)
2062 space_map_t *sm = msp->ms_sm;
2064 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
2065 dmu_object_info_t doi;
2066 uint64_t vdev_blocksize = 1ULL << msp->ms_group->mg_vd->vdev_ashift;
2068 ASSERT(MUTEX_HELD(&msp->ms_lock));
2069 ASSERT(msp->ms_loaded);
2072 * Use the ms_size_tree range tree, which is ordered by size, to
2073 * obtain the largest segment in the free tree. We always condense
2074 * metaslabs that are empty and metaslabs for which a condense
2075 * request has been made.
2077 rs = avl_last(&msp->ms_size_tree);
2078 if (rs == NULL || msp->ms_condense_wanted)
2082 * Calculate the number of 64-bit entries this segment would
2083 * require when written to disk. If this single segment would be
2084 * larger on-disk than the entire current on-disk structure, then
2085 * clearly condensing will increase the on-disk structure size.
2087 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
2088 entries = size / (MIN(size, SM_RUN_MAX));
2089 segsz = entries * sizeof (uint64_t);
2091 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
2092 object_size = space_map_length(msp->ms_sm);
2094 dmu_object_info_from_db(sm->sm_dbuf, &doi);
2095 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
2097 return (segsz <= object_size &&
2098 object_size >= (optimal_size * zfs_condense_pct / 100) &&
2099 object_size > zfs_metaslab_condense_block_threshold * record_size);
2103 * Condense the on-disk space map representation to its minimized form.
2104 * The minimized form consists of a small number of allocations followed by
2105 * the entries of the free range tree.
2108 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
2110 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2111 range_tree_t *condense_tree;
2112 space_map_t *sm = msp->ms_sm;
2115 ASSERT(MUTEX_HELD(&msp->ms_lock));
2116 ASSERT3U(spa_sync_pass(spa), ==, 1);
2117 ASSERT(msp->ms_loaded);
2120 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2121 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2122 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2123 msp->ms_group->mg_vd->vdev_spa->spa_name,
2124 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
2125 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2127 msp->ms_condense_wanted = B_FALSE;
2130 * Create an range tree that is 100% allocated. We remove segments
2131 * that have been freed in this txg, any deferred frees that exist,
2132 * and any allocation in the future. Removing segments should be
2133 * a relatively inexpensive operation since we expect these trees to
2134 * have a small number of nodes.
2136 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
2137 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2140 * Remove what's been freed in this txg from the condense_tree.
2141 * Since we're in sync_pass 1, we know that all the frees from
2142 * this txg are in the freeingtree.
2144 range_tree_walk(msp->ms_freeingtree, range_tree_remove, condense_tree);
2146 for (t = 0; t < TXG_DEFER_SIZE; t++) {
2147 range_tree_walk(msp->ms_defertree[t],
2148 range_tree_remove, condense_tree);
2151 for (t = 1; t < TXG_CONCURRENT_STATES; t++) {
2152 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
2153 range_tree_remove, condense_tree);
2157 * We're about to drop the metaslab's lock thus allowing
2158 * other consumers to change it's content. Set the
2159 * metaslab's ms_condensing flag to ensure that
2160 * allocations on this metaslab do not occur while we're
2161 * in the middle of committing it to disk. This is only critical
2162 * for the ms_tree as all other range trees use per txg
2163 * views of their content.
2165 msp->ms_condensing = B_TRUE;
2167 mutex_exit(&msp->ms_lock);
2168 space_map_truncate(sm, tx);
2169 mutex_enter(&msp->ms_lock);
2172 * While we would ideally like to create a space map representation
2173 * that consists only of allocation records, doing so can be
2174 * prohibitively expensive because the in-core free tree can be
2175 * large, and therefore computationally expensive to subtract
2176 * from the condense_tree. Instead we sync out two trees, a cheap
2177 * allocation only tree followed by the in-core free tree. While not
2178 * optimal, this is typically close to optimal, and much cheaper to
2181 space_map_write(sm, condense_tree, SM_ALLOC, tx);
2182 range_tree_vacate(condense_tree, NULL, NULL);
2183 range_tree_destroy(condense_tree);
2185 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
2186 msp->ms_condensing = B_FALSE;
2190 * Write a metaslab to disk in the context of the specified transaction group.
2193 metaslab_sync(metaslab_t *msp, uint64_t txg)
2195 metaslab_group_t *mg = msp->ms_group;
2196 vdev_t *vd = mg->mg_vd;
2197 spa_t *spa = vd->vdev_spa;
2198 objset_t *mos = spa_meta_objset(spa);
2199 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
2201 uint64_t object = space_map_object(msp->ms_sm);
2203 ASSERT(!vd->vdev_ishole);
2206 * This metaslab has just been added so there's no work to do now.
2208 if (msp->ms_freeingtree == NULL) {
2209 ASSERT3P(alloctree, ==, NULL);
2213 ASSERT3P(alloctree, !=, NULL);
2214 ASSERT3P(msp->ms_freeingtree, !=, NULL);
2215 ASSERT3P(msp->ms_freedtree, !=, NULL);
2218 * Normally, we don't want to process a metaslab if there
2219 * are no allocations or frees to perform. However, if the metaslab
2220 * is being forced to condense we need to let it through.
2222 if (range_tree_space(alloctree) == 0 &&
2223 range_tree_space(msp->ms_freeingtree) == 0 &&
2224 !msp->ms_condense_wanted)
2228 * The only state that can actually be changing concurrently with
2229 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2230 * be modifying this txg's alloctree, freeingtree, freedtree, or
2231 * space_map_phys_t. Therefore, we only hold ms_lock to satify
2232 * space map ASSERTs. We drop it whenever we call into the DMU,
2233 * because the DMU can call down to us (e.g. via zio_free()) at
2237 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2239 if (msp->ms_sm == NULL) {
2240 uint64_t new_object;
2242 new_object = space_map_alloc(mos, tx);
2243 VERIFY3U(new_object, !=, 0);
2245 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2246 msp->ms_start, msp->ms_size, vd->vdev_ashift,
2248 ASSERT(msp->ms_sm != NULL);
2251 mutex_enter(&msp->ms_lock);
2254 * Note: metaslab_condense() clears the space map's histogram.
2255 * Therefore we must verify and remove this histogram before
2258 metaslab_group_histogram_verify(mg);
2259 metaslab_class_histogram_verify(mg->mg_class);
2260 metaslab_group_histogram_remove(mg, msp);
2262 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
2263 metaslab_should_condense(msp)) {
2264 metaslab_condense(msp, txg, tx);
2266 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
2267 space_map_write(msp->ms_sm, msp->ms_freeingtree, SM_FREE, tx);
2270 if (msp->ms_loaded) {
2274 * When the space map is loaded, we have an accruate
2275 * histogram in the range tree. This gives us an opportunity
2276 * to bring the space map's histogram up-to-date so we clear
2277 * it first before updating it.
2279 space_map_histogram_clear(msp->ms_sm);
2280 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
2283 * Since we've cleared the histogram we need to add back
2284 * any free space that has already been processed, plus
2285 * any deferred space. This allows the on-disk histogram
2286 * to accurately reflect all free space even if some space
2287 * is not yet available for allocation (i.e. deferred).
2289 space_map_histogram_add(msp->ms_sm, msp->ms_freedtree, tx);
2292 * Add back any deferred free space that has not been
2293 * added back into the in-core free tree yet. This will
2294 * ensure that we don't end up with a space map histogram
2295 * that is completely empty unless the metaslab is fully
2298 for (t = 0; t < TXG_DEFER_SIZE; t++) {
2299 space_map_histogram_add(msp->ms_sm,
2300 msp->ms_defertree[t], tx);
2305 * Always add the free space from this sync pass to the space
2306 * map histogram. We want to make sure that the on-disk histogram
2307 * accounts for all free space. If the space map is not loaded,
2308 * then we will lose some accuracy but will correct it the next
2309 * time we load the space map.
2311 space_map_histogram_add(msp->ms_sm, msp->ms_freeingtree, tx);
2313 metaslab_group_histogram_add(mg, msp);
2314 metaslab_group_histogram_verify(mg);
2315 metaslab_class_histogram_verify(mg->mg_class);
2318 * For sync pass 1, we avoid traversing this txg's free range tree
2319 * and instead will just swap the pointers for freeingtree and
2320 * freedtree. We can safely do this since the freed_tree is
2321 * guaranteed to be empty on the initial pass.
2323 if (spa_sync_pass(spa) == 1) {
2324 range_tree_swap(&msp->ms_freeingtree, &msp->ms_freedtree);
2326 range_tree_vacate(msp->ms_freeingtree,
2327 range_tree_add, msp->ms_freedtree);
2329 range_tree_vacate(alloctree, NULL, NULL);
2331 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2332 ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
2333 ASSERT0(range_tree_space(msp->ms_freeingtree));
2335 mutex_exit(&msp->ms_lock);
2337 if (object != space_map_object(msp->ms_sm)) {
2338 object = space_map_object(msp->ms_sm);
2339 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2340 msp->ms_id, sizeof (uint64_t), &object, tx);
2346 * Called after a transaction group has completely synced to mark
2347 * all of the metaslab's free space as usable.
2350 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2352 metaslab_group_t *mg = msp->ms_group;
2353 vdev_t *vd = mg->mg_vd;
2354 spa_t *spa = vd->vdev_spa;
2355 range_tree_t **defer_tree;
2356 int64_t alloc_delta, defer_delta;
2357 uint64_t free_space;
2358 boolean_t defer_allowed = B_TRUE;
2361 ASSERT(!vd->vdev_ishole);
2363 mutex_enter(&msp->ms_lock);
2366 * If this metaslab is just becoming available, initialize its
2367 * range trees and add its capacity to the vdev.
2369 if (msp->ms_freedtree == NULL) {
2370 for (t = 0; t < TXG_SIZE; t++) {
2371 ASSERT(msp->ms_alloctree[t] == NULL);
2373 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2377 ASSERT3P(msp->ms_freeingtree, ==, NULL);
2378 msp->ms_freeingtree = range_tree_create(NULL, msp,
2381 ASSERT3P(msp->ms_freedtree, ==, NULL);
2382 msp->ms_freedtree = range_tree_create(NULL, msp,
2385 for (t = 0; t < TXG_DEFER_SIZE; t++) {
2386 ASSERT(msp->ms_defertree[t] == NULL);
2388 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2392 vdev_space_update(vd, 0, 0, msp->ms_size);
2395 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2397 free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2398 metaslab_class_get_alloc(spa_normal_class(spa));
2399 if (free_space <= spa_get_slop_space(spa)) {
2400 defer_allowed = B_FALSE;
2404 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2405 if (defer_allowed) {
2406 defer_delta = range_tree_space(msp->ms_freedtree) -
2407 range_tree_space(*defer_tree);
2409 defer_delta -= range_tree_space(*defer_tree);
2412 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2415 * If there's a metaslab_load() in progress, wait for it to complete
2416 * so that we have a consistent view of the in-core space map.
2418 metaslab_load_wait(msp);
2421 * Move the frees from the defer_tree back to the free
2422 * range tree (if it's loaded). Swap the freed_tree and the
2423 * defer_tree -- this is safe to do because we've just emptied out
2426 range_tree_vacate(*defer_tree,
2427 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2428 if (defer_allowed) {
2429 range_tree_swap(&msp->ms_freedtree, defer_tree);
2431 range_tree_vacate(msp->ms_freedtree,
2432 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2435 space_map_update(msp->ms_sm);
2437 msp->ms_deferspace += defer_delta;
2438 ASSERT3S(msp->ms_deferspace, >=, 0);
2439 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2440 if (msp->ms_deferspace != 0) {
2442 * Keep syncing this metaslab until all deferred frees
2443 * are back in circulation.
2445 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2449 * Calculate the new weights before unloading any metaslabs.
2450 * This will give us the most accurate weighting.
2452 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2455 * If the metaslab is loaded and we've not tried to load or allocate
2456 * from it in 'metaslab_unload_delay' txgs, then unload it.
2458 if (msp->ms_loaded &&
2459 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2461 for (t = 1; t < TXG_CONCURRENT_STATES; t++) {
2462 VERIFY0(range_tree_space(
2463 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2466 if (!metaslab_debug_unload)
2467 metaslab_unload(msp);
2470 mutex_exit(&msp->ms_lock);
2474 metaslab_sync_reassess(metaslab_group_t *mg)
2476 metaslab_group_alloc_update(mg);
2477 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2480 * Preload the next potential metaslabs
2482 metaslab_group_preload(mg);
2486 metaslab_distance(metaslab_t *msp, dva_t *dva)
2488 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2489 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2490 uint64_t start = msp->ms_id;
2492 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2493 return (1ULL << 63);
2496 return ((start - offset) << ms_shift);
2498 return ((offset - start) << ms_shift);
2503 * ==========================================================================
2504 * Metaslab allocation tracing facility
2505 * ==========================================================================
2507 #ifdef _METASLAB_TRACING
2508 kstat_t *metaslab_trace_ksp;
2509 kstat_named_t metaslab_trace_over_limit;
2512 metaslab_alloc_trace_init(void)
2514 ASSERT(metaslab_alloc_trace_cache == NULL);
2515 metaslab_alloc_trace_cache = kmem_cache_create(
2516 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
2517 0, NULL, NULL, NULL, NULL, NULL, 0);
2518 metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
2519 "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
2520 if (metaslab_trace_ksp != NULL) {
2521 metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
2522 kstat_named_init(&metaslab_trace_over_limit,
2523 "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
2524 kstat_install(metaslab_trace_ksp);
2529 metaslab_alloc_trace_fini(void)
2531 if (metaslab_trace_ksp != NULL) {
2532 kstat_delete(metaslab_trace_ksp);
2533 metaslab_trace_ksp = NULL;
2535 kmem_cache_destroy(metaslab_alloc_trace_cache);
2536 metaslab_alloc_trace_cache = NULL;
2540 * Add an allocation trace element to the allocation tracing list.
2543 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
2544 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset)
2546 metaslab_alloc_trace_t *mat;
2548 if (!metaslab_trace_enabled)
2552 * When the tracing list reaches its maximum we remove
2553 * the second element in the list before adding a new one.
2554 * By removing the second element we preserve the original
2555 * entry as a clue to what allocations steps have already been
2558 if (zal->zal_size == metaslab_trace_max_entries) {
2559 metaslab_alloc_trace_t *mat_next;
2561 panic("too many entries in allocation list");
2563 atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
2565 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
2566 list_remove(&zal->zal_list, mat_next);
2567 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
2570 mat = kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
2571 list_link_init(&mat->mat_list_node);
2574 mat->mat_size = psize;
2575 mat->mat_dva_id = dva_id;
2576 mat->mat_offset = offset;
2577 mat->mat_weight = 0;
2580 mat->mat_weight = msp->ms_weight;
2583 * The list is part of the zio so locking is not required. Only
2584 * a single thread will perform allocations for a given zio.
2586 list_insert_tail(&zal->zal_list, mat);
2589 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
2593 metaslab_trace_init(zio_alloc_list_t *zal)
2595 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
2596 offsetof(metaslab_alloc_trace_t, mat_list_node));
2601 metaslab_trace_fini(zio_alloc_list_t *zal)
2603 metaslab_alloc_trace_t *mat;
2605 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
2606 kmem_cache_free(metaslab_alloc_trace_cache, mat);
2607 list_destroy(&zal->zal_list);
2612 #define metaslab_trace_add(zal, mg, msp, psize, id, off)
2615 metaslab_alloc_trace_init(void)
2620 metaslab_alloc_trace_fini(void)
2625 metaslab_trace_init(zio_alloc_list_t *zal)
2630 metaslab_trace_fini(zio_alloc_list_t *zal)
2634 #endif /* _METASLAB_TRACING */
2637 * ==========================================================================
2638 * Metaslab block operations
2639 * ==========================================================================
2643 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags)
2645 metaslab_group_t *mg;
2647 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2648 flags & METASLAB_DONT_THROTTLE)
2651 mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2652 if (!mg->mg_class->mc_alloc_throttle_enabled)
2655 (void) refcount_add(&mg->mg_alloc_queue_depth, tag);
2659 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags)
2661 metaslab_group_t *mg;
2663 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2664 flags & METASLAB_DONT_THROTTLE)
2667 mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2668 if (!mg->mg_class->mc_alloc_throttle_enabled)
2671 (void) refcount_remove(&mg->mg_alloc_queue_depth, tag);
2675 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag)
2678 const dva_t *dva = bp->blk_dva;
2679 int ndvas = BP_GET_NDVAS(bp);
2682 for (d = 0; d < ndvas; d++) {
2683 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
2684 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2685 VERIFY(refcount_not_held(&mg->mg_alloc_queue_depth, tag));
2691 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2694 range_tree_t *rt = msp->ms_tree;
2695 metaslab_class_t *mc = msp->ms_group->mg_class;
2697 VERIFY(!msp->ms_condensing);
2699 start = mc->mc_ops->msop_alloc(msp, size);
2700 if (start != -1ULL) {
2701 metaslab_group_t *mg = msp->ms_group;
2702 vdev_t *vd = mg->mg_vd;
2704 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2705 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2706 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2707 range_tree_remove(rt, start, size);
2709 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2710 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2712 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size);
2714 /* Track the last successful allocation */
2715 msp->ms_alloc_txg = txg;
2716 metaslab_verify_space(msp, txg);
2720 * Now that we've attempted the allocation we need to update the
2721 * metaslab's maximum block size since it may have changed.
2723 msp->ms_max_size = metaslab_block_maxsize(msp);
2728 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
2729 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2731 metaslab_t *msp = NULL;
2733 uint64_t offset = -1ULL;
2734 uint64_t activation_weight;
2735 uint64_t target_distance;
2738 activation_weight = METASLAB_WEIGHT_PRIMARY;
2739 for (i = 0; i < d; i++) {
2740 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2741 activation_weight = METASLAB_WEIGHT_SECONDARY;
2746 search = kmem_alloc(sizeof (*search), KM_SLEEP);
2747 search->ms_weight = UINT64_MAX;
2748 search->ms_start = 0;
2750 boolean_t was_active;
2751 avl_tree_t *t = &mg->mg_metaslab_tree;
2754 mutex_enter(&mg->mg_lock);
2757 * Find the metaslab with the highest weight that is less
2758 * than what we've already tried. In the common case, this
2759 * means that we will examine each metaslab at most once.
2760 * Note that concurrent callers could reorder metaslabs
2761 * by activation/passivation once we have dropped the mg_lock.
2762 * If a metaslab is activated by another thread, and we fail
2763 * to allocate from the metaslab we have selected, we may
2764 * not try the newly-activated metaslab, and instead activate
2765 * another metaslab. This is not optimal, but generally
2766 * does not cause any problems (a possible exception being
2767 * if every metaslab is completely full except for the
2768 * the newly-activated metaslab which we fail to examine).
2770 msp = avl_find(t, search, &idx);
2772 msp = avl_nearest(t, idx, AVL_AFTER);
2773 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
2775 if (!metaslab_should_allocate(msp, asize)) {
2776 metaslab_trace_add(zal, mg, msp, asize, d,
2782 * If the selected metaslab is condensing, skip it.
2784 if (msp->ms_condensing)
2787 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2788 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2791 target_distance = min_distance +
2792 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2795 for (i = 0; i < d; i++) {
2796 if (metaslab_distance(msp, &dva[i]) <
2803 mutex_exit(&mg->mg_lock);
2805 kmem_free(search, sizeof (*search));
2808 search->ms_weight = msp->ms_weight;
2809 search->ms_start = msp->ms_start + 1;
2811 mutex_enter(&msp->ms_lock);
2814 * Ensure that the metaslab we have selected is still
2815 * capable of handling our request. It's possible that
2816 * another thread may have changed the weight while we
2817 * were blocked on the metaslab lock. We check the
2818 * active status first to see if we need to reselect
2821 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
2822 mutex_exit(&msp->ms_lock);
2826 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2827 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2828 metaslab_passivate(msp,
2829 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2830 mutex_exit(&msp->ms_lock);
2834 if (metaslab_activate(msp, activation_weight) != 0) {
2835 mutex_exit(&msp->ms_lock);
2838 msp->ms_selected_txg = txg;
2841 * Now that we have the lock, recheck to see if we should
2842 * continue to use this metaslab for this allocation. The
2843 * the metaslab is now loaded so metaslab_should_allocate() can
2844 * accurately determine if the allocation attempt should
2847 if (!metaslab_should_allocate(msp, asize)) {
2848 /* Passivate this metaslab and select a new one. */
2849 metaslab_trace_add(zal, mg, msp, asize, d,
2856 * If this metaslab is currently condensing then pick again as
2857 * we can't manipulate this metaslab until it's committed
2860 if (msp->ms_condensing) {
2861 metaslab_trace_add(zal, mg, msp, asize, d,
2863 mutex_exit(&msp->ms_lock);
2867 offset = metaslab_block_alloc(msp, asize, txg);
2868 metaslab_trace_add(zal, mg, msp, asize, d, offset);
2870 if (offset != -1ULL) {
2871 /* Proactively passivate the metaslab, if needed */
2872 metaslab_segment_may_passivate(msp);
2876 ASSERT(msp->ms_loaded);
2879 * We were unable to allocate from this metaslab so determine
2880 * a new weight for this metaslab. Now that we have loaded
2881 * the metaslab we can provide a better hint to the metaslab
2884 * For space-based metaslabs, we use the maximum block size.
2885 * This information is only available when the metaslab
2886 * is loaded and is more accurate than the generic free
2887 * space weight that was calculated by metaslab_weight().
2888 * This information allows us to quickly compare the maximum
2889 * available allocation in the metaslab to the allocation
2890 * size being requested.
2892 * For segment-based metaslabs, determine the new weight
2893 * based on the highest bucket in the range tree. We
2894 * explicitly use the loaded segment weight (i.e. the range
2895 * tree histogram) since it contains the space that is
2896 * currently available for allocation and is accurate
2897 * even within a sync pass.
2899 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
2900 uint64_t weight = metaslab_block_maxsize(msp);
2901 WEIGHT_SET_SPACEBASED(weight);
2902 metaslab_passivate(msp, weight);
2904 metaslab_passivate(msp,
2905 metaslab_weight_from_range_tree(msp));
2909 * We have just failed an allocation attempt, check
2910 * that metaslab_should_allocate() agrees. Otherwise,
2911 * we may end up in an infinite loop retrying the same
2914 ASSERT(!metaslab_should_allocate(msp, asize));
2915 mutex_exit(&msp->ms_lock);
2917 mutex_exit(&msp->ms_lock);
2918 kmem_free(search, sizeof (*search));
2923 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
2924 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2927 ASSERT(mg->mg_initialized);
2929 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
2930 min_distance, dva, d);
2932 mutex_enter(&mg->mg_lock);
2933 if (offset == -1ULL) {
2934 mg->mg_failed_allocations++;
2935 metaslab_trace_add(zal, mg, NULL, asize, d,
2936 TRACE_GROUP_FAILURE);
2937 if (asize == SPA_GANGBLOCKSIZE) {
2939 * This metaslab group was unable to allocate
2940 * the minimum gang block size so it must be out of
2941 * space. We must notify the allocation throttle
2942 * to start skipping allocation attempts to this
2943 * metaslab group until more space becomes available.
2944 * Note: this failure cannot be caused by the
2945 * allocation throttle since the allocation throttle
2946 * is only responsible for skipping devices and
2947 * not failing block allocations.
2949 mg->mg_no_free_space = B_TRUE;
2952 mg->mg_allocations++;
2953 mutex_exit(&mg->mg_lock);
2958 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
2959 * on the same vdev as an existing DVA of this BP, then try to allocate it
2960 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
2963 int ditto_same_vdev_distance_shift = 3;
2966 * Allocate a block for the specified i/o.
2969 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2970 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
2971 zio_alloc_list_t *zal)
2973 metaslab_group_t *mg, *fast_mg, *rotor;
2975 boolean_t try_hard = B_FALSE;
2977 ASSERT(!DVA_IS_VALID(&dva[d]));
2980 * For testing, make some blocks above a certain size be gang blocks.
2982 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) {
2983 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG);
2984 return (SET_ERROR(ENOSPC));
2988 * Start at the rotor and loop through all mgs until we find something.
2989 * Note that there's no locking on mc_rotor or mc_aliquot because
2990 * nothing actually breaks if we miss a few updates -- we just won't
2991 * allocate quite as evenly. It all balances out over time.
2993 * If we are doing ditto or log blocks, try to spread them across
2994 * consecutive vdevs. If we're forced to reuse a vdev before we've
2995 * allocated all of our ditto blocks, then try and spread them out on
2996 * that vdev as much as possible. If it turns out to not be possible,
2997 * gradually lower our standards until anything becomes acceptable.
2998 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2999 * gives us hope of containing our fault domains to something we're
3000 * able to reason about. Otherwise, any two top-level vdev failures
3001 * will guarantee the loss of data. With consecutive allocation,
3002 * only two adjacent top-level vdev failures will result in data loss.
3004 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3005 * ourselves on the same vdev as our gang block header. That
3006 * way, we can hope for locality in vdev_cache, plus it makes our
3007 * fault domains something tractable.
3010 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3013 * It's possible the vdev we're using as the hint no
3014 * longer exists (i.e. removed). Consult the rotor when
3020 if (flags & METASLAB_HINTBP_AVOID &&
3021 mg->mg_next != NULL)
3026 } else if (d != 0) {
3027 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3028 mg = vd->vdev_mg->mg_next;
3029 } else if (flags & METASLAB_FASTWRITE) {
3030 mg = fast_mg = mc->mc_rotor;
3033 if (fast_mg->mg_vd->vdev_pending_fastwrite <
3034 mg->mg_vd->vdev_pending_fastwrite)
3036 } while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor);
3043 * If the hint put us into the wrong metaslab class, or into a
3044 * metaslab group that has been passivated, just follow the rotor.
3046 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3052 boolean_t allocatable;
3054 uint64_t distance, asize;
3056 ASSERT(mg->mg_activation_count == 1);
3060 * Don't allocate from faulted devices.
3063 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
3064 allocatable = vdev_allocatable(vd);
3065 spa_config_exit(spa, SCL_ZIO, FTAG);
3067 allocatable = vdev_allocatable(vd);
3071 * Determine if the selected metaslab group is eligible
3072 * for allocations. If we're ganging then don't allow
3073 * this metaslab group to skip allocations since that would
3074 * inadvertently return ENOSPC and suspend the pool
3075 * even though space is still available.
3077 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
3078 allocatable = metaslab_group_allocatable(mg, rotor,
3083 metaslab_trace_add(zal, mg, NULL, psize, d,
3084 TRACE_NOT_ALLOCATABLE);
3088 ASSERT(mg->mg_initialized);
3091 * Avoid writing single-copy data to a failing,
3092 * non-redundant vdev, unless we've already tried all
3095 if ((vd->vdev_stat.vs_write_errors > 0 ||
3096 vd->vdev_state < VDEV_STATE_HEALTHY) &&
3097 d == 0 && !try_hard && vd->vdev_children == 0) {
3098 metaslab_trace_add(zal, mg, NULL, psize, d,
3103 ASSERT(mg->mg_class == mc);
3106 * If we don't need to try hard, then require that the
3107 * block be 1/8th of the device away from any other DVAs
3108 * in this BP. If we are trying hard, allow any offset
3109 * to be used (distance=0).
3113 distance = vd->vdev_asize >>
3114 ditto_same_vdev_distance_shift;
3115 if (distance <= (1ULL << vd->vdev_ms_shift))
3119 asize = vdev_psize_to_asize(vd, psize);
3120 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3122 offset = metaslab_group_alloc(mg, zal, asize, txg, distance,
3125 if (offset != -1ULL) {
3127 * If we've just selected this metaslab group,
3128 * figure out whether the corresponding vdev is
3129 * over- or under-used relative to the pool,
3130 * and set an allocation bias to even it out.
3132 * Bias is also used to compensate for unequally
3133 * sized vdevs so that space is allocated fairly.
3135 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3136 vdev_stat_t *vs = &vd->vdev_stat;
3137 int64_t vs_free = vs->vs_space - vs->vs_alloc;
3138 int64_t mc_free = mc->mc_space - mc->mc_alloc;
3142 * Calculate how much more or less we should
3143 * try to allocate from this device during
3144 * this iteration around the rotor.
3146 * This basically introduces a zero-centered
3147 * bias towards the devices with the most
3148 * free space, while compensating for vdev
3152 * vdev V1 = 16M/128M
3153 * vdev V2 = 16M/128M
3154 * ratio(V1) = 100% ratio(V2) = 100%
3156 * vdev V1 = 16M/128M
3157 * vdev V2 = 64M/128M
3158 * ratio(V1) = 127% ratio(V2) = 72%
3160 * vdev V1 = 16M/128M
3161 * vdev V2 = 64M/512M
3162 * ratio(V1) = 40% ratio(V2) = 160%
3164 ratio = (vs_free * mc->mc_alloc_groups * 100) /
3166 mg->mg_bias = ((ratio - 100) *
3167 (int64_t)mg->mg_aliquot) / 100;
3168 } else if (!metaslab_bias_enabled) {
3172 if ((flags & METASLAB_FASTWRITE) ||
3173 atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3174 mg->mg_aliquot + mg->mg_bias) {
3175 mc->mc_rotor = mg->mg_next;
3179 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3180 DVA_SET_OFFSET(&dva[d], offset);
3181 DVA_SET_GANG(&dva[d],
3182 ((flags & METASLAB_GANG_HEADER) ? 1 : 0));
3183 DVA_SET_ASIZE(&dva[d], asize);
3185 if (flags & METASLAB_FASTWRITE) {
3186 atomic_add_64(&vd->vdev_pending_fastwrite,
3193 mc->mc_rotor = mg->mg_next;
3195 } while ((mg = mg->mg_next) != rotor);
3198 * If we haven't tried hard, do so now.
3205 bzero(&dva[d], sizeof (dva_t));
3207 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC);
3208 return (SET_ERROR(ENOSPC));
3212 * Free the block represented by DVA in the context of the specified
3213 * transaction group.
3216 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
3218 uint64_t vdev = DVA_GET_VDEV(dva);
3219 uint64_t offset = DVA_GET_OFFSET(dva);
3220 uint64_t size = DVA_GET_ASIZE(dva);
3224 if (txg > spa_freeze_txg(spa))
3227 if ((vd = vdev_lookup_top(spa, vdev)) == NULL || !DVA_IS_VALID(dva) ||
3228 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3229 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
3230 (u_longlong_t)vdev, (u_longlong_t)offset,
3231 (u_longlong_t)size);
3235 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3237 if (DVA_GET_GANG(dva))
3238 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3240 mutex_enter(&msp->ms_lock);
3243 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
3246 VERIFY(!msp->ms_condensing);
3247 VERIFY3U(offset, >=, msp->ms_start);
3248 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3249 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
3251 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3252 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3253 range_tree_add(msp->ms_tree, offset, size);
3254 msp->ms_max_size = metaslab_block_maxsize(msp);
3256 VERIFY3U(txg, ==, spa->spa_syncing_txg);
3257 if (range_tree_space(msp->ms_freeingtree) == 0)
3258 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3259 range_tree_add(msp->ms_freeingtree, offset, size);
3262 mutex_exit(&msp->ms_lock);
3266 * Intent log support: upon opening the pool after a crash, notify the SPA
3267 * of blocks that the intent log has allocated for immediate write, but
3268 * which are still considered free by the SPA because the last transaction
3269 * group didn't commit yet.
3272 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3274 uint64_t vdev = DVA_GET_VDEV(dva);
3275 uint64_t offset = DVA_GET_OFFSET(dva);
3276 uint64_t size = DVA_GET_ASIZE(dva);
3281 ASSERT(DVA_IS_VALID(dva));
3283 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3284 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
3285 return (SET_ERROR(ENXIO));
3287 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3289 if (DVA_GET_GANG(dva))
3290 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3292 mutex_enter(&msp->ms_lock);
3294 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3295 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
3297 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
3298 error = SET_ERROR(ENOENT);
3300 if (error || txg == 0) { /* txg == 0 indicates dry run */
3301 mutex_exit(&msp->ms_lock);
3305 VERIFY(!msp->ms_condensing);
3306 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3307 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3308 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
3309 range_tree_remove(msp->ms_tree, offset, size);
3311 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
3312 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
3313 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3314 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
3317 mutex_exit(&msp->ms_lock);
3323 * Reserve some allocation slots. The reservation system must be called
3324 * before we call into the allocator. If there aren't any available slots
3325 * then the I/O will be throttled until an I/O completes and its slots are
3326 * freed up. The function returns true if it was successful in placing
3330 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio,
3333 uint64_t available_slots = 0;
3334 uint64_t reserved_slots;
3335 boolean_t slot_reserved = B_FALSE;
3337 ASSERT(mc->mc_alloc_throttle_enabled);
3338 mutex_enter(&mc->mc_lock);
3340 reserved_slots = refcount_count(&mc->mc_alloc_slots);
3341 if (reserved_slots < mc->mc_alloc_max_slots)
3342 available_slots = mc->mc_alloc_max_slots - reserved_slots;
3344 if (slots <= available_slots || GANG_ALLOCATION(flags)) {
3348 * We reserve the slots individually so that we can unreserve
3349 * them individually when an I/O completes.
3351 for (d = 0; d < slots; d++) {
3352 reserved_slots = refcount_add(&mc->mc_alloc_slots, zio);
3354 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3355 slot_reserved = B_TRUE;
3358 mutex_exit(&mc->mc_lock);
3359 return (slot_reserved);
3363 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio)
3367 ASSERT(mc->mc_alloc_throttle_enabled);
3368 mutex_enter(&mc->mc_lock);
3369 for (d = 0; d < slots; d++) {
3370 (void) refcount_remove(&mc->mc_alloc_slots, zio);
3372 mutex_exit(&mc->mc_lock);
3376 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
3377 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
3378 zio_alloc_list_t *zal, zio_t *zio)
3380 dva_t *dva = bp->blk_dva;
3381 dva_t *hintdva = hintbp->blk_dva;
3384 ASSERT(bp->blk_birth == 0);
3385 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
3387 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3389 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
3390 spa_config_exit(spa, SCL_ALLOC, FTAG);
3391 return (SET_ERROR(ENOSPC));
3394 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
3395 ASSERT(BP_GET_NDVAS(bp) == 0);
3396 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
3397 ASSERT3P(zal, !=, NULL);
3399 for (d = 0; d < ndvas; d++) {
3400 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
3403 for (d--; d >= 0; d--) {
3404 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
3405 metaslab_group_alloc_decrement(spa,
3406 DVA_GET_VDEV(&dva[d]), zio, flags);
3407 bzero(&dva[d], sizeof (dva_t));
3409 spa_config_exit(spa, SCL_ALLOC, FTAG);
3413 * Update the metaslab group's queue depth
3414 * based on the newly allocated dva.
3416 metaslab_group_alloc_increment(spa,
3417 DVA_GET_VDEV(&dva[d]), zio, flags);
3422 ASSERT(BP_GET_NDVAS(bp) == ndvas);
3424 spa_config_exit(spa, SCL_ALLOC, FTAG);
3426 BP_SET_BIRTH(bp, txg, 0);
3432 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
3434 const dva_t *dva = bp->blk_dva;
3435 int d, ndvas = BP_GET_NDVAS(bp);
3437 ASSERT(!BP_IS_HOLE(bp));
3438 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
3440 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
3442 for (d = 0; d < ndvas; d++)
3443 metaslab_free_dva(spa, &dva[d], txg, now);
3445 spa_config_exit(spa, SCL_FREE, FTAG);
3449 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
3451 const dva_t *dva = bp->blk_dva;
3452 int ndvas = BP_GET_NDVAS(bp);
3455 ASSERT(!BP_IS_HOLE(bp));
3459 * First do a dry run to make sure all DVAs are claimable,
3460 * so we don't have to unwind from partial failures below.
3462 if ((error = metaslab_claim(spa, bp, 0)) != 0)
3466 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3468 for (d = 0; d < ndvas; d++)
3469 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
3472 spa_config_exit(spa, SCL_ALLOC, FTAG);
3474 ASSERT(error == 0 || txg == 0);
3480 metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp)
3482 const dva_t *dva = bp->blk_dva;
3483 int ndvas = BP_GET_NDVAS(bp);
3484 uint64_t psize = BP_GET_PSIZE(bp);
3488 ASSERT(!BP_IS_HOLE(bp));
3489 ASSERT(!BP_IS_EMBEDDED(bp));
3492 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3494 for (d = 0; d < ndvas; d++) {
3495 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
3497 atomic_add_64(&vd->vdev_pending_fastwrite, psize);
3500 spa_config_exit(spa, SCL_VDEV, FTAG);
3504 metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp)
3506 const dva_t *dva = bp->blk_dva;
3507 int ndvas = BP_GET_NDVAS(bp);
3508 uint64_t psize = BP_GET_PSIZE(bp);
3512 ASSERT(!BP_IS_HOLE(bp));
3513 ASSERT(!BP_IS_EMBEDDED(bp));
3516 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3518 for (d = 0; d < ndvas; d++) {
3519 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
3521 ASSERT3U(vd->vdev_pending_fastwrite, >=, psize);
3522 atomic_sub_64(&vd->vdev_pending_fastwrite, psize);
3525 spa_config_exit(spa, SCL_VDEV, FTAG);
3529 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
3533 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3536 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3537 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3538 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
3539 vdev_t *vd = vdev_lookup_top(spa, vdev);
3540 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
3541 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
3542 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3545 range_tree_verify(msp->ms_tree, offset, size);
3547 range_tree_verify(msp->ms_freeingtree, offset, size);
3548 range_tree_verify(msp->ms_freedtree, offset, size);
3549 for (j = 0; j < TXG_DEFER_SIZE; j++)
3550 range_tree_verify(msp->ms_defertree[j], offset, size);
3552 spa_config_exit(spa, SCL_VDEV, FTAG);
3555 #if defined(_KERNEL) && defined(HAVE_SPL)
3557 module_param(metaslab_aliquot, ulong, 0644);
3558 MODULE_PARM_DESC(metaslab_aliquot,
3559 "allocation granularity (a.k.a. stripe size)");
3561 module_param(metaslab_debug_load, int, 0644);
3562 MODULE_PARM_DESC(metaslab_debug_load,
3563 "load all metaslabs when pool is first opened");
3565 module_param(metaslab_debug_unload, int, 0644);
3566 MODULE_PARM_DESC(metaslab_debug_unload,
3567 "prevent metaslabs from being unloaded");
3569 module_param(metaslab_preload_enabled, int, 0644);
3570 MODULE_PARM_DESC(metaslab_preload_enabled,
3571 "preload potential metaslabs during reassessment");
3573 module_param(zfs_mg_noalloc_threshold, int, 0644);
3574 MODULE_PARM_DESC(zfs_mg_noalloc_threshold,
3575 "percentage of free space for metaslab group to allow allocation");
3577 module_param(zfs_mg_fragmentation_threshold, int, 0644);
3578 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold,
3579 "fragmentation for metaslab group to allow allocation");
3581 module_param(zfs_metaslab_fragmentation_threshold, int, 0644);
3582 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold,
3583 "fragmentation for metaslab to allow allocation");
3585 module_param(metaslab_fragmentation_factor_enabled, int, 0644);
3586 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled,
3587 "use the fragmentation metric to prefer less fragmented metaslabs");
3589 module_param(metaslab_lba_weighting_enabled, int, 0644);
3590 MODULE_PARM_DESC(metaslab_lba_weighting_enabled,
3591 "prefer metaslabs with lower LBAs");
3593 module_param(metaslab_bias_enabled, int, 0644);
3594 MODULE_PARM_DESC(metaslab_bias_enabled,
3595 "enable metaslab group biasing");
3597 module_param(zfs_metaslab_segment_weight_enabled, int, 0644);
3598 MODULE_PARM_DESC(zfs_metaslab_segment_weight_enabled,
3599 "enable segment-based metaslab selection");
3601 module_param(zfs_metaslab_switch_threshold, int, 0644);
3602 MODULE_PARM_DESC(zfs_metaslab_switch_threshold,
3603 "segment-based metaslab selection maximum buckets before switching");
3604 #endif /* _KERNEL && HAVE_SPL */