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, 2014 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 SYSCTL_DECL(_vfs_zfs);
38 SYSCTL_NODE(_vfs_zfs, OID_AUTO, metaslab, CTLFLAG_RW, 0, "ZFS metaslab");
41 * Allow allocations to switch to gang blocks quickly. We do this to
42 * avoid having to load lots of space_maps in a given txg. There are,
43 * however, some cases where we want to avoid "fast" ganging and instead
44 * we want to do an exhaustive search of all metaslabs on this device.
45 * Currently we don't allow any gang, slog, or dump device related allocations
48 #define CAN_FASTGANG(flags) \
49 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
50 METASLAB_GANG_AVOID)))
52 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
53 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
54 #define METASLAB_ACTIVE_MASK \
55 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
57 uint64_t metaslab_aliquot = 512ULL << 10;
58 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
59 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, gang_bang, CTLFLAG_RWTUN,
60 &metaslab_gang_bang, 0,
61 "Force gang block allocation for blocks larger than or equal to this value");
64 * The in-core space map representation is more compact than its on-disk form.
65 * The zfs_condense_pct determines how much more compact the in-core
66 * space_map representation must be before we compact it on-disk.
67 * Values should be greater than or equal to 100.
69 int zfs_condense_pct = 200;
70 SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN,
72 "Condense on-disk spacemap when it is more than this many percents"
73 " of in-memory counterpart");
76 * Condensing a metaslab is not guaranteed to actually reduce the amount of
77 * space used on disk. In particular, a space map uses data in increments of
78 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
79 * same number of blocks after condensing. Since the goal of condensing is to
80 * reduce the number of IOPs required to read the space map, we only want to
81 * condense when we can be sure we will reduce the number of blocks used by the
82 * space map. Unfortunately, we cannot precisely compute whether or not this is
83 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
84 * we apply the following heuristic: do not condense a spacemap unless the
85 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
88 int zfs_metaslab_condense_block_threshold = 4;
91 * The zfs_mg_noalloc_threshold defines which metaslab groups should
92 * be eligible for allocation. The value is defined as a percentage of
93 * free space. Metaslab groups that have more free space than
94 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
95 * a metaslab group's free space is less than or equal to the
96 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
97 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
98 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
99 * groups are allowed to accept allocations. Gang blocks are always
100 * eligible to allocate on any metaslab group. The default value of 0 means
101 * no metaslab group will be excluded based on this criterion.
103 int zfs_mg_noalloc_threshold = 0;
104 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN,
105 &zfs_mg_noalloc_threshold, 0,
106 "Percentage of metaslab group size that should be free"
107 " to make it eligible for allocation");
110 * Metaslab groups are considered eligible for allocations if their
111 * fragmenation metric (measured as a percentage) is less than or equal to
112 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
113 * then it will be skipped unless all metaslab groups within the metaslab
114 * class have also crossed this threshold.
116 int zfs_mg_fragmentation_threshold = 85;
117 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_fragmentation_threshold, CTLFLAG_RWTUN,
118 &zfs_mg_fragmentation_threshold, 0,
119 "Percentage of metaslab group size that should be considered "
120 "eligible for allocations unless all metaslab groups within the metaslab class "
121 "have also crossed this threshold");
124 * Allow metaslabs to keep their active state as long as their fragmentation
125 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
126 * active metaslab that exceeds this threshold will no longer keep its active
127 * status allowing better metaslabs to be selected.
129 int zfs_metaslab_fragmentation_threshold = 70;
130 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_threshold, CTLFLAG_RWTUN,
131 &zfs_metaslab_fragmentation_threshold, 0,
132 "Maximum percentage of metaslab fragmentation level to keep their active state");
135 * When set will load all metaslabs when pool is first opened.
137 int metaslab_debug_load = 0;
138 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN,
139 &metaslab_debug_load, 0,
140 "Load all metaslabs when pool is first opened");
143 * When set will prevent metaslabs from being unloaded.
145 int metaslab_debug_unload = 0;
146 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN,
147 &metaslab_debug_unload, 0,
148 "Prevent metaslabs from being unloaded");
151 * Minimum size which forces the dynamic allocator to change
152 * it's allocation strategy. Once the space map cannot satisfy
153 * an allocation of this size then it switches to using more
154 * aggressive strategy (i.e search by size rather than offset).
156 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
157 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN,
158 &metaslab_df_alloc_threshold, 0,
159 "Minimum size which forces the dynamic allocator to change it's allocation strategy");
162 * The minimum free space, in percent, which must be available
163 * in a space map to continue allocations in a first-fit fashion.
164 * Once the space_map's free space drops below this level we dynamically
165 * switch to using best-fit allocations.
167 int metaslab_df_free_pct = 4;
168 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN,
169 &metaslab_df_free_pct, 0,
170 "The minimum free space, in percent, which must be available in a "
171 "space map to continue allocations in a first-fit fashion");
174 * A metaslab is considered "free" if it contains a contiguous
175 * segment which is greater than metaslab_min_alloc_size.
177 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
178 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN,
179 &metaslab_min_alloc_size, 0,
180 "A metaslab is considered \"free\" if it contains a contiguous "
181 "segment which is greater than vfs.zfs.metaslab.min_alloc_size");
184 * Percentage of all cpus that can be used by the metaslab taskq.
186 int metaslab_load_pct = 50;
187 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN,
188 &metaslab_load_pct, 0,
189 "Percentage of cpus that can be used by the metaslab taskq");
192 * Determines how many txgs a metaslab may remain loaded without having any
193 * allocations from it. As long as a metaslab continues to be used we will
196 int metaslab_unload_delay = TXG_SIZE * 2;
197 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN,
198 &metaslab_unload_delay, 0,
199 "Number of TXGs that an unused metaslab can be kept in memory");
202 * Max number of metaslabs per group to preload.
204 int metaslab_preload_limit = SPA_DVAS_PER_BP;
205 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
206 &metaslab_preload_limit, 0,
207 "Max number of metaslabs per group to preload");
210 * Enable/disable preloading of metaslab.
212 boolean_t metaslab_preload_enabled = B_TRUE;
213 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
214 &metaslab_preload_enabled, 0,
215 "Max number of metaslabs per group to preload");
218 * Enable/disable fragmentation weighting on metaslabs.
220 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
221 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_factor_enabled, CTLFLAG_RWTUN,
222 &metaslab_fragmentation_factor_enabled, 0,
223 "Enable fragmentation weighting on metaslabs");
226 * Enable/disable lba weighting (i.e. outer tracks are given preference).
228 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
229 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, lba_weighting_enabled, CTLFLAG_RWTUN,
230 &metaslab_lba_weighting_enabled, 0,
231 "Enable LBA weighting (i.e. outer tracks are given preference)");
234 * Enable/disable metaslab group biasing.
236 boolean_t metaslab_bias_enabled = B_TRUE;
237 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, bias_enabled, CTLFLAG_RWTUN,
238 &metaslab_bias_enabled, 0,
239 "Enable metaslab group biasing");
241 static uint64_t metaslab_fragmentation(metaslab_t *);
244 * ==========================================================================
246 * ==========================================================================
249 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
251 metaslab_class_t *mc;
253 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
263 metaslab_class_destroy(metaslab_class_t *mc)
265 ASSERT(mc->mc_rotor == NULL);
266 ASSERT(mc->mc_alloc == 0);
267 ASSERT(mc->mc_deferred == 0);
268 ASSERT(mc->mc_space == 0);
269 ASSERT(mc->mc_dspace == 0);
271 kmem_free(mc, sizeof (metaslab_class_t));
275 metaslab_class_validate(metaslab_class_t *mc)
277 metaslab_group_t *mg;
281 * Must hold one of the spa_config locks.
283 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
284 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
286 if ((mg = mc->mc_rotor) == NULL)
291 ASSERT(vd->vdev_mg != NULL);
292 ASSERT3P(vd->vdev_top, ==, vd);
293 ASSERT3P(mg->mg_class, ==, mc);
294 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
295 } while ((mg = mg->mg_next) != mc->mc_rotor);
301 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
302 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
304 atomic_add_64(&mc->mc_alloc, alloc_delta);
305 atomic_add_64(&mc->mc_deferred, defer_delta);
306 atomic_add_64(&mc->mc_space, space_delta);
307 atomic_add_64(&mc->mc_dspace, dspace_delta);
311 metaslab_class_minblocksize_update(metaslab_class_t *mc)
313 metaslab_group_t *mg;
315 uint64_t minashift = UINT64_MAX;
317 if ((mg = mc->mc_rotor) == NULL) {
318 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
324 if (vd->vdev_ashift < minashift)
325 minashift = vd->vdev_ashift;
326 } while ((mg = mg->mg_next) != mc->mc_rotor);
328 mc->mc_minblocksize = 1ULL << minashift;
332 metaslab_class_get_alloc(metaslab_class_t *mc)
334 return (mc->mc_alloc);
338 metaslab_class_get_deferred(metaslab_class_t *mc)
340 return (mc->mc_deferred);
344 metaslab_class_get_space(metaslab_class_t *mc)
346 return (mc->mc_space);
350 metaslab_class_get_dspace(metaslab_class_t *mc)
352 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
356 metaslab_class_get_minblocksize(metaslab_class_t *mc)
358 return (mc->mc_minblocksize);
362 metaslab_class_histogram_verify(metaslab_class_t *mc)
364 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
368 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
371 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
374 for (int c = 0; c < rvd->vdev_children; c++) {
375 vdev_t *tvd = rvd->vdev_child[c];
376 metaslab_group_t *mg = tvd->vdev_mg;
379 * Skip any holes, uninitialized top-levels, or
380 * vdevs that are not in this metalab class.
382 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
383 mg->mg_class != mc) {
387 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
388 mc_hist[i] += mg->mg_histogram[i];
391 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
392 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
394 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
398 * Calculate the metaslab class's fragmentation metric. The metric
399 * is weighted based on the space contribution of each metaslab group.
400 * The return value will be a number between 0 and 100 (inclusive), or
401 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
402 * zfs_frag_table for more information about the metric.
405 metaslab_class_fragmentation(metaslab_class_t *mc)
407 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
408 uint64_t fragmentation = 0;
410 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
412 for (int c = 0; c < rvd->vdev_children; c++) {
413 vdev_t *tvd = rvd->vdev_child[c];
414 metaslab_group_t *mg = tvd->vdev_mg;
417 * Skip any holes, uninitialized top-levels, or
418 * vdevs that are not in this metalab class.
420 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
421 mg->mg_class != mc) {
426 * If a metaslab group does not contain a fragmentation
427 * metric then just bail out.
429 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
430 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
431 return (ZFS_FRAG_INVALID);
435 * Determine how much this metaslab_group is contributing
436 * to the overall pool fragmentation metric.
438 fragmentation += mg->mg_fragmentation *
439 metaslab_group_get_space(mg);
441 fragmentation /= metaslab_class_get_space(mc);
443 ASSERT3U(fragmentation, <=, 100);
444 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
445 return (fragmentation);
449 * Calculate the amount of expandable space that is available in
450 * this metaslab class. If a device is expanded then its expandable
451 * space will be the amount of allocatable space that is currently not
452 * part of this metaslab class.
455 metaslab_class_expandable_space(metaslab_class_t *mc)
457 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
460 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
461 for (int c = 0; c < rvd->vdev_children; c++) {
462 vdev_t *tvd = rvd->vdev_child[c];
463 metaslab_group_t *mg = tvd->vdev_mg;
465 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
466 mg->mg_class != mc) {
470 space += tvd->vdev_max_asize - tvd->vdev_asize;
472 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
477 * ==========================================================================
479 * ==========================================================================
482 metaslab_compare(const void *x1, const void *x2)
484 const metaslab_t *m1 = x1;
485 const metaslab_t *m2 = x2;
487 if (m1->ms_weight < m2->ms_weight)
489 if (m1->ms_weight > m2->ms_weight)
493 * If the weights are identical, use the offset to force uniqueness.
495 if (m1->ms_start < m2->ms_start)
497 if (m1->ms_start > m2->ms_start)
500 ASSERT3P(m1, ==, m2);
506 * Update the allocatable flag and the metaslab group's capacity.
507 * The allocatable flag is set to true if the capacity is below
508 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
509 * from allocatable to non-allocatable or vice versa then the metaslab
510 * group's class is updated to reflect the transition.
513 metaslab_group_alloc_update(metaslab_group_t *mg)
515 vdev_t *vd = mg->mg_vd;
516 metaslab_class_t *mc = mg->mg_class;
517 vdev_stat_t *vs = &vd->vdev_stat;
518 boolean_t was_allocatable;
520 ASSERT(vd == vd->vdev_top);
522 mutex_enter(&mg->mg_lock);
523 was_allocatable = mg->mg_allocatable;
525 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
529 * A metaslab group is considered allocatable if it has plenty
530 * of free space or is not heavily fragmented. We only take
531 * fragmentation into account if the metaslab group has a valid
532 * fragmentation metric (i.e. a value between 0 and 100).
534 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
535 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
536 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
539 * The mc_alloc_groups maintains a count of the number of
540 * groups in this metaslab class that are still above the
541 * zfs_mg_noalloc_threshold. This is used by the allocating
542 * threads to determine if they should avoid allocations to
543 * a given group. The allocator will avoid allocations to a group
544 * if that group has reached or is below the zfs_mg_noalloc_threshold
545 * and there are still other groups that are above the threshold.
546 * When a group transitions from allocatable to non-allocatable or
547 * vice versa we update the metaslab class to reflect that change.
548 * When the mc_alloc_groups value drops to 0 that means that all
549 * groups have reached the zfs_mg_noalloc_threshold making all groups
550 * eligible for allocations. This effectively means that all devices
551 * are balanced again.
553 if (was_allocatable && !mg->mg_allocatable)
554 mc->mc_alloc_groups--;
555 else if (!was_allocatable && mg->mg_allocatable)
556 mc->mc_alloc_groups++;
558 mutex_exit(&mg->mg_lock);
562 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
564 metaslab_group_t *mg;
566 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
567 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
568 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
569 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
572 mg->mg_activation_count = 0;
574 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
575 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
581 metaslab_group_destroy(metaslab_group_t *mg)
583 ASSERT(mg->mg_prev == NULL);
584 ASSERT(mg->mg_next == NULL);
586 * We may have gone below zero with the activation count
587 * either because we never activated in the first place or
588 * because we're done, and possibly removing the vdev.
590 ASSERT(mg->mg_activation_count <= 0);
592 taskq_destroy(mg->mg_taskq);
593 avl_destroy(&mg->mg_metaslab_tree);
594 mutex_destroy(&mg->mg_lock);
595 kmem_free(mg, sizeof (metaslab_group_t));
599 metaslab_group_activate(metaslab_group_t *mg)
601 metaslab_class_t *mc = mg->mg_class;
602 metaslab_group_t *mgprev, *mgnext;
604 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
606 ASSERT(mc->mc_rotor != mg);
607 ASSERT(mg->mg_prev == NULL);
608 ASSERT(mg->mg_next == NULL);
609 ASSERT(mg->mg_activation_count <= 0);
611 if (++mg->mg_activation_count <= 0)
614 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
615 metaslab_group_alloc_update(mg);
617 if ((mgprev = mc->mc_rotor) == NULL) {
621 mgnext = mgprev->mg_next;
622 mg->mg_prev = mgprev;
623 mg->mg_next = mgnext;
624 mgprev->mg_next = mg;
625 mgnext->mg_prev = mg;
628 metaslab_class_minblocksize_update(mc);
632 metaslab_group_passivate(metaslab_group_t *mg)
634 metaslab_class_t *mc = mg->mg_class;
635 metaslab_group_t *mgprev, *mgnext;
637 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
639 if (--mg->mg_activation_count != 0) {
640 ASSERT(mc->mc_rotor != mg);
641 ASSERT(mg->mg_prev == NULL);
642 ASSERT(mg->mg_next == NULL);
643 ASSERT(mg->mg_activation_count < 0);
647 taskq_wait(mg->mg_taskq);
648 metaslab_group_alloc_update(mg);
650 mgprev = mg->mg_prev;
651 mgnext = mg->mg_next;
656 mc->mc_rotor = mgnext;
657 mgprev->mg_next = mgnext;
658 mgnext->mg_prev = mgprev;
663 metaslab_class_minblocksize_update(mc);
667 metaslab_group_get_space(metaslab_group_t *mg)
669 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
673 metaslab_group_histogram_verify(metaslab_group_t *mg)
676 vdev_t *vd = mg->mg_vd;
677 uint64_t ashift = vd->vdev_ashift;
680 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
683 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
686 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
687 SPACE_MAP_HISTOGRAM_SIZE + ashift);
689 for (int m = 0; m < vd->vdev_ms_count; m++) {
690 metaslab_t *msp = vd->vdev_ms[m];
692 if (msp->ms_sm == NULL)
695 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
696 mg_hist[i + ashift] +=
697 msp->ms_sm->sm_phys->smp_histogram[i];
700 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
701 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
703 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
707 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
709 metaslab_class_t *mc = mg->mg_class;
710 uint64_t ashift = mg->mg_vd->vdev_ashift;
712 ASSERT(MUTEX_HELD(&msp->ms_lock));
713 if (msp->ms_sm == NULL)
716 mutex_enter(&mg->mg_lock);
717 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
718 mg->mg_histogram[i + ashift] +=
719 msp->ms_sm->sm_phys->smp_histogram[i];
720 mc->mc_histogram[i + ashift] +=
721 msp->ms_sm->sm_phys->smp_histogram[i];
723 mutex_exit(&mg->mg_lock);
727 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
729 metaslab_class_t *mc = mg->mg_class;
730 uint64_t ashift = mg->mg_vd->vdev_ashift;
732 ASSERT(MUTEX_HELD(&msp->ms_lock));
733 if (msp->ms_sm == NULL)
736 mutex_enter(&mg->mg_lock);
737 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
738 ASSERT3U(mg->mg_histogram[i + ashift], >=,
739 msp->ms_sm->sm_phys->smp_histogram[i]);
740 ASSERT3U(mc->mc_histogram[i + ashift], >=,
741 msp->ms_sm->sm_phys->smp_histogram[i]);
743 mg->mg_histogram[i + ashift] -=
744 msp->ms_sm->sm_phys->smp_histogram[i];
745 mc->mc_histogram[i + ashift] -=
746 msp->ms_sm->sm_phys->smp_histogram[i];
748 mutex_exit(&mg->mg_lock);
752 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
754 ASSERT(msp->ms_group == NULL);
755 mutex_enter(&mg->mg_lock);
758 avl_add(&mg->mg_metaslab_tree, msp);
759 mutex_exit(&mg->mg_lock);
761 mutex_enter(&msp->ms_lock);
762 metaslab_group_histogram_add(mg, msp);
763 mutex_exit(&msp->ms_lock);
767 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
769 mutex_enter(&msp->ms_lock);
770 metaslab_group_histogram_remove(mg, msp);
771 mutex_exit(&msp->ms_lock);
773 mutex_enter(&mg->mg_lock);
774 ASSERT(msp->ms_group == mg);
775 avl_remove(&mg->mg_metaslab_tree, msp);
776 msp->ms_group = NULL;
777 mutex_exit(&mg->mg_lock);
781 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
784 * Although in principle the weight can be any value, in
785 * practice we do not use values in the range [1, 511].
787 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
788 ASSERT(MUTEX_HELD(&msp->ms_lock));
790 mutex_enter(&mg->mg_lock);
791 ASSERT(msp->ms_group == mg);
792 avl_remove(&mg->mg_metaslab_tree, msp);
793 msp->ms_weight = weight;
794 avl_add(&mg->mg_metaslab_tree, msp);
795 mutex_exit(&mg->mg_lock);
799 * Calculate the fragmentation for a given metaslab group. We can use
800 * a simple average here since all metaslabs within the group must have
801 * the same size. The return value will be a value between 0 and 100
802 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
803 * group have a fragmentation metric.
806 metaslab_group_fragmentation(metaslab_group_t *mg)
808 vdev_t *vd = mg->mg_vd;
809 uint64_t fragmentation = 0;
810 uint64_t valid_ms = 0;
812 for (int m = 0; m < vd->vdev_ms_count; m++) {
813 metaslab_t *msp = vd->vdev_ms[m];
815 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
819 fragmentation += msp->ms_fragmentation;
822 if (valid_ms <= vd->vdev_ms_count / 2)
823 return (ZFS_FRAG_INVALID);
825 fragmentation /= valid_ms;
826 ASSERT3U(fragmentation, <=, 100);
827 return (fragmentation);
831 * Determine if a given metaslab group should skip allocations. A metaslab
832 * group should avoid allocations if its free capacity is less than the
833 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
834 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
835 * that can still handle allocations.
838 metaslab_group_allocatable(metaslab_group_t *mg)
840 vdev_t *vd = mg->mg_vd;
841 spa_t *spa = vd->vdev_spa;
842 metaslab_class_t *mc = mg->mg_class;
845 * We use two key metrics to determine if a metaslab group is
846 * considered allocatable -- free space and fragmentation. If
847 * the free space is greater than the free space threshold and
848 * the fragmentation is less than the fragmentation threshold then
849 * consider the group allocatable. There are two case when we will
850 * not consider these key metrics. The first is if the group is
851 * associated with a slog device and the second is if all groups
852 * in this metaslab class have already been consider ineligible
855 return ((mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
856 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
857 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)) ||
858 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
862 * ==========================================================================
863 * Range tree callbacks
864 * ==========================================================================
868 * Comparison function for the private size-ordered tree. Tree is sorted
869 * by size, larger sizes at the end of the tree.
872 metaslab_rangesize_compare(const void *x1, const void *x2)
874 const range_seg_t *r1 = x1;
875 const range_seg_t *r2 = x2;
876 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
877 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
879 if (rs_size1 < rs_size2)
881 if (rs_size1 > rs_size2)
884 if (r1->rs_start < r2->rs_start)
887 if (r1->rs_start > r2->rs_start)
894 * Create any block allocator specific components. The current allocators
895 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
898 metaslab_rt_create(range_tree_t *rt, void *arg)
900 metaslab_t *msp = arg;
902 ASSERT3P(rt->rt_arg, ==, msp);
903 ASSERT(msp->ms_tree == NULL);
905 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
906 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
910 * Destroy the block allocator specific components.
913 metaslab_rt_destroy(range_tree_t *rt, void *arg)
915 metaslab_t *msp = arg;
917 ASSERT3P(rt->rt_arg, ==, msp);
918 ASSERT3P(msp->ms_tree, ==, rt);
919 ASSERT0(avl_numnodes(&msp->ms_size_tree));
921 avl_destroy(&msp->ms_size_tree);
925 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
927 metaslab_t *msp = arg;
929 ASSERT3P(rt->rt_arg, ==, msp);
930 ASSERT3P(msp->ms_tree, ==, rt);
931 VERIFY(!msp->ms_condensing);
932 avl_add(&msp->ms_size_tree, rs);
936 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
938 metaslab_t *msp = arg;
940 ASSERT3P(rt->rt_arg, ==, msp);
941 ASSERT3P(msp->ms_tree, ==, rt);
942 VERIFY(!msp->ms_condensing);
943 avl_remove(&msp->ms_size_tree, rs);
947 metaslab_rt_vacate(range_tree_t *rt, void *arg)
949 metaslab_t *msp = arg;
951 ASSERT3P(rt->rt_arg, ==, msp);
952 ASSERT3P(msp->ms_tree, ==, rt);
955 * Normally one would walk the tree freeing nodes along the way.
956 * Since the nodes are shared with the range trees we can avoid
957 * walking all nodes and just reinitialize the avl tree. The nodes
958 * will be freed by the range tree, so we don't want to free them here.
960 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
961 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
964 static range_tree_ops_t metaslab_rt_ops = {
973 * ==========================================================================
974 * Metaslab block operations
975 * ==========================================================================
979 * Return the maximum contiguous segment within the metaslab.
982 metaslab_block_maxsize(metaslab_t *msp)
984 avl_tree_t *t = &msp->ms_size_tree;
987 if (t == NULL || (rs = avl_last(t)) == NULL)
990 return (rs->rs_end - rs->rs_start);
994 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
997 range_tree_t *rt = msp->ms_tree;
999 VERIFY(!msp->ms_condensing);
1001 start = msp->ms_ops->msop_alloc(msp, size);
1002 if (start != -1ULL) {
1003 vdev_t *vd = msp->ms_group->mg_vd;
1005 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
1006 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1007 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
1008 range_tree_remove(rt, start, size);
1014 * ==========================================================================
1015 * Common allocator routines
1016 * ==========================================================================
1020 * This is a helper function that can be used by the allocator to find
1021 * a suitable block to allocate. This will search the specified AVL
1022 * tree looking for a block that matches the specified criteria.
1025 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1028 range_seg_t *rs, rsearch;
1031 rsearch.rs_start = *cursor;
1032 rsearch.rs_end = *cursor + size;
1034 rs = avl_find(t, &rsearch, &where);
1036 rs = avl_nearest(t, where, AVL_AFTER);
1038 while (rs != NULL) {
1039 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1041 if (offset + size <= rs->rs_end) {
1042 *cursor = offset + size;
1045 rs = AVL_NEXT(t, rs);
1049 * If we know we've searched the whole map (*cursor == 0), give up.
1050 * Otherwise, reset the cursor to the beginning and try again.
1056 return (metaslab_block_picker(t, cursor, size, align));
1060 * ==========================================================================
1061 * The first-fit block allocator
1062 * ==========================================================================
1065 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1068 * Find the largest power of 2 block size that evenly divides the
1069 * requested size. This is used to try to allocate blocks with similar
1070 * alignment from the same area of the metaslab (i.e. same cursor
1071 * bucket) but it does not guarantee that other allocations sizes
1072 * may exist in the same region.
1074 uint64_t align = size & -size;
1075 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1076 avl_tree_t *t = &msp->ms_tree->rt_root;
1078 return (metaslab_block_picker(t, cursor, size, align));
1081 static metaslab_ops_t metaslab_ff_ops = {
1086 * ==========================================================================
1087 * Dynamic block allocator -
1088 * Uses the first fit allocation scheme until space get low and then
1089 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1090 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1091 * ==========================================================================
1094 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1097 * Find the largest power of 2 block size that evenly divides the
1098 * requested size. This is used to try to allocate blocks with similar
1099 * alignment from the same area of the metaslab (i.e. same cursor
1100 * bucket) but it does not guarantee that other allocations sizes
1101 * may exist in the same region.
1103 uint64_t align = size & -size;
1104 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1105 range_tree_t *rt = msp->ms_tree;
1106 avl_tree_t *t = &rt->rt_root;
1107 uint64_t max_size = metaslab_block_maxsize(msp);
1108 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1110 ASSERT(MUTEX_HELD(&msp->ms_lock));
1111 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1113 if (max_size < size)
1117 * If we're running low on space switch to using the size
1118 * sorted AVL tree (best-fit).
1120 if (max_size < metaslab_df_alloc_threshold ||
1121 free_pct < metaslab_df_free_pct) {
1122 t = &msp->ms_size_tree;
1126 return (metaslab_block_picker(t, cursor, size, 1ULL));
1129 static metaslab_ops_t metaslab_df_ops = {
1134 * ==========================================================================
1135 * Cursor fit block allocator -
1136 * Select the largest region in the metaslab, set the cursor to the beginning
1137 * of the range and the cursor_end to the end of the range. As allocations
1138 * are made advance the cursor. Continue allocating from the cursor until
1139 * the range is exhausted and then find a new range.
1140 * ==========================================================================
1143 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1145 range_tree_t *rt = msp->ms_tree;
1146 avl_tree_t *t = &msp->ms_size_tree;
1147 uint64_t *cursor = &msp->ms_lbas[0];
1148 uint64_t *cursor_end = &msp->ms_lbas[1];
1149 uint64_t offset = 0;
1151 ASSERT(MUTEX_HELD(&msp->ms_lock));
1152 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1154 ASSERT3U(*cursor_end, >=, *cursor);
1156 if ((*cursor + size) > *cursor_end) {
1159 rs = avl_last(&msp->ms_size_tree);
1160 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1163 *cursor = rs->rs_start;
1164 *cursor_end = rs->rs_end;
1173 static metaslab_ops_t metaslab_cf_ops = {
1178 * ==========================================================================
1179 * New dynamic fit allocator -
1180 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1181 * contiguous blocks. If no region is found then just use the largest segment
1183 * ==========================================================================
1187 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1188 * to request from the allocator.
1190 uint64_t metaslab_ndf_clump_shift = 4;
1193 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1195 avl_tree_t *t = &msp->ms_tree->rt_root;
1197 range_seg_t *rs, rsearch;
1198 uint64_t hbit = highbit64(size);
1199 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1200 uint64_t max_size = metaslab_block_maxsize(msp);
1202 ASSERT(MUTEX_HELD(&msp->ms_lock));
1203 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1205 if (max_size < size)
1208 rsearch.rs_start = *cursor;
1209 rsearch.rs_end = *cursor + size;
1211 rs = avl_find(t, &rsearch, &where);
1212 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1213 t = &msp->ms_size_tree;
1215 rsearch.rs_start = 0;
1216 rsearch.rs_end = MIN(max_size,
1217 1ULL << (hbit + metaslab_ndf_clump_shift));
1218 rs = avl_find(t, &rsearch, &where);
1220 rs = avl_nearest(t, where, AVL_AFTER);
1224 if ((rs->rs_end - rs->rs_start) >= size) {
1225 *cursor = rs->rs_start + size;
1226 return (rs->rs_start);
1231 static metaslab_ops_t metaslab_ndf_ops = {
1235 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1238 * ==========================================================================
1240 * ==========================================================================
1244 * Wait for any in-progress metaslab loads to complete.
1247 metaslab_load_wait(metaslab_t *msp)
1249 ASSERT(MUTEX_HELD(&msp->ms_lock));
1251 while (msp->ms_loading) {
1252 ASSERT(!msp->ms_loaded);
1253 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1258 metaslab_load(metaslab_t *msp)
1262 ASSERT(MUTEX_HELD(&msp->ms_lock));
1263 ASSERT(!msp->ms_loaded);
1264 ASSERT(!msp->ms_loading);
1266 msp->ms_loading = B_TRUE;
1269 * If the space map has not been allocated yet, then treat
1270 * all the space in the metaslab as free and add it to the
1273 if (msp->ms_sm != NULL)
1274 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1276 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1278 msp->ms_loaded = (error == 0);
1279 msp->ms_loading = B_FALSE;
1281 if (msp->ms_loaded) {
1282 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1283 range_tree_walk(msp->ms_defertree[t],
1284 range_tree_remove, msp->ms_tree);
1287 cv_broadcast(&msp->ms_load_cv);
1292 metaslab_unload(metaslab_t *msp)
1294 ASSERT(MUTEX_HELD(&msp->ms_lock));
1295 range_tree_vacate(msp->ms_tree, NULL, NULL);
1296 msp->ms_loaded = B_FALSE;
1297 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1301 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1304 vdev_t *vd = mg->mg_vd;
1305 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1309 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1310 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1311 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1313 ms->ms_start = id << vd->vdev_ms_shift;
1314 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1317 * We only open space map objects that already exist. All others
1318 * will be opened when we finally allocate an object for it.
1321 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1322 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1325 kmem_free(ms, sizeof (metaslab_t));
1329 ASSERT(ms->ms_sm != NULL);
1333 * We create the main range tree here, but we don't create the
1334 * alloctree and freetree until metaslab_sync_done(). This serves
1335 * two purposes: it allows metaslab_sync_done() to detect the
1336 * addition of new space; and for debugging, it ensures that we'd
1337 * data fault on any attempt to use this metaslab before it's ready.
1339 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1340 metaslab_group_add(mg, ms);
1342 ms->ms_fragmentation = metaslab_fragmentation(ms);
1343 ms->ms_ops = mg->mg_class->mc_ops;
1346 * If we're opening an existing pool (txg == 0) or creating
1347 * a new one (txg == TXG_INITIAL), all space is available now.
1348 * If we're adding space to an existing pool, the new space
1349 * does not become available until after this txg has synced.
1351 if (txg <= TXG_INITIAL)
1352 metaslab_sync_done(ms, 0);
1355 * If metaslab_debug_load is set and we're initializing a metaslab
1356 * that has an allocated space_map object then load the its space
1357 * map so that can verify frees.
1359 if (metaslab_debug_load && ms->ms_sm != NULL) {
1360 mutex_enter(&ms->ms_lock);
1361 VERIFY0(metaslab_load(ms));
1362 mutex_exit(&ms->ms_lock);
1366 vdev_dirty(vd, 0, NULL, txg);
1367 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1376 metaslab_fini(metaslab_t *msp)
1378 metaslab_group_t *mg = msp->ms_group;
1380 metaslab_group_remove(mg, msp);
1382 mutex_enter(&msp->ms_lock);
1384 VERIFY(msp->ms_group == NULL);
1385 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1387 space_map_close(msp->ms_sm);
1389 metaslab_unload(msp);
1390 range_tree_destroy(msp->ms_tree);
1392 for (int t = 0; t < TXG_SIZE; t++) {
1393 range_tree_destroy(msp->ms_alloctree[t]);
1394 range_tree_destroy(msp->ms_freetree[t]);
1397 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1398 range_tree_destroy(msp->ms_defertree[t]);
1401 ASSERT0(msp->ms_deferspace);
1403 mutex_exit(&msp->ms_lock);
1404 cv_destroy(&msp->ms_load_cv);
1405 mutex_destroy(&msp->ms_lock);
1407 kmem_free(msp, sizeof (metaslab_t));
1410 #define FRAGMENTATION_TABLE_SIZE 17
1413 * This table defines a segment size based fragmentation metric that will
1414 * allow each metaslab to derive its own fragmentation value. This is done
1415 * by calculating the space in each bucket of the spacemap histogram and
1416 * multiplying that by the fragmetation metric in this table. Doing
1417 * this for all buckets and dividing it by the total amount of free
1418 * space in this metaslab (i.e. the total free space in all buckets) gives
1419 * us the fragmentation metric. This means that a high fragmentation metric
1420 * equates to most of the free space being comprised of small segments.
1421 * Conversely, if the metric is low, then most of the free space is in
1422 * large segments. A 10% change in fragmentation equates to approximately
1423 * double the number of segments.
1425 * This table defines 0% fragmented space using 16MB segments. Testing has
1426 * shown that segments that are greater than or equal to 16MB do not suffer
1427 * from drastic performance problems. Using this value, we derive the rest
1428 * of the table. Since the fragmentation value is never stored on disk, it
1429 * is possible to change these calculations in the future.
1431 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1451 * Calclate the metaslab's fragmentation metric. A return value
1452 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1453 * not support this metric. Otherwise, the return value should be in the
1457 metaslab_fragmentation(metaslab_t *msp)
1459 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1460 uint64_t fragmentation = 0;
1462 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1463 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1465 if (!feature_enabled)
1466 return (ZFS_FRAG_INVALID);
1469 * A null space map means that the entire metaslab is free
1470 * and thus is not fragmented.
1472 if (msp->ms_sm == NULL)
1476 * If this metaslab's space_map has not been upgraded, flag it
1477 * so that we upgrade next time we encounter it.
1479 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1480 uint64_t txg = spa_syncing_txg(spa);
1481 vdev_t *vd = msp->ms_group->mg_vd;
1483 if (spa_writeable(spa)) {
1484 msp->ms_condense_wanted = B_TRUE;
1485 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1486 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1487 "msp %p, vd %p", txg, msp, vd);
1489 return (ZFS_FRAG_INVALID);
1492 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1494 uint8_t shift = msp->ms_sm->sm_shift;
1495 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1496 FRAGMENTATION_TABLE_SIZE - 1);
1498 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1501 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1504 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1505 fragmentation += space * zfs_frag_table[idx];
1509 fragmentation /= total;
1510 ASSERT3U(fragmentation, <=, 100);
1511 return (fragmentation);
1515 * Compute a weight -- a selection preference value -- for the given metaslab.
1516 * This is based on the amount of free space, the level of fragmentation,
1517 * the LBA range, and whether the metaslab is loaded.
1520 metaslab_weight(metaslab_t *msp)
1522 metaslab_group_t *mg = msp->ms_group;
1523 vdev_t *vd = mg->mg_vd;
1524 uint64_t weight, space;
1526 ASSERT(MUTEX_HELD(&msp->ms_lock));
1529 * This vdev is in the process of being removed so there is nothing
1530 * for us to do here.
1532 if (vd->vdev_removing) {
1533 ASSERT0(space_map_allocated(msp->ms_sm));
1534 ASSERT0(vd->vdev_ms_shift);
1539 * The baseline weight is the metaslab's free space.
1541 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1543 msp->ms_fragmentation = metaslab_fragmentation(msp);
1544 if (metaslab_fragmentation_factor_enabled &&
1545 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1547 * Use the fragmentation information to inversely scale
1548 * down the baseline weight. We need to ensure that we
1549 * don't exclude this metaslab completely when it's 100%
1550 * fragmented. To avoid this we reduce the fragmented value
1553 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1556 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1557 * this metaslab again. The fragmentation metric may have
1558 * decreased the space to something smaller than
1559 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1560 * so that we can consume any remaining space.
1562 if (space > 0 && space < SPA_MINBLOCKSIZE)
1563 space = SPA_MINBLOCKSIZE;
1568 * Modern disks have uniform bit density and constant angular velocity.
1569 * Therefore, the outer recording zones are faster (higher bandwidth)
1570 * than the inner zones by the ratio of outer to inner track diameter,
1571 * which is typically around 2:1. We account for this by assigning
1572 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1573 * In effect, this means that we'll select the metaslab with the most
1574 * free bandwidth rather than simply the one with the most free space.
1576 if (metaslab_lba_weighting_enabled) {
1577 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1578 ASSERT(weight >= space && weight <= 2 * space);
1582 * If this metaslab is one we're actively using, adjust its
1583 * weight to make it preferable to any inactive metaslab so
1584 * we'll polish it off. If the fragmentation on this metaslab
1585 * has exceed our threshold, then don't mark it active.
1587 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1588 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1589 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1596 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1598 ASSERT(MUTEX_HELD(&msp->ms_lock));
1600 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1601 metaslab_load_wait(msp);
1602 if (!msp->ms_loaded) {
1603 int error = metaslab_load(msp);
1605 metaslab_group_sort(msp->ms_group, msp, 0);
1610 metaslab_group_sort(msp->ms_group, msp,
1611 msp->ms_weight | activation_weight);
1613 ASSERT(msp->ms_loaded);
1614 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1620 metaslab_passivate(metaslab_t *msp, uint64_t size)
1623 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1624 * this metaslab again. In that case, it had better be empty,
1625 * or we would be leaving space on the table.
1627 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1628 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1629 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1633 metaslab_preload(void *arg)
1635 metaslab_t *msp = arg;
1636 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1638 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1640 mutex_enter(&msp->ms_lock);
1641 metaslab_load_wait(msp);
1642 if (!msp->ms_loaded)
1643 (void) metaslab_load(msp);
1646 * Set the ms_access_txg value so that we don't unload it right away.
1648 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1649 mutex_exit(&msp->ms_lock);
1653 metaslab_group_preload(metaslab_group_t *mg)
1655 spa_t *spa = mg->mg_vd->vdev_spa;
1657 avl_tree_t *t = &mg->mg_metaslab_tree;
1660 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1661 taskq_wait(mg->mg_taskq);
1665 mutex_enter(&mg->mg_lock);
1667 * Load the next potential metaslabs
1670 while (msp != NULL) {
1671 metaslab_t *msp_next = AVL_NEXT(t, msp);
1674 * We preload only the maximum number of metaslabs specified
1675 * by metaslab_preload_limit. If a metaslab is being forced
1676 * to condense then we preload it too. This will ensure
1677 * that force condensing happens in the next txg.
1679 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
1685 * We must drop the metaslab group lock here to preserve
1686 * lock ordering with the ms_lock (when grabbing both
1687 * the mg_lock and the ms_lock, the ms_lock must be taken
1688 * first). As a result, it is possible that the ordering
1689 * of the metaslabs within the avl tree may change before
1690 * we reacquire the lock. The metaslab cannot be removed from
1691 * the tree while we're in syncing context so it is safe to
1692 * drop the mg_lock here. If the metaslabs are reordered
1693 * nothing will break -- we just may end up loading a
1694 * less than optimal one.
1696 mutex_exit(&mg->mg_lock);
1697 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1698 msp, TQ_SLEEP) != 0);
1699 mutex_enter(&mg->mg_lock);
1702 mutex_exit(&mg->mg_lock);
1706 * Determine if the space map's on-disk footprint is past our tolerance
1707 * for inefficiency. We would like to use the following criteria to make
1710 * 1. The size of the space map object should not dramatically increase as a
1711 * result of writing out the free space range tree.
1713 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1714 * times the size than the free space range tree representation
1715 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1717 * 3. The on-disk size of the space map should actually decrease.
1719 * Checking the first condition is tricky since we don't want to walk
1720 * the entire AVL tree calculating the estimated on-disk size. Instead we
1721 * use the size-ordered range tree in the metaslab and calculate the
1722 * size required to write out the largest segment in our free tree. If the
1723 * size required to represent that segment on disk is larger than the space
1724 * map object then we avoid condensing this map.
1726 * To determine the second criterion we use a best-case estimate and assume
1727 * each segment can be represented on-disk as a single 64-bit entry. We refer
1728 * to this best-case estimate as the space map's minimal form.
1730 * Unfortunately, we cannot compute the on-disk size of the space map in this
1731 * context because we cannot accurately compute the effects of compression, etc.
1732 * Instead, we apply the heuristic described in the block comment for
1733 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1734 * is greater than a threshold number of blocks.
1737 metaslab_should_condense(metaslab_t *msp)
1739 space_map_t *sm = msp->ms_sm;
1741 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
1742 dmu_object_info_t doi;
1743 uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
1745 ASSERT(MUTEX_HELD(&msp->ms_lock));
1746 ASSERT(msp->ms_loaded);
1749 * Use the ms_size_tree range tree, which is ordered by size, to
1750 * obtain the largest segment in the free tree. We always condense
1751 * metaslabs that are empty and metaslabs for which a condense
1752 * request has been made.
1754 rs = avl_last(&msp->ms_size_tree);
1755 if (rs == NULL || msp->ms_condense_wanted)
1759 * Calculate the number of 64-bit entries this segment would
1760 * require when written to disk. If this single segment would be
1761 * larger on-disk than the entire current on-disk structure, then
1762 * clearly condensing will increase the on-disk structure size.
1764 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1765 entries = size / (MIN(size, SM_RUN_MAX));
1766 segsz = entries * sizeof (uint64_t);
1768 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
1769 object_size = space_map_length(msp->ms_sm);
1771 dmu_object_info_from_db(sm->sm_dbuf, &doi);
1772 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
1774 return (segsz <= object_size &&
1775 object_size >= (optimal_size * zfs_condense_pct / 100) &&
1776 object_size > zfs_metaslab_condense_block_threshold * record_size);
1780 * Condense the on-disk space map representation to its minimized form.
1781 * The minimized form consists of a small number of allocations followed by
1782 * the entries of the free range tree.
1785 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1787 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1788 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1789 range_tree_t *condense_tree;
1790 space_map_t *sm = msp->ms_sm;
1792 ASSERT(MUTEX_HELD(&msp->ms_lock));
1793 ASSERT3U(spa_sync_pass(spa), ==, 1);
1794 ASSERT(msp->ms_loaded);
1797 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1798 "smp size %llu, segments %lu, forcing condense=%s", txg,
1799 msp->ms_id, msp, space_map_length(msp->ms_sm),
1800 avl_numnodes(&msp->ms_tree->rt_root),
1801 msp->ms_condense_wanted ? "TRUE" : "FALSE");
1803 msp->ms_condense_wanted = B_FALSE;
1806 * Create an range tree that is 100% allocated. We remove segments
1807 * that have been freed in this txg, any deferred frees that exist,
1808 * and any allocation in the future. Removing segments should be
1809 * a relatively inexpensive operation since we expect these trees to
1810 * have a small number of nodes.
1812 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1813 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1816 * Remove what's been freed in this txg from the condense_tree.
1817 * Since we're in sync_pass 1, we know that all the frees from
1818 * this txg are in the freetree.
1820 range_tree_walk(freetree, range_tree_remove, condense_tree);
1822 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1823 range_tree_walk(msp->ms_defertree[t],
1824 range_tree_remove, condense_tree);
1827 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1828 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1829 range_tree_remove, condense_tree);
1833 * We're about to drop the metaslab's lock thus allowing
1834 * other consumers to change it's content. Set the
1835 * metaslab's ms_condensing flag to ensure that
1836 * allocations on this metaslab do not occur while we're
1837 * in the middle of committing it to disk. This is only critical
1838 * for the ms_tree as all other range trees use per txg
1839 * views of their content.
1841 msp->ms_condensing = B_TRUE;
1843 mutex_exit(&msp->ms_lock);
1844 space_map_truncate(sm, tx);
1845 mutex_enter(&msp->ms_lock);
1848 * While we would ideally like to create a space_map representation
1849 * that consists only of allocation records, doing so can be
1850 * prohibitively expensive because the in-core free tree can be
1851 * large, and therefore computationally expensive to subtract
1852 * from the condense_tree. Instead we sync out two trees, a cheap
1853 * allocation only tree followed by the in-core free tree. While not
1854 * optimal, this is typically close to optimal, and much cheaper to
1857 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1858 range_tree_vacate(condense_tree, NULL, NULL);
1859 range_tree_destroy(condense_tree);
1861 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1862 msp->ms_condensing = B_FALSE;
1866 * Write a metaslab to disk in the context of the specified transaction group.
1869 metaslab_sync(metaslab_t *msp, uint64_t txg)
1871 metaslab_group_t *mg = msp->ms_group;
1872 vdev_t *vd = mg->mg_vd;
1873 spa_t *spa = vd->vdev_spa;
1874 objset_t *mos = spa_meta_objset(spa);
1875 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1876 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1877 range_tree_t **freed_tree =
1878 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1880 uint64_t object = space_map_object(msp->ms_sm);
1882 ASSERT(!vd->vdev_ishole);
1885 * This metaslab has just been added so there's no work to do now.
1887 if (*freetree == NULL) {
1888 ASSERT3P(alloctree, ==, NULL);
1892 ASSERT3P(alloctree, !=, NULL);
1893 ASSERT3P(*freetree, !=, NULL);
1894 ASSERT3P(*freed_tree, !=, NULL);
1897 * Normally, we don't want to process a metaslab if there
1898 * are no allocations or frees to perform. However, if the metaslab
1899 * is being forced to condense we need to let it through.
1901 if (range_tree_space(alloctree) == 0 &&
1902 range_tree_space(*freetree) == 0 &&
1903 !msp->ms_condense_wanted)
1907 * The only state that can actually be changing concurrently with
1908 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1909 * be modifying this txg's alloctree, freetree, freed_tree, or
1910 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1911 * space_map ASSERTs. We drop it whenever we call into the DMU,
1912 * because the DMU can call down to us (e.g. via zio_free()) at
1916 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1918 if (msp->ms_sm == NULL) {
1919 uint64_t new_object;
1921 new_object = space_map_alloc(mos, tx);
1922 VERIFY3U(new_object, !=, 0);
1924 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1925 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1927 ASSERT(msp->ms_sm != NULL);
1930 mutex_enter(&msp->ms_lock);
1933 * Note: metaslab_condense() clears the space_map's histogram.
1934 * Therefore we must verify and remove this histogram before
1937 metaslab_group_histogram_verify(mg);
1938 metaslab_class_histogram_verify(mg->mg_class);
1939 metaslab_group_histogram_remove(mg, msp);
1941 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1942 metaslab_should_condense(msp)) {
1943 metaslab_condense(msp, txg, tx);
1945 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1946 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1949 if (msp->ms_loaded) {
1951 * When the space map is loaded, we have an accruate
1952 * histogram in the range tree. This gives us an opportunity
1953 * to bring the space map's histogram up-to-date so we clear
1954 * it first before updating it.
1956 space_map_histogram_clear(msp->ms_sm);
1957 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1960 * Since the space map is not loaded we simply update the
1961 * exisiting histogram with what was freed in this txg. This
1962 * means that the on-disk histogram may not have an accurate
1963 * view of the free space but it's close enough to allow
1964 * us to make allocation decisions.
1966 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1968 metaslab_group_histogram_add(mg, msp);
1969 metaslab_group_histogram_verify(mg);
1970 metaslab_class_histogram_verify(mg->mg_class);
1973 * For sync pass 1, we avoid traversing this txg's free range tree
1974 * and instead will just swap the pointers for freetree and
1975 * freed_tree. We can safely do this since the freed_tree is
1976 * guaranteed to be empty on the initial pass.
1978 if (spa_sync_pass(spa) == 1) {
1979 range_tree_swap(freetree, freed_tree);
1981 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1983 range_tree_vacate(alloctree, NULL, NULL);
1985 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1986 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1988 mutex_exit(&msp->ms_lock);
1990 if (object != space_map_object(msp->ms_sm)) {
1991 object = space_map_object(msp->ms_sm);
1992 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1993 msp->ms_id, sizeof (uint64_t), &object, tx);
1999 * Called after a transaction group has completely synced to mark
2000 * all of the metaslab's free space as usable.
2003 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2005 metaslab_group_t *mg = msp->ms_group;
2006 vdev_t *vd = mg->mg_vd;
2007 range_tree_t **freed_tree;
2008 range_tree_t **defer_tree;
2009 int64_t alloc_delta, defer_delta;
2011 ASSERT(!vd->vdev_ishole);
2013 mutex_enter(&msp->ms_lock);
2016 * If this metaslab is just becoming available, initialize its
2017 * alloctrees, freetrees, and defertree and add its capacity to
2020 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
2021 for (int t = 0; t < TXG_SIZE; t++) {
2022 ASSERT(msp->ms_alloctree[t] == NULL);
2023 ASSERT(msp->ms_freetree[t] == NULL);
2025 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2027 msp->ms_freetree[t] = range_tree_create(NULL, msp,
2031 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2032 ASSERT(msp->ms_defertree[t] == NULL);
2034 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2038 vdev_space_update(vd, 0, 0, msp->ms_size);
2041 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2042 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2044 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2045 defer_delta = range_tree_space(*freed_tree) -
2046 range_tree_space(*defer_tree);
2048 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2050 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2051 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2054 * If there's a metaslab_load() in progress, wait for it to complete
2055 * so that we have a consistent view of the in-core space map.
2057 metaslab_load_wait(msp);
2060 * Move the frees from the defer_tree back to the free
2061 * range tree (if it's loaded). Swap the freed_tree and the
2062 * defer_tree -- this is safe to do because we've just emptied out
2065 range_tree_vacate(*defer_tree,
2066 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2067 range_tree_swap(freed_tree, defer_tree);
2069 space_map_update(msp->ms_sm);
2071 msp->ms_deferspace += defer_delta;
2072 ASSERT3S(msp->ms_deferspace, >=, 0);
2073 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2074 if (msp->ms_deferspace != 0) {
2076 * Keep syncing this metaslab until all deferred frees
2077 * are back in circulation.
2079 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2082 if (msp->ms_loaded && msp->ms_access_txg < txg) {
2083 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2084 VERIFY0(range_tree_space(
2085 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2088 if (!metaslab_debug_unload)
2089 metaslab_unload(msp);
2092 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2093 mutex_exit(&msp->ms_lock);
2097 metaslab_sync_reassess(metaslab_group_t *mg)
2099 metaslab_group_alloc_update(mg);
2100 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2103 * Preload the next potential metaslabs
2105 metaslab_group_preload(mg);
2109 metaslab_distance(metaslab_t *msp, dva_t *dva)
2111 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2112 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2113 uint64_t start = msp->ms_id;
2115 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2116 return (1ULL << 63);
2119 return ((start - offset) << ms_shift);
2121 return ((offset - start) << ms_shift);
2126 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
2127 uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2129 spa_t *spa = mg->mg_vd->vdev_spa;
2130 metaslab_t *msp = NULL;
2131 uint64_t offset = -1ULL;
2132 avl_tree_t *t = &mg->mg_metaslab_tree;
2133 uint64_t activation_weight;
2134 uint64_t target_distance;
2137 activation_weight = METASLAB_WEIGHT_PRIMARY;
2138 for (i = 0; i < d; i++) {
2139 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2140 activation_weight = METASLAB_WEIGHT_SECONDARY;
2146 boolean_t was_active;
2148 mutex_enter(&mg->mg_lock);
2149 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
2150 if (msp->ms_weight < asize) {
2151 spa_dbgmsg(spa, "%s: failed to meet weight "
2152 "requirement: vdev %llu, txg %llu, mg %p, "
2153 "msp %p, psize %llu, asize %llu, "
2154 "weight %llu", spa_name(spa),
2155 mg->mg_vd->vdev_id, txg,
2156 mg, msp, psize, asize, msp->ms_weight);
2157 mutex_exit(&mg->mg_lock);
2162 * If the selected metaslab is condensing, skip it.
2164 if (msp->ms_condensing)
2167 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2168 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2171 target_distance = min_distance +
2172 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2175 for (i = 0; i < d; i++)
2176 if (metaslab_distance(msp, &dva[i]) <
2182 mutex_exit(&mg->mg_lock);
2186 mutex_enter(&msp->ms_lock);
2189 * Ensure that the metaslab we have selected is still
2190 * capable of handling our request. It's possible that
2191 * another thread may have changed the weight while we
2192 * were blocked on the metaslab lock.
2194 if (msp->ms_weight < asize || (was_active &&
2195 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
2196 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
2197 mutex_exit(&msp->ms_lock);
2201 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2202 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2203 metaslab_passivate(msp,
2204 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2205 mutex_exit(&msp->ms_lock);
2209 if (metaslab_activate(msp, activation_weight) != 0) {
2210 mutex_exit(&msp->ms_lock);
2215 * If this metaslab is currently condensing then pick again as
2216 * we can't manipulate this metaslab until it's committed
2219 if (msp->ms_condensing) {
2220 mutex_exit(&msp->ms_lock);
2224 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
2227 metaslab_passivate(msp, metaslab_block_maxsize(msp));
2228 mutex_exit(&msp->ms_lock);
2231 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2232 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2234 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
2235 msp->ms_access_txg = txg + metaslab_unload_delay;
2237 mutex_exit(&msp->ms_lock);
2243 * Allocate a block for the specified i/o.
2246 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2247 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
2249 metaslab_group_t *mg, *rotor;
2253 int zio_lock = B_FALSE;
2254 boolean_t allocatable;
2255 uint64_t offset = -1ULL;
2259 ASSERT(!DVA_IS_VALID(&dva[d]));
2262 * For testing, make some blocks above a certain size be gang blocks.
2264 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
2265 return (SET_ERROR(ENOSPC));
2268 * Start at the rotor and loop through all mgs until we find something.
2269 * Note that there's no locking on mc_rotor or mc_aliquot because
2270 * nothing actually breaks if we miss a few updates -- we just won't
2271 * allocate quite as evenly. It all balances out over time.
2273 * If we are doing ditto or log blocks, try to spread them across
2274 * consecutive vdevs. If we're forced to reuse a vdev before we've
2275 * allocated all of our ditto blocks, then try and spread them out on
2276 * that vdev as much as possible. If it turns out to not be possible,
2277 * gradually lower our standards until anything becomes acceptable.
2278 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2279 * gives us hope of containing our fault domains to something we're
2280 * able to reason about. Otherwise, any two top-level vdev failures
2281 * will guarantee the loss of data. With consecutive allocation,
2282 * only two adjacent top-level vdev failures will result in data loss.
2284 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2285 * ourselves on the same vdev as our gang block header. That
2286 * way, we can hope for locality in vdev_cache, plus it makes our
2287 * fault domains something tractable.
2290 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
2293 * It's possible the vdev we're using as the hint no
2294 * longer exists (i.e. removed). Consult the rotor when
2300 if (flags & METASLAB_HINTBP_AVOID &&
2301 mg->mg_next != NULL)
2306 } else if (d != 0) {
2307 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
2308 mg = vd->vdev_mg->mg_next;
2314 * If the hint put us into the wrong metaslab class, or into a
2315 * metaslab group that has been passivated, just follow the rotor.
2317 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2324 ASSERT(mg->mg_activation_count == 1);
2329 * Don't allocate from faulted devices.
2332 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
2333 allocatable = vdev_allocatable(vd);
2334 spa_config_exit(spa, SCL_ZIO, FTAG);
2336 allocatable = vdev_allocatable(vd);
2340 * Determine if the selected metaslab group is eligible
2341 * for allocations. If we're ganging or have requested
2342 * an allocation for the smallest gang block size
2343 * then we don't want to avoid allocating to the this
2344 * metaslab group. If we're in this condition we should
2345 * try to allocate from any device possible so that we
2346 * don't inadvertently return ENOSPC and suspend the pool
2347 * even though space is still available.
2349 if (allocatable && CAN_FASTGANG(flags) &&
2350 psize > SPA_GANGBLOCKSIZE)
2351 allocatable = metaslab_group_allocatable(mg);
2357 * Avoid writing single-copy data to a failing vdev
2358 * unless the user instructs us that it is okay.
2360 if ((vd->vdev_stat.vs_write_errors > 0 ||
2361 vd->vdev_state < VDEV_STATE_HEALTHY) &&
2362 d == 0 && dshift == 3 && vd->vdev_children == 0) {
2367 ASSERT(mg->mg_class == mc);
2369 distance = vd->vdev_asize >> dshift;
2370 if (distance <= (1ULL << vd->vdev_ms_shift))
2375 asize = vdev_psize_to_asize(vd, psize);
2376 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
2378 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
2380 if (offset != -1ULL) {
2382 * If we've just selected this metaslab group,
2383 * figure out whether the corresponding vdev is
2384 * over- or under-used relative to the pool,
2385 * and set an allocation bias to even it out.
2387 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
2388 vdev_stat_t *vs = &vd->vdev_stat;
2391 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
2392 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2395 * Calculate how much more or less we should
2396 * try to allocate from this device during
2397 * this iteration around the rotor.
2398 * For example, if a device is 80% full
2399 * and the pool is 20% full then we should
2400 * reduce allocations by 60% on this device.
2402 * mg_bias = (20 - 80) * 512K / 100 = -307K
2404 * This reduces allocations by 307K for this
2407 mg->mg_bias = ((cu - vu) *
2408 (int64_t)mg->mg_aliquot) / 100;
2409 } else if (!metaslab_bias_enabled) {
2413 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2414 mg->mg_aliquot + mg->mg_bias) {
2415 mc->mc_rotor = mg->mg_next;
2419 DVA_SET_VDEV(&dva[d], vd->vdev_id);
2420 DVA_SET_OFFSET(&dva[d], offset);
2421 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2422 DVA_SET_ASIZE(&dva[d], asize);
2427 mc->mc_rotor = mg->mg_next;
2429 } while ((mg = mg->mg_next) != rotor);
2433 ASSERT(dshift < 64);
2437 if (!allocatable && !zio_lock) {
2443 bzero(&dva[d], sizeof (dva_t));
2445 return (SET_ERROR(ENOSPC));
2449 * Free the block represented by DVA in the context of the specified
2450 * transaction group.
2453 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2455 uint64_t vdev = DVA_GET_VDEV(dva);
2456 uint64_t offset = DVA_GET_OFFSET(dva);
2457 uint64_t size = DVA_GET_ASIZE(dva);
2461 ASSERT(DVA_IS_VALID(dva));
2463 if (txg > spa_freeze_txg(spa))
2466 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2467 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2468 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2469 (u_longlong_t)vdev, (u_longlong_t)offset);
2474 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2476 if (DVA_GET_GANG(dva))
2477 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2479 mutex_enter(&msp->ms_lock);
2482 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2485 VERIFY(!msp->ms_condensing);
2486 VERIFY3U(offset, >=, msp->ms_start);
2487 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2488 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2490 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2491 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2492 range_tree_add(msp->ms_tree, offset, size);
2494 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2495 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2496 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2500 mutex_exit(&msp->ms_lock);
2504 * Intent log support: upon opening the pool after a crash, notify the SPA
2505 * of blocks that the intent log has allocated for immediate write, but
2506 * which are still considered free by the SPA because the last transaction
2507 * group didn't commit yet.
2510 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2512 uint64_t vdev = DVA_GET_VDEV(dva);
2513 uint64_t offset = DVA_GET_OFFSET(dva);
2514 uint64_t size = DVA_GET_ASIZE(dva);
2519 ASSERT(DVA_IS_VALID(dva));
2521 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2522 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2523 return (SET_ERROR(ENXIO));
2525 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2527 if (DVA_GET_GANG(dva))
2528 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2530 mutex_enter(&msp->ms_lock);
2532 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2533 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2535 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2536 error = SET_ERROR(ENOENT);
2538 if (error || txg == 0) { /* txg == 0 indicates dry run */
2539 mutex_exit(&msp->ms_lock);
2543 VERIFY(!msp->ms_condensing);
2544 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2545 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2546 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2547 range_tree_remove(msp->ms_tree, offset, size);
2549 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2550 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2551 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2552 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2555 mutex_exit(&msp->ms_lock);
2561 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2562 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2564 dva_t *dva = bp->blk_dva;
2565 dva_t *hintdva = hintbp->blk_dva;
2568 ASSERT(bp->blk_birth == 0);
2569 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2571 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2573 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2574 spa_config_exit(spa, SCL_ALLOC, FTAG);
2575 return (SET_ERROR(ENOSPC));
2578 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2579 ASSERT(BP_GET_NDVAS(bp) == 0);
2580 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2582 for (int d = 0; d < ndvas; d++) {
2583 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2586 for (d--; d >= 0; d--) {
2587 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2588 bzero(&dva[d], sizeof (dva_t));
2590 spa_config_exit(spa, SCL_ALLOC, FTAG);
2595 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2597 spa_config_exit(spa, SCL_ALLOC, FTAG);
2599 BP_SET_BIRTH(bp, txg, txg);
2605 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2607 const dva_t *dva = bp->blk_dva;
2608 int ndvas = BP_GET_NDVAS(bp);
2610 ASSERT(!BP_IS_HOLE(bp));
2611 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2613 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2615 for (int d = 0; d < ndvas; d++)
2616 metaslab_free_dva(spa, &dva[d], txg, now);
2618 spa_config_exit(spa, SCL_FREE, FTAG);
2622 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2624 const dva_t *dva = bp->blk_dva;
2625 int ndvas = BP_GET_NDVAS(bp);
2628 ASSERT(!BP_IS_HOLE(bp));
2632 * First do a dry run to make sure all DVAs are claimable,
2633 * so we don't have to unwind from partial failures below.
2635 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2639 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2641 for (int d = 0; d < ndvas; d++)
2642 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2645 spa_config_exit(spa, SCL_ALLOC, FTAG);
2647 ASSERT(error == 0 || txg == 0);
2653 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2655 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2658 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2659 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2660 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2661 vdev_t *vd = vdev_lookup_top(spa, vdev);
2662 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2663 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2664 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2667 range_tree_verify(msp->ms_tree, offset, size);
2669 for (int j = 0; j < TXG_SIZE; j++)
2670 range_tree_verify(msp->ms_freetree[j], offset, size);
2671 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2672 range_tree_verify(msp->ms_defertree[j], offset, size);
2674 spa_config_exit(spa, SCL_VDEV, FTAG);