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 TUNABLE_QUAD("vfs.zfs.metaslab.gang_bang", &metaslab_gang_bang);
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 TUNABLE_INT("vfs.zfs.condense_pct", &zfs_condense_pct);
72 SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN,
74 "Condense on-disk spacemap when it is more than this many percents"
75 " of in-memory counterpart");
78 * Condensing a metaslab is not guaranteed to actually reduce the amount of
79 * space used on disk. In particular, a space map uses data in increments of
80 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
81 * same number of blocks after condensing. Since the goal of condensing is to
82 * reduce the number of IOPs required to read the space map, we only want to
83 * condense when we can be sure we will reduce the number of blocks used by the
84 * space map. Unfortunately, we cannot precisely compute whether or not this is
85 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
86 * we apply the following heuristic: do not condense a spacemap unless the
87 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
90 int zfs_metaslab_condense_block_threshold = 4;
93 * The zfs_mg_noalloc_threshold defines which metaslab groups should
94 * be eligible for allocation. The value is defined as a percentage of
95 * free space. Metaslab groups that have more free space than
96 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
97 * a metaslab group's free space is less than or equal to the
98 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
99 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
100 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
101 * groups are allowed to accept allocations. Gang blocks are always
102 * eligible to allocate on any metaslab group. The default value of 0 means
103 * no metaslab group will be excluded based on this criterion.
105 int zfs_mg_noalloc_threshold = 0;
106 TUNABLE_INT("vfs.zfs.mg_noalloc_threshold", &zfs_mg_noalloc_threshold);
107 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN,
108 &zfs_mg_noalloc_threshold, 0,
109 "Percentage of metaslab group size that should be free"
110 " to make it eligible for allocation");
113 * Metaslab groups are considered eligible for allocations if their
114 * fragmenation metric (measured as a percentage) is less than or equal to
115 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
116 * then it will be skipped unless all metaslab groups within the metaslab
117 * class have also crossed this threshold.
119 int zfs_mg_fragmentation_threshold = 85;
120 TUNABLE_INT("vfs.zfs.mg_fragmentation_threshold", &zfs_mg_fragmentation_threshold);
121 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_fragmentation_threshold, CTLFLAG_RWTUN,
122 &zfs_mg_fragmentation_threshold, 0,
123 "Percentage of metaslab group size that should be considered "
124 "eligible for allocations unless all metaslab groups within the metaslab class "
125 "have also crossed this threshold");
128 * Allow metaslabs to keep their active state as long as their fragmentation
129 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
130 * active metaslab that exceeds this threshold will no longer keep its active
131 * status allowing better metaslabs to be selected.
133 int zfs_metaslab_fragmentation_threshold = 70;
134 TUNABLE_INT("vfs.zfs.metaslab.fragmentation_threshold",
135 &zfs_metaslab_fragmentation_threshold);
136 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_threshold, CTLFLAG_RWTUN,
137 &zfs_metaslab_fragmentation_threshold, 0,
138 "Maximum percentage of metaslab fragmentation level to keep their active state");
141 * When set will load all metaslabs when pool is first opened.
143 int metaslab_debug_load = 0;
144 TUNABLE_INT("vfs.zfs.metaslab.debug_load", &metaslab_debug_load);
145 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN,
146 &metaslab_debug_load, 0,
147 "Load all metaslabs when pool is first opened");
150 * When set will prevent metaslabs from being unloaded.
152 int metaslab_debug_unload = 0;
153 TUNABLE_INT("vfs.zfs.metaslab.debug_unload", &metaslab_debug_unload);
154 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN,
155 &metaslab_debug_unload, 0,
156 "Prevent metaslabs from being unloaded");
159 * Minimum size which forces the dynamic allocator to change
160 * it's allocation strategy. Once the space map cannot satisfy
161 * an allocation of this size then it switches to using more
162 * aggressive strategy (i.e search by size rather than offset).
164 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
165 TUNABLE_QUAD("vfs.zfs.metaslab.df_alloc_threshold",
166 &metaslab_df_alloc_threshold);
167 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN,
168 &metaslab_df_alloc_threshold, 0,
169 "Minimum size which forces the dynamic allocator to change it's allocation strategy");
172 * The minimum free space, in percent, which must be available
173 * in a space map to continue allocations in a first-fit fashion.
174 * Once the space_map's free space drops below this level we dynamically
175 * switch to using best-fit allocations.
177 int metaslab_df_free_pct = 4;
178 TUNABLE_INT("vfs.zfs.metaslab.df_free_pct", &metaslab_df_free_pct);
179 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN,
180 &metaslab_df_free_pct, 0,
181 "The minimum free space, in percent, which must be available in a space map to continue allocations in a first-fit fashion");
184 * A metaslab is considered "free" if it contains a contiguous
185 * segment which is greater than metaslab_min_alloc_size.
187 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
188 TUNABLE_QUAD("vfs.zfs.metaslab.min_alloc_size",
189 &metaslab_min_alloc_size);
190 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN,
191 &metaslab_min_alloc_size, 0,
192 "A metaslab is considered \"free\" if it contains a contiguous segment which is greater than vfs.zfs.metaslab.min_alloc_size");
195 * Percentage of all cpus that can be used by the metaslab taskq.
197 int metaslab_load_pct = 50;
198 TUNABLE_INT("vfs.zfs.metaslab.load_pct", &metaslab_load_pct);
199 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN,
200 &metaslab_load_pct, 0,
201 "Percentage of cpus that can be used by the metaslab taskq");
204 * Determines how many txgs a metaslab may remain loaded without having any
205 * allocations from it. As long as a metaslab continues to be used we will
208 int metaslab_unload_delay = TXG_SIZE * 2;
209 TUNABLE_INT("vfs.zfs.metaslab.unload_delay", &metaslab_unload_delay);
210 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN,
211 &metaslab_unload_delay, 0,
212 "Number of TXGs that an unused metaslab can be kept in memory");
215 * Max number of metaslabs per group to preload.
217 int metaslab_preload_limit = SPA_DVAS_PER_BP;
218 TUNABLE_INT("vfs.zfs.metaslab.preload_limit", &metaslab_preload_limit);
219 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
220 &metaslab_preload_limit, 0,
221 "Max number of metaslabs per group to preload");
224 * Enable/disable preloading of metaslab.
226 boolean_t metaslab_preload_enabled = B_TRUE;
227 TUNABLE_INT("vfs.zfs.metaslab.preload_enabled", &metaslab_preload_enabled);
228 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
229 &metaslab_preload_enabled, 0,
230 "Max number of metaslabs per group to preload");
233 * Enable/disable fragmentation weighting on metaslabs.
235 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
236 TUNABLE_INT("vfs.zfs.metaslab_fragmentation_factor_enabled",
237 &metaslab_fragmentation_factor_enabled);
238 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_factor_enabled, CTLFLAG_RWTUN,
239 &metaslab_fragmentation_factor_enabled, 0,
240 "Enable fragmentation weighting on metaslabs");
243 * Enable/disable lba weighting (i.e. outer tracks are given preference).
245 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
246 TUNABLE_INT("vfs.zfs.metaslab.lba_weighting_enabled",
247 &metaslab_lba_weighting_enabled);
248 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, lba_weighting_enabled, CTLFLAG_RWTUN,
249 &metaslab_lba_weighting_enabled, 0,
250 "Enable LBA weighting (i.e. outer tracks are given preference)");
253 * Enable/disable metaslab group biasing.
255 boolean_t metaslab_bias_enabled = B_TRUE;
256 TUNABLE_INT("vfs.zfs.metaslab.bias_enabled",
257 &metaslab_bias_enabled);
258 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, bias_enabled, CTLFLAG_RWTUN,
259 &metaslab_bias_enabled, 0,
260 "Enable metaslab group biasing");
262 static uint64_t metaslab_fragmentation(metaslab_t *);
265 * ==========================================================================
267 * ==========================================================================
270 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
272 metaslab_class_t *mc;
274 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
284 metaslab_class_destroy(metaslab_class_t *mc)
286 ASSERT(mc->mc_rotor == NULL);
287 ASSERT(mc->mc_alloc == 0);
288 ASSERT(mc->mc_deferred == 0);
289 ASSERT(mc->mc_space == 0);
290 ASSERT(mc->mc_dspace == 0);
292 kmem_free(mc, sizeof (metaslab_class_t));
296 metaslab_class_validate(metaslab_class_t *mc)
298 metaslab_group_t *mg;
302 * Must hold one of the spa_config locks.
304 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
305 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
307 if ((mg = mc->mc_rotor) == NULL)
312 ASSERT(vd->vdev_mg != NULL);
313 ASSERT3P(vd->vdev_top, ==, vd);
314 ASSERT3P(mg->mg_class, ==, mc);
315 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
316 } while ((mg = mg->mg_next) != mc->mc_rotor);
322 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
323 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
325 atomic_add_64(&mc->mc_alloc, alloc_delta);
326 atomic_add_64(&mc->mc_deferred, defer_delta);
327 atomic_add_64(&mc->mc_space, space_delta);
328 atomic_add_64(&mc->mc_dspace, dspace_delta);
332 metaslab_class_minblocksize_update(metaslab_class_t *mc)
334 metaslab_group_t *mg;
336 uint64_t minashift = UINT64_MAX;
338 if ((mg = mc->mc_rotor) == NULL) {
339 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
345 if (vd->vdev_ashift < minashift)
346 minashift = vd->vdev_ashift;
347 } while ((mg = mg->mg_next) != mc->mc_rotor);
349 mc->mc_minblocksize = 1ULL << minashift;
353 metaslab_class_get_alloc(metaslab_class_t *mc)
355 return (mc->mc_alloc);
359 metaslab_class_get_deferred(metaslab_class_t *mc)
361 return (mc->mc_deferred);
365 metaslab_class_get_space(metaslab_class_t *mc)
367 return (mc->mc_space);
371 metaslab_class_get_dspace(metaslab_class_t *mc)
373 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
377 metaslab_class_get_minblocksize(metaslab_class_t *mc)
379 return (mc->mc_minblocksize);
383 metaslab_class_histogram_verify(metaslab_class_t *mc)
385 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
389 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
392 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
395 for (int c = 0; c < rvd->vdev_children; c++) {
396 vdev_t *tvd = rvd->vdev_child[c];
397 metaslab_group_t *mg = tvd->vdev_mg;
400 * Skip any holes, uninitialized top-levels, or
401 * vdevs that are not in this metalab class.
403 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
404 mg->mg_class != mc) {
408 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
409 mc_hist[i] += mg->mg_histogram[i];
412 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
413 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
415 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
419 * Calculate the metaslab class's fragmentation metric. The metric
420 * is weighted based on the space contribution of each metaslab group.
421 * The return value will be a number between 0 and 100 (inclusive), or
422 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
423 * zfs_frag_table for more information about the metric.
426 metaslab_class_fragmentation(metaslab_class_t *mc)
428 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
429 uint64_t fragmentation = 0;
431 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
433 for (int c = 0; c < rvd->vdev_children; c++) {
434 vdev_t *tvd = rvd->vdev_child[c];
435 metaslab_group_t *mg = tvd->vdev_mg;
438 * Skip any holes, uninitialized top-levels, or
439 * vdevs that are not in this metalab class.
441 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
442 mg->mg_class != mc) {
447 * If a metaslab group does not contain a fragmentation
448 * metric then just bail out.
450 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
451 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
452 return (ZFS_FRAG_INVALID);
456 * Determine how much this metaslab_group is contributing
457 * to the overall pool fragmentation metric.
459 fragmentation += mg->mg_fragmentation *
460 metaslab_group_get_space(mg);
462 fragmentation /= metaslab_class_get_space(mc);
464 ASSERT3U(fragmentation, <=, 100);
465 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
466 return (fragmentation);
470 * Calculate the amount of expandable space that is available in
471 * this metaslab class. If a device is expanded then its expandable
472 * space will be the amount of allocatable space that is currently not
473 * part of this metaslab class.
476 metaslab_class_expandable_space(metaslab_class_t *mc)
478 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
481 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
482 for (int c = 0; c < rvd->vdev_children; c++) {
483 vdev_t *tvd = rvd->vdev_child[c];
484 metaslab_group_t *mg = tvd->vdev_mg;
486 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
487 mg->mg_class != mc) {
491 space += tvd->vdev_max_asize - tvd->vdev_asize;
493 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
498 * ==========================================================================
500 * ==========================================================================
503 metaslab_compare(const void *x1, const void *x2)
505 const metaslab_t *m1 = x1;
506 const metaslab_t *m2 = x2;
508 if (m1->ms_weight < m2->ms_weight)
510 if (m1->ms_weight > m2->ms_weight)
514 * If the weights are identical, use the offset to force uniqueness.
516 if (m1->ms_start < m2->ms_start)
518 if (m1->ms_start > m2->ms_start)
521 ASSERT3P(m1, ==, m2);
527 * Update the allocatable flag and the metaslab group's capacity.
528 * The allocatable flag is set to true if the capacity is below
529 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
530 * from allocatable to non-allocatable or vice versa then the metaslab
531 * group's class is updated to reflect the transition.
534 metaslab_group_alloc_update(metaslab_group_t *mg)
536 vdev_t *vd = mg->mg_vd;
537 metaslab_class_t *mc = mg->mg_class;
538 vdev_stat_t *vs = &vd->vdev_stat;
539 boolean_t was_allocatable;
541 ASSERT(vd == vd->vdev_top);
543 mutex_enter(&mg->mg_lock);
544 was_allocatable = mg->mg_allocatable;
546 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
550 * A metaslab group is considered allocatable if it has plenty
551 * of free space or is not heavily fragmented. We only take
552 * fragmentation into account if the metaslab group has a valid
553 * fragmentation metric (i.e. a value between 0 and 100).
555 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
556 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
557 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
560 * The mc_alloc_groups maintains a count of the number of
561 * groups in this metaslab class that are still above the
562 * zfs_mg_noalloc_threshold. This is used by the allocating
563 * threads to determine if they should avoid allocations to
564 * a given group. The allocator will avoid allocations to a group
565 * if that group has reached or is below the zfs_mg_noalloc_threshold
566 * and there are still other groups that are above the threshold.
567 * When a group transitions from allocatable to non-allocatable or
568 * vice versa we update the metaslab class to reflect that change.
569 * When the mc_alloc_groups value drops to 0 that means that all
570 * groups have reached the zfs_mg_noalloc_threshold making all groups
571 * eligible for allocations. This effectively means that all devices
572 * are balanced again.
574 if (was_allocatable && !mg->mg_allocatable)
575 mc->mc_alloc_groups--;
576 else if (!was_allocatable && mg->mg_allocatable)
577 mc->mc_alloc_groups++;
579 mutex_exit(&mg->mg_lock);
583 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
585 metaslab_group_t *mg;
587 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
588 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
589 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
590 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
593 mg->mg_activation_count = 0;
595 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
596 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
602 metaslab_group_destroy(metaslab_group_t *mg)
604 ASSERT(mg->mg_prev == NULL);
605 ASSERT(mg->mg_next == NULL);
607 * We may have gone below zero with the activation count
608 * either because we never activated in the first place or
609 * because we're done, and possibly removing the vdev.
611 ASSERT(mg->mg_activation_count <= 0);
613 taskq_destroy(mg->mg_taskq);
614 avl_destroy(&mg->mg_metaslab_tree);
615 mutex_destroy(&mg->mg_lock);
616 kmem_free(mg, sizeof (metaslab_group_t));
620 metaslab_group_activate(metaslab_group_t *mg)
622 metaslab_class_t *mc = mg->mg_class;
623 metaslab_group_t *mgprev, *mgnext;
625 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
627 ASSERT(mc->mc_rotor != mg);
628 ASSERT(mg->mg_prev == NULL);
629 ASSERT(mg->mg_next == NULL);
630 ASSERT(mg->mg_activation_count <= 0);
632 if (++mg->mg_activation_count <= 0)
635 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
636 metaslab_group_alloc_update(mg);
638 if ((mgprev = mc->mc_rotor) == NULL) {
642 mgnext = mgprev->mg_next;
643 mg->mg_prev = mgprev;
644 mg->mg_next = mgnext;
645 mgprev->mg_next = mg;
646 mgnext->mg_prev = mg;
649 metaslab_class_minblocksize_update(mc);
653 metaslab_group_passivate(metaslab_group_t *mg)
655 metaslab_class_t *mc = mg->mg_class;
656 metaslab_group_t *mgprev, *mgnext;
658 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
660 if (--mg->mg_activation_count != 0) {
661 ASSERT(mc->mc_rotor != mg);
662 ASSERT(mg->mg_prev == NULL);
663 ASSERT(mg->mg_next == NULL);
664 ASSERT(mg->mg_activation_count < 0);
668 taskq_wait(mg->mg_taskq);
669 metaslab_group_alloc_update(mg);
671 mgprev = mg->mg_prev;
672 mgnext = mg->mg_next;
677 mc->mc_rotor = mgnext;
678 mgprev->mg_next = mgnext;
679 mgnext->mg_prev = mgprev;
684 metaslab_class_minblocksize_update(mc);
688 metaslab_group_get_space(metaslab_group_t *mg)
690 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
694 metaslab_group_histogram_verify(metaslab_group_t *mg)
697 vdev_t *vd = mg->mg_vd;
698 uint64_t ashift = vd->vdev_ashift;
701 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
704 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
707 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
708 SPACE_MAP_HISTOGRAM_SIZE + ashift);
710 for (int m = 0; m < vd->vdev_ms_count; m++) {
711 metaslab_t *msp = vd->vdev_ms[m];
713 if (msp->ms_sm == NULL)
716 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
717 mg_hist[i + ashift] +=
718 msp->ms_sm->sm_phys->smp_histogram[i];
721 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
722 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
724 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
728 metaslab_group_histogram_add(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 mg->mg_histogram[i + ashift] +=
740 msp->ms_sm->sm_phys->smp_histogram[i];
741 mc->mc_histogram[i + ashift] +=
742 msp->ms_sm->sm_phys->smp_histogram[i];
744 mutex_exit(&mg->mg_lock);
748 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
750 metaslab_class_t *mc = mg->mg_class;
751 uint64_t ashift = mg->mg_vd->vdev_ashift;
753 ASSERT(MUTEX_HELD(&msp->ms_lock));
754 if (msp->ms_sm == NULL)
757 mutex_enter(&mg->mg_lock);
758 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
759 ASSERT3U(mg->mg_histogram[i + ashift], >=,
760 msp->ms_sm->sm_phys->smp_histogram[i]);
761 ASSERT3U(mc->mc_histogram[i + ashift], >=,
762 msp->ms_sm->sm_phys->smp_histogram[i]);
764 mg->mg_histogram[i + ashift] -=
765 msp->ms_sm->sm_phys->smp_histogram[i];
766 mc->mc_histogram[i + ashift] -=
767 msp->ms_sm->sm_phys->smp_histogram[i];
769 mutex_exit(&mg->mg_lock);
773 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
775 ASSERT(msp->ms_group == NULL);
776 mutex_enter(&mg->mg_lock);
779 avl_add(&mg->mg_metaslab_tree, msp);
780 mutex_exit(&mg->mg_lock);
782 mutex_enter(&msp->ms_lock);
783 metaslab_group_histogram_add(mg, msp);
784 mutex_exit(&msp->ms_lock);
788 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
790 mutex_enter(&msp->ms_lock);
791 metaslab_group_histogram_remove(mg, msp);
792 mutex_exit(&msp->ms_lock);
794 mutex_enter(&mg->mg_lock);
795 ASSERT(msp->ms_group == mg);
796 avl_remove(&mg->mg_metaslab_tree, msp);
797 msp->ms_group = NULL;
798 mutex_exit(&mg->mg_lock);
802 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
805 * Although in principle the weight can be any value, in
806 * practice we do not use values in the range [1, 511].
808 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
809 ASSERT(MUTEX_HELD(&msp->ms_lock));
811 mutex_enter(&mg->mg_lock);
812 ASSERT(msp->ms_group == mg);
813 avl_remove(&mg->mg_metaslab_tree, msp);
814 msp->ms_weight = weight;
815 avl_add(&mg->mg_metaslab_tree, msp);
816 mutex_exit(&mg->mg_lock);
820 * Calculate the fragmentation for a given metaslab group. We can use
821 * a simple average here since all metaslabs within the group must have
822 * the same size. The return value will be a value between 0 and 100
823 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
824 * group have a fragmentation metric.
827 metaslab_group_fragmentation(metaslab_group_t *mg)
829 vdev_t *vd = mg->mg_vd;
830 uint64_t fragmentation = 0;
831 uint64_t valid_ms = 0;
833 for (int m = 0; m < vd->vdev_ms_count; m++) {
834 metaslab_t *msp = vd->vdev_ms[m];
836 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
840 fragmentation += msp->ms_fragmentation;
843 if (valid_ms <= vd->vdev_ms_count / 2)
844 return (ZFS_FRAG_INVALID);
846 fragmentation /= valid_ms;
847 ASSERT3U(fragmentation, <=, 100);
848 return (fragmentation);
852 * Determine if a given metaslab group should skip allocations. A metaslab
853 * group should avoid allocations if its free capacity is less than the
854 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
855 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
856 * that can still handle allocations.
859 metaslab_group_allocatable(metaslab_group_t *mg)
861 vdev_t *vd = mg->mg_vd;
862 spa_t *spa = vd->vdev_spa;
863 metaslab_class_t *mc = mg->mg_class;
866 * We use two key metrics to determine if a metaslab group is
867 * considered allocatable -- free space and fragmentation. If
868 * the free space is greater than the free space threshold and
869 * the fragmentation is less than the fragmentation threshold then
870 * consider the group allocatable. There are two case when we will
871 * not consider these key metrics. The first is if the group is
872 * associated with a slog device and the second is if all groups
873 * in this metaslab class have already been consider ineligible
876 return ((mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
877 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
878 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)) ||
879 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
883 * ==========================================================================
884 * Range tree callbacks
885 * ==========================================================================
889 * Comparison function for the private size-ordered tree. Tree is sorted
890 * by size, larger sizes at the end of the tree.
893 metaslab_rangesize_compare(const void *x1, const void *x2)
895 const range_seg_t *r1 = x1;
896 const range_seg_t *r2 = x2;
897 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
898 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
900 if (rs_size1 < rs_size2)
902 if (rs_size1 > rs_size2)
905 if (r1->rs_start < r2->rs_start)
908 if (r1->rs_start > r2->rs_start)
915 * Create any block allocator specific components. The current allocators
916 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
919 metaslab_rt_create(range_tree_t *rt, void *arg)
921 metaslab_t *msp = arg;
923 ASSERT3P(rt->rt_arg, ==, msp);
924 ASSERT(msp->ms_tree == NULL);
926 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
927 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
931 * Destroy the block allocator specific components.
934 metaslab_rt_destroy(range_tree_t *rt, void *arg)
936 metaslab_t *msp = arg;
938 ASSERT3P(rt->rt_arg, ==, msp);
939 ASSERT3P(msp->ms_tree, ==, rt);
940 ASSERT0(avl_numnodes(&msp->ms_size_tree));
942 avl_destroy(&msp->ms_size_tree);
946 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
948 metaslab_t *msp = arg;
950 ASSERT3P(rt->rt_arg, ==, msp);
951 ASSERT3P(msp->ms_tree, ==, rt);
952 VERIFY(!msp->ms_condensing);
953 avl_add(&msp->ms_size_tree, rs);
957 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
959 metaslab_t *msp = arg;
961 ASSERT3P(rt->rt_arg, ==, msp);
962 ASSERT3P(msp->ms_tree, ==, rt);
963 VERIFY(!msp->ms_condensing);
964 avl_remove(&msp->ms_size_tree, rs);
968 metaslab_rt_vacate(range_tree_t *rt, void *arg)
970 metaslab_t *msp = arg;
972 ASSERT3P(rt->rt_arg, ==, msp);
973 ASSERT3P(msp->ms_tree, ==, rt);
976 * Normally one would walk the tree freeing nodes along the way.
977 * Since the nodes are shared with the range trees we can avoid
978 * walking all nodes and just reinitialize the avl tree. The nodes
979 * will be freed by the range tree, so we don't want to free them here.
981 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
982 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
985 static range_tree_ops_t metaslab_rt_ops = {
994 * ==========================================================================
995 * Metaslab block operations
996 * ==========================================================================
1000 * Return the maximum contiguous segment within the metaslab.
1003 metaslab_block_maxsize(metaslab_t *msp)
1005 avl_tree_t *t = &msp->ms_size_tree;
1008 if (t == NULL || (rs = avl_last(t)) == NULL)
1011 return (rs->rs_end - rs->rs_start);
1015 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
1018 range_tree_t *rt = msp->ms_tree;
1020 VERIFY(!msp->ms_condensing);
1022 start = msp->ms_ops->msop_alloc(msp, size);
1023 if (start != -1ULL) {
1024 vdev_t *vd = msp->ms_group->mg_vd;
1026 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
1027 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1028 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
1029 range_tree_remove(rt, start, size);
1035 * ==========================================================================
1036 * Common allocator routines
1037 * ==========================================================================
1041 * This is a helper function that can be used by the allocator to find
1042 * a suitable block to allocate. This will search the specified AVL
1043 * tree looking for a block that matches the specified criteria.
1046 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1049 range_seg_t *rs, rsearch;
1052 rsearch.rs_start = *cursor;
1053 rsearch.rs_end = *cursor + size;
1055 rs = avl_find(t, &rsearch, &where);
1057 rs = avl_nearest(t, where, AVL_AFTER);
1059 while (rs != NULL) {
1060 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1062 if (offset + size <= rs->rs_end) {
1063 *cursor = offset + size;
1066 rs = AVL_NEXT(t, rs);
1070 * If we know we've searched the whole map (*cursor == 0), give up.
1071 * Otherwise, reset the cursor to the beginning and try again.
1077 return (metaslab_block_picker(t, cursor, size, align));
1081 * ==========================================================================
1082 * The first-fit block allocator
1083 * ==========================================================================
1086 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1089 * Find the largest power of 2 block size that evenly divides the
1090 * requested size. This is used to try to allocate blocks with similar
1091 * alignment from the same area of the metaslab (i.e. same cursor
1092 * bucket) but it does not guarantee that other allocations sizes
1093 * may exist in the same region.
1095 uint64_t align = size & -size;
1096 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1097 avl_tree_t *t = &msp->ms_tree->rt_root;
1099 return (metaslab_block_picker(t, cursor, size, align));
1102 static metaslab_ops_t metaslab_ff_ops = {
1107 * ==========================================================================
1108 * Dynamic block allocator -
1109 * Uses the first fit allocation scheme until space get low and then
1110 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1111 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1112 * ==========================================================================
1115 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1118 * Find the largest power of 2 block size that evenly divides the
1119 * requested size. This is used to try to allocate blocks with similar
1120 * alignment from the same area of the metaslab (i.e. same cursor
1121 * bucket) but it does not guarantee that other allocations sizes
1122 * may exist in the same region.
1124 uint64_t align = size & -size;
1125 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1126 range_tree_t *rt = msp->ms_tree;
1127 avl_tree_t *t = &rt->rt_root;
1128 uint64_t max_size = metaslab_block_maxsize(msp);
1129 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1131 ASSERT(MUTEX_HELD(&msp->ms_lock));
1132 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1134 if (max_size < size)
1138 * If we're running low on space switch to using the size
1139 * sorted AVL tree (best-fit).
1141 if (max_size < metaslab_df_alloc_threshold ||
1142 free_pct < metaslab_df_free_pct) {
1143 t = &msp->ms_size_tree;
1147 return (metaslab_block_picker(t, cursor, size, 1ULL));
1150 static metaslab_ops_t metaslab_df_ops = {
1155 * ==========================================================================
1156 * Cursor fit block allocator -
1157 * Select the largest region in the metaslab, set the cursor to the beginning
1158 * of the range and the cursor_end to the end of the range. As allocations
1159 * are made advance the cursor. Continue allocating from the cursor until
1160 * the range is exhausted and then find a new range.
1161 * ==========================================================================
1164 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1166 range_tree_t *rt = msp->ms_tree;
1167 avl_tree_t *t = &msp->ms_size_tree;
1168 uint64_t *cursor = &msp->ms_lbas[0];
1169 uint64_t *cursor_end = &msp->ms_lbas[1];
1170 uint64_t offset = 0;
1172 ASSERT(MUTEX_HELD(&msp->ms_lock));
1173 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1175 ASSERT3U(*cursor_end, >=, *cursor);
1177 if ((*cursor + size) > *cursor_end) {
1180 rs = avl_last(&msp->ms_size_tree);
1181 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1184 *cursor = rs->rs_start;
1185 *cursor_end = rs->rs_end;
1194 static metaslab_ops_t metaslab_cf_ops = {
1199 * ==========================================================================
1200 * New dynamic fit allocator -
1201 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1202 * contiguous blocks. If no region is found then just use the largest segment
1204 * ==========================================================================
1208 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1209 * to request from the allocator.
1211 uint64_t metaslab_ndf_clump_shift = 4;
1214 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1216 avl_tree_t *t = &msp->ms_tree->rt_root;
1218 range_seg_t *rs, rsearch;
1219 uint64_t hbit = highbit64(size);
1220 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1221 uint64_t max_size = metaslab_block_maxsize(msp);
1223 ASSERT(MUTEX_HELD(&msp->ms_lock));
1224 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1226 if (max_size < size)
1229 rsearch.rs_start = *cursor;
1230 rsearch.rs_end = *cursor + size;
1232 rs = avl_find(t, &rsearch, &where);
1233 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1234 t = &msp->ms_size_tree;
1236 rsearch.rs_start = 0;
1237 rsearch.rs_end = MIN(max_size,
1238 1ULL << (hbit + metaslab_ndf_clump_shift));
1239 rs = avl_find(t, &rsearch, &where);
1241 rs = avl_nearest(t, where, AVL_AFTER);
1245 if ((rs->rs_end - rs->rs_start) >= size) {
1246 *cursor = rs->rs_start + size;
1247 return (rs->rs_start);
1252 static metaslab_ops_t metaslab_ndf_ops = {
1256 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1259 * ==========================================================================
1261 * ==========================================================================
1265 * Wait for any in-progress metaslab loads to complete.
1268 metaslab_load_wait(metaslab_t *msp)
1270 ASSERT(MUTEX_HELD(&msp->ms_lock));
1272 while (msp->ms_loading) {
1273 ASSERT(!msp->ms_loaded);
1274 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1279 metaslab_load(metaslab_t *msp)
1283 ASSERT(MUTEX_HELD(&msp->ms_lock));
1284 ASSERT(!msp->ms_loaded);
1285 ASSERT(!msp->ms_loading);
1287 msp->ms_loading = B_TRUE;
1290 * If the space map has not been allocated yet, then treat
1291 * all the space in the metaslab as free and add it to the
1294 if (msp->ms_sm != NULL)
1295 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1297 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1299 msp->ms_loaded = (error == 0);
1300 msp->ms_loading = B_FALSE;
1302 if (msp->ms_loaded) {
1303 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1304 range_tree_walk(msp->ms_defertree[t],
1305 range_tree_remove, msp->ms_tree);
1308 cv_broadcast(&msp->ms_load_cv);
1313 metaslab_unload(metaslab_t *msp)
1315 ASSERT(MUTEX_HELD(&msp->ms_lock));
1316 range_tree_vacate(msp->ms_tree, NULL, NULL);
1317 msp->ms_loaded = B_FALSE;
1318 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1322 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1325 vdev_t *vd = mg->mg_vd;
1326 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1330 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1331 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1332 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1334 ms->ms_start = id << vd->vdev_ms_shift;
1335 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1338 * We only open space map objects that already exist. All others
1339 * will be opened when we finally allocate an object for it.
1342 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1343 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1346 kmem_free(ms, sizeof (metaslab_t));
1350 ASSERT(ms->ms_sm != NULL);
1354 * We create the main range tree here, but we don't create the
1355 * alloctree and freetree until metaslab_sync_done(). This serves
1356 * two purposes: it allows metaslab_sync_done() to detect the
1357 * addition of new space; and for debugging, it ensures that we'd
1358 * data fault on any attempt to use this metaslab before it's ready.
1360 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1361 metaslab_group_add(mg, ms);
1363 ms->ms_fragmentation = metaslab_fragmentation(ms);
1364 ms->ms_ops = mg->mg_class->mc_ops;
1367 * If we're opening an existing pool (txg == 0) or creating
1368 * a new one (txg == TXG_INITIAL), all space is available now.
1369 * If we're adding space to an existing pool, the new space
1370 * does not become available until after this txg has synced.
1372 if (txg <= TXG_INITIAL)
1373 metaslab_sync_done(ms, 0);
1376 * If metaslab_debug_load is set and we're initializing a metaslab
1377 * that has an allocated space_map object then load the its space
1378 * map so that can verify frees.
1380 if (metaslab_debug_load && ms->ms_sm != NULL) {
1381 mutex_enter(&ms->ms_lock);
1382 VERIFY0(metaslab_load(ms));
1383 mutex_exit(&ms->ms_lock);
1387 vdev_dirty(vd, 0, NULL, txg);
1388 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1397 metaslab_fini(metaslab_t *msp)
1399 metaslab_group_t *mg = msp->ms_group;
1401 metaslab_group_remove(mg, msp);
1403 mutex_enter(&msp->ms_lock);
1405 VERIFY(msp->ms_group == NULL);
1406 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1408 space_map_close(msp->ms_sm);
1410 metaslab_unload(msp);
1411 range_tree_destroy(msp->ms_tree);
1413 for (int t = 0; t < TXG_SIZE; t++) {
1414 range_tree_destroy(msp->ms_alloctree[t]);
1415 range_tree_destroy(msp->ms_freetree[t]);
1418 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1419 range_tree_destroy(msp->ms_defertree[t]);
1422 ASSERT0(msp->ms_deferspace);
1424 mutex_exit(&msp->ms_lock);
1425 cv_destroy(&msp->ms_load_cv);
1426 mutex_destroy(&msp->ms_lock);
1428 kmem_free(msp, sizeof (metaslab_t));
1431 #define FRAGMENTATION_TABLE_SIZE 17
1434 * This table defines a segment size based fragmentation metric that will
1435 * allow each metaslab to derive its own fragmentation value. This is done
1436 * by calculating the space in each bucket of the spacemap histogram and
1437 * multiplying that by the fragmetation metric in this table. Doing
1438 * this for all buckets and dividing it by the total amount of free
1439 * space in this metaslab (i.e. the total free space in all buckets) gives
1440 * us the fragmentation metric. This means that a high fragmentation metric
1441 * equates to most of the free space being comprised of small segments.
1442 * Conversely, if the metric is low, then most of the free space is in
1443 * large segments. A 10% change in fragmentation equates to approximately
1444 * double the number of segments.
1446 * This table defines 0% fragmented space using 16MB segments. Testing has
1447 * shown that segments that are greater than or equal to 16MB do not suffer
1448 * from drastic performance problems. Using this value, we derive the rest
1449 * of the table. Since the fragmentation value is never stored on disk, it
1450 * is possible to change these calculations in the future.
1452 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1472 * Calclate the metaslab's fragmentation metric. A return value
1473 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1474 * not support this metric. Otherwise, the return value should be in the
1478 metaslab_fragmentation(metaslab_t *msp)
1480 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1481 uint64_t fragmentation = 0;
1483 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1484 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1486 if (!feature_enabled)
1487 return (ZFS_FRAG_INVALID);
1490 * A null space map means that the entire metaslab is free
1491 * and thus is not fragmented.
1493 if (msp->ms_sm == NULL)
1497 * If this metaslab's space_map has not been upgraded, flag it
1498 * so that we upgrade next time we encounter it.
1500 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1501 uint64_t txg = spa_syncing_txg(spa);
1502 vdev_t *vd = msp->ms_group->mg_vd;
1504 if (spa_writeable(spa)) {
1505 msp->ms_condense_wanted = B_TRUE;
1506 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1507 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1508 "msp %p, vd %p", txg, msp, vd);
1510 return (ZFS_FRAG_INVALID);
1513 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1515 uint8_t shift = msp->ms_sm->sm_shift;
1516 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1517 FRAGMENTATION_TABLE_SIZE - 1);
1519 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1522 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1525 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1526 fragmentation += space * zfs_frag_table[idx];
1530 fragmentation /= total;
1531 ASSERT3U(fragmentation, <=, 100);
1532 return (fragmentation);
1536 * Compute a weight -- a selection preference value -- for the given metaslab.
1537 * This is based on the amount of free space, the level of fragmentation,
1538 * the LBA range, and whether the metaslab is loaded.
1541 metaslab_weight(metaslab_t *msp)
1543 metaslab_group_t *mg = msp->ms_group;
1544 vdev_t *vd = mg->mg_vd;
1545 uint64_t weight, space;
1547 ASSERT(MUTEX_HELD(&msp->ms_lock));
1550 * This vdev is in the process of being removed so there is nothing
1551 * for us to do here.
1553 if (vd->vdev_removing) {
1554 ASSERT0(space_map_allocated(msp->ms_sm));
1555 ASSERT0(vd->vdev_ms_shift);
1560 * The baseline weight is the metaslab's free space.
1562 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1564 msp->ms_fragmentation = metaslab_fragmentation(msp);
1565 if (metaslab_fragmentation_factor_enabled &&
1566 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1568 * Use the fragmentation information to inversely scale
1569 * down the baseline weight. We need to ensure that we
1570 * don't exclude this metaslab completely when it's 100%
1571 * fragmented. To avoid this we reduce the fragmented value
1574 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1577 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1578 * this metaslab again. The fragmentation metric may have
1579 * decreased the space to something smaller than
1580 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1581 * so that we can consume any remaining space.
1583 if (space > 0 && space < SPA_MINBLOCKSIZE)
1584 space = SPA_MINBLOCKSIZE;
1589 * Modern disks have uniform bit density and constant angular velocity.
1590 * Therefore, the outer recording zones are faster (higher bandwidth)
1591 * than the inner zones by the ratio of outer to inner track diameter,
1592 * which is typically around 2:1. We account for this by assigning
1593 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1594 * In effect, this means that we'll select the metaslab with the most
1595 * free bandwidth rather than simply the one with the most free space.
1597 if (metaslab_lba_weighting_enabled) {
1598 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1599 ASSERT(weight >= space && weight <= 2 * space);
1603 * If this metaslab is one we're actively using, adjust its
1604 * weight to make it preferable to any inactive metaslab so
1605 * we'll polish it off. If the fragmentation on this metaslab
1606 * has exceed our threshold, then don't mark it active.
1608 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1609 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1610 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1617 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1619 ASSERT(MUTEX_HELD(&msp->ms_lock));
1621 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1622 metaslab_load_wait(msp);
1623 if (!msp->ms_loaded) {
1624 int error = metaslab_load(msp);
1626 metaslab_group_sort(msp->ms_group, msp, 0);
1631 metaslab_group_sort(msp->ms_group, msp,
1632 msp->ms_weight | activation_weight);
1634 ASSERT(msp->ms_loaded);
1635 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1641 metaslab_passivate(metaslab_t *msp, uint64_t size)
1644 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1645 * this metaslab again. In that case, it had better be empty,
1646 * or we would be leaving space on the table.
1648 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1649 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1650 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1654 metaslab_preload(void *arg)
1656 metaslab_t *msp = arg;
1657 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1659 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1661 mutex_enter(&msp->ms_lock);
1662 metaslab_load_wait(msp);
1663 if (!msp->ms_loaded)
1664 (void) metaslab_load(msp);
1667 * Set the ms_access_txg value so that we don't unload it right away.
1669 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1670 mutex_exit(&msp->ms_lock);
1674 metaslab_group_preload(metaslab_group_t *mg)
1676 spa_t *spa = mg->mg_vd->vdev_spa;
1678 avl_tree_t *t = &mg->mg_metaslab_tree;
1681 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1682 taskq_wait(mg->mg_taskq);
1686 mutex_enter(&mg->mg_lock);
1688 * Load the next potential metaslabs
1691 while (msp != NULL) {
1692 metaslab_t *msp_next = AVL_NEXT(t, msp);
1695 * We preload only the maximum number of metaslabs specified
1696 * by metaslab_preload_limit. If a metaslab is being forced
1697 * to condense then we preload it too. This will ensure
1698 * that force condensing happens in the next txg.
1700 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
1706 * We must drop the metaslab group lock here to preserve
1707 * lock ordering with the ms_lock (when grabbing both
1708 * the mg_lock and the ms_lock, the ms_lock must be taken
1709 * first). As a result, it is possible that the ordering
1710 * of the metaslabs within the avl tree may change before
1711 * we reacquire the lock. The metaslab cannot be removed from
1712 * the tree while we're in syncing context so it is safe to
1713 * drop the mg_lock here. If the metaslabs are reordered
1714 * nothing will break -- we just may end up loading a
1715 * less than optimal one.
1717 mutex_exit(&mg->mg_lock);
1718 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1719 msp, TQ_SLEEP) != 0);
1720 mutex_enter(&mg->mg_lock);
1723 mutex_exit(&mg->mg_lock);
1727 * Determine if the space map's on-disk footprint is past our tolerance
1728 * for inefficiency. We would like to use the following criteria to make
1731 * 1. The size of the space map object should not dramatically increase as a
1732 * result of writing out the free space range tree.
1734 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1735 * times the size than the free space range tree representation
1736 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1738 * 3. The on-disk size of the space map should actually decrease.
1740 * Checking the first condition is tricky since we don't want to walk
1741 * the entire AVL tree calculating the estimated on-disk size. Instead we
1742 * use the size-ordered range tree in the metaslab and calculate the
1743 * size required to write out the largest segment in our free tree. If the
1744 * size required to represent that segment on disk is larger than the space
1745 * map object then we avoid condensing this map.
1747 * To determine the second criterion we use a best-case estimate and assume
1748 * each segment can be represented on-disk as a single 64-bit entry. We refer
1749 * to this best-case estimate as the space map's minimal form.
1751 * Unfortunately, we cannot compute the on-disk size of the space map in this
1752 * context because we cannot accurately compute the effects of compression, etc.
1753 * Instead, we apply the heuristic described in the block comment for
1754 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1755 * is greater than a threshold number of blocks.
1758 metaslab_should_condense(metaslab_t *msp)
1760 space_map_t *sm = msp->ms_sm;
1762 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
1763 dmu_object_info_t doi;
1764 uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
1766 ASSERT(MUTEX_HELD(&msp->ms_lock));
1767 ASSERT(msp->ms_loaded);
1770 * Use the ms_size_tree range tree, which is ordered by size, to
1771 * obtain the largest segment in the free tree. We always condense
1772 * metaslabs that are empty and metaslabs for which a condense
1773 * request has been made.
1775 rs = avl_last(&msp->ms_size_tree);
1776 if (rs == NULL || msp->ms_condense_wanted)
1780 * Calculate the number of 64-bit entries this segment would
1781 * require when written to disk. If this single segment would be
1782 * larger on-disk than the entire current on-disk structure, then
1783 * clearly condensing will increase the on-disk structure size.
1785 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1786 entries = size / (MIN(size, SM_RUN_MAX));
1787 segsz = entries * sizeof (uint64_t);
1789 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
1790 object_size = space_map_length(msp->ms_sm);
1792 dmu_object_info_from_db(sm->sm_dbuf, &doi);
1793 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
1795 return (segsz <= object_size &&
1796 object_size >= (optimal_size * zfs_condense_pct / 100) &&
1797 object_size > zfs_metaslab_condense_block_threshold * record_size);
1801 * Condense the on-disk space map representation to its minimized form.
1802 * The minimized form consists of a small number of allocations followed by
1803 * the entries of the free range tree.
1806 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1808 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1809 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1810 range_tree_t *condense_tree;
1811 space_map_t *sm = msp->ms_sm;
1813 ASSERT(MUTEX_HELD(&msp->ms_lock));
1814 ASSERT3U(spa_sync_pass(spa), ==, 1);
1815 ASSERT(msp->ms_loaded);
1818 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1819 "smp size %llu, segments %lu, forcing condense=%s", txg,
1820 msp->ms_id, msp, space_map_length(msp->ms_sm),
1821 avl_numnodes(&msp->ms_tree->rt_root),
1822 msp->ms_condense_wanted ? "TRUE" : "FALSE");
1824 msp->ms_condense_wanted = B_FALSE;
1827 * Create an range tree that is 100% allocated. We remove segments
1828 * that have been freed in this txg, any deferred frees that exist,
1829 * and any allocation in the future. Removing segments should be
1830 * a relatively inexpensive operation since we expect these trees to
1831 * have a small number of nodes.
1833 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1834 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1837 * Remove what's been freed in this txg from the condense_tree.
1838 * Since we're in sync_pass 1, we know that all the frees from
1839 * this txg are in the freetree.
1841 range_tree_walk(freetree, range_tree_remove, condense_tree);
1843 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1844 range_tree_walk(msp->ms_defertree[t],
1845 range_tree_remove, condense_tree);
1848 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1849 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1850 range_tree_remove, condense_tree);
1854 * We're about to drop the metaslab's lock thus allowing
1855 * other consumers to change it's content. Set the
1856 * metaslab's ms_condensing flag to ensure that
1857 * allocations on this metaslab do not occur while we're
1858 * in the middle of committing it to disk. This is only critical
1859 * for the ms_tree as all other range trees use per txg
1860 * views of their content.
1862 msp->ms_condensing = B_TRUE;
1864 mutex_exit(&msp->ms_lock);
1865 space_map_truncate(sm, tx);
1866 mutex_enter(&msp->ms_lock);
1869 * While we would ideally like to create a space_map representation
1870 * that consists only of allocation records, doing so can be
1871 * prohibitively expensive because the in-core free tree can be
1872 * large, and therefore computationally expensive to subtract
1873 * from the condense_tree. Instead we sync out two trees, a cheap
1874 * allocation only tree followed by the in-core free tree. While not
1875 * optimal, this is typically close to optimal, and much cheaper to
1878 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1879 range_tree_vacate(condense_tree, NULL, NULL);
1880 range_tree_destroy(condense_tree);
1882 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1883 msp->ms_condensing = B_FALSE;
1887 * Write a metaslab to disk in the context of the specified transaction group.
1890 metaslab_sync(metaslab_t *msp, uint64_t txg)
1892 metaslab_group_t *mg = msp->ms_group;
1893 vdev_t *vd = mg->mg_vd;
1894 spa_t *spa = vd->vdev_spa;
1895 objset_t *mos = spa_meta_objset(spa);
1896 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1897 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1898 range_tree_t **freed_tree =
1899 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1901 uint64_t object = space_map_object(msp->ms_sm);
1903 ASSERT(!vd->vdev_ishole);
1906 * This metaslab has just been added so there's no work to do now.
1908 if (*freetree == NULL) {
1909 ASSERT3P(alloctree, ==, NULL);
1913 ASSERT3P(alloctree, !=, NULL);
1914 ASSERT3P(*freetree, !=, NULL);
1915 ASSERT3P(*freed_tree, !=, NULL);
1918 * Normally, we don't want to process a metaslab if there
1919 * are no allocations or frees to perform. However, if the metaslab
1920 * is being forced to condense we need to let it through.
1922 if (range_tree_space(alloctree) == 0 &&
1923 range_tree_space(*freetree) == 0 &&
1924 !msp->ms_condense_wanted)
1928 * The only state that can actually be changing concurrently with
1929 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1930 * be modifying this txg's alloctree, freetree, freed_tree, or
1931 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1932 * space_map ASSERTs. We drop it whenever we call into the DMU,
1933 * because the DMU can call down to us (e.g. via zio_free()) at
1937 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1939 if (msp->ms_sm == NULL) {
1940 uint64_t new_object;
1942 new_object = space_map_alloc(mos, tx);
1943 VERIFY3U(new_object, !=, 0);
1945 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1946 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1948 ASSERT(msp->ms_sm != NULL);
1951 mutex_enter(&msp->ms_lock);
1954 * Note: metaslab_condense() clears the space_map's histogram.
1955 * Therefore we must verify and remove this histogram before
1958 metaslab_group_histogram_verify(mg);
1959 metaslab_class_histogram_verify(mg->mg_class);
1960 metaslab_group_histogram_remove(mg, msp);
1962 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1963 metaslab_should_condense(msp)) {
1964 metaslab_condense(msp, txg, tx);
1966 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1967 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1970 if (msp->ms_loaded) {
1972 * When the space map is loaded, we have an accruate
1973 * histogram in the range tree. This gives us an opportunity
1974 * to bring the space map's histogram up-to-date so we clear
1975 * it first before updating it.
1977 space_map_histogram_clear(msp->ms_sm);
1978 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1981 * Since the space map is not loaded we simply update the
1982 * exisiting histogram with what was freed in this txg. This
1983 * means that the on-disk histogram may not have an accurate
1984 * view of the free space but it's close enough to allow
1985 * us to make allocation decisions.
1987 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1989 metaslab_group_histogram_add(mg, msp);
1990 metaslab_group_histogram_verify(mg);
1991 metaslab_class_histogram_verify(mg->mg_class);
1994 * For sync pass 1, we avoid traversing this txg's free range tree
1995 * and instead will just swap the pointers for freetree and
1996 * freed_tree. We can safely do this since the freed_tree is
1997 * guaranteed to be empty on the initial pass.
1999 if (spa_sync_pass(spa) == 1) {
2000 range_tree_swap(freetree, freed_tree);
2002 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
2004 range_tree_vacate(alloctree, NULL, NULL);
2006 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2007 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2009 mutex_exit(&msp->ms_lock);
2011 if (object != space_map_object(msp->ms_sm)) {
2012 object = space_map_object(msp->ms_sm);
2013 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2014 msp->ms_id, sizeof (uint64_t), &object, tx);
2020 * Called after a transaction group has completely synced to mark
2021 * all of the metaslab's free space as usable.
2024 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2026 metaslab_group_t *mg = msp->ms_group;
2027 vdev_t *vd = mg->mg_vd;
2028 range_tree_t **freed_tree;
2029 range_tree_t **defer_tree;
2030 int64_t alloc_delta, defer_delta;
2032 ASSERT(!vd->vdev_ishole);
2034 mutex_enter(&msp->ms_lock);
2037 * If this metaslab is just becoming available, initialize its
2038 * alloctrees, freetrees, and defertree and add its capacity to
2041 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
2042 for (int t = 0; t < TXG_SIZE; t++) {
2043 ASSERT(msp->ms_alloctree[t] == NULL);
2044 ASSERT(msp->ms_freetree[t] == NULL);
2046 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2048 msp->ms_freetree[t] = range_tree_create(NULL, msp,
2052 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2053 ASSERT(msp->ms_defertree[t] == NULL);
2055 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2059 vdev_space_update(vd, 0, 0, msp->ms_size);
2062 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2063 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2065 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2066 defer_delta = range_tree_space(*freed_tree) -
2067 range_tree_space(*defer_tree);
2069 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2071 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2072 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2075 * If there's a metaslab_load() in progress, wait for it to complete
2076 * so that we have a consistent view of the in-core space map.
2078 metaslab_load_wait(msp);
2081 * Move the frees from the defer_tree back to the free
2082 * range tree (if it's loaded). Swap the freed_tree and the
2083 * defer_tree -- this is safe to do because we've just emptied out
2086 range_tree_vacate(*defer_tree,
2087 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2088 range_tree_swap(freed_tree, defer_tree);
2090 space_map_update(msp->ms_sm);
2092 msp->ms_deferspace += defer_delta;
2093 ASSERT3S(msp->ms_deferspace, >=, 0);
2094 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2095 if (msp->ms_deferspace != 0) {
2097 * Keep syncing this metaslab until all deferred frees
2098 * are back in circulation.
2100 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2103 if (msp->ms_loaded && msp->ms_access_txg < txg) {
2104 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2105 VERIFY0(range_tree_space(
2106 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2109 if (!metaslab_debug_unload)
2110 metaslab_unload(msp);
2113 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2114 mutex_exit(&msp->ms_lock);
2118 metaslab_sync_reassess(metaslab_group_t *mg)
2120 metaslab_group_alloc_update(mg);
2121 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2124 * Preload the next potential metaslabs
2126 metaslab_group_preload(mg);
2130 metaslab_distance(metaslab_t *msp, dva_t *dva)
2132 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2133 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2134 uint64_t start = msp->ms_id;
2136 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2137 return (1ULL << 63);
2140 return ((start - offset) << ms_shift);
2142 return ((offset - start) << ms_shift);
2147 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
2148 uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2150 spa_t *spa = mg->mg_vd->vdev_spa;
2151 metaslab_t *msp = NULL;
2152 uint64_t offset = -1ULL;
2153 avl_tree_t *t = &mg->mg_metaslab_tree;
2154 uint64_t activation_weight;
2155 uint64_t target_distance;
2158 activation_weight = METASLAB_WEIGHT_PRIMARY;
2159 for (i = 0; i < d; i++) {
2160 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2161 activation_weight = METASLAB_WEIGHT_SECONDARY;
2167 boolean_t was_active;
2169 mutex_enter(&mg->mg_lock);
2170 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
2171 if (msp->ms_weight < asize) {
2172 spa_dbgmsg(spa, "%s: failed to meet weight "
2173 "requirement: vdev %llu, txg %llu, mg %p, "
2174 "msp %p, psize %llu, asize %llu, "
2175 "weight %llu", spa_name(spa),
2176 mg->mg_vd->vdev_id, txg,
2177 mg, msp, psize, asize, msp->ms_weight);
2178 mutex_exit(&mg->mg_lock);
2183 * If the selected metaslab is condensing, skip it.
2185 if (msp->ms_condensing)
2188 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2189 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2192 target_distance = min_distance +
2193 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2196 for (i = 0; i < d; i++)
2197 if (metaslab_distance(msp, &dva[i]) <
2203 mutex_exit(&mg->mg_lock);
2207 mutex_enter(&msp->ms_lock);
2210 * Ensure that the metaslab we have selected is still
2211 * capable of handling our request. It's possible that
2212 * another thread may have changed the weight while we
2213 * were blocked on the metaslab lock.
2215 if (msp->ms_weight < asize || (was_active &&
2216 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
2217 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
2218 mutex_exit(&msp->ms_lock);
2222 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2223 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2224 metaslab_passivate(msp,
2225 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2226 mutex_exit(&msp->ms_lock);
2230 if (metaslab_activate(msp, activation_weight) != 0) {
2231 mutex_exit(&msp->ms_lock);
2236 * If this metaslab is currently condensing then pick again as
2237 * we can't manipulate this metaslab until it's committed
2240 if (msp->ms_condensing) {
2241 mutex_exit(&msp->ms_lock);
2245 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
2248 metaslab_passivate(msp, metaslab_block_maxsize(msp));
2249 mutex_exit(&msp->ms_lock);
2252 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2253 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2255 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
2256 msp->ms_access_txg = txg + metaslab_unload_delay;
2258 mutex_exit(&msp->ms_lock);
2264 * Allocate a block for the specified i/o.
2267 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2268 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
2270 metaslab_group_t *mg, *rotor;
2274 int zio_lock = B_FALSE;
2275 boolean_t allocatable;
2276 uint64_t offset = -1ULL;
2280 ASSERT(!DVA_IS_VALID(&dva[d]));
2283 * For testing, make some blocks above a certain size be gang blocks.
2285 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
2286 return (SET_ERROR(ENOSPC));
2289 * Start at the rotor and loop through all mgs until we find something.
2290 * Note that there's no locking on mc_rotor or mc_aliquot because
2291 * nothing actually breaks if we miss a few updates -- we just won't
2292 * allocate quite as evenly. It all balances out over time.
2294 * If we are doing ditto or log blocks, try to spread them across
2295 * consecutive vdevs. If we're forced to reuse a vdev before we've
2296 * allocated all of our ditto blocks, then try and spread them out on
2297 * that vdev as much as possible. If it turns out to not be possible,
2298 * gradually lower our standards until anything becomes acceptable.
2299 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2300 * gives us hope of containing our fault domains to something we're
2301 * able to reason about. Otherwise, any two top-level vdev failures
2302 * will guarantee the loss of data. With consecutive allocation,
2303 * only two adjacent top-level vdev failures will result in data loss.
2305 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2306 * ourselves on the same vdev as our gang block header. That
2307 * way, we can hope for locality in vdev_cache, plus it makes our
2308 * fault domains something tractable.
2311 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
2314 * It's possible the vdev we're using as the hint no
2315 * longer exists (i.e. removed). Consult the rotor when
2321 if (flags & METASLAB_HINTBP_AVOID &&
2322 mg->mg_next != NULL)
2327 } else if (d != 0) {
2328 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
2329 mg = vd->vdev_mg->mg_next;
2335 * If the hint put us into the wrong metaslab class, or into a
2336 * metaslab group that has been passivated, just follow the rotor.
2338 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2345 ASSERT(mg->mg_activation_count == 1);
2350 * Don't allocate from faulted devices.
2353 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
2354 allocatable = vdev_allocatable(vd);
2355 spa_config_exit(spa, SCL_ZIO, FTAG);
2357 allocatable = vdev_allocatable(vd);
2361 * Determine if the selected metaslab group is eligible
2362 * for allocations. If we're ganging or have requested
2363 * an allocation for the smallest gang block size
2364 * then we don't want to avoid allocating to the this
2365 * metaslab group. If we're in this condition we should
2366 * try to allocate from any device possible so that we
2367 * don't inadvertently return ENOSPC and suspend the pool
2368 * even though space is still available.
2370 if (allocatable && CAN_FASTGANG(flags) &&
2371 psize > SPA_GANGBLOCKSIZE)
2372 allocatable = metaslab_group_allocatable(mg);
2378 * Avoid writing single-copy data to a failing vdev
2379 * unless the user instructs us that it is okay.
2381 if ((vd->vdev_stat.vs_write_errors > 0 ||
2382 vd->vdev_state < VDEV_STATE_HEALTHY) &&
2383 d == 0 && dshift == 3 && vd->vdev_children == 0) {
2388 ASSERT(mg->mg_class == mc);
2390 distance = vd->vdev_asize >> dshift;
2391 if (distance <= (1ULL << vd->vdev_ms_shift))
2396 asize = vdev_psize_to_asize(vd, psize);
2397 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
2399 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
2401 if (offset != -1ULL) {
2403 * If we've just selected this metaslab group,
2404 * figure out whether the corresponding vdev is
2405 * over- or under-used relative to the pool,
2406 * and set an allocation bias to even it out.
2408 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
2409 vdev_stat_t *vs = &vd->vdev_stat;
2412 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
2413 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2416 * Calculate how much more or less we should
2417 * try to allocate from this device during
2418 * this iteration around the rotor.
2419 * For example, if a device is 80% full
2420 * and the pool is 20% full then we should
2421 * reduce allocations by 60% on this device.
2423 * mg_bias = (20 - 80) * 512K / 100 = -307K
2425 * This reduces allocations by 307K for this
2428 mg->mg_bias = ((cu - vu) *
2429 (int64_t)mg->mg_aliquot) / 100;
2430 } else if (!metaslab_bias_enabled) {
2434 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2435 mg->mg_aliquot + mg->mg_bias) {
2436 mc->mc_rotor = mg->mg_next;
2440 DVA_SET_VDEV(&dva[d], vd->vdev_id);
2441 DVA_SET_OFFSET(&dva[d], offset);
2442 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2443 DVA_SET_ASIZE(&dva[d], asize);
2448 mc->mc_rotor = mg->mg_next;
2450 } while ((mg = mg->mg_next) != rotor);
2454 ASSERT(dshift < 64);
2458 if (!allocatable && !zio_lock) {
2464 bzero(&dva[d], sizeof (dva_t));
2466 return (SET_ERROR(ENOSPC));
2470 * Free the block represented by DVA in the context of the specified
2471 * transaction group.
2474 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2476 uint64_t vdev = DVA_GET_VDEV(dva);
2477 uint64_t offset = DVA_GET_OFFSET(dva);
2478 uint64_t size = DVA_GET_ASIZE(dva);
2482 ASSERT(DVA_IS_VALID(dva));
2484 if (txg > spa_freeze_txg(spa))
2487 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2488 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2489 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2490 (u_longlong_t)vdev, (u_longlong_t)offset);
2495 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2497 if (DVA_GET_GANG(dva))
2498 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2500 mutex_enter(&msp->ms_lock);
2503 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2506 VERIFY(!msp->ms_condensing);
2507 VERIFY3U(offset, >=, msp->ms_start);
2508 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2509 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2511 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2512 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2513 range_tree_add(msp->ms_tree, offset, size);
2515 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2516 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2517 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2521 mutex_exit(&msp->ms_lock);
2525 * Intent log support: upon opening the pool after a crash, notify the SPA
2526 * of blocks that the intent log has allocated for immediate write, but
2527 * which are still considered free by the SPA because the last transaction
2528 * group didn't commit yet.
2531 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2533 uint64_t vdev = DVA_GET_VDEV(dva);
2534 uint64_t offset = DVA_GET_OFFSET(dva);
2535 uint64_t size = DVA_GET_ASIZE(dva);
2540 ASSERT(DVA_IS_VALID(dva));
2542 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2543 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2544 return (SET_ERROR(ENXIO));
2546 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2548 if (DVA_GET_GANG(dva))
2549 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2551 mutex_enter(&msp->ms_lock);
2553 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2554 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2556 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2557 error = SET_ERROR(ENOENT);
2559 if (error || txg == 0) { /* txg == 0 indicates dry run */
2560 mutex_exit(&msp->ms_lock);
2564 VERIFY(!msp->ms_condensing);
2565 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2566 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2567 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2568 range_tree_remove(msp->ms_tree, offset, size);
2570 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2571 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2572 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2573 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2576 mutex_exit(&msp->ms_lock);
2582 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2583 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2585 dva_t *dva = bp->blk_dva;
2586 dva_t *hintdva = hintbp->blk_dva;
2589 ASSERT(bp->blk_birth == 0);
2590 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2592 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2594 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2595 spa_config_exit(spa, SCL_ALLOC, FTAG);
2596 return (SET_ERROR(ENOSPC));
2599 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2600 ASSERT(BP_GET_NDVAS(bp) == 0);
2601 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2603 for (int d = 0; d < ndvas; d++) {
2604 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2607 for (d--; d >= 0; d--) {
2608 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2609 bzero(&dva[d], sizeof (dva_t));
2611 spa_config_exit(spa, SCL_ALLOC, FTAG);
2616 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2618 spa_config_exit(spa, SCL_ALLOC, FTAG);
2620 BP_SET_BIRTH(bp, txg, txg);
2626 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2628 const dva_t *dva = bp->blk_dva;
2629 int ndvas = BP_GET_NDVAS(bp);
2631 ASSERT(!BP_IS_HOLE(bp));
2632 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2634 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2636 for (int d = 0; d < ndvas; d++)
2637 metaslab_free_dva(spa, &dva[d], txg, now);
2639 spa_config_exit(spa, SCL_FREE, FTAG);
2643 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2645 const dva_t *dva = bp->blk_dva;
2646 int ndvas = BP_GET_NDVAS(bp);
2649 ASSERT(!BP_IS_HOLE(bp));
2653 * First do a dry run to make sure all DVAs are claimable,
2654 * so we don't have to unwind from partial failures below.
2656 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2660 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2662 for (int d = 0; d < ndvas; d++)
2663 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2666 spa_config_exit(spa, SCL_ALLOC, FTAG);
2668 ASSERT(error == 0 || txg == 0);
2674 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2676 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2679 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2680 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2681 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2682 vdev_t *vd = vdev_lookup_top(spa, vdev);
2683 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2684 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2685 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2688 range_tree_verify(msp->ms_tree, offset, size);
2690 for (int j = 0; j < TXG_SIZE; j++)
2691 range_tree_verify(msp->ms_freetree[j], offset, size);
2692 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2693 range_tree_verify(msp->ms_defertree[j], offset, size);
2695 spa_config_exit(spa, SCL_VDEV, FTAG);