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) {
416 space += tvd->vdev_max_asize - tvd->vdev_asize;
418 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
423 metaslab_compare(const void *x1, const void *x2)
425 const metaslab_t *m1 = (const metaslab_t *)x1;
426 const metaslab_t *m2 = (const metaslab_t *)x2;
428 int cmp = AVL_CMP(m2->ms_weight, m1->ms_weight);
432 IMPLY(AVL_CMP(m1->ms_start, m2->ms_start) == 0, m1 == m2);
434 return (AVL_CMP(m1->ms_start, m2->ms_start));
438 * Verify that the space accounting on disk matches the in-core range_trees.
441 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
443 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
444 uint64_t allocated = 0;
446 uint64_t sm_free_space, msp_free_space;
449 ASSERT(MUTEX_HELD(&msp->ms_lock));
451 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
455 * We can only verify the metaslab space when we're called
456 * from syncing context with a loaded metaslab that has an allocated
457 * space map. Calling this in non-syncing context does not
458 * provide a consistent view of the metaslab since we're performing
459 * allocations in the future.
461 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
465 sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) -
466 space_map_alloc_delta(msp->ms_sm);
469 * Account for future allocations since we would have already
470 * deducted that space from the ms_freetree.
472 for (t = 0; t < TXG_CONCURRENT_STATES; t++) {
474 range_tree_space(msp->ms_alloctree[(txg + t) & TXG_MASK]);
476 freed = range_tree_space(msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK]);
478 msp_free_space = range_tree_space(msp->ms_tree) + allocated +
479 msp->ms_deferspace + freed;
481 VERIFY3U(sm_free_space, ==, msp_free_space);
485 * ==========================================================================
487 * ==========================================================================
490 * Update the allocatable flag and the metaslab group's capacity.
491 * The allocatable flag is set to true if the capacity is below
492 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
493 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
494 * transitions from allocatable to non-allocatable or vice versa then the
495 * metaslab group's class is updated to reflect the transition.
498 metaslab_group_alloc_update(metaslab_group_t *mg)
500 vdev_t *vd = mg->mg_vd;
501 metaslab_class_t *mc = mg->mg_class;
502 vdev_stat_t *vs = &vd->vdev_stat;
503 boolean_t was_allocatable;
504 boolean_t was_initialized;
506 ASSERT(vd == vd->vdev_top);
508 mutex_enter(&mg->mg_lock);
509 was_allocatable = mg->mg_allocatable;
510 was_initialized = mg->mg_initialized;
512 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
515 mutex_enter(&mc->mc_lock);
518 * If the metaslab group was just added then it won't
519 * have any space until we finish syncing out this txg.
520 * At that point we will consider it initialized and available
521 * for allocations. We also don't consider non-activated
522 * metaslab groups (e.g. vdevs that are in the middle of being removed)
523 * to be initialized, because they can't be used for allocation.
525 mg->mg_initialized = metaslab_group_initialized(mg);
526 if (!was_initialized && mg->mg_initialized) {
528 } else if (was_initialized && !mg->mg_initialized) {
529 ASSERT3U(mc->mc_groups, >, 0);
532 if (mg->mg_initialized)
533 mg->mg_no_free_space = B_FALSE;
536 * A metaslab group is considered allocatable if it has plenty
537 * of free space or is not heavily fragmented. We only take
538 * fragmentation into account if the metaslab group has a valid
539 * fragmentation metric (i.e. a value between 0 and 100).
541 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
542 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
543 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
544 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
547 * The mc_alloc_groups maintains a count of the number of
548 * groups in this metaslab class that are still above the
549 * zfs_mg_noalloc_threshold. This is used by the allocating
550 * threads to determine if they should avoid allocations to
551 * a given group. The allocator will avoid allocations to a group
552 * if that group has reached or is below the zfs_mg_noalloc_threshold
553 * and there are still other groups that are above the threshold.
554 * When a group transitions from allocatable to non-allocatable or
555 * vice versa we update the metaslab class to reflect that change.
556 * When the mc_alloc_groups value drops to 0 that means that all
557 * groups have reached the zfs_mg_noalloc_threshold making all groups
558 * eligible for allocations. This effectively means that all devices
559 * are balanced again.
561 if (was_allocatable && !mg->mg_allocatable)
562 mc->mc_alloc_groups--;
563 else if (!was_allocatable && mg->mg_allocatable)
564 mc->mc_alloc_groups++;
565 mutex_exit(&mc->mc_lock);
567 mutex_exit(&mg->mg_lock);
571 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
573 metaslab_group_t *mg;
575 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
576 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
577 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
578 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
581 mg->mg_activation_count = 0;
582 mg->mg_initialized = B_FALSE;
583 mg->mg_no_free_space = B_TRUE;
584 refcount_create_tracked(&mg->mg_alloc_queue_depth);
586 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
587 maxclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT | TASKQ_DYNAMIC);
593 metaslab_group_destroy(metaslab_group_t *mg)
595 ASSERT(mg->mg_prev == NULL);
596 ASSERT(mg->mg_next == NULL);
598 * We may have gone below zero with the activation count
599 * either because we never activated in the first place or
600 * because we're done, and possibly removing the vdev.
602 ASSERT(mg->mg_activation_count <= 0);
604 taskq_destroy(mg->mg_taskq);
605 avl_destroy(&mg->mg_metaslab_tree);
606 mutex_destroy(&mg->mg_lock);
607 refcount_destroy(&mg->mg_alloc_queue_depth);
608 kmem_free(mg, sizeof (metaslab_group_t));
612 metaslab_group_activate(metaslab_group_t *mg)
614 metaslab_class_t *mc = mg->mg_class;
615 metaslab_group_t *mgprev, *mgnext;
617 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
619 ASSERT(mc->mc_rotor != mg);
620 ASSERT(mg->mg_prev == NULL);
621 ASSERT(mg->mg_next == NULL);
622 ASSERT(mg->mg_activation_count <= 0);
624 if (++mg->mg_activation_count <= 0)
627 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
628 metaslab_group_alloc_update(mg);
630 if ((mgprev = mc->mc_rotor) == NULL) {
634 mgnext = mgprev->mg_next;
635 mg->mg_prev = mgprev;
636 mg->mg_next = mgnext;
637 mgprev->mg_next = mg;
638 mgnext->mg_prev = mg;
644 metaslab_group_passivate(metaslab_group_t *mg)
646 metaslab_class_t *mc = mg->mg_class;
647 metaslab_group_t *mgprev, *mgnext;
649 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
651 if (--mg->mg_activation_count != 0) {
652 ASSERT(mc->mc_rotor != mg);
653 ASSERT(mg->mg_prev == NULL);
654 ASSERT(mg->mg_next == NULL);
655 ASSERT(mg->mg_activation_count < 0);
659 taskq_wait_outstanding(mg->mg_taskq, 0);
660 metaslab_group_alloc_update(mg);
662 mgprev = mg->mg_prev;
663 mgnext = mg->mg_next;
668 mc->mc_rotor = mgnext;
669 mgprev->mg_next = mgnext;
670 mgnext->mg_prev = mgprev;
678 metaslab_group_initialized(metaslab_group_t *mg)
680 vdev_t *vd = mg->mg_vd;
681 vdev_stat_t *vs = &vd->vdev_stat;
683 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
687 metaslab_group_get_space(metaslab_group_t *mg)
689 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
693 metaslab_group_histogram_verify(metaslab_group_t *mg)
696 vdev_t *vd = mg->mg_vd;
697 uint64_t ashift = vd->vdev_ashift;
700 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
703 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
706 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
707 SPACE_MAP_HISTOGRAM_SIZE + ashift);
709 for (m = 0; m < vd->vdev_ms_count; m++) {
710 metaslab_t *msp = vd->vdev_ms[m];
712 if (msp->ms_sm == NULL)
715 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
716 mg_hist[i + ashift] +=
717 msp->ms_sm->sm_phys->smp_histogram[i];
720 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
721 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
723 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
727 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
729 metaslab_class_t *mc = mg->mg_class;
730 uint64_t ashift = mg->mg_vd->vdev_ashift;
733 ASSERT(MUTEX_HELD(&msp->ms_lock));
734 if (msp->ms_sm == NULL)
737 mutex_enter(&mg->mg_lock);
738 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
739 mg->mg_histogram[i + ashift] +=
740 msp->ms_sm->sm_phys->smp_histogram[i];
741 mc->mc_histogram[i + ashift] +=
742 msp->ms_sm->sm_phys->smp_histogram[i];
744 mutex_exit(&mg->mg_lock);
748 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
750 metaslab_class_t *mc = mg->mg_class;
751 uint64_t ashift = mg->mg_vd->vdev_ashift;
754 ASSERT(MUTEX_HELD(&msp->ms_lock));
755 if (msp->ms_sm == NULL)
758 mutex_enter(&mg->mg_lock);
759 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
760 ASSERT3U(mg->mg_histogram[i + ashift], >=,
761 msp->ms_sm->sm_phys->smp_histogram[i]);
762 ASSERT3U(mc->mc_histogram[i + ashift], >=,
763 msp->ms_sm->sm_phys->smp_histogram[i]);
765 mg->mg_histogram[i + ashift] -=
766 msp->ms_sm->sm_phys->smp_histogram[i];
767 mc->mc_histogram[i + ashift] -=
768 msp->ms_sm->sm_phys->smp_histogram[i];
770 mutex_exit(&mg->mg_lock);
774 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
776 ASSERT(msp->ms_group == NULL);
777 mutex_enter(&mg->mg_lock);
780 avl_add(&mg->mg_metaslab_tree, msp);
781 mutex_exit(&mg->mg_lock);
783 mutex_enter(&msp->ms_lock);
784 metaslab_group_histogram_add(mg, msp);
785 mutex_exit(&msp->ms_lock);
789 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
791 mutex_enter(&msp->ms_lock);
792 metaslab_group_histogram_remove(mg, msp);
793 mutex_exit(&msp->ms_lock);
795 mutex_enter(&mg->mg_lock);
796 ASSERT(msp->ms_group == mg);
797 avl_remove(&mg->mg_metaslab_tree, msp);
798 msp->ms_group = NULL;
799 mutex_exit(&mg->mg_lock);
803 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
806 * Although in principle the weight can be any value, in
807 * practice we do not use values in the range [1, 511].
809 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
810 ASSERT(MUTEX_HELD(&msp->ms_lock));
812 mutex_enter(&mg->mg_lock);
813 ASSERT(msp->ms_group == mg);
814 avl_remove(&mg->mg_metaslab_tree, msp);
815 msp->ms_weight = weight;
816 avl_add(&mg->mg_metaslab_tree, msp);
817 mutex_exit(&mg->mg_lock);
821 * Calculate the fragmentation for a given metaslab group. We can use
822 * a simple average here since all metaslabs within the group must have
823 * the same size. The return value will be a value between 0 and 100
824 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
825 * group have a fragmentation metric.
828 metaslab_group_fragmentation(metaslab_group_t *mg)
830 vdev_t *vd = mg->mg_vd;
831 uint64_t fragmentation = 0;
832 uint64_t valid_ms = 0;
835 for (m = 0; m < vd->vdev_ms_count; m++) {
836 metaslab_t *msp = vd->vdev_ms[m];
838 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
842 fragmentation += msp->ms_fragmentation;
845 if (valid_ms <= vd->vdev_ms_count / 2)
846 return (ZFS_FRAG_INVALID);
848 fragmentation /= valid_ms;
849 ASSERT3U(fragmentation, <=, 100);
850 return (fragmentation);
854 * Determine if a given metaslab group should skip allocations. A metaslab
855 * group should avoid allocations if its free capacity is less than the
856 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
857 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
858 * that can still handle allocations. If the allocation throttle is enabled
859 * then we skip allocations to devices that have reached their maximum
860 * allocation queue depth unless the selected metaslab group is the only
861 * eligible group remaining.
864 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
867 spa_t *spa = mg->mg_vd->vdev_spa;
868 metaslab_class_t *mc = mg->mg_class;
871 * We can only consider skipping this metaslab group if it's
872 * in the normal metaslab class and there are other metaslab
873 * groups to select from. Otherwise, we always consider it eligible
876 if (mc != spa_normal_class(spa) || mc->mc_groups <= 1)
880 * If the metaslab group's mg_allocatable flag is set (see comments
881 * in metaslab_group_alloc_update() for more information) and
882 * the allocation throttle is disabled then allow allocations to this
883 * device. However, if the allocation throttle is enabled then
884 * check if we have reached our allocation limit (mg_alloc_queue_depth)
885 * to determine if we should allow allocations to this metaslab group.
886 * If all metaslab groups are no longer considered allocatable
887 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
888 * gang block size then we allow allocations on this metaslab group
889 * regardless of the mg_allocatable or throttle settings.
891 if (mg->mg_allocatable) {
892 metaslab_group_t *mgp;
894 uint64_t qmax = mg->mg_max_alloc_queue_depth;
896 if (!mc->mc_alloc_throttle_enabled)
900 * If this metaslab group does not have any free space, then
901 * there is no point in looking further.
903 if (mg->mg_no_free_space)
906 qdepth = refcount_count(&mg->mg_alloc_queue_depth);
909 * If this metaslab group is below its qmax or it's
910 * the only allocatable metasable group, then attempt
911 * to allocate from it.
913 if (qdepth < qmax || mc->mc_alloc_groups == 1)
915 ASSERT3U(mc->mc_alloc_groups, >, 1);
918 * Since this metaslab group is at or over its qmax, we
919 * need to determine if there are metaslab groups after this
920 * one that might be able to handle this allocation. This is
921 * racy since we can't hold the locks for all metaslab
922 * groups at the same time when we make this check.
924 for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
925 qmax = mgp->mg_max_alloc_queue_depth;
927 qdepth = refcount_count(&mgp->mg_alloc_queue_depth);
930 * If there is another metaslab group that
931 * might be able to handle the allocation, then
932 * we return false so that we skip this group.
934 if (qdepth < qmax && !mgp->mg_no_free_space)
939 * We didn't find another group to handle the allocation
940 * so we can't skip this metaslab group even though
941 * we are at or over our qmax.
945 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
952 * ==========================================================================
953 * Range tree callbacks
954 * ==========================================================================
958 * Comparison function for the private size-ordered tree. Tree is sorted
959 * by size, larger sizes at the end of the tree.
962 metaslab_rangesize_compare(const void *x1, const void *x2)
964 const range_seg_t *r1 = x1;
965 const range_seg_t *r2 = x2;
966 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
967 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
969 int cmp = AVL_CMP(rs_size1, rs_size2);
973 return (AVL_CMP(r1->rs_start, r2->rs_start));
977 * Create any block allocator specific components. The current allocators
978 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
981 metaslab_rt_create(range_tree_t *rt, void *arg)
983 metaslab_t *msp = arg;
985 ASSERT3P(rt->rt_arg, ==, msp);
986 ASSERT(msp->ms_tree == NULL);
988 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
989 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
993 * Destroy the block allocator specific components.
996 metaslab_rt_destroy(range_tree_t *rt, void *arg)
998 metaslab_t *msp = arg;
1000 ASSERT3P(rt->rt_arg, ==, msp);
1001 ASSERT3P(msp->ms_tree, ==, rt);
1002 ASSERT0(avl_numnodes(&msp->ms_size_tree));
1004 avl_destroy(&msp->ms_size_tree);
1008 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1010 metaslab_t *msp = arg;
1012 ASSERT3P(rt->rt_arg, ==, msp);
1013 ASSERT3P(msp->ms_tree, ==, rt);
1014 VERIFY(!msp->ms_condensing);
1015 avl_add(&msp->ms_size_tree, rs);
1019 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1021 metaslab_t *msp = arg;
1023 ASSERT3P(rt->rt_arg, ==, msp);
1024 ASSERT3P(msp->ms_tree, ==, rt);
1025 VERIFY(!msp->ms_condensing);
1026 avl_remove(&msp->ms_size_tree, rs);
1030 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1032 metaslab_t *msp = arg;
1034 ASSERT3P(rt->rt_arg, ==, msp);
1035 ASSERT3P(msp->ms_tree, ==, rt);
1038 * Normally one would walk the tree freeing nodes along the way.
1039 * Since the nodes are shared with the range trees we can avoid
1040 * walking all nodes and just reinitialize the avl tree. The nodes
1041 * will be freed by the range tree, so we don't want to free them here.
1043 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
1044 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1047 static range_tree_ops_t metaslab_rt_ops = {
1049 metaslab_rt_destroy,
1056 * ==========================================================================
1057 * Common allocator routines
1058 * ==========================================================================
1062 * Return the maximum contiguous segment within the metaslab.
1065 metaslab_block_maxsize(metaslab_t *msp)
1067 avl_tree_t *t = &msp->ms_size_tree;
1070 if (t == NULL || (rs = avl_last(t)) == NULL)
1073 return (rs->rs_end - rs->rs_start);
1076 static range_seg_t *
1077 metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
1079 range_seg_t *rs, rsearch;
1082 rsearch.rs_start = start;
1083 rsearch.rs_end = start + size;
1085 rs = avl_find(t, &rsearch, &where);
1087 rs = avl_nearest(t, where, AVL_AFTER);
1093 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
1094 defined(WITH_DF_BLOCK_ALLOCATOR) || \
1095 defined(WITH_CF_BLOCK_ALLOCATOR)
1097 * This is a helper function that can be used by the allocator to find
1098 * a suitable block to allocate. This will search the specified AVL
1099 * tree looking for a block that matches the specified criteria.
1102 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1105 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1107 while (rs != NULL) {
1108 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1110 if (offset + size <= rs->rs_end) {
1111 *cursor = offset + size;
1114 rs = AVL_NEXT(t, rs);
1118 * If we know we've searched the whole map (*cursor == 0), give up.
1119 * Otherwise, reset the cursor to the beginning and try again.
1125 return (metaslab_block_picker(t, cursor, size, align));
1127 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
1129 #if defined(WITH_FF_BLOCK_ALLOCATOR)
1131 * ==========================================================================
1132 * The first-fit block allocator
1133 * ==========================================================================
1136 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1139 * Find the largest power of 2 block size that evenly divides the
1140 * requested size. This is used to try to allocate blocks with similar
1141 * alignment from the same area of the metaslab (i.e. same cursor
1142 * bucket) but it does not guarantee that other allocations sizes
1143 * may exist in the same region.
1145 uint64_t align = size & -size;
1146 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1147 avl_tree_t *t = &msp->ms_tree->rt_root;
1149 return (metaslab_block_picker(t, cursor, size, align));
1152 static metaslab_ops_t metaslab_ff_ops = {
1156 metaslab_ops_t *zfs_metaslab_ops = &metaslab_ff_ops;
1157 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1159 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1161 * ==========================================================================
1162 * Dynamic block allocator -
1163 * Uses the first fit allocation scheme until space get low and then
1164 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1165 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1166 * ==========================================================================
1169 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1172 * Find the largest power of 2 block size that evenly divides the
1173 * requested size. This is used to try to allocate blocks with similar
1174 * alignment from the same area of the metaslab (i.e. same cursor
1175 * bucket) but it does not guarantee that other allocations sizes
1176 * may exist in the same region.
1178 uint64_t align = size & -size;
1179 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1180 range_tree_t *rt = msp->ms_tree;
1181 avl_tree_t *t = &rt->rt_root;
1182 uint64_t max_size = metaslab_block_maxsize(msp);
1183 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1185 ASSERT(MUTEX_HELD(&msp->ms_lock));
1186 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1188 if (max_size < size)
1192 * If we're running low on space switch to using the size
1193 * sorted AVL tree (best-fit).
1195 if (max_size < metaslab_df_alloc_threshold ||
1196 free_pct < metaslab_df_free_pct) {
1197 t = &msp->ms_size_tree;
1201 return (metaslab_block_picker(t, cursor, size, 1ULL));
1204 static metaslab_ops_t metaslab_df_ops = {
1208 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1209 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1211 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1213 * ==========================================================================
1214 * Cursor fit block allocator -
1215 * Select the largest region in the metaslab, set the cursor to the beginning
1216 * of the range and the cursor_end to the end of the range. As allocations
1217 * are made advance the cursor. Continue allocating from the cursor until
1218 * the range is exhausted and then find a new range.
1219 * ==========================================================================
1222 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1224 range_tree_t *rt = msp->ms_tree;
1225 avl_tree_t *t = &msp->ms_size_tree;
1226 uint64_t *cursor = &msp->ms_lbas[0];
1227 uint64_t *cursor_end = &msp->ms_lbas[1];
1228 uint64_t offset = 0;
1230 ASSERT(MUTEX_HELD(&msp->ms_lock));
1231 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1233 ASSERT3U(*cursor_end, >=, *cursor);
1235 if ((*cursor + size) > *cursor_end) {
1238 rs = avl_last(&msp->ms_size_tree);
1239 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1242 *cursor = rs->rs_start;
1243 *cursor_end = rs->rs_end;
1252 static metaslab_ops_t metaslab_cf_ops = {
1256 metaslab_ops_t *zfs_metaslab_ops = &metaslab_cf_ops;
1257 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1259 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1261 * ==========================================================================
1262 * New dynamic fit allocator -
1263 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1264 * contiguous blocks. If no region is found then just use the largest segment
1266 * ==========================================================================
1270 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1271 * to request from the allocator.
1273 uint64_t metaslab_ndf_clump_shift = 4;
1276 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1278 avl_tree_t *t = &msp->ms_tree->rt_root;
1280 range_seg_t *rs, rsearch;
1281 uint64_t hbit = highbit64(size);
1282 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1283 uint64_t max_size = metaslab_block_maxsize(msp);
1285 ASSERT(MUTEX_HELD(&msp->ms_lock));
1286 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1288 if (max_size < size)
1291 rsearch.rs_start = *cursor;
1292 rsearch.rs_end = *cursor + size;
1294 rs = avl_find(t, &rsearch, &where);
1295 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1296 t = &msp->ms_size_tree;
1298 rsearch.rs_start = 0;
1299 rsearch.rs_end = MIN(max_size,
1300 1ULL << (hbit + metaslab_ndf_clump_shift));
1301 rs = avl_find(t, &rsearch, &where);
1303 rs = avl_nearest(t, where, AVL_AFTER);
1307 if ((rs->rs_end - rs->rs_start) >= size) {
1308 *cursor = rs->rs_start + size;
1309 return (rs->rs_start);
1314 static metaslab_ops_t metaslab_ndf_ops = {
1318 metaslab_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops;
1319 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1323 * ==========================================================================
1325 * ==========================================================================
1329 * Wait for any in-progress metaslab loads to complete.
1332 metaslab_load_wait(metaslab_t *msp)
1334 ASSERT(MUTEX_HELD(&msp->ms_lock));
1336 while (msp->ms_loading) {
1337 ASSERT(!msp->ms_loaded);
1338 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1343 metaslab_load(metaslab_t *msp)
1347 boolean_t success = B_FALSE;
1349 ASSERT(MUTEX_HELD(&msp->ms_lock));
1350 ASSERT(!msp->ms_loaded);
1351 ASSERT(!msp->ms_loading);
1353 msp->ms_loading = B_TRUE;
1356 * If the space map has not been allocated yet, then treat
1357 * all the space in the metaslab as free and add it to the
1360 if (msp->ms_sm != NULL)
1361 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1363 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1365 success = (error == 0);
1366 msp->ms_loading = B_FALSE;
1369 ASSERT3P(msp->ms_group, !=, NULL);
1370 msp->ms_loaded = B_TRUE;
1372 for (t = 0; t < TXG_DEFER_SIZE; t++) {
1373 range_tree_walk(msp->ms_defertree[t],
1374 range_tree_remove, msp->ms_tree);
1376 msp->ms_max_size = metaslab_block_maxsize(msp);
1378 cv_broadcast(&msp->ms_load_cv);
1383 metaslab_unload(metaslab_t *msp)
1385 ASSERT(MUTEX_HELD(&msp->ms_lock));
1386 range_tree_vacate(msp->ms_tree, NULL, NULL);
1387 msp->ms_loaded = B_FALSE;
1388 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1389 msp->ms_max_size = 0;
1393 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1396 vdev_t *vd = mg->mg_vd;
1397 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1401 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1402 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1403 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1405 ms->ms_start = id << vd->vdev_ms_shift;
1406 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1409 * We only open space map objects that already exist. All others
1410 * will be opened when we finally allocate an object for it.
1413 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1414 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1417 kmem_free(ms, sizeof (metaslab_t));
1421 ASSERT(ms->ms_sm != NULL);
1425 * We create the main range tree here, but we don't create the
1426 * alloctree and freetree until metaslab_sync_done(). This serves
1427 * two purposes: it allows metaslab_sync_done() to detect the
1428 * addition of new space; and for debugging, it ensures that we'd
1429 * data fault on any attempt to use this metaslab before it's ready.
1431 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1432 metaslab_group_add(mg, ms);
1434 metaslab_set_fragmentation(ms);
1437 * If we're opening an existing pool (txg == 0) or creating
1438 * a new one (txg == TXG_INITIAL), all space is available now.
1439 * If we're adding space to an existing pool, the new space
1440 * does not become available until after this txg has synced.
1441 * The metaslab's weight will also be initialized when we sync
1442 * out this txg. This ensures that we don't attempt to allocate
1443 * from it before we have initialized it completely.
1445 if (txg <= TXG_INITIAL)
1446 metaslab_sync_done(ms, 0);
1449 * If metaslab_debug_load is set and we're initializing a metaslab
1450 * that has an allocated space map object then load the its space
1451 * map so that can verify frees.
1453 if (metaslab_debug_load && ms->ms_sm != NULL) {
1454 mutex_enter(&ms->ms_lock);
1455 VERIFY0(metaslab_load(ms));
1456 mutex_exit(&ms->ms_lock);
1460 vdev_dirty(vd, 0, NULL, txg);
1461 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1470 metaslab_fini(metaslab_t *msp)
1474 metaslab_group_t *mg = msp->ms_group;
1476 metaslab_group_remove(mg, msp);
1478 mutex_enter(&msp->ms_lock);
1479 VERIFY(msp->ms_group == NULL);
1480 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1482 space_map_close(msp->ms_sm);
1484 metaslab_unload(msp);
1485 range_tree_destroy(msp->ms_tree);
1487 for (t = 0; t < TXG_SIZE; t++) {
1488 range_tree_destroy(msp->ms_alloctree[t]);
1489 range_tree_destroy(msp->ms_freetree[t]);
1492 for (t = 0; t < TXG_DEFER_SIZE; t++) {
1493 range_tree_destroy(msp->ms_defertree[t]);
1496 ASSERT0(msp->ms_deferspace);
1498 mutex_exit(&msp->ms_lock);
1499 cv_destroy(&msp->ms_load_cv);
1500 mutex_destroy(&msp->ms_lock);
1502 kmem_free(msp, sizeof (metaslab_t));
1505 #define FRAGMENTATION_TABLE_SIZE 17
1508 * This table defines a segment size based fragmentation metric that will
1509 * allow each metaslab to derive its own fragmentation value. This is done
1510 * by calculating the space in each bucket of the spacemap histogram and
1511 * multiplying that by the fragmetation metric in this table. Doing
1512 * this for all buckets and dividing it by the total amount of free
1513 * space in this metaslab (i.e. the total free space in all buckets) gives
1514 * us the fragmentation metric. This means that a high fragmentation metric
1515 * equates to most of the free space being comprised of small segments.
1516 * Conversely, if the metric is low, then most of the free space is in
1517 * large segments. A 10% change in fragmentation equates to approximately
1518 * double the number of segments.
1520 * This table defines 0% fragmented space using 16MB segments. Testing has
1521 * shown that segments that are greater than or equal to 16MB do not suffer
1522 * from drastic performance problems. Using this value, we derive the rest
1523 * of the table. Since the fragmentation value is never stored on disk, it
1524 * is possible to change these calculations in the future.
1526 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1546 * Calclate the metaslab's fragmentation metric. A return value
1547 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1548 * not support this metric. Otherwise, the return value should be in the
1552 metaslab_set_fragmentation(metaslab_t *msp)
1554 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1555 uint64_t fragmentation = 0;
1557 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1558 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1561 if (!feature_enabled) {
1562 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1567 * A null space map means that the entire metaslab is free
1568 * and thus is not fragmented.
1570 if (msp->ms_sm == NULL) {
1571 msp->ms_fragmentation = 0;
1576 * If this metaslab's space map has not been upgraded, flag it
1577 * so that we upgrade next time we encounter it.
1579 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1580 vdev_t *vd = msp->ms_group->mg_vd;
1582 if (spa_writeable(vd->vdev_spa)) {
1583 uint64_t txg = spa_syncing_txg(spa);
1585 msp->ms_condense_wanted = B_TRUE;
1586 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1587 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1588 "msp %p, vd %p", txg, msp, vd);
1590 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1594 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1596 uint8_t shift = msp->ms_sm->sm_shift;
1598 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1599 FRAGMENTATION_TABLE_SIZE - 1);
1601 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1604 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1607 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1608 fragmentation += space * zfs_frag_table[idx];
1612 fragmentation /= total;
1613 ASSERT3U(fragmentation, <=, 100);
1615 msp->ms_fragmentation = fragmentation;
1619 * Compute a weight -- a selection preference value -- for the given metaslab.
1620 * This is based on the amount of free space, the level of fragmentation,
1621 * the LBA range, and whether the metaslab is loaded.
1624 metaslab_space_weight(metaslab_t *msp)
1626 metaslab_group_t *mg = msp->ms_group;
1627 vdev_t *vd = mg->mg_vd;
1628 uint64_t weight, space;
1630 ASSERT(MUTEX_HELD(&msp->ms_lock));
1631 ASSERT(!vd->vdev_removing);
1634 * The baseline weight is the metaslab's free space.
1636 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1638 if (metaslab_fragmentation_factor_enabled &&
1639 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1641 * Use the fragmentation information to inversely scale
1642 * down the baseline weight. We need to ensure that we
1643 * don't exclude this metaslab completely when it's 100%
1644 * fragmented. To avoid this we reduce the fragmented value
1647 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1650 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1651 * this metaslab again. The fragmentation metric may have
1652 * decreased the space to something smaller than
1653 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1654 * so that we can consume any remaining space.
1656 if (space > 0 && space < SPA_MINBLOCKSIZE)
1657 space = SPA_MINBLOCKSIZE;
1662 * Modern disks have uniform bit density and constant angular velocity.
1663 * Therefore, the outer recording zones are faster (higher bandwidth)
1664 * than the inner zones by the ratio of outer to inner track diameter,
1665 * which is typically around 2:1. We account for this by assigning
1666 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1667 * In effect, this means that we'll select the metaslab with the most
1668 * free bandwidth rather than simply the one with the most free space.
1670 if (!vd->vdev_nonrot && metaslab_lba_weighting_enabled) {
1671 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1672 ASSERT(weight >= space && weight <= 2 * space);
1676 * If this metaslab is one we're actively using, adjust its
1677 * weight to make it preferable to any inactive metaslab so
1678 * we'll polish it off. If the fragmentation on this metaslab
1679 * has exceed our threshold, then don't mark it active.
1681 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1682 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1683 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1686 WEIGHT_SET_SPACEBASED(weight);
1691 * Return the weight of the specified metaslab, according to the segment-based
1692 * weighting algorithm. The metaslab must be loaded. This function can
1693 * be called within a sync pass since it relies only on the metaslab's
1694 * range tree which is always accurate when the metaslab is loaded.
1697 metaslab_weight_from_range_tree(metaslab_t *msp)
1699 uint64_t weight = 0;
1700 uint32_t segments = 0;
1703 ASSERT(msp->ms_loaded);
1705 for (i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT; i--) {
1706 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
1707 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1710 segments += msp->ms_tree->rt_histogram[i];
1713 * The range tree provides more precision than the space map
1714 * and must be downgraded so that all values fit within the
1715 * space map's histogram. This allows us to compare loaded
1716 * vs. unloaded metaslabs to determine which metaslab is
1717 * considered "best".
1722 if (segments != 0) {
1723 WEIGHT_SET_COUNT(weight, segments);
1724 WEIGHT_SET_INDEX(weight, i);
1725 WEIGHT_SET_ACTIVE(weight, 0);
1733 * Calculate the weight based on the on-disk histogram. This should only
1734 * be called after a sync pass has completely finished since the on-disk
1735 * information is updated in metaslab_sync().
1738 metaslab_weight_from_spacemap(metaslab_t *msp)
1740 uint64_t weight = 0;
1743 for (i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
1744 if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) {
1745 WEIGHT_SET_COUNT(weight,
1746 msp->ms_sm->sm_phys->smp_histogram[i]);
1747 WEIGHT_SET_INDEX(weight, i +
1748 msp->ms_sm->sm_shift);
1749 WEIGHT_SET_ACTIVE(weight, 0);
1757 * Compute a segment-based weight for the specified metaslab. The weight
1758 * is determined by highest bucket in the histogram. The information
1759 * for the highest bucket is encoded into the weight value.
1762 metaslab_segment_weight(metaslab_t *msp)
1764 metaslab_group_t *mg = msp->ms_group;
1765 uint64_t weight = 0;
1766 uint8_t shift = mg->mg_vd->vdev_ashift;
1768 ASSERT(MUTEX_HELD(&msp->ms_lock));
1771 * The metaslab is completely free.
1773 if (space_map_allocated(msp->ms_sm) == 0) {
1774 int idx = highbit64(msp->ms_size) - 1;
1775 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1777 if (idx < max_idx) {
1778 WEIGHT_SET_COUNT(weight, 1ULL);
1779 WEIGHT_SET_INDEX(weight, idx);
1781 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
1782 WEIGHT_SET_INDEX(weight, max_idx);
1784 WEIGHT_SET_ACTIVE(weight, 0);
1785 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
1790 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
1793 * If the metaslab is fully allocated then just make the weight 0.
1795 if (space_map_allocated(msp->ms_sm) == msp->ms_size)
1798 * If the metaslab is already loaded, then use the range tree to
1799 * determine the weight. Otherwise, we rely on the space map information
1800 * to generate the weight.
1802 if (msp->ms_loaded) {
1803 weight = metaslab_weight_from_range_tree(msp);
1805 weight = metaslab_weight_from_spacemap(msp);
1809 * If the metaslab was active the last time we calculated its weight
1810 * then keep it active. We want to consume the entire region that
1811 * is associated with this weight.
1813 if (msp->ms_activation_weight != 0 && weight != 0)
1814 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
1819 * Determine if we should attempt to allocate from this metaslab. If the
1820 * metaslab has a maximum size then we can quickly determine if the desired
1821 * allocation size can be satisfied. Otherwise, if we're using segment-based
1822 * weighting then we can determine the maximum allocation that this metaslab
1823 * can accommodate based on the index encoded in the weight. If we're using
1824 * space-based weights then rely on the entire weight (excluding the weight
1828 metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
1830 boolean_t should_allocate;
1832 if (msp->ms_max_size != 0)
1833 return (msp->ms_max_size >= asize);
1835 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
1837 * The metaslab segment weight indicates segments in the
1838 * range [2^i, 2^(i+1)), where i is the index in the weight.
1839 * Since the asize might be in the middle of the range, we
1840 * should attempt the allocation if asize < 2^(i+1).
1842 should_allocate = (asize <
1843 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1845 should_allocate = (asize <=
1846 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
1848 return (should_allocate);
1851 metaslab_weight(metaslab_t *msp)
1853 vdev_t *vd = msp->ms_group->mg_vd;
1854 spa_t *spa = vd->vdev_spa;
1857 ASSERT(MUTEX_HELD(&msp->ms_lock));
1860 * This vdev is in the process of being removed so there is nothing
1861 * for us to do here.
1863 if (vd->vdev_removing) {
1864 ASSERT0(space_map_allocated(msp->ms_sm));
1865 ASSERT0(vd->vdev_ms_shift);
1869 metaslab_set_fragmentation(msp);
1872 * Update the maximum size if the metaslab is loaded. This will
1873 * ensure that we get an accurate maximum size if newly freed space
1874 * has been added back into the free tree.
1877 msp->ms_max_size = metaslab_block_maxsize(msp);
1880 * Segment-based weighting requires space map histogram support.
1882 if (zfs_metaslab_segment_weight_enabled &&
1883 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
1884 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
1885 sizeof (space_map_phys_t))) {
1886 weight = metaslab_segment_weight(msp);
1888 weight = metaslab_space_weight(msp);
1894 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1896 ASSERT(MUTEX_HELD(&msp->ms_lock));
1898 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1899 metaslab_load_wait(msp);
1900 if (!msp->ms_loaded) {
1901 int error = metaslab_load(msp);
1903 metaslab_group_sort(msp->ms_group, msp, 0);
1908 msp->ms_activation_weight = msp->ms_weight;
1909 metaslab_group_sort(msp->ms_group, msp,
1910 msp->ms_weight | activation_weight);
1912 ASSERT(msp->ms_loaded);
1913 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1919 metaslab_passivate(metaslab_t *msp, uint64_t weight)
1921 ASSERTV(uint64_t size = weight & ~METASLAB_WEIGHT_TYPE);
1924 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1925 * this metaslab again. In that case, it had better be empty,
1926 * or we would be leaving space on the table.
1928 ASSERT(size >= SPA_MINBLOCKSIZE ||
1929 range_tree_space(msp->ms_tree) == 0);
1930 ASSERT0(weight & METASLAB_ACTIVE_MASK);
1932 msp->ms_activation_weight = 0;
1933 metaslab_group_sort(msp->ms_group, msp, weight);
1934 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1938 * Segment-based metaslabs are activated once and remain active until
1939 * we either fail an allocation attempt (similar to space-based metaslabs)
1940 * or have exhausted the free space in zfs_metaslab_switch_threshold
1941 * buckets since the metaslab was activated. This function checks to see
1942 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
1943 * metaslab and passivates it proactively. This will allow us to select a
1944 * metaslab with a larger contiguous region, if any, remaining within this
1945 * metaslab group. If we're in sync pass > 1, then we continue using this
1946 * metaslab so that we don't dirty more block and cause more sync passes.
1949 metaslab_segment_may_passivate(metaslab_t *msp)
1951 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1953 int activation_idx, current_idx;
1955 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
1959 * Since we are in the middle of a sync pass, the most accurate
1960 * information that is accessible to us is the in-core range tree
1961 * histogram; calculate the new weight based on that information.
1963 weight = metaslab_weight_from_range_tree(msp);
1964 activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
1965 current_idx = WEIGHT_GET_INDEX(weight);
1967 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
1968 metaslab_passivate(msp, weight);
1972 metaslab_preload(void *arg)
1974 metaslab_t *msp = arg;
1975 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1976 fstrans_cookie_t cookie = spl_fstrans_mark();
1978 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1980 mutex_enter(&msp->ms_lock);
1981 metaslab_load_wait(msp);
1982 if (!msp->ms_loaded)
1983 (void) metaslab_load(msp);
1984 msp->ms_selected_txg = spa_syncing_txg(spa);
1985 mutex_exit(&msp->ms_lock);
1986 spl_fstrans_unmark(cookie);
1990 metaslab_group_preload(metaslab_group_t *mg)
1992 spa_t *spa = mg->mg_vd->vdev_spa;
1994 avl_tree_t *t = &mg->mg_metaslab_tree;
1997 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1998 taskq_wait_outstanding(mg->mg_taskq, 0);
2002 mutex_enter(&mg->mg_lock);
2004 * Load the next potential metaslabs
2006 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2008 * We preload only the maximum number of metaslabs specified
2009 * by metaslab_preload_limit. If a metaslab is being forced
2010 * to condense then we preload it too. This will ensure
2011 * that force condensing happens in the next txg.
2013 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2017 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2018 msp, TQ_SLEEP) != TASKQID_INVALID);
2020 mutex_exit(&mg->mg_lock);
2024 * Determine if the space map's on-disk footprint is past our tolerance
2025 * for inefficiency. We would like to use the following criteria to make
2028 * 1. The size of the space map object should not dramatically increase as a
2029 * result of writing out the free space range tree.
2031 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2032 * times the size than the free space range tree representation
2033 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
2035 * 3. The on-disk size of the space map should actually decrease.
2037 * Checking the first condition is tricky since we don't want to walk
2038 * the entire AVL tree calculating the estimated on-disk size. Instead we
2039 * use the size-ordered range tree in the metaslab and calculate the
2040 * size required to write out the largest segment in our free tree. If the
2041 * size required to represent that segment on disk is larger than the space
2042 * map object then we avoid condensing this map.
2044 * To determine the second criterion we use a best-case estimate and assume
2045 * each segment can be represented on-disk as a single 64-bit entry. We refer
2046 * to this best-case estimate as the space map's minimal form.
2048 * Unfortunately, we cannot compute the on-disk size of the space map in this
2049 * context because we cannot accurately compute the effects of compression, etc.
2050 * Instead, we apply the heuristic described in the block comment for
2051 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2052 * is greater than a threshold number of blocks.
2055 metaslab_should_condense(metaslab_t *msp)
2057 space_map_t *sm = msp->ms_sm;
2059 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
2060 dmu_object_info_t doi;
2061 uint64_t vdev_blocksize = 1ULL << msp->ms_group->mg_vd->vdev_ashift;
2063 ASSERT(MUTEX_HELD(&msp->ms_lock));
2064 ASSERT(msp->ms_loaded);
2067 * Use the ms_size_tree range tree, which is ordered by size, to
2068 * obtain the largest segment in the free tree. We always condense
2069 * metaslabs that are empty and metaslabs for which a condense
2070 * request has been made.
2072 rs = avl_last(&msp->ms_size_tree);
2073 if (rs == NULL || msp->ms_condense_wanted)
2077 * Calculate the number of 64-bit entries this segment would
2078 * require when written to disk. If this single segment would be
2079 * larger on-disk than the entire current on-disk structure, then
2080 * clearly condensing will increase the on-disk structure size.
2082 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
2083 entries = size / (MIN(size, SM_RUN_MAX));
2084 segsz = entries * sizeof (uint64_t);
2086 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
2087 object_size = space_map_length(msp->ms_sm);
2089 dmu_object_info_from_db(sm->sm_dbuf, &doi);
2090 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
2092 return (segsz <= object_size &&
2093 object_size >= (optimal_size * zfs_condense_pct / 100) &&
2094 object_size > zfs_metaslab_condense_block_threshold * record_size);
2098 * Condense the on-disk space map representation to its minimized form.
2099 * The minimized form consists of a small number of allocations followed by
2100 * the entries of the free range tree.
2103 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
2105 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2106 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
2107 range_tree_t *condense_tree;
2108 space_map_t *sm = msp->ms_sm;
2111 ASSERT(MUTEX_HELD(&msp->ms_lock));
2112 ASSERT3U(spa_sync_pass(spa), ==, 1);
2113 ASSERT(msp->ms_loaded);
2116 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2117 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2118 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2119 msp->ms_group->mg_vd->vdev_spa->spa_name,
2120 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
2121 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2123 msp->ms_condense_wanted = B_FALSE;
2126 * Create an range tree that is 100% allocated. We remove segments
2127 * that have been freed in this txg, any deferred frees that exist,
2128 * and any allocation in the future. Removing segments should be
2129 * a relatively inexpensive operation since we expect these trees to
2130 * have a small number of nodes.
2132 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
2133 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2136 * Remove what's been freed in this txg from the condense_tree.
2137 * Since we're in sync_pass 1, we know that all the frees from
2138 * this txg are in the freetree.
2140 range_tree_walk(freetree, range_tree_remove, condense_tree);
2142 for (t = 0; t < TXG_DEFER_SIZE; t++) {
2143 range_tree_walk(msp->ms_defertree[t],
2144 range_tree_remove, condense_tree);
2147 for (t = 1; t < TXG_CONCURRENT_STATES; t++) {
2148 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
2149 range_tree_remove, condense_tree);
2153 * We're about to drop the metaslab's lock thus allowing
2154 * other consumers to change it's content. Set the
2155 * metaslab's ms_condensing flag to ensure that
2156 * allocations on this metaslab do not occur while we're
2157 * in the middle of committing it to disk. This is only critical
2158 * for the ms_tree as all other range trees use per txg
2159 * views of their content.
2161 msp->ms_condensing = B_TRUE;
2163 mutex_exit(&msp->ms_lock);
2164 space_map_truncate(sm, tx);
2165 mutex_enter(&msp->ms_lock);
2168 * While we would ideally like to create a space map representation
2169 * that consists only of allocation records, doing so can be
2170 * prohibitively expensive because the in-core free tree can be
2171 * large, and therefore computationally expensive to subtract
2172 * from the condense_tree. Instead we sync out two trees, a cheap
2173 * allocation only tree followed by the in-core free tree. While not
2174 * optimal, this is typically close to optimal, and much cheaper to
2177 space_map_write(sm, condense_tree, SM_ALLOC, tx);
2178 range_tree_vacate(condense_tree, NULL, NULL);
2179 range_tree_destroy(condense_tree);
2181 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
2182 msp->ms_condensing = B_FALSE;
2186 * Write a metaslab to disk in the context of the specified transaction group.
2189 metaslab_sync(metaslab_t *msp, uint64_t txg)
2191 metaslab_group_t *mg = msp->ms_group;
2192 vdev_t *vd = mg->mg_vd;
2193 spa_t *spa = vd->vdev_spa;
2194 objset_t *mos = spa_meta_objset(spa);
2195 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
2196 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
2197 range_tree_t **freed_tree =
2198 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2200 uint64_t object = space_map_object(msp->ms_sm);
2202 ASSERT(!vd->vdev_ishole);
2205 * This metaslab has just been added so there's no work to do now.
2207 if (*freetree == NULL) {
2208 ASSERT3P(alloctree, ==, NULL);
2212 ASSERT3P(alloctree, !=, NULL);
2213 ASSERT3P(*freetree, !=, NULL);
2214 ASSERT3P(*freed_tree, !=, NULL);
2217 * Normally, we don't want to process a metaslab if there
2218 * are no allocations or frees to perform. However, if the metaslab
2219 * is being forced to condense we need to let it through.
2221 if (range_tree_space(alloctree) == 0 &&
2222 range_tree_space(*freetree) == 0 &&
2223 !msp->ms_condense_wanted)
2227 * The only state that can actually be changing concurrently with
2228 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2229 * be modifying this txg's alloctree, freetree, freed_tree, or
2230 * space_map_phys_t. Therefore, we only hold ms_lock to satify
2231 * space map ASSERTs. We drop it whenever we call into the DMU,
2232 * because the DMU can call down to us (e.g. via zio_free()) at
2236 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2238 if (msp->ms_sm == NULL) {
2239 uint64_t new_object;
2241 new_object = space_map_alloc(mos, tx);
2242 VERIFY3U(new_object, !=, 0);
2244 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2245 msp->ms_start, msp->ms_size, vd->vdev_ashift,
2247 ASSERT(msp->ms_sm != NULL);
2250 mutex_enter(&msp->ms_lock);
2253 * Note: metaslab_condense() clears the space map's histogram.
2254 * Therefore we must verify and remove this histogram before
2257 metaslab_group_histogram_verify(mg);
2258 metaslab_class_histogram_verify(mg->mg_class);
2259 metaslab_group_histogram_remove(mg, msp);
2261 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
2262 metaslab_should_condense(msp)) {
2263 metaslab_condense(msp, txg, tx);
2265 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
2266 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
2269 if (msp->ms_loaded) {
2273 * When the space map is loaded, we have an accruate
2274 * histogram in the range tree. This gives us an opportunity
2275 * to bring the space map's histogram up-to-date so we clear
2276 * it first before updating it.
2278 space_map_histogram_clear(msp->ms_sm);
2279 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
2282 * Since we've cleared the histogram we need to add back
2283 * any free space that has already been processed, plus
2284 * any deferred space. This allows the on-disk histogram
2285 * to accurately reflect all free space even if some space
2286 * is not yet available for allocation (i.e. deferred).
2288 space_map_histogram_add(msp->ms_sm, *freed_tree, tx);
2291 * Add back any deferred free space that has not been
2292 * added back into the in-core free tree yet. This will
2293 * ensure that we don't end up with a space map histogram
2294 * that is completely empty unless the metaslab is fully
2297 for (t = 0; t < TXG_DEFER_SIZE; t++) {
2298 space_map_histogram_add(msp->ms_sm,
2299 msp->ms_defertree[t], tx);
2304 * Always add the free space from this sync pass to the space
2305 * map histogram. We want to make sure that the on-disk histogram
2306 * accounts for all free space. If the space map is not loaded,
2307 * then we will lose some accuracy but will correct it the next
2308 * time we load the space map.
2310 space_map_histogram_add(msp->ms_sm, *freetree, tx);
2312 metaslab_group_histogram_add(mg, msp);
2313 metaslab_group_histogram_verify(mg);
2314 metaslab_class_histogram_verify(mg->mg_class);
2317 * For sync pass 1, we avoid traversing this txg's free range tree
2318 * and instead will just swap the pointers for freetree and
2319 * freed_tree. We can safely do this since the freed_tree is
2320 * guaranteed to be empty on the initial pass.
2322 if (spa_sync_pass(spa) == 1) {
2323 range_tree_swap(freetree, freed_tree);
2325 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
2327 range_tree_vacate(alloctree, NULL, NULL);
2329 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2330 ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
2331 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2333 mutex_exit(&msp->ms_lock);
2335 if (object != space_map_object(msp->ms_sm)) {
2336 object = space_map_object(msp->ms_sm);
2337 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2338 msp->ms_id, sizeof (uint64_t), &object, tx);
2344 * Called after a transaction group has completely synced to mark
2345 * all of the metaslab's free space as usable.
2348 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2350 metaslab_group_t *mg = msp->ms_group;
2351 vdev_t *vd = mg->mg_vd;
2352 spa_t *spa = vd->vdev_spa;
2353 range_tree_t **freed_tree;
2354 range_tree_t **defer_tree;
2355 int64_t alloc_delta, defer_delta;
2356 uint64_t free_space;
2357 boolean_t defer_allowed = B_TRUE;
2360 ASSERT(!vd->vdev_ishole);
2362 mutex_enter(&msp->ms_lock);
2365 * If this metaslab is just becoming available, initialize its
2366 * alloctrees, freetrees, and defertree and add its capacity to
2369 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
2370 for (t = 0; t < TXG_SIZE; t++) {
2371 ASSERT(msp->ms_alloctree[t] == NULL);
2372 ASSERT(msp->ms_freetree[t] == NULL);
2374 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2376 msp->ms_freetree[t] = range_tree_create(NULL, msp,
2380 for (t = 0; t < TXG_DEFER_SIZE; t++) {
2381 ASSERT(msp->ms_defertree[t] == NULL);
2383 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2387 vdev_space_update(vd, 0, 0, msp->ms_size);
2390 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2391 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2393 free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2394 metaslab_class_get_alloc(spa_normal_class(spa));
2395 if (free_space <= spa_get_slop_space(spa)) {
2396 defer_allowed = B_FALSE;
2400 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2401 if (defer_allowed) {
2402 defer_delta = range_tree_space(*freed_tree) -
2403 range_tree_space(*defer_tree);
2405 defer_delta -= range_tree_space(*defer_tree);
2408 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2410 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2411 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2414 * If there's a metaslab_load() in progress, wait for it to complete
2415 * so that we have a consistent view of the in-core space map.
2417 metaslab_load_wait(msp);
2420 * Move the frees from the defer_tree back to the free
2421 * range tree (if it's loaded). Swap the freed_tree and the
2422 * defer_tree -- this is safe to do because we've just emptied out
2425 range_tree_vacate(*defer_tree,
2426 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2427 if (defer_allowed) {
2428 range_tree_swap(freed_tree, defer_tree);
2430 range_tree_vacate(*freed_tree,
2431 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2434 space_map_update(msp->ms_sm);
2436 msp->ms_deferspace += defer_delta;
2437 ASSERT3S(msp->ms_deferspace, >=, 0);
2438 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2439 if (msp->ms_deferspace != 0) {
2441 * Keep syncing this metaslab until all deferred frees
2442 * are back in circulation.
2444 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2448 * Calculate the new weights before unloading any metaslabs.
2449 * This will give us the most accurate weighting.
2451 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2454 * If the metaslab is loaded and we've not tried to load or allocate
2455 * from it in 'metaslab_unload_delay' txgs, then unload it.
2457 if (msp->ms_loaded &&
2458 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2460 for (t = 1; t < TXG_CONCURRENT_STATES; t++) {
2461 VERIFY0(range_tree_space(
2462 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2465 if (!metaslab_debug_unload)
2466 metaslab_unload(msp);
2469 mutex_exit(&msp->ms_lock);
2473 metaslab_sync_reassess(metaslab_group_t *mg)
2475 metaslab_group_alloc_update(mg);
2476 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2479 * Preload the next potential metaslabs
2481 metaslab_group_preload(mg);
2485 metaslab_distance(metaslab_t *msp, dva_t *dva)
2487 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2488 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2489 uint64_t start = msp->ms_id;
2491 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2492 return (1ULL << 63);
2495 return ((start - offset) << ms_shift);
2497 return ((offset - start) << ms_shift);
2502 * ==========================================================================
2503 * Metaslab allocation tracing facility
2504 * ==========================================================================
2506 #ifdef _METASLAB_TRACING
2507 kstat_t *metaslab_trace_ksp;
2508 kstat_named_t metaslab_trace_over_limit;
2511 metaslab_alloc_trace_init(void)
2513 ASSERT(metaslab_alloc_trace_cache == NULL);
2514 metaslab_alloc_trace_cache = kmem_cache_create(
2515 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
2516 0, NULL, NULL, NULL, NULL, NULL, 0);
2517 metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
2518 "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
2519 if (metaslab_trace_ksp != NULL) {
2520 metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
2521 kstat_named_init(&metaslab_trace_over_limit,
2522 "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
2523 kstat_install(metaslab_trace_ksp);
2528 metaslab_alloc_trace_fini(void)
2530 if (metaslab_trace_ksp != NULL) {
2531 kstat_delete(metaslab_trace_ksp);
2532 metaslab_trace_ksp = NULL;
2534 kmem_cache_destroy(metaslab_alloc_trace_cache);
2535 metaslab_alloc_trace_cache = NULL;
2539 * Add an allocation trace element to the allocation tracing list.
2542 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
2543 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset)
2545 metaslab_alloc_trace_t *mat;
2547 if (!metaslab_trace_enabled)
2551 * When the tracing list reaches its maximum we remove
2552 * the second element in the list before adding a new one.
2553 * By removing the second element we preserve the original
2554 * entry as a clue to what allocations steps have already been
2557 if (zal->zal_size == metaslab_trace_max_entries) {
2558 metaslab_alloc_trace_t *mat_next;
2560 panic("too many entries in allocation list");
2562 atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
2564 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
2565 list_remove(&zal->zal_list, mat_next);
2566 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
2569 mat = kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
2570 list_link_init(&mat->mat_list_node);
2573 mat->mat_size = psize;
2574 mat->mat_dva_id = dva_id;
2575 mat->mat_offset = offset;
2576 mat->mat_weight = 0;
2579 mat->mat_weight = msp->ms_weight;
2582 * The list is part of the zio so locking is not required. Only
2583 * a single thread will perform allocations for a given zio.
2585 list_insert_tail(&zal->zal_list, mat);
2588 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
2592 metaslab_trace_init(zio_alloc_list_t *zal)
2594 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
2595 offsetof(metaslab_alloc_trace_t, mat_list_node));
2600 metaslab_trace_fini(zio_alloc_list_t *zal)
2602 metaslab_alloc_trace_t *mat;
2604 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
2605 kmem_cache_free(metaslab_alloc_trace_cache, mat);
2606 list_destroy(&zal->zal_list);
2611 #define metaslab_trace_add(zal, mg, msp, psize, id, off)
2614 metaslab_alloc_trace_init(void)
2619 metaslab_alloc_trace_fini(void)
2624 metaslab_trace_init(zio_alloc_list_t *zal)
2629 metaslab_trace_fini(zio_alloc_list_t *zal)
2633 #endif /* _METASLAB_TRACING */
2636 * ==========================================================================
2637 * Metaslab block operations
2638 * ==========================================================================
2642 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags)
2644 metaslab_group_t *mg;
2646 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2647 flags & METASLAB_DONT_THROTTLE)
2650 mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2651 if (!mg->mg_class->mc_alloc_throttle_enabled)
2654 (void) refcount_add(&mg->mg_alloc_queue_depth, tag);
2658 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags)
2660 metaslab_group_t *mg;
2662 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2663 flags & METASLAB_DONT_THROTTLE)
2666 mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2667 if (!mg->mg_class->mc_alloc_throttle_enabled)
2670 (void) refcount_remove(&mg->mg_alloc_queue_depth, tag);
2674 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag)
2677 const dva_t *dva = bp->blk_dva;
2678 int ndvas = BP_GET_NDVAS(bp);
2681 for (d = 0; d < ndvas; d++) {
2682 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
2683 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2684 VERIFY(refcount_not_held(&mg->mg_alloc_queue_depth, tag));
2690 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2693 range_tree_t *rt = msp->ms_tree;
2694 metaslab_class_t *mc = msp->ms_group->mg_class;
2696 VERIFY(!msp->ms_condensing);
2698 start = mc->mc_ops->msop_alloc(msp, size);
2699 if (start != -1ULL) {
2700 metaslab_group_t *mg = msp->ms_group;
2701 vdev_t *vd = mg->mg_vd;
2703 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2704 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2705 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2706 range_tree_remove(rt, start, size);
2708 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2709 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2711 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size);
2713 /* Track the last successful allocation */
2714 msp->ms_alloc_txg = txg;
2715 metaslab_verify_space(msp, txg);
2719 * Now that we've attempted the allocation we need to update the
2720 * metaslab's maximum block size since it may have changed.
2722 msp->ms_max_size = metaslab_block_maxsize(msp);
2727 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
2728 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2730 metaslab_t *msp = NULL;
2732 uint64_t offset = -1ULL;
2733 uint64_t activation_weight;
2734 uint64_t target_distance;
2737 activation_weight = METASLAB_WEIGHT_PRIMARY;
2738 for (i = 0; i < d; i++) {
2739 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2740 activation_weight = METASLAB_WEIGHT_SECONDARY;
2745 search = kmem_alloc(sizeof (*search), KM_SLEEP);
2746 search->ms_weight = UINT64_MAX;
2747 search->ms_start = 0;
2749 boolean_t was_active;
2750 avl_tree_t *t = &mg->mg_metaslab_tree;
2753 mutex_enter(&mg->mg_lock);
2756 * Find the metaslab with the highest weight that is less
2757 * than what we've already tried. In the common case, this
2758 * means that we will examine each metaslab at most once.
2759 * Note that concurrent callers could reorder metaslabs
2760 * by activation/passivation once we have dropped the mg_lock.
2761 * If a metaslab is activated by another thread, and we fail
2762 * to allocate from the metaslab we have selected, we may
2763 * not try the newly-activated metaslab, and instead activate
2764 * another metaslab. This is not optimal, but generally
2765 * does not cause any problems (a possible exception being
2766 * if every metaslab is completely full except for the
2767 * the newly-activated metaslab which we fail to examine).
2769 msp = avl_find(t, search, &idx);
2771 msp = avl_nearest(t, idx, AVL_AFTER);
2772 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
2774 if (!metaslab_should_allocate(msp, asize)) {
2775 metaslab_trace_add(zal, mg, msp, asize, d,
2781 * If the selected metaslab is condensing, skip it.
2783 if (msp->ms_condensing)
2786 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2787 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2790 target_distance = min_distance +
2791 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2794 for (i = 0; i < d; i++) {
2795 if (metaslab_distance(msp, &dva[i]) <
2802 mutex_exit(&mg->mg_lock);
2804 kmem_free(search, sizeof (*search));
2807 search->ms_weight = msp->ms_weight;
2808 search->ms_start = msp->ms_start + 1;
2810 mutex_enter(&msp->ms_lock);
2813 * Ensure that the metaslab we have selected is still
2814 * capable of handling our request. It's possible that
2815 * another thread may have changed the weight while we
2816 * were blocked on the metaslab lock. We check the
2817 * active status first to see if we need to reselect
2820 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
2821 mutex_exit(&msp->ms_lock);
2825 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2826 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2827 metaslab_passivate(msp,
2828 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2829 mutex_exit(&msp->ms_lock);
2833 if (metaslab_activate(msp, activation_weight) != 0) {
2834 mutex_exit(&msp->ms_lock);
2837 msp->ms_selected_txg = txg;
2840 * Now that we have the lock, recheck to see if we should
2841 * continue to use this metaslab for this allocation. The
2842 * the metaslab is now loaded so metaslab_should_allocate() can
2843 * accurately determine if the allocation attempt should
2846 if (!metaslab_should_allocate(msp, asize)) {
2847 /* Passivate this metaslab and select a new one. */
2848 metaslab_trace_add(zal, mg, msp, asize, d,
2855 * If this metaslab is currently condensing then pick again as
2856 * we can't manipulate this metaslab until it's committed
2859 if (msp->ms_condensing) {
2860 metaslab_trace_add(zal, mg, msp, asize, d,
2862 mutex_exit(&msp->ms_lock);
2866 offset = metaslab_block_alloc(msp, asize, txg);
2867 metaslab_trace_add(zal, mg, msp, asize, d, offset);
2869 if (offset != -1ULL) {
2870 /* Proactively passivate the metaslab, if needed */
2871 metaslab_segment_may_passivate(msp);
2875 ASSERT(msp->ms_loaded);
2878 * We were unable to allocate from this metaslab so determine
2879 * a new weight for this metaslab. Now that we have loaded
2880 * the metaslab we can provide a better hint to the metaslab
2883 * For space-based metaslabs, we use the maximum block size.
2884 * This information is only available when the metaslab
2885 * is loaded and is more accurate than the generic free
2886 * space weight that was calculated by metaslab_weight().
2887 * This information allows us to quickly compare the maximum
2888 * available allocation in the metaslab to the allocation
2889 * size being requested.
2891 * For segment-based metaslabs, determine the new weight
2892 * based on the highest bucket in the range tree. We
2893 * explicitly use the loaded segment weight (i.e. the range
2894 * tree histogram) since it contains the space that is
2895 * currently available for allocation and is accurate
2896 * even within a sync pass.
2898 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
2899 uint64_t weight = metaslab_block_maxsize(msp);
2900 WEIGHT_SET_SPACEBASED(weight);
2901 metaslab_passivate(msp, weight);
2903 metaslab_passivate(msp,
2904 metaslab_weight_from_range_tree(msp));
2908 * We have just failed an allocation attempt, check
2909 * that metaslab_should_allocate() agrees. Otherwise,
2910 * we may end up in an infinite loop retrying the same
2913 ASSERT(!metaslab_should_allocate(msp, asize));
2914 mutex_exit(&msp->ms_lock);
2916 mutex_exit(&msp->ms_lock);
2917 kmem_free(search, sizeof (*search));
2922 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
2923 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2926 ASSERT(mg->mg_initialized);
2928 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
2929 min_distance, dva, d);
2931 mutex_enter(&mg->mg_lock);
2932 if (offset == -1ULL) {
2933 mg->mg_failed_allocations++;
2934 metaslab_trace_add(zal, mg, NULL, asize, d,
2935 TRACE_GROUP_FAILURE);
2936 if (asize == SPA_GANGBLOCKSIZE) {
2938 * This metaslab group was unable to allocate
2939 * the minimum gang block size so it must be out of
2940 * space. We must notify the allocation throttle
2941 * to start skipping allocation attempts to this
2942 * metaslab group until more space becomes available.
2943 * Note: this failure cannot be caused by the
2944 * allocation throttle since the allocation throttle
2945 * is only responsible for skipping devices and
2946 * not failing block allocations.
2948 mg->mg_no_free_space = B_TRUE;
2951 mg->mg_allocations++;
2952 mutex_exit(&mg->mg_lock);
2957 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
2958 * on the same vdev as an existing DVA of this BP, then try to allocate it
2959 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
2962 int ditto_same_vdev_distance_shift = 3;
2965 * Allocate a block for the specified i/o.
2968 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2969 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
2970 zio_alloc_list_t *zal)
2972 metaslab_group_t *mg, *fast_mg, *rotor;
2974 boolean_t try_hard = B_FALSE;
2976 ASSERT(!DVA_IS_VALID(&dva[d]));
2979 * For testing, make some blocks above a certain size be gang blocks.
2981 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) {
2982 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG);
2983 return (SET_ERROR(ENOSPC));
2987 * Start at the rotor and loop through all mgs until we find something.
2988 * Note that there's no locking on mc_rotor or mc_aliquot because
2989 * nothing actually breaks if we miss a few updates -- we just won't
2990 * allocate quite as evenly. It all balances out over time.
2992 * If we are doing ditto or log blocks, try to spread them across
2993 * consecutive vdevs. If we're forced to reuse a vdev before we've
2994 * allocated all of our ditto blocks, then try and spread them out on
2995 * that vdev as much as possible. If it turns out to not be possible,
2996 * gradually lower our standards until anything becomes acceptable.
2997 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2998 * gives us hope of containing our fault domains to something we're
2999 * able to reason about. Otherwise, any two top-level vdev failures
3000 * will guarantee the loss of data. With consecutive allocation,
3001 * only two adjacent top-level vdev failures will result in data loss.
3003 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3004 * ourselves on the same vdev as our gang block header. That
3005 * way, we can hope for locality in vdev_cache, plus it makes our
3006 * fault domains something tractable.
3009 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3012 * It's possible the vdev we're using as the hint no
3013 * longer exists (i.e. removed). Consult the rotor when
3019 if (flags & METASLAB_HINTBP_AVOID &&
3020 mg->mg_next != NULL)
3025 } else if (d != 0) {
3026 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3027 mg = vd->vdev_mg->mg_next;
3028 } else if (flags & METASLAB_FASTWRITE) {
3029 mg = fast_mg = mc->mc_rotor;
3032 if (fast_mg->mg_vd->vdev_pending_fastwrite <
3033 mg->mg_vd->vdev_pending_fastwrite)
3035 } while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor);
3042 * If the hint put us into the wrong metaslab class, or into a
3043 * metaslab group that has been passivated, just follow the rotor.
3045 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3051 boolean_t allocatable;
3053 uint64_t distance, asize;
3055 ASSERT(mg->mg_activation_count == 1);
3059 * Don't allocate from faulted devices.
3062 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
3063 allocatable = vdev_allocatable(vd);
3064 spa_config_exit(spa, SCL_ZIO, FTAG);
3066 allocatable = vdev_allocatable(vd);
3070 * Determine if the selected metaslab group is eligible
3071 * for allocations. If we're ganging then don't allow
3072 * this metaslab group to skip allocations since that would
3073 * inadvertently return ENOSPC and suspend the pool
3074 * even though space is still available.
3076 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
3077 allocatable = metaslab_group_allocatable(mg, rotor,
3082 metaslab_trace_add(zal, mg, NULL, psize, d,
3083 TRACE_NOT_ALLOCATABLE);
3087 ASSERT(mg->mg_initialized);
3090 * Avoid writing single-copy data to a failing,
3091 * non-redundant vdev, unless we've already tried all
3094 if ((vd->vdev_stat.vs_write_errors > 0 ||
3095 vd->vdev_state < VDEV_STATE_HEALTHY) &&
3096 d == 0 && !try_hard && vd->vdev_children == 0) {
3097 metaslab_trace_add(zal, mg, NULL, psize, d,
3102 ASSERT(mg->mg_class == mc);
3105 * If we don't need to try hard, then require that the
3106 * block be 1/8th of the device away from any other DVAs
3107 * in this BP. If we are trying hard, allow any offset
3108 * to be used (distance=0).
3112 distance = vd->vdev_asize >>
3113 ditto_same_vdev_distance_shift;
3114 if (distance <= (1ULL << vd->vdev_ms_shift))
3118 asize = vdev_psize_to_asize(vd, psize);
3119 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3121 offset = metaslab_group_alloc(mg, zal, asize, txg, distance,
3124 if (offset != -1ULL) {
3126 * If we've just selected this metaslab group,
3127 * figure out whether the corresponding vdev is
3128 * over- or under-used relative to the pool,
3129 * and set an allocation bias to even it out.
3131 * Bias is also used to compensate for unequally
3132 * sized vdevs so that space is allocated fairly.
3134 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3135 vdev_stat_t *vs = &vd->vdev_stat;
3136 int64_t vs_free = vs->vs_space - vs->vs_alloc;
3137 int64_t mc_free = mc->mc_space - mc->mc_alloc;
3141 * Calculate how much more or less we should
3142 * try to allocate from this device during
3143 * this iteration around the rotor.
3145 * This basically introduces a zero-centered
3146 * bias towards the devices with the most
3147 * free space, while compensating for vdev
3151 * vdev V1 = 16M/128M
3152 * vdev V2 = 16M/128M
3153 * ratio(V1) = 100% ratio(V2) = 100%
3155 * vdev V1 = 16M/128M
3156 * vdev V2 = 64M/128M
3157 * ratio(V1) = 127% ratio(V2) = 72%
3159 * vdev V1 = 16M/128M
3160 * vdev V2 = 64M/512M
3161 * ratio(V1) = 40% ratio(V2) = 160%
3163 ratio = (vs_free * mc->mc_alloc_groups * 100) /
3165 mg->mg_bias = ((ratio - 100) *
3166 (int64_t)mg->mg_aliquot) / 100;
3167 } else if (!metaslab_bias_enabled) {
3171 if ((flags & METASLAB_FASTWRITE) ||
3172 atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3173 mg->mg_aliquot + mg->mg_bias) {
3174 mc->mc_rotor = mg->mg_next;
3178 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3179 DVA_SET_OFFSET(&dva[d], offset);
3180 DVA_SET_GANG(&dva[d],
3181 ((flags & METASLAB_GANG_HEADER) ? 1 : 0));
3182 DVA_SET_ASIZE(&dva[d], asize);
3184 if (flags & METASLAB_FASTWRITE) {
3185 atomic_add_64(&vd->vdev_pending_fastwrite,
3192 mc->mc_rotor = mg->mg_next;
3194 } while ((mg = mg->mg_next) != rotor);
3197 * If we haven't tried hard, do so now.
3204 bzero(&dva[d], sizeof (dva_t));
3206 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC);
3207 return (SET_ERROR(ENOSPC));
3211 * Free the block represented by DVA in the context of the specified
3212 * transaction group.
3215 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
3217 uint64_t vdev = DVA_GET_VDEV(dva);
3218 uint64_t offset = DVA_GET_OFFSET(dva);
3219 uint64_t size = DVA_GET_ASIZE(dva);
3223 if (txg > spa_freeze_txg(spa))
3226 if ((vd = vdev_lookup_top(spa, vdev)) == NULL || !DVA_IS_VALID(dva) ||
3227 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3228 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
3229 (u_longlong_t)vdev, (u_longlong_t)offset,
3230 (u_longlong_t)size);
3234 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3236 if (DVA_GET_GANG(dva))
3237 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3239 mutex_enter(&msp->ms_lock);
3242 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
3245 VERIFY(!msp->ms_condensing);
3246 VERIFY3U(offset, >=, msp->ms_start);
3247 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3248 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
3250 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3251 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3252 range_tree_add(msp->ms_tree, offset, size);
3253 msp->ms_max_size = metaslab_block_maxsize(msp);
3255 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
3256 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3257 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
3261 mutex_exit(&msp->ms_lock);
3265 * Intent log support: upon opening the pool after a crash, notify the SPA
3266 * of blocks that the intent log has allocated for immediate write, but
3267 * which are still considered free by the SPA because the last transaction
3268 * group didn't commit yet.
3271 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3273 uint64_t vdev = DVA_GET_VDEV(dva);
3274 uint64_t offset = DVA_GET_OFFSET(dva);
3275 uint64_t size = DVA_GET_ASIZE(dva);
3280 ASSERT(DVA_IS_VALID(dva));
3282 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3283 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
3284 return (SET_ERROR(ENXIO));
3286 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3288 if (DVA_GET_GANG(dva))
3289 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3291 mutex_enter(&msp->ms_lock);
3293 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3294 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
3296 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
3297 error = SET_ERROR(ENOENT);
3299 if (error || txg == 0) { /* txg == 0 indicates dry run */
3300 mutex_exit(&msp->ms_lock);
3304 VERIFY(!msp->ms_condensing);
3305 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3306 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3307 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
3308 range_tree_remove(msp->ms_tree, offset, size);
3310 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
3311 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
3312 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3313 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
3316 mutex_exit(&msp->ms_lock);
3322 * Reserve some allocation slots. The reservation system must be called
3323 * before we call into the allocator. If there aren't any available slots
3324 * then the I/O will be throttled until an I/O completes and its slots are
3325 * freed up. The function returns true if it was successful in placing
3329 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio,
3332 uint64_t available_slots = 0;
3333 uint64_t reserved_slots;
3334 boolean_t slot_reserved = B_FALSE;
3336 ASSERT(mc->mc_alloc_throttle_enabled);
3337 mutex_enter(&mc->mc_lock);
3339 reserved_slots = refcount_count(&mc->mc_alloc_slots);
3340 if (reserved_slots < mc->mc_alloc_max_slots)
3341 available_slots = mc->mc_alloc_max_slots - reserved_slots;
3343 if (slots <= available_slots || GANG_ALLOCATION(flags)) {
3347 * We reserve the slots individually so that we can unreserve
3348 * them individually when an I/O completes.
3350 for (d = 0; d < slots; d++) {
3351 reserved_slots = refcount_add(&mc->mc_alloc_slots, zio);
3353 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3354 slot_reserved = B_TRUE;
3357 mutex_exit(&mc->mc_lock);
3358 return (slot_reserved);
3362 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio)
3366 ASSERT(mc->mc_alloc_throttle_enabled);
3367 mutex_enter(&mc->mc_lock);
3368 for (d = 0; d < slots; d++) {
3369 (void) refcount_remove(&mc->mc_alloc_slots, zio);
3371 mutex_exit(&mc->mc_lock);
3375 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
3376 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
3377 zio_alloc_list_t *zal, zio_t *zio)
3379 dva_t *dva = bp->blk_dva;
3380 dva_t *hintdva = hintbp->blk_dva;
3383 ASSERT(bp->blk_birth == 0);
3384 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
3386 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3388 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
3389 spa_config_exit(spa, SCL_ALLOC, FTAG);
3390 return (SET_ERROR(ENOSPC));
3393 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
3394 ASSERT(BP_GET_NDVAS(bp) == 0);
3395 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
3396 ASSERT3P(zal, !=, NULL);
3398 for (d = 0; d < ndvas; d++) {
3399 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
3402 for (d--; d >= 0; d--) {
3403 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
3404 metaslab_group_alloc_decrement(spa,
3405 DVA_GET_VDEV(&dva[d]), zio, flags);
3406 bzero(&dva[d], sizeof (dva_t));
3408 spa_config_exit(spa, SCL_ALLOC, FTAG);
3412 * Update the metaslab group's queue depth
3413 * based on the newly allocated dva.
3415 metaslab_group_alloc_increment(spa,
3416 DVA_GET_VDEV(&dva[d]), zio, flags);
3421 ASSERT(BP_GET_NDVAS(bp) == ndvas);
3423 spa_config_exit(spa, SCL_ALLOC, FTAG);
3425 BP_SET_BIRTH(bp, txg, 0);
3431 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
3433 const dva_t *dva = bp->blk_dva;
3434 int d, ndvas = BP_GET_NDVAS(bp);
3436 ASSERT(!BP_IS_HOLE(bp));
3437 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
3439 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
3441 for (d = 0; d < ndvas; d++)
3442 metaslab_free_dva(spa, &dva[d], txg, now);
3444 spa_config_exit(spa, SCL_FREE, FTAG);
3448 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
3450 const dva_t *dva = bp->blk_dva;
3451 int ndvas = BP_GET_NDVAS(bp);
3454 ASSERT(!BP_IS_HOLE(bp));
3458 * First do a dry run to make sure all DVAs are claimable,
3459 * so we don't have to unwind from partial failures below.
3461 if ((error = metaslab_claim(spa, bp, 0)) != 0)
3465 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3467 for (d = 0; d < ndvas; d++)
3468 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
3471 spa_config_exit(spa, SCL_ALLOC, FTAG);
3473 ASSERT(error == 0 || txg == 0);
3479 metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp)
3481 const dva_t *dva = bp->blk_dva;
3482 int ndvas = BP_GET_NDVAS(bp);
3483 uint64_t psize = BP_GET_PSIZE(bp);
3487 ASSERT(!BP_IS_HOLE(bp));
3488 ASSERT(!BP_IS_EMBEDDED(bp));
3491 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3493 for (d = 0; d < ndvas; d++) {
3494 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
3496 atomic_add_64(&vd->vdev_pending_fastwrite, psize);
3499 spa_config_exit(spa, SCL_VDEV, FTAG);
3503 metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp)
3505 const dva_t *dva = bp->blk_dva;
3506 int ndvas = BP_GET_NDVAS(bp);
3507 uint64_t psize = BP_GET_PSIZE(bp);
3511 ASSERT(!BP_IS_HOLE(bp));
3512 ASSERT(!BP_IS_EMBEDDED(bp));
3515 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3517 for (d = 0; d < ndvas; d++) {
3518 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
3520 ASSERT3U(vd->vdev_pending_fastwrite, >=, psize);
3521 atomic_sub_64(&vd->vdev_pending_fastwrite, psize);
3524 spa_config_exit(spa, SCL_VDEV, FTAG);
3528 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
3532 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3535 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3536 for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3537 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
3538 vdev_t *vd = vdev_lookup_top(spa, vdev);
3539 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
3540 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
3541 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3544 range_tree_verify(msp->ms_tree, offset, size);
3546 for (j = 0; j < TXG_SIZE; j++)
3547 range_tree_verify(msp->ms_freetree[j], offset, size);
3548 for (j = 0; j < TXG_DEFER_SIZE; j++)
3549 range_tree_verify(msp->ms_defertree[j], offset, size);
3551 spa_config_exit(spa, SCL_VDEV, FTAG);
3554 #if defined(_KERNEL) && defined(HAVE_SPL)
3556 module_param(metaslab_aliquot, ulong, 0644);
3557 MODULE_PARM_DESC(metaslab_aliquot,
3558 "allocation granularity (a.k.a. stripe size)");
3560 module_param(metaslab_debug_load, int, 0644);
3561 MODULE_PARM_DESC(metaslab_debug_load,
3562 "load all metaslabs when pool is first opened");
3564 module_param(metaslab_debug_unload, int, 0644);
3565 MODULE_PARM_DESC(metaslab_debug_unload,
3566 "prevent metaslabs from being unloaded");
3568 module_param(metaslab_preload_enabled, int, 0644);
3569 MODULE_PARM_DESC(metaslab_preload_enabled,
3570 "preload potential metaslabs during reassessment");
3572 module_param(zfs_mg_noalloc_threshold, int, 0644);
3573 MODULE_PARM_DESC(zfs_mg_noalloc_threshold,
3574 "percentage of free space for metaslab group to allow allocation");
3576 module_param(zfs_mg_fragmentation_threshold, int, 0644);
3577 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold,
3578 "fragmentation for metaslab group to allow allocation");
3580 module_param(zfs_metaslab_fragmentation_threshold, int, 0644);
3581 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold,
3582 "fragmentation for metaslab to allow allocation");
3584 module_param(metaslab_fragmentation_factor_enabled, int, 0644);
3585 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled,
3586 "use the fragmentation metric to prefer less fragmented metaslabs");
3588 module_param(metaslab_lba_weighting_enabled, int, 0644);
3589 MODULE_PARM_DESC(metaslab_lba_weighting_enabled,
3590 "prefer metaslabs with lower LBAs");
3592 module_param(metaslab_bias_enabled, int, 0644);
3593 MODULE_PARM_DESC(metaslab_bias_enabled,
3594 "enable metaslab group biasing");
3596 module_param(zfs_metaslab_segment_weight_enabled, int, 0644);
3597 MODULE_PARM_DESC(zfs_metaslab_segment_weight_enabled,
3598 "enable segment-based metaslab selection");
3600 module_param(zfs_metaslab_switch_threshold, int, 0644);
3601 MODULE_PARM_DESC(zfs_metaslab_switch_threshold,
3602 "segment-based metaslab selection maximum buckets before switching");
3603 #endif /* _KERNEL && HAVE_SPL */