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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/zfs_context.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
38 SYSCTL_DECL(_vfs_zfs);
39 SYSCTL_NODE(_vfs_zfs, OID_AUTO, metaslab, CTLFLAG_RW, 0, "ZFS metaslab");
42 * Allow allocations to switch to gang blocks quickly. We do this to
43 * avoid having to load lots of space_maps in a given txg. There are,
44 * however, some cases where we want to avoid "fast" ganging and instead
45 * we want to do an exhaustive search of all metaslabs on this device.
46 * Currently we don't allow any gang, slog, or dump device related allocations
49 #define CAN_FASTGANG(flags) \
50 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
51 METASLAB_GANG_AVOID)))
53 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
54 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
55 #define METASLAB_ACTIVE_MASK \
56 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
58 uint64_t metaslab_aliquot = 512ULL << 10;
59 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
60 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, gang_bang, CTLFLAG_RWTUN,
61 &metaslab_gang_bang, 0,
62 "Force gang block allocation for blocks larger than or equal to this value");
65 * The in-core space map representation is more compact than its on-disk form.
66 * The zfs_condense_pct determines how much more compact the in-core
67 * space_map representation must be before we compact it on-disk.
68 * Values should be greater than or equal to 100.
70 int zfs_condense_pct = 200;
71 SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN,
73 "Condense on-disk spacemap when it is more than this many percents"
74 " of in-memory counterpart");
77 * Condensing a metaslab is not guaranteed to actually reduce the amount of
78 * space used on disk. In particular, a space map uses data in increments of
79 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
80 * same number of blocks after condensing. Since the goal of condensing is to
81 * reduce the number of IOPs required to read the space map, we only want to
82 * condense when we can be sure we will reduce the number of blocks used by the
83 * space map. Unfortunately, we cannot precisely compute whether or not this is
84 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
85 * we apply the following heuristic: do not condense a spacemap unless the
86 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
89 int zfs_metaslab_condense_block_threshold = 4;
92 * The zfs_mg_noalloc_threshold defines which metaslab groups should
93 * be eligible for allocation. The value is defined as a percentage of
94 * free space. Metaslab groups that have more free space than
95 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
96 * a metaslab group's free space is less than or equal to the
97 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
98 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
99 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
100 * groups are allowed to accept allocations. Gang blocks are always
101 * eligible to allocate on any metaslab group. The default value of 0 means
102 * no metaslab group will be excluded based on this criterion.
104 int zfs_mg_noalloc_threshold = 0;
105 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN,
106 &zfs_mg_noalloc_threshold, 0,
107 "Percentage of metaslab group size that should be free"
108 " to make it eligible for allocation");
111 * Metaslab groups are considered eligible for allocations if their
112 * fragmenation metric (measured as a percentage) is less than or equal to
113 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
114 * then it will be skipped unless all metaslab groups within the metaslab
115 * class have also crossed this threshold.
117 int zfs_mg_fragmentation_threshold = 85;
118 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_fragmentation_threshold, CTLFLAG_RWTUN,
119 &zfs_mg_fragmentation_threshold, 0,
120 "Percentage of metaslab group size that should be considered "
121 "eligible for allocations unless all metaslab groups within the metaslab class "
122 "have also crossed this threshold");
125 * Allow metaslabs to keep their active state as long as their fragmentation
126 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
127 * active metaslab that exceeds this threshold will no longer keep its active
128 * status allowing better metaslabs to be selected.
130 int zfs_metaslab_fragmentation_threshold = 70;
131 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_threshold, CTLFLAG_RWTUN,
132 &zfs_metaslab_fragmentation_threshold, 0,
133 "Maximum percentage of metaslab fragmentation level to keep their active state");
136 * When set will load all metaslabs when pool is first opened.
138 int metaslab_debug_load = 0;
139 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN,
140 &metaslab_debug_load, 0,
141 "Load all metaslabs when pool is first opened");
144 * When set will prevent metaslabs from being unloaded.
146 int metaslab_debug_unload = 0;
147 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN,
148 &metaslab_debug_unload, 0,
149 "Prevent metaslabs from being unloaded");
152 * Minimum size which forces the dynamic allocator to change
153 * it's allocation strategy. Once the space map cannot satisfy
154 * an allocation of this size then it switches to using more
155 * aggressive strategy (i.e search by size rather than offset).
157 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
158 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN,
159 &metaslab_df_alloc_threshold, 0,
160 "Minimum size which forces the dynamic allocator to change it's allocation strategy");
163 * The minimum free space, in percent, which must be available
164 * in a space map to continue allocations in a first-fit fashion.
165 * Once the space_map's free space drops below this level we dynamically
166 * switch to using best-fit allocations.
168 int metaslab_df_free_pct = 4;
169 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN,
170 &metaslab_df_free_pct, 0,
171 "The minimum free space, in percent, which must be available in a "
172 "space map to continue allocations in a first-fit fashion");
175 * A metaslab is considered "free" if it contains a contiguous
176 * segment which is greater than metaslab_min_alloc_size.
178 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
179 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN,
180 &metaslab_min_alloc_size, 0,
181 "A metaslab is considered \"free\" if it contains a contiguous "
182 "segment which is greater than vfs.zfs.metaslab.min_alloc_size");
185 * Percentage of all cpus that can be used by the metaslab taskq.
187 int metaslab_load_pct = 50;
188 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN,
189 &metaslab_load_pct, 0,
190 "Percentage of cpus that can be used by the metaslab taskq");
193 * Determines how many txgs a metaslab may remain loaded without having any
194 * allocations from it. As long as a metaslab continues to be used we will
197 int metaslab_unload_delay = TXG_SIZE * 2;
198 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN,
199 &metaslab_unload_delay, 0,
200 "Number of TXGs that an unused metaslab can be kept in memory");
203 * Max number of metaslabs per group to preload.
205 int metaslab_preload_limit = SPA_DVAS_PER_BP;
206 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
207 &metaslab_preload_limit, 0,
208 "Max number of metaslabs per group to preload");
211 * Enable/disable preloading of metaslab.
213 boolean_t metaslab_preload_enabled = B_TRUE;
214 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
215 &metaslab_preload_enabled, 0,
216 "Max number of metaslabs per group to preload");
219 * Enable/disable fragmentation weighting on metaslabs.
221 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
222 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_factor_enabled, CTLFLAG_RWTUN,
223 &metaslab_fragmentation_factor_enabled, 0,
224 "Enable fragmentation weighting on metaslabs");
227 * Enable/disable lba weighting (i.e. outer tracks are given preference).
229 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
230 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, lba_weighting_enabled, CTLFLAG_RWTUN,
231 &metaslab_lba_weighting_enabled, 0,
232 "Enable LBA weighting (i.e. outer tracks are given preference)");
235 * Enable/disable metaslab group biasing.
237 boolean_t metaslab_bias_enabled = B_TRUE;
238 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, bias_enabled, CTLFLAG_RWTUN,
239 &metaslab_bias_enabled, 0,
240 "Enable metaslab group biasing");
242 static uint64_t metaslab_fragmentation(metaslab_t *);
245 * ==========================================================================
247 * ==========================================================================
250 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
252 metaslab_class_t *mc;
254 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
264 metaslab_class_destroy(metaslab_class_t *mc)
266 ASSERT(mc->mc_rotor == NULL);
267 ASSERT(mc->mc_alloc == 0);
268 ASSERT(mc->mc_deferred == 0);
269 ASSERT(mc->mc_space == 0);
270 ASSERT(mc->mc_dspace == 0);
272 kmem_free(mc, sizeof (metaslab_class_t));
276 metaslab_class_validate(metaslab_class_t *mc)
278 metaslab_group_t *mg;
282 * Must hold one of the spa_config locks.
284 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
285 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
287 if ((mg = mc->mc_rotor) == NULL)
292 ASSERT(vd->vdev_mg != NULL);
293 ASSERT3P(vd->vdev_top, ==, vd);
294 ASSERT3P(mg->mg_class, ==, mc);
295 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
296 } while ((mg = mg->mg_next) != mc->mc_rotor);
302 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
303 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
305 atomic_add_64(&mc->mc_alloc, alloc_delta);
306 atomic_add_64(&mc->mc_deferred, defer_delta);
307 atomic_add_64(&mc->mc_space, space_delta);
308 atomic_add_64(&mc->mc_dspace, dspace_delta);
312 metaslab_class_minblocksize_update(metaslab_class_t *mc)
314 metaslab_group_t *mg;
316 uint64_t minashift = UINT64_MAX;
318 if ((mg = mc->mc_rotor) == NULL) {
319 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
325 if (vd->vdev_ashift < minashift)
326 minashift = vd->vdev_ashift;
327 } while ((mg = mg->mg_next) != mc->mc_rotor);
329 mc->mc_minblocksize = 1ULL << minashift;
333 metaslab_class_get_alloc(metaslab_class_t *mc)
335 return (mc->mc_alloc);
339 metaslab_class_get_deferred(metaslab_class_t *mc)
341 return (mc->mc_deferred);
345 metaslab_class_get_space(metaslab_class_t *mc)
347 return (mc->mc_space);
351 metaslab_class_get_dspace(metaslab_class_t *mc)
353 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
357 metaslab_class_get_minblocksize(metaslab_class_t *mc)
359 return (mc->mc_minblocksize);
363 metaslab_class_histogram_verify(metaslab_class_t *mc)
365 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
369 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
372 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
375 for (int c = 0; c < rvd->vdev_children; c++) {
376 vdev_t *tvd = rvd->vdev_child[c];
377 metaslab_group_t *mg = tvd->vdev_mg;
380 * Skip any holes, uninitialized top-levels, or
381 * vdevs that are not in this metalab class.
383 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
384 mg->mg_class != mc) {
388 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
389 mc_hist[i] += mg->mg_histogram[i];
392 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
393 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
395 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
399 * Calculate the metaslab class's fragmentation metric. The metric
400 * is weighted based on the space contribution of each metaslab group.
401 * The return value will be a number between 0 and 100 (inclusive), or
402 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
403 * zfs_frag_table for more information about the metric.
406 metaslab_class_fragmentation(metaslab_class_t *mc)
408 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
409 uint64_t fragmentation = 0;
411 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
413 for (int c = 0; c < rvd->vdev_children; c++) {
414 vdev_t *tvd = rvd->vdev_child[c];
415 metaslab_group_t *mg = tvd->vdev_mg;
418 * Skip any holes, uninitialized top-levels, or
419 * vdevs that are not in this metalab class.
421 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
422 mg->mg_class != mc) {
427 * If a metaslab group does not contain a fragmentation
428 * metric then just bail out.
430 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
431 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
432 return (ZFS_FRAG_INVALID);
436 * Determine how much this metaslab_group is contributing
437 * to the overall pool fragmentation metric.
439 fragmentation += mg->mg_fragmentation *
440 metaslab_group_get_space(mg);
442 fragmentation /= metaslab_class_get_space(mc);
444 ASSERT3U(fragmentation, <=, 100);
445 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
446 return (fragmentation);
450 * Calculate the amount of expandable space that is available in
451 * this metaslab class. If a device is expanded then its expandable
452 * space will be the amount of allocatable space that is currently not
453 * part of this metaslab class.
456 metaslab_class_expandable_space(metaslab_class_t *mc)
458 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
461 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
462 for (int c = 0; c < rvd->vdev_children; c++) {
463 vdev_t *tvd = rvd->vdev_child[c];
464 metaslab_group_t *mg = tvd->vdev_mg;
466 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
467 mg->mg_class != mc) {
471 space += tvd->vdev_max_asize - tvd->vdev_asize;
473 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
478 * ==========================================================================
480 * ==========================================================================
483 metaslab_compare(const void *x1, const void *x2)
485 const metaslab_t *m1 = x1;
486 const metaslab_t *m2 = x2;
488 if (m1->ms_weight < m2->ms_weight)
490 if (m1->ms_weight > m2->ms_weight)
494 * If the weights are identical, use the offset to force uniqueness.
496 if (m1->ms_start < m2->ms_start)
498 if (m1->ms_start > m2->ms_start)
501 ASSERT3P(m1, ==, m2);
507 * Update the allocatable flag and the metaslab group's capacity.
508 * The allocatable flag is set to true if the capacity is below
509 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
510 * from allocatable to non-allocatable or vice versa then the metaslab
511 * group's class is updated to reflect the transition.
514 metaslab_group_alloc_update(metaslab_group_t *mg)
516 vdev_t *vd = mg->mg_vd;
517 metaslab_class_t *mc = mg->mg_class;
518 vdev_stat_t *vs = &vd->vdev_stat;
519 boolean_t was_allocatable;
521 ASSERT(vd == vd->vdev_top);
523 mutex_enter(&mg->mg_lock);
524 was_allocatable = mg->mg_allocatable;
526 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
530 * A metaslab group is considered allocatable if it has plenty
531 * of free space or is not heavily fragmented. We only take
532 * fragmentation into account if the metaslab group has a valid
533 * fragmentation metric (i.e. a value between 0 and 100).
535 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
536 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
537 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
540 * The mc_alloc_groups maintains a count of the number of
541 * groups in this metaslab class that are still above the
542 * zfs_mg_noalloc_threshold. This is used by the allocating
543 * threads to determine if they should avoid allocations to
544 * a given group. The allocator will avoid allocations to a group
545 * if that group has reached or is below the zfs_mg_noalloc_threshold
546 * and there are still other groups that are above the threshold.
547 * When a group transitions from allocatable to non-allocatable or
548 * vice versa we update the metaslab class to reflect that change.
549 * When the mc_alloc_groups value drops to 0 that means that all
550 * groups have reached the zfs_mg_noalloc_threshold making all groups
551 * eligible for allocations. This effectively means that all devices
552 * are balanced again.
554 if (was_allocatable && !mg->mg_allocatable)
555 mc->mc_alloc_groups--;
556 else if (!was_allocatable && mg->mg_allocatable)
557 mc->mc_alloc_groups++;
559 mutex_exit(&mg->mg_lock);
563 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
565 metaslab_group_t *mg;
567 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
568 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
569 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
570 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
573 mg->mg_activation_count = 0;
575 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
576 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
582 metaslab_group_destroy(metaslab_group_t *mg)
584 ASSERT(mg->mg_prev == NULL);
585 ASSERT(mg->mg_next == NULL);
587 * We may have gone below zero with the activation count
588 * either because we never activated in the first place or
589 * because we're done, and possibly removing the vdev.
591 ASSERT(mg->mg_activation_count <= 0);
593 taskq_destroy(mg->mg_taskq);
594 avl_destroy(&mg->mg_metaslab_tree);
595 mutex_destroy(&mg->mg_lock);
596 kmem_free(mg, sizeof (metaslab_group_t));
600 metaslab_group_activate(metaslab_group_t *mg)
602 metaslab_class_t *mc = mg->mg_class;
603 metaslab_group_t *mgprev, *mgnext;
605 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
607 ASSERT(mc->mc_rotor != mg);
608 ASSERT(mg->mg_prev == NULL);
609 ASSERT(mg->mg_next == NULL);
610 ASSERT(mg->mg_activation_count <= 0);
612 if (++mg->mg_activation_count <= 0)
615 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
616 metaslab_group_alloc_update(mg);
618 if ((mgprev = mc->mc_rotor) == NULL) {
622 mgnext = mgprev->mg_next;
623 mg->mg_prev = mgprev;
624 mg->mg_next = mgnext;
625 mgprev->mg_next = mg;
626 mgnext->mg_prev = mg;
629 metaslab_class_minblocksize_update(mc);
633 metaslab_group_passivate(metaslab_group_t *mg)
635 metaslab_class_t *mc = mg->mg_class;
636 metaslab_group_t *mgprev, *mgnext;
638 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
640 if (--mg->mg_activation_count != 0) {
641 ASSERT(mc->mc_rotor != mg);
642 ASSERT(mg->mg_prev == NULL);
643 ASSERT(mg->mg_next == NULL);
644 ASSERT(mg->mg_activation_count < 0);
648 taskq_wait(mg->mg_taskq);
649 metaslab_group_alloc_update(mg);
651 mgprev = mg->mg_prev;
652 mgnext = mg->mg_next;
657 mc->mc_rotor = mgnext;
658 mgprev->mg_next = mgnext;
659 mgnext->mg_prev = mgprev;
664 metaslab_class_minblocksize_update(mc);
668 metaslab_group_get_space(metaslab_group_t *mg)
670 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
674 metaslab_group_histogram_verify(metaslab_group_t *mg)
677 vdev_t *vd = mg->mg_vd;
678 uint64_t ashift = vd->vdev_ashift;
681 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
684 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
687 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
688 SPACE_MAP_HISTOGRAM_SIZE + ashift);
690 for (int m = 0; m < vd->vdev_ms_count; m++) {
691 metaslab_t *msp = vd->vdev_ms[m];
693 if (msp->ms_sm == NULL)
696 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
697 mg_hist[i + ashift] +=
698 msp->ms_sm->sm_phys->smp_histogram[i];
701 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
702 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
704 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
708 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
710 metaslab_class_t *mc = mg->mg_class;
711 uint64_t ashift = mg->mg_vd->vdev_ashift;
713 ASSERT(MUTEX_HELD(&msp->ms_lock));
714 if (msp->ms_sm == NULL)
717 mutex_enter(&mg->mg_lock);
718 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
719 mg->mg_histogram[i + ashift] +=
720 msp->ms_sm->sm_phys->smp_histogram[i];
721 mc->mc_histogram[i + ashift] +=
722 msp->ms_sm->sm_phys->smp_histogram[i];
724 mutex_exit(&mg->mg_lock);
728 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
730 metaslab_class_t *mc = mg->mg_class;
731 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 (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
739 ASSERT3U(mg->mg_histogram[i + ashift], >=,
740 msp->ms_sm->sm_phys->smp_histogram[i]);
741 ASSERT3U(mc->mc_histogram[i + ashift], >=,
742 msp->ms_sm->sm_phys->smp_histogram[i]);
744 mg->mg_histogram[i + ashift] -=
745 msp->ms_sm->sm_phys->smp_histogram[i];
746 mc->mc_histogram[i + ashift] -=
747 msp->ms_sm->sm_phys->smp_histogram[i];
749 mutex_exit(&mg->mg_lock);
753 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
755 ASSERT(msp->ms_group == NULL);
756 mutex_enter(&mg->mg_lock);
759 avl_add(&mg->mg_metaslab_tree, msp);
760 mutex_exit(&mg->mg_lock);
762 mutex_enter(&msp->ms_lock);
763 metaslab_group_histogram_add(mg, msp);
764 mutex_exit(&msp->ms_lock);
768 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
770 mutex_enter(&msp->ms_lock);
771 metaslab_group_histogram_remove(mg, msp);
772 mutex_exit(&msp->ms_lock);
774 mutex_enter(&mg->mg_lock);
775 ASSERT(msp->ms_group == mg);
776 avl_remove(&mg->mg_metaslab_tree, msp);
777 msp->ms_group = NULL;
778 mutex_exit(&mg->mg_lock);
782 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
785 * Although in principle the weight can be any value, in
786 * practice we do not use values in the range [1, 511].
788 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
789 ASSERT(MUTEX_HELD(&msp->ms_lock));
791 mutex_enter(&mg->mg_lock);
792 ASSERT(msp->ms_group == mg);
793 avl_remove(&mg->mg_metaslab_tree, msp);
794 msp->ms_weight = weight;
795 avl_add(&mg->mg_metaslab_tree, msp);
796 mutex_exit(&mg->mg_lock);
800 * Calculate the fragmentation for a given metaslab group. We can use
801 * a simple average here since all metaslabs within the group must have
802 * the same size. The return value will be a value between 0 and 100
803 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
804 * group have a fragmentation metric.
807 metaslab_group_fragmentation(metaslab_group_t *mg)
809 vdev_t *vd = mg->mg_vd;
810 uint64_t fragmentation = 0;
811 uint64_t valid_ms = 0;
813 for (int m = 0; m < vd->vdev_ms_count; m++) {
814 metaslab_t *msp = vd->vdev_ms[m];
816 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
820 fragmentation += msp->ms_fragmentation;
823 if (valid_ms <= vd->vdev_ms_count / 2)
824 return (ZFS_FRAG_INVALID);
826 fragmentation /= valid_ms;
827 ASSERT3U(fragmentation, <=, 100);
828 return (fragmentation);
832 * Determine if a given metaslab group should skip allocations. A metaslab
833 * group should avoid allocations if its free capacity is less than the
834 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
835 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
836 * that can still handle allocations.
839 metaslab_group_allocatable(metaslab_group_t *mg)
841 vdev_t *vd = mg->mg_vd;
842 spa_t *spa = vd->vdev_spa;
843 metaslab_class_t *mc = mg->mg_class;
846 * We use two key metrics to determine if a metaslab group is
847 * considered allocatable -- free space and fragmentation. If
848 * the free space is greater than the free space threshold and
849 * the fragmentation is less than the fragmentation threshold then
850 * consider the group allocatable. There are two case when we will
851 * not consider these key metrics. The first is if the group is
852 * associated with a slog device and the second is if all groups
853 * in this metaslab class have already been consider ineligible
856 return ((mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
857 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
858 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)) ||
859 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
863 * ==========================================================================
864 * Range tree callbacks
865 * ==========================================================================
869 * Comparison function for the private size-ordered tree. Tree is sorted
870 * by size, larger sizes at the end of the tree.
873 metaslab_rangesize_compare(const void *x1, const void *x2)
875 const range_seg_t *r1 = x1;
876 const range_seg_t *r2 = x2;
877 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
878 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
880 if (rs_size1 < rs_size2)
882 if (rs_size1 > rs_size2)
885 if (r1->rs_start < r2->rs_start)
888 if (r1->rs_start > r2->rs_start)
895 * Create any block allocator specific components. The current allocators
896 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
899 metaslab_rt_create(range_tree_t *rt, void *arg)
901 metaslab_t *msp = arg;
903 ASSERT3P(rt->rt_arg, ==, msp);
904 ASSERT(msp->ms_tree == NULL);
906 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
907 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
911 * Destroy the block allocator specific components.
914 metaslab_rt_destroy(range_tree_t *rt, void *arg)
916 metaslab_t *msp = arg;
918 ASSERT3P(rt->rt_arg, ==, msp);
919 ASSERT3P(msp->ms_tree, ==, rt);
920 ASSERT0(avl_numnodes(&msp->ms_size_tree));
922 avl_destroy(&msp->ms_size_tree);
926 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
928 metaslab_t *msp = arg;
930 ASSERT3P(rt->rt_arg, ==, msp);
931 ASSERT3P(msp->ms_tree, ==, rt);
932 VERIFY(!msp->ms_condensing);
933 avl_add(&msp->ms_size_tree, rs);
937 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
939 metaslab_t *msp = arg;
941 ASSERT3P(rt->rt_arg, ==, msp);
942 ASSERT3P(msp->ms_tree, ==, rt);
943 VERIFY(!msp->ms_condensing);
944 avl_remove(&msp->ms_size_tree, rs);
948 metaslab_rt_vacate(range_tree_t *rt, void *arg)
950 metaslab_t *msp = arg;
952 ASSERT3P(rt->rt_arg, ==, msp);
953 ASSERT3P(msp->ms_tree, ==, rt);
956 * Normally one would walk the tree freeing nodes along the way.
957 * Since the nodes are shared with the range trees we can avoid
958 * walking all nodes and just reinitialize the avl tree. The nodes
959 * will be freed by the range tree, so we don't want to free them here.
961 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
962 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
965 static range_tree_ops_t metaslab_rt_ops = {
974 * ==========================================================================
975 * Metaslab block operations
976 * ==========================================================================
980 * Return the maximum contiguous segment within the metaslab.
983 metaslab_block_maxsize(metaslab_t *msp)
985 avl_tree_t *t = &msp->ms_size_tree;
988 if (t == NULL || (rs = avl_last(t)) == NULL)
991 return (rs->rs_end - rs->rs_start);
995 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
998 range_tree_t *rt = msp->ms_tree;
1000 VERIFY(!msp->ms_condensing);
1002 start = msp->ms_ops->msop_alloc(msp, size);
1003 if (start != -1ULL) {
1004 vdev_t *vd = msp->ms_group->mg_vd;
1006 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
1007 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1008 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
1009 range_tree_remove(rt, start, size);
1015 * ==========================================================================
1016 * Common allocator routines
1017 * ==========================================================================
1021 * This is a helper function that can be used by the allocator to find
1022 * a suitable block to allocate. This will search the specified AVL
1023 * tree looking for a block that matches the specified criteria.
1026 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1029 range_seg_t *rs, rsearch;
1032 rsearch.rs_start = *cursor;
1033 rsearch.rs_end = *cursor + size;
1035 rs = avl_find(t, &rsearch, &where);
1037 rs = avl_nearest(t, where, AVL_AFTER);
1039 while (rs != NULL) {
1040 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1042 if (offset + size <= rs->rs_end) {
1043 *cursor = offset + size;
1046 rs = AVL_NEXT(t, rs);
1050 * If we know we've searched the whole map (*cursor == 0), give up.
1051 * Otherwise, reset the cursor to the beginning and try again.
1057 return (metaslab_block_picker(t, cursor, size, align));
1061 * ==========================================================================
1062 * The first-fit block allocator
1063 * ==========================================================================
1066 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1069 * Find the largest power of 2 block size that evenly divides the
1070 * requested size. This is used to try to allocate blocks with similar
1071 * alignment from the same area of the metaslab (i.e. same cursor
1072 * bucket) but it does not guarantee that other allocations sizes
1073 * may exist in the same region.
1075 uint64_t align = size & -size;
1076 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1077 avl_tree_t *t = &msp->ms_tree->rt_root;
1079 return (metaslab_block_picker(t, cursor, size, align));
1082 static metaslab_ops_t metaslab_ff_ops = {
1087 * ==========================================================================
1088 * Dynamic block allocator -
1089 * Uses the first fit allocation scheme until space get low and then
1090 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1091 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1092 * ==========================================================================
1095 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1098 * Find the largest power of 2 block size that evenly divides the
1099 * requested size. This is used to try to allocate blocks with similar
1100 * alignment from the same area of the metaslab (i.e. same cursor
1101 * bucket) but it does not guarantee that other allocations sizes
1102 * may exist in the same region.
1104 uint64_t align = size & -size;
1105 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1106 range_tree_t *rt = msp->ms_tree;
1107 avl_tree_t *t = &rt->rt_root;
1108 uint64_t max_size = metaslab_block_maxsize(msp);
1109 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1111 ASSERT(MUTEX_HELD(&msp->ms_lock));
1112 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1114 if (max_size < size)
1118 * If we're running low on space switch to using the size
1119 * sorted AVL tree (best-fit).
1121 if (max_size < metaslab_df_alloc_threshold ||
1122 free_pct < metaslab_df_free_pct) {
1123 t = &msp->ms_size_tree;
1127 return (metaslab_block_picker(t, cursor, size, 1ULL));
1130 static metaslab_ops_t metaslab_df_ops = {
1135 * ==========================================================================
1136 * Cursor fit block allocator -
1137 * Select the largest region in the metaslab, set the cursor to the beginning
1138 * of the range and the cursor_end to the end of the range. As allocations
1139 * are made advance the cursor. Continue allocating from the cursor until
1140 * the range is exhausted and then find a new range.
1141 * ==========================================================================
1144 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1146 range_tree_t *rt = msp->ms_tree;
1147 avl_tree_t *t = &msp->ms_size_tree;
1148 uint64_t *cursor = &msp->ms_lbas[0];
1149 uint64_t *cursor_end = &msp->ms_lbas[1];
1150 uint64_t offset = 0;
1152 ASSERT(MUTEX_HELD(&msp->ms_lock));
1153 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1155 ASSERT3U(*cursor_end, >=, *cursor);
1157 if ((*cursor + size) > *cursor_end) {
1160 rs = avl_last(&msp->ms_size_tree);
1161 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1164 *cursor = rs->rs_start;
1165 *cursor_end = rs->rs_end;
1174 static metaslab_ops_t metaslab_cf_ops = {
1179 * ==========================================================================
1180 * New dynamic fit allocator -
1181 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1182 * contiguous blocks. If no region is found then just use the largest segment
1184 * ==========================================================================
1188 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1189 * to request from the allocator.
1191 uint64_t metaslab_ndf_clump_shift = 4;
1194 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1196 avl_tree_t *t = &msp->ms_tree->rt_root;
1198 range_seg_t *rs, rsearch;
1199 uint64_t hbit = highbit64(size);
1200 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1201 uint64_t max_size = metaslab_block_maxsize(msp);
1203 ASSERT(MUTEX_HELD(&msp->ms_lock));
1204 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1206 if (max_size < size)
1209 rsearch.rs_start = *cursor;
1210 rsearch.rs_end = *cursor + size;
1212 rs = avl_find(t, &rsearch, &where);
1213 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1214 t = &msp->ms_size_tree;
1216 rsearch.rs_start = 0;
1217 rsearch.rs_end = MIN(max_size,
1218 1ULL << (hbit + metaslab_ndf_clump_shift));
1219 rs = avl_find(t, &rsearch, &where);
1221 rs = avl_nearest(t, where, AVL_AFTER);
1225 if ((rs->rs_end - rs->rs_start) >= size) {
1226 *cursor = rs->rs_start + size;
1227 return (rs->rs_start);
1232 static metaslab_ops_t metaslab_ndf_ops = {
1236 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1239 * ==========================================================================
1241 * ==========================================================================
1245 * Wait for any in-progress metaslab loads to complete.
1248 metaslab_load_wait(metaslab_t *msp)
1250 ASSERT(MUTEX_HELD(&msp->ms_lock));
1252 while (msp->ms_loading) {
1253 ASSERT(!msp->ms_loaded);
1254 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1259 metaslab_load(metaslab_t *msp)
1263 ASSERT(MUTEX_HELD(&msp->ms_lock));
1264 ASSERT(!msp->ms_loaded);
1265 ASSERT(!msp->ms_loading);
1267 msp->ms_loading = B_TRUE;
1270 * If the space map has not been allocated yet, then treat
1271 * all the space in the metaslab as free and add it to the
1274 if (msp->ms_sm != NULL)
1275 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1277 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1279 msp->ms_loaded = (error == 0);
1280 msp->ms_loading = B_FALSE;
1282 if (msp->ms_loaded) {
1283 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1284 range_tree_walk(msp->ms_defertree[t],
1285 range_tree_remove, msp->ms_tree);
1288 cv_broadcast(&msp->ms_load_cv);
1293 metaslab_unload(metaslab_t *msp)
1295 ASSERT(MUTEX_HELD(&msp->ms_lock));
1296 range_tree_vacate(msp->ms_tree, NULL, NULL);
1297 msp->ms_loaded = B_FALSE;
1298 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1302 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1305 vdev_t *vd = mg->mg_vd;
1306 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1310 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1311 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1312 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1314 ms->ms_start = id << vd->vdev_ms_shift;
1315 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1318 * We only open space map objects that already exist. All others
1319 * will be opened when we finally allocate an object for it.
1322 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1323 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1326 kmem_free(ms, sizeof (metaslab_t));
1330 ASSERT(ms->ms_sm != NULL);
1334 * We create the main range tree here, but we don't create the
1335 * alloctree and freetree until metaslab_sync_done(). This serves
1336 * two purposes: it allows metaslab_sync_done() to detect the
1337 * addition of new space; and for debugging, it ensures that we'd
1338 * data fault on any attempt to use this metaslab before it's ready.
1340 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1341 metaslab_group_add(mg, ms);
1343 ms->ms_fragmentation = metaslab_fragmentation(ms);
1344 ms->ms_ops = mg->mg_class->mc_ops;
1347 * If we're opening an existing pool (txg == 0) or creating
1348 * a new one (txg == TXG_INITIAL), all space is available now.
1349 * If we're adding space to an existing pool, the new space
1350 * does not become available until after this txg has synced.
1352 if (txg <= TXG_INITIAL)
1353 metaslab_sync_done(ms, 0);
1356 * If metaslab_debug_load is set and we're initializing a metaslab
1357 * that has an allocated space_map object then load the its space
1358 * map so that can verify frees.
1360 if (metaslab_debug_load && ms->ms_sm != NULL) {
1361 mutex_enter(&ms->ms_lock);
1362 VERIFY0(metaslab_load(ms));
1363 mutex_exit(&ms->ms_lock);
1367 vdev_dirty(vd, 0, NULL, txg);
1368 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1377 metaslab_fini(metaslab_t *msp)
1379 metaslab_group_t *mg = msp->ms_group;
1381 metaslab_group_remove(mg, msp);
1383 mutex_enter(&msp->ms_lock);
1385 VERIFY(msp->ms_group == NULL);
1386 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1388 space_map_close(msp->ms_sm);
1390 metaslab_unload(msp);
1391 range_tree_destroy(msp->ms_tree);
1393 for (int t = 0; t < TXG_SIZE; t++) {
1394 range_tree_destroy(msp->ms_alloctree[t]);
1395 range_tree_destroy(msp->ms_freetree[t]);
1398 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1399 range_tree_destroy(msp->ms_defertree[t]);
1402 ASSERT0(msp->ms_deferspace);
1404 mutex_exit(&msp->ms_lock);
1405 cv_destroy(&msp->ms_load_cv);
1406 mutex_destroy(&msp->ms_lock);
1408 kmem_free(msp, sizeof (metaslab_t));
1411 #define FRAGMENTATION_TABLE_SIZE 17
1414 * This table defines a segment size based fragmentation metric that will
1415 * allow each metaslab to derive its own fragmentation value. This is done
1416 * by calculating the space in each bucket of the spacemap histogram and
1417 * multiplying that by the fragmetation metric in this table. Doing
1418 * this for all buckets and dividing it by the total amount of free
1419 * space in this metaslab (i.e. the total free space in all buckets) gives
1420 * us the fragmentation metric. This means that a high fragmentation metric
1421 * equates to most of the free space being comprised of small segments.
1422 * Conversely, if the metric is low, then most of the free space is in
1423 * large segments. A 10% change in fragmentation equates to approximately
1424 * double the number of segments.
1426 * This table defines 0% fragmented space using 16MB segments. Testing has
1427 * shown that segments that are greater than or equal to 16MB do not suffer
1428 * from drastic performance problems. Using this value, we derive the rest
1429 * of the table. Since the fragmentation value is never stored on disk, it
1430 * is possible to change these calculations in the future.
1432 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1452 * Calclate the metaslab's fragmentation metric. A return value
1453 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1454 * not support this metric. Otherwise, the return value should be in the
1458 metaslab_fragmentation(metaslab_t *msp)
1460 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1461 uint64_t fragmentation = 0;
1463 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1464 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1466 if (!feature_enabled)
1467 return (ZFS_FRAG_INVALID);
1470 * A null space map means that the entire metaslab is free
1471 * and thus is not fragmented.
1473 if (msp->ms_sm == NULL)
1477 * If this metaslab's space_map has not been upgraded, flag it
1478 * so that we upgrade next time we encounter it.
1480 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1481 uint64_t txg = spa_syncing_txg(spa);
1482 vdev_t *vd = msp->ms_group->mg_vd;
1484 if (spa_writeable(spa)) {
1485 msp->ms_condense_wanted = B_TRUE;
1486 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1487 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1488 "msp %p, vd %p", txg, msp, vd);
1490 return (ZFS_FRAG_INVALID);
1493 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1495 uint8_t shift = msp->ms_sm->sm_shift;
1496 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1497 FRAGMENTATION_TABLE_SIZE - 1);
1499 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1502 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1505 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1506 fragmentation += space * zfs_frag_table[idx];
1510 fragmentation /= total;
1511 ASSERT3U(fragmentation, <=, 100);
1512 return (fragmentation);
1516 * Compute a weight -- a selection preference value -- for the given metaslab.
1517 * This is based on the amount of free space, the level of fragmentation,
1518 * the LBA range, and whether the metaslab is loaded.
1521 metaslab_weight(metaslab_t *msp)
1523 metaslab_group_t *mg = msp->ms_group;
1524 vdev_t *vd = mg->mg_vd;
1525 uint64_t weight, space;
1527 ASSERT(MUTEX_HELD(&msp->ms_lock));
1530 * This vdev is in the process of being removed so there is nothing
1531 * for us to do here.
1533 if (vd->vdev_removing) {
1534 ASSERT0(space_map_allocated(msp->ms_sm));
1535 ASSERT0(vd->vdev_ms_shift);
1540 * The baseline weight is the metaslab's free space.
1542 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1544 msp->ms_fragmentation = metaslab_fragmentation(msp);
1545 if (metaslab_fragmentation_factor_enabled &&
1546 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1548 * Use the fragmentation information to inversely scale
1549 * down the baseline weight. We need to ensure that we
1550 * don't exclude this metaslab completely when it's 100%
1551 * fragmented. To avoid this we reduce the fragmented value
1554 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1557 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1558 * this metaslab again. The fragmentation metric may have
1559 * decreased the space to something smaller than
1560 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1561 * so that we can consume any remaining space.
1563 if (space > 0 && space < SPA_MINBLOCKSIZE)
1564 space = SPA_MINBLOCKSIZE;
1569 * Modern disks have uniform bit density and constant angular velocity.
1570 * Therefore, the outer recording zones are faster (higher bandwidth)
1571 * than the inner zones by the ratio of outer to inner track diameter,
1572 * which is typically around 2:1. We account for this by assigning
1573 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1574 * In effect, this means that we'll select the metaslab with the most
1575 * free bandwidth rather than simply the one with the most free space.
1577 if (metaslab_lba_weighting_enabled) {
1578 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1579 ASSERT(weight >= space && weight <= 2 * space);
1583 * If this metaslab is one we're actively using, adjust its
1584 * weight to make it preferable to any inactive metaslab so
1585 * we'll polish it off. If the fragmentation on this metaslab
1586 * has exceed our threshold, then don't mark it active.
1588 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1589 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1590 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1597 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1599 ASSERT(MUTEX_HELD(&msp->ms_lock));
1601 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1602 metaslab_load_wait(msp);
1603 if (!msp->ms_loaded) {
1604 int error = metaslab_load(msp);
1606 metaslab_group_sort(msp->ms_group, msp, 0);
1611 metaslab_group_sort(msp->ms_group, msp,
1612 msp->ms_weight | activation_weight);
1614 ASSERT(msp->ms_loaded);
1615 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1621 metaslab_passivate(metaslab_t *msp, uint64_t size)
1624 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1625 * this metaslab again. In that case, it had better be empty,
1626 * or we would be leaving space on the table.
1628 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1629 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1630 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1634 metaslab_preload(void *arg)
1636 metaslab_t *msp = arg;
1637 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1639 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1641 mutex_enter(&msp->ms_lock);
1642 metaslab_load_wait(msp);
1643 if (!msp->ms_loaded)
1644 (void) metaslab_load(msp);
1647 * Set the ms_access_txg value so that we don't unload it right away.
1649 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1650 mutex_exit(&msp->ms_lock);
1654 metaslab_group_preload(metaslab_group_t *mg)
1656 spa_t *spa = mg->mg_vd->vdev_spa;
1658 avl_tree_t *t = &mg->mg_metaslab_tree;
1661 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1662 taskq_wait(mg->mg_taskq);
1666 mutex_enter(&mg->mg_lock);
1668 * Load the next potential metaslabs
1671 while (msp != NULL) {
1672 metaslab_t *msp_next = AVL_NEXT(t, msp);
1675 * We preload only the maximum number of metaslabs specified
1676 * by metaslab_preload_limit. If a metaslab is being forced
1677 * to condense then we preload it too. This will ensure
1678 * that force condensing happens in the next txg.
1680 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
1686 * We must drop the metaslab group lock here to preserve
1687 * lock ordering with the ms_lock (when grabbing both
1688 * the mg_lock and the ms_lock, the ms_lock must be taken
1689 * first). As a result, it is possible that the ordering
1690 * of the metaslabs within the avl tree may change before
1691 * we reacquire the lock. The metaslab cannot be removed from
1692 * the tree while we're in syncing context so it is safe to
1693 * drop the mg_lock here. If the metaslabs are reordered
1694 * nothing will break -- we just may end up loading a
1695 * less than optimal one.
1697 mutex_exit(&mg->mg_lock);
1698 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1699 msp, TQ_SLEEP) != 0);
1700 mutex_enter(&mg->mg_lock);
1703 mutex_exit(&mg->mg_lock);
1707 * Determine if the space map's on-disk footprint is past our tolerance
1708 * for inefficiency. We would like to use the following criteria to make
1711 * 1. The size of the space map object should not dramatically increase as a
1712 * result of writing out the free space range tree.
1714 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1715 * times the size than the free space range tree representation
1716 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1718 * 3. The on-disk size of the space map should actually decrease.
1720 * Checking the first condition is tricky since we don't want to walk
1721 * the entire AVL tree calculating the estimated on-disk size. Instead we
1722 * use the size-ordered range tree in the metaslab and calculate the
1723 * size required to write out the largest segment in our free tree. If the
1724 * size required to represent that segment on disk is larger than the space
1725 * map object then we avoid condensing this map.
1727 * To determine the second criterion we use a best-case estimate and assume
1728 * each segment can be represented on-disk as a single 64-bit entry. We refer
1729 * to this best-case estimate as the space map's minimal form.
1731 * Unfortunately, we cannot compute the on-disk size of the space map in this
1732 * context because we cannot accurately compute the effects of compression, etc.
1733 * Instead, we apply the heuristic described in the block comment for
1734 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1735 * is greater than a threshold number of blocks.
1738 metaslab_should_condense(metaslab_t *msp)
1740 space_map_t *sm = msp->ms_sm;
1742 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
1743 dmu_object_info_t doi;
1744 uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
1746 ASSERT(MUTEX_HELD(&msp->ms_lock));
1747 ASSERT(msp->ms_loaded);
1750 * Use the ms_size_tree range tree, which is ordered by size, to
1751 * obtain the largest segment in the free tree. We always condense
1752 * metaslabs that are empty and metaslabs for which a condense
1753 * request has been made.
1755 rs = avl_last(&msp->ms_size_tree);
1756 if (rs == NULL || msp->ms_condense_wanted)
1760 * Calculate the number of 64-bit entries this segment would
1761 * require when written to disk. If this single segment would be
1762 * larger on-disk than the entire current on-disk structure, then
1763 * clearly condensing will increase the on-disk structure size.
1765 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1766 entries = size / (MIN(size, SM_RUN_MAX));
1767 segsz = entries * sizeof (uint64_t);
1769 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
1770 object_size = space_map_length(msp->ms_sm);
1772 dmu_object_info_from_db(sm->sm_dbuf, &doi);
1773 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
1775 return (segsz <= object_size &&
1776 object_size >= (optimal_size * zfs_condense_pct / 100) &&
1777 object_size > zfs_metaslab_condense_block_threshold * record_size);
1781 * Condense the on-disk space map representation to its minimized form.
1782 * The minimized form consists of a small number of allocations followed by
1783 * the entries of the free range tree.
1786 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1788 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1789 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1790 range_tree_t *condense_tree;
1791 space_map_t *sm = msp->ms_sm;
1793 ASSERT(MUTEX_HELD(&msp->ms_lock));
1794 ASSERT3U(spa_sync_pass(spa), ==, 1);
1795 ASSERT(msp->ms_loaded);
1798 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
1799 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
1800 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
1801 msp->ms_group->mg_vd->vdev_spa->spa_name,
1802 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
1803 msp->ms_condense_wanted ? "TRUE" : "FALSE");
1805 msp->ms_condense_wanted = B_FALSE;
1808 * Create an range tree that is 100% allocated. We remove segments
1809 * that have been freed in this txg, any deferred frees that exist,
1810 * and any allocation in the future. Removing segments should be
1811 * a relatively inexpensive operation since we expect these trees to
1812 * have a small number of nodes.
1814 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1815 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1818 * Remove what's been freed in this txg from the condense_tree.
1819 * Since we're in sync_pass 1, we know that all the frees from
1820 * this txg are in the freetree.
1822 range_tree_walk(freetree, range_tree_remove, condense_tree);
1824 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1825 range_tree_walk(msp->ms_defertree[t],
1826 range_tree_remove, condense_tree);
1829 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1830 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1831 range_tree_remove, condense_tree);
1835 * We're about to drop the metaslab's lock thus allowing
1836 * other consumers to change it's content. Set the
1837 * metaslab's ms_condensing flag to ensure that
1838 * allocations on this metaslab do not occur while we're
1839 * in the middle of committing it to disk. This is only critical
1840 * for the ms_tree as all other range trees use per txg
1841 * views of their content.
1843 msp->ms_condensing = B_TRUE;
1845 mutex_exit(&msp->ms_lock);
1846 space_map_truncate(sm, tx);
1847 mutex_enter(&msp->ms_lock);
1850 * While we would ideally like to create a space_map representation
1851 * that consists only of allocation records, doing so can be
1852 * prohibitively expensive because the in-core free tree can be
1853 * large, and therefore computationally expensive to subtract
1854 * from the condense_tree. Instead we sync out two trees, a cheap
1855 * allocation only tree followed by the in-core free tree. While not
1856 * optimal, this is typically close to optimal, and much cheaper to
1859 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1860 range_tree_vacate(condense_tree, NULL, NULL);
1861 range_tree_destroy(condense_tree);
1863 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1864 msp->ms_condensing = B_FALSE;
1868 * Write a metaslab to disk in the context of the specified transaction group.
1871 metaslab_sync(metaslab_t *msp, uint64_t txg)
1873 metaslab_group_t *mg = msp->ms_group;
1874 vdev_t *vd = mg->mg_vd;
1875 spa_t *spa = vd->vdev_spa;
1876 objset_t *mos = spa_meta_objset(spa);
1877 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1878 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1879 range_tree_t **freed_tree =
1880 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1882 uint64_t object = space_map_object(msp->ms_sm);
1884 ASSERT(!vd->vdev_ishole);
1887 * This metaslab has just been added so there's no work to do now.
1889 if (*freetree == NULL) {
1890 ASSERT3P(alloctree, ==, NULL);
1894 ASSERT3P(alloctree, !=, NULL);
1895 ASSERT3P(*freetree, !=, NULL);
1896 ASSERT3P(*freed_tree, !=, NULL);
1899 * Normally, we don't want to process a metaslab if there
1900 * are no allocations or frees to perform. However, if the metaslab
1901 * is being forced to condense we need to let it through.
1903 if (range_tree_space(alloctree) == 0 &&
1904 range_tree_space(*freetree) == 0 &&
1905 !msp->ms_condense_wanted)
1909 * The only state that can actually be changing concurrently with
1910 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1911 * be modifying this txg's alloctree, freetree, freed_tree, or
1912 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1913 * space_map ASSERTs. We drop it whenever we call into the DMU,
1914 * because the DMU can call down to us (e.g. via zio_free()) at
1918 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1920 if (msp->ms_sm == NULL) {
1921 uint64_t new_object;
1923 new_object = space_map_alloc(mos, tx);
1924 VERIFY3U(new_object, !=, 0);
1926 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1927 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1929 ASSERT(msp->ms_sm != NULL);
1932 mutex_enter(&msp->ms_lock);
1935 * Note: metaslab_condense() clears the space_map's histogram.
1936 * Therefore we must verify and remove this histogram before
1939 metaslab_group_histogram_verify(mg);
1940 metaslab_class_histogram_verify(mg->mg_class);
1941 metaslab_group_histogram_remove(mg, msp);
1943 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1944 metaslab_should_condense(msp)) {
1945 metaslab_condense(msp, txg, tx);
1947 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1948 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1951 if (msp->ms_loaded) {
1953 * When the space map is loaded, we have an accruate
1954 * histogram in the range tree. This gives us an opportunity
1955 * to bring the space map's histogram up-to-date so we clear
1956 * it first before updating it.
1958 space_map_histogram_clear(msp->ms_sm);
1959 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1962 * Since the space map is not loaded we simply update the
1963 * exisiting histogram with what was freed in this txg. This
1964 * means that the on-disk histogram may not have an accurate
1965 * view of the free space but it's close enough to allow
1966 * us to make allocation decisions.
1968 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1970 metaslab_group_histogram_add(mg, msp);
1971 metaslab_group_histogram_verify(mg);
1972 metaslab_class_histogram_verify(mg->mg_class);
1975 * For sync pass 1, we avoid traversing this txg's free range tree
1976 * and instead will just swap the pointers for freetree and
1977 * freed_tree. We can safely do this since the freed_tree is
1978 * guaranteed to be empty on the initial pass.
1980 if (spa_sync_pass(spa) == 1) {
1981 range_tree_swap(freetree, freed_tree);
1983 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1985 range_tree_vacate(alloctree, NULL, NULL);
1987 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1988 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1990 mutex_exit(&msp->ms_lock);
1992 if (object != space_map_object(msp->ms_sm)) {
1993 object = space_map_object(msp->ms_sm);
1994 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1995 msp->ms_id, sizeof (uint64_t), &object, tx);
2001 * Called after a transaction group has completely synced to mark
2002 * all of the metaslab's free space as usable.
2005 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2007 metaslab_group_t *mg = msp->ms_group;
2008 vdev_t *vd = mg->mg_vd;
2009 range_tree_t **freed_tree;
2010 range_tree_t **defer_tree;
2011 int64_t alloc_delta, defer_delta;
2013 ASSERT(!vd->vdev_ishole);
2015 mutex_enter(&msp->ms_lock);
2018 * If this metaslab is just becoming available, initialize its
2019 * alloctrees, freetrees, and defertree and add its capacity to
2022 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
2023 for (int t = 0; t < TXG_SIZE; t++) {
2024 ASSERT(msp->ms_alloctree[t] == NULL);
2025 ASSERT(msp->ms_freetree[t] == NULL);
2027 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2029 msp->ms_freetree[t] = range_tree_create(NULL, msp,
2033 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2034 ASSERT(msp->ms_defertree[t] == NULL);
2036 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2040 vdev_space_update(vd, 0, 0, msp->ms_size);
2043 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2044 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2046 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2047 defer_delta = range_tree_space(*freed_tree) -
2048 range_tree_space(*defer_tree);
2050 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2052 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2053 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2056 * If there's a metaslab_load() in progress, wait for it to complete
2057 * so that we have a consistent view of the in-core space map.
2059 metaslab_load_wait(msp);
2062 * Move the frees from the defer_tree back to the free
2063 * range tree (if it's loaded). Swap the freed_tree and the
2064 * defer_tree -- this is safe to do because we've just emptied out
2067 range_tree_vacate(*defer_tree,
2068 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2069 range_tree_swap(freed_tree, defer_tree);
2071 space_map_update(msp->ms_sm);
2073 msp->ms_deferspace += defer_delta;
2074 ASSERT3S(msp->ms_deferspace, >=, 0);
2075 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2076 if (msp->ms_deferspace != 0) {
2078 * Keep syncing this metaslab until all deferred frees
2079 * are back in circulation.
2081 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2084 if (msp->ms_loaded && msp->ms_access_txg < txg) {
2085 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2086 VERIFY0(range_tree_space(
2087 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2090 if (!metaslab_debug_unload)
2091 metaslab_unload(msp);
2094 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2095 mutex_exit(&msp->ms_lock);
2099 metaslab_sync_reassess(metaslab_group_t *mg)
2101 metaslab_group_alloc_update(mg);
2102 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2105 * Preload the next potential metaslabs
2107 metaslab_group_preload(mg);
2111 metaslab_distance(metaslab_t *msp, dva_t *dva)
2113 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2114 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2115 uint64_t start = msp->ms_id;
2117 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2118 return (1ULL << 63);
2121 return ((start - offset) << ms_shift);
2123 return ((offset - start) << ms_shift);
2128 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
2129 uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2131 spa_t *spa = mg->mg_vd->vdev_spa;
2132 metaslab_t *msp = NULL;
2133 uint64_t offset = -1ULL;
2134 avl_tree_t *t = &mg->mg_metaslab_tree;
2135 uint64_t activation_weight;
2136 uint64_t target_distance;
2139 activation_weight = METASLAB_WEIGHT_PRIMARY;
2140 for (i = 0; i < d; i++) {
2141 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2142 activation_weight = METASLAB_WEIGHT_SECONDARY;
2148 boolean_t was_active;
2150 mutex_enter(&mg->mg_lock);
2151 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
2152 if (msp->ms_weight < asize) {
2153 spa_dbgmsg(spa, "%s: failed to meet weight "
2154 "requirement: vdev %llu, txg %llu, mg %p, "
2155 "msp %p, psize %llu, asize %llu, "
2156 "weight %llu", spa_name(spa),
2157 mg->mg_vd->vdev_id, txg,
2158 mg, msp, psize, asize, msp->ms_weight);
2159 mutex_exit(&mg->mg_lock);
2164 * If the selected metaslab is condensing, skip it.
2166 if (msp->ms_condensing)
2169 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2170 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2173 target_distance = min_distance +
2174 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2177 for (i = 0; i < d; i++)
2178 if (metaslab_distance(msp, &dva[i]) <
2184 mutex_exit(&mg->mg_lock);
2188 mutex_enter(&msp->ms_lock);
2191 * Ensure that the metaslab we have selected is still
2192 * capable of handling our request. It's possible that
2193 * another thread may have changed the weight while we
2194 * were blocked on the metaslab lock.
2196 if (msp->ms_weight < asize || (was_active &&
2197 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
2198 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
2199 mutex_exit(&msp->ms_lock);
2203 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2204 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2205 metaslab_passivate(msp,
2206 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2207 mutex_exit(&msp->ms_lock);
2211 if (metaslab_activate(msp, activation_weight) != 0) {
2212 mutex_exit(&msp->ms_lock);
2217 * If this metaslab is currently condensing then pick again as
2218 * we can't manipulate this metaslab until it's committed
2221 if (msp->ms_condensing) {
2222 mutex_exit(&msp->ms_lock);
2226 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
2229 metaslab_passivate(msp, metaslab_block_maxsize(msp));
2230 mutex_exit(&msp->ms_lock);
2233 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2234 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2236 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
2237 msp->ms_access_txg = txg + metaslab_unload_delay;
2239 mutex_exit(&msp->ms_lock);
2245 * Allocate a block for the specified i/o.
2248 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2249 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
2251 metaslab_group_t *mg, *rotor;
2255 int zio_lock = B_FALSE;
2256 boolean_t allocatable;
2257 uint64_t offset = -1ULL;
2261 ASSERT(!DVA_IS_VALID(&dva[d]));
2264 * For testing, make some blocks above a certain size be gang blocks.
2266 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
2267 return (SET_ERROR(ENOSPC));
2270 * Start at the rotor and loop through all mgs until we find something.
2271 * Note that there's no locking on mc_rotor or mc_aliquot because
2272 * nothing actually breaks if we miss a few updates -- we just won't
2273 * allocate quite as evenly. It all balances out over time.
2275 * If we are doing ditto or log blocks, try to spread them across
2276 * consecutive vdevs. If we're forced to reuse a vdev before we've
2277 * allocated all of our ditto blocks, then try and spread them out on
2278 * that vdev as much as possible. If it turns out to not be possible,
2279 * gradually lower our standards until anything becomes acceptable.
2280 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2281 * gives us hope of containing our fault domains to something we're
2282 * able to reason about. Otherwise, any two top-level vdev failures
2283 * will guarantee the loss of data. With consecutive allocation,
2284 * only two adjacent top-level vdev failures will result in data loss.
2286 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2287 * ourselves on the same vdev as our gang block header. That
2288 * way, we can hope for locality in vdev_cache, plus it makes our
2289 * fault domains something tractable.
2292 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
2295 * It's possible the vdev we're using as the hint no
2296 * longer exists (i.e. removed). Consult the rotor when
2302 if (flags & METASLAB_HINTBP_AVOID &&
2303 mg->mg_next != NULL)
2308 } else if (d != 0) {
2309 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
2310 mg = vd->vdev_mg->mg_next;
2316 * If the hint put us into the wrong metaslab class, or into a
2317 * metaslab group that has been passivated, just follow the rotor.
2319 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2326 ASSERT(mg->mg_activation_count == 1);
2331 * Don't allocate from faulted devices.
2334 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
2335 allocatable = vdev_allocatable(vd);
2336 spa_config_exit(spa, SCL_ZIO, FTAG);
2338 allocatable = vdev_allocatable(vd);
2342 * Determine if the selected metaslab group is eligible
2343 * for allocations. If we're ganging or have requested
2344 * an allocation for the smallest gang block size
2345 * then we don't want to avoid allocating to the this
2346 * metaslab group. If we're in this condition we should
2347 * try to allocate from any device possible so that we
2348 * don't inadvertently return ENOSPC and suspend the pool
2349 * even though space is still available.
2351 if (allocatable && CAN_FASTGANG(flags) &&
2352 psize > SPA_GANGBLOCKSIZE)
2353 allocatable = metaslab_group_allocatable(mg);
2359 * Avoid writing single-copy data to a failing vdev
2360 * unless the user instructs us that it is okay.
2362 if ((vd->vdev_stat.vs_write_errors > 0 ||
2363 vd->vdev_state < VDEV_STATE_HEALTHY) &&
2364 d == 0 && dshift == 3 && vd->vdev_children == 0) {
2369 ASSERT(mg->mg_class == mc);
2371 distance = vd->vdev_asize >> dshift;
2372 if (distance <= (1ULL << vd->vdev_ms_shift))
2377 asize = vdev_psize_to_asize(vd, psize);
2378 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
2380 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
2382 if (offset != -1ULL) {
2384 * If we've just selected this metaslab group,
2385 * figure out whether the corresponding vdev is
2386 * over- or under-used relative to the pool,
2387 * and set an allocation bias to even it out.
2389 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
2390 vdev_stat_t *vs = &vd->vdev_stat;
2393 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
2394 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2397 * Calculate how much more or less we should
2398 * try to allocate from this device during
2399 * this iteration around the rotor.
2400 * For example, if a device is 80% full
2401 * and the pool is 20% full then we should
2402 * reduce allocations by 60% on this device.
2404 * mg_bias = (20 - 80) * 512K / 100 = -307K
2406 * This reduces allocations by 307K for this
2409 mg->mg_bias = ((cu - vu) *
2410 (int64_t)mg->mg_aliquot) / 100;
2411 } else if (!metaslab_bias_enabled) {
2415 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2416 mg->mg_aliquot + mg->mg_bias) {
2417 mc->mc_rotor = mg->mg_next;
2421 DVA_SET_VDEV(&dva[d], vd->vdev_id);
2422 DVA_SET_OFFSET(&dva[d], offset);
2423 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2424 DVA_SET_ASIZE(&dva[d], asize);
2429 mc->mc_rotor = mg->mg_next;
2431 } while ((mg = mg->mg_next) != rotor);
2435 ASSERT(dshift < 64);
2439 if (!allocatable && !zio_lock) {
2445 bzero(&dva[d], sizeof (dva_t));
2447 return (SET_ERROR(ENOSPC));
2451 * Free the block represented by DVA in the context of the specified
2452 * transaction group.
2455 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2457 uint64_t vdev = DVA_GET_VDEV(dva);
2458 uint64_t offset = DVA_GET_OFFSET(dva);
2459 uint64_t size = DVA_GET_ASIZE(dva);
2463 ASSERT(DVA_IS_VALID(dva));
2465 if (txg > spa_freeze_txg(spa))
2468 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2469 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2470 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2471 (u_longlong_t)vdev, (u_longlong_t)offset);
2476 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2478 if (DVA_GET_GANG(dva))
2479 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2481 mutex_enter(&msp->ms_lock);
2484 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2487 VERIFY(!msp->ms_condensing);
2488 VERIFY3U(offset, >=, msp->ms_start);
2489 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2490 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2492 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2493 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2494 range_tree_add(msp->ms_tree, offset, size);
2496 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2497 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2498 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2502 mutex_exit(&msp->ms_lock);
2506 * Intent log support: upon opening the pool after a crash, notify the SPA
2507 * of blocks that the intent log has allocated for immediate write, but
2508 * which are still considered free by the SPA because the last transaction
2509 * group didn't commit yet.
2512 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2514 uint64_t vdev = DVA_GET_VDEV(dva);
2515 uint64_t offset = DVA_GET_OFFSET(dva);
2516 uint64_t size = DVA_GET_ASIZE(dva);
2521 ASSERT(DVA_IS_VALID(dva));
2523 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2524 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2525 return (SET_ERROR(ENXIO));
2527 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2529 if (DVA_GET_GANG(dva))
2530 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2532 mutex_enter(&msp->ms_lock);
2534 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2535 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2537 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2538 error = SET_ERROR(ENOENT);
2540 if (error || txg == 0) { /* txg == 0 indicates dry run */
2541 mutex_exit(&msp->ms_lock);
2545 VERIFY(!msp->ms_condensing);
2546 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2547 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2548 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2549 range_tree_remove(msp->ms_tree, offset, size);
2551 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2552 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2553 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2554 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2557 mutex_exit(&msp->ms_lock);
2563 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2564 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2566 dva_t *dva = bp->blk_dva;
2567 dva_t *hintdva = hintbp->blk_dva;
2570 ASSERT(bp->blk_birth == 0);
2571 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2573 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2575 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2576 spa_config_exit(spa, SCL_ALLOC, FTAG);
2577 return (SET_ERROR(ENOSPC));
2580 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2581 ASSERT(BP_GET_NDVAS(bp) == 0);
2582 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2584 for (int d = 0; d < ndvas; d++) {
2585 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2588 for (d--; d >= 0; d--) {
2589 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2590 bzero(&dva[d], sizeof (dva_t));
2592 spa_config_exit(spa, SCL_ALLOC, FTAG);
2597 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2599 spa_config_exit(spa, SCL_ALLOC, FTAG);
2601 BP_SET_BIRTH(bp, txg, txg);
2607 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2609 const dva_t *dva = bp->blk_dva;
2610 int ndvas = BP_GET_NDVAS(bp);
2612 ASSERT(!BP_IS_HOLE(bp));
2613 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2615 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2617 for (int d = 0; d < ndvas; d++)
2618 metaslab_free_dva(spa, &dva[d], txg, now);
2620 spa_config_exit(spa, SCL_FREE, FTAG);
2624 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2626 const dva_t *dva = bp->blk_dva;
2627 int ndvas = BP_GET_NDVAS(bp);
2630 ASSERT(!BP_IS_HOLE(bp));
2634 * First do a dry run to make sure all DVAs are claimable,
2635 * so we don't have to unwind from partial failures below.
2637 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2641 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2643 for (int d = 0; d < ndvas; d++)
2644 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2647 spa_config_exit(spa, SCL_ALLOC, FTAG);
2649 ASSERT(error == 0 || txg == 0);
2655 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2657 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2660 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2661 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2662 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2663 vdev_t *vd = vdev_lookup_top(spa, vdev);
2664 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2665 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2666 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2669 range_tree_verify(msp->ms_tree, offset, size);
2671 for (int j = 0; j < TXG_SIZE; j++)
2672 range_tree_verify(msp->ms_freetree[j], offset, size);
2673 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2674 range_tree_verify(msp->ms_defertree[j], offset, size);
2676 spa_config_exit(spa, SCL_VDEV, FTAG);