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 TUNABLE_QUAD("vfs.zfs.metaslab.gang_bang", &metaslab_gang_bang);
61 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, gang_bang, CTLFLAG_RWTUN,
62 &metaslab_gang_bang, 0,
63 "Force gang block allocation for blocks larger than or equal to this value");
66 * The in-core space map representation is more compact than its on-disk form.
67 * The zfs_condense_pct determines how much more compact the in-core
68 * space_map representation must be before we compact it on-disk.
69 * Values should be greater than or equal to 100.
71 int zfs_condense_pct = 200;
72 TUNABLE_INT("vfs.zfs.condense_pct", &zfs_condense_pct);
73 SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN,
75 "Condense on-disk spacemap when it is more than this many percents"
76 " of in-memory counterpart");
79 * Condensing a metaslab is not guaranteed to actually reduce the amount of
80 * space used on disk. In particular, a space map uses data in increments of
81 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
82 * same number of blocks after condensing. Since the goal of condensing is to
83 * reduce the number of IOPs required to read the space map, we only want to
84 * condense when we can be sure we will reduce the number of blocks used by the
85 * space map. Unfortunately, we cannot precisely compute whether or not this is
86 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
87 * we apply the following heuristic: do not condense a spacemap unless the
88 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
91 int zfs_metaslab_condense_block_threshold = 4;
94 * The zfs_mg_noalloc_threshold defines which metaslab groups should
95 * be eligible for allocation. The value is defined as a percentage of
96 * free space. Metaslab groups that have more free space than
97 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
98 * a metaslab group's free space is less than or equal to the
99 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
100 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
101 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
102 * groups are allowed to accept allocations. Gang blocks are always
103 * eligible to allocate on any metaslab group. The default value of 0 means
104 * no metaslab group will be excluded based on this criterion.
106 int zfs_mg_noalloc_threshold = 0;
107 TUNABLE_INT("vfs.zfs.mg_noalloc_threshold", &zfs_mg_noalloc_threshold);
108 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN,
109 &zfs_mg_noalloc_threshold, 0,
110 "Percentage of metaslab group size that should be free"
111 " to make it eligible for allocation");
114 * Metaslab groups are considered eligible for allocations if their
115 * fragmenation metric (measured as a percentage) is less than or equal to
116 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
117 * then it will be skipped unless all metaslab groups within the metaslab
118 * class have also crossed this threshold.
120 int zfs_mg_fragmentation_threshold = 85;
121 TUNABLE_INT("vfs.zfs.mg_fragmentation_threshold", &zfs_mg_fragmentation_threshold);
122 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_fragmentation_threshold, CTLFLAG_RWTUN,
123 &zfs_mg_fragmentation_threshold, 0,
124 "Percentage of metaslab group size that should be considered "
125 "eligible for allocations unless all metaslab groups within the metaslab class "
126 "have also crossed this threshold");
129 * Allow metaslabs to keep their active state as long as their fragmentation
130 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
131 * active metaslab that exceeds this threshold will no longer keep its active
132 * status allowing better metaslabs to be selected.
134 int zfs_metaslab_fragmentation_threshold = 70;
135 TUNABLE_INT("vfs.zfs.metaslab.fragmentation_threshold",
136 &zfs_metaslab_fragmentation_threshold);
137 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_threshold, CTLFLAG_RWTUN,
138 &zfs_metaslab_fragmentation_threshold, 0,
139 "Maximum percentage of metaslab fragmentation level to keep their active state");
142 * When set will load all metaslabs when pool is first opened.
144 int metaslab_debug_load = 0;
145 TUNABLE_INT("vfs.zfs.metaslab.debug_load", &metaslab_debug_load);
146 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN,
147 &metaslab_debug_load, 0,
148 "Load all metaslabs when pool is first opened");
151 * When set will prevent metaslabs from being unloaded.
153 int metaslab_debug_unload = 0;
154 TUNABLE_INT("vfs.zfs.metaslab.debug_unload", &metaslab_debug_unload);
155 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN,
156 &metaslab_debug_unload, 0,
157 "Prevent metaslabs from being unloaded");
160 * Minimum size which forces the dynamic allocator to change
161 * it's allocation strategy. Once the space map cannot satisfy
162 * an allocation of this size then it switches to using more
163 * aggressive strategy (i.e search by size rather than offset).
165 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
166 TUNABLE_QUAD("vfs.zfs.metaslab.df_alloc_threshold",
167 &metaslab_df_alloc_threshold);
168 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN,
169 &metaslab_df_alloc_threshold, 0,
170 "Minimum size which forces the dynamic allocator to change it's allocation strategy");
173 * The minimum free space, in percent, which must be available
174 * in a space map to continue allocations in a first-fit fashion.
175 * Once the space_map's free space drops below this level we dynamically
176 * switch to using best-fit allocations.
178 int metaslab_df_free_pct = 4;
179 TUNABLE_INT("vfs.zfs.metaslab.df_free_pct", &metaslab_df_free_pct);
180 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN,
181 &metaslab_df_free_pct, 0,
182 "The minimum free space, in percent, which must be available in a space map to continue allocations in a first-fit fashion");
185 * A metaslab is considered "free" if it contains a contiguous
186 * segment which is greater than metaslab_min_alloc_size.
188 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
189 TUNABLE_QUAD("vfs.zfs.metaslab.min_alloc_size",
190 &metaslab_min_alloc_size);
191 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN,
192 &metaslab_min_alloc_size, 0,
193 "A metaslab is considered \"free\" if it contains a contiguous segment which is greater than vfs.zfs.metaslab.min_alloc_size");
196 * Percentage of all cpus that can be used by the metaslab taskq.
198 int metaslab_load_pct = 50;
199 TUNABLE_INT("vfs.zfs.metaslab.load_pct", &metaslab_load_pct);
200 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN,
201 &metaslab_load_pct, 0,
202 "Percentage of cpus that can be used by the metaslab taskq");
205 * Determines how many txgs a metaslab may remain loaded without having any
206 * allocations from it. As long as a metaslab continues to be used we will
209 int metaslab_unload_delay = TXG_SIZE * 2;
210 TUNABLE_INT("vfs.zfs.metaslab.unload_delay", &metaslab_unload_delay);
211 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN,
212 &metaslab_unload_delay, 0,
213 "Number of TXGs that an unused metaslab can be kept in memory");
216 * Max number of metaslabs per group to preload.
218 int metaslab_preload_limit = SPA_DVAS_PER_BP;
219 TUNABLE_INT("vfs.zfs.metaslab.preload_limit", &metaslab_preload_limit);
220 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
221 &metaslab_preload_limit, 0,
222 "Max number of metaslabs per group to preload");
225 * Enable/disable preloading of metaslab.
227 boolean_t metaslab_preload_enabled = B_TRUE;
228 TUNABLE_INT("vfs.zfs.metaslab.preload_enabled", &metaslab_preload_enabled);
229 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
230 &metaslab_preload_enabled, 0,
231 "Max number of metaslabs per group to preload");
234 * Enable/disable fragmentation weighting on metaslabs.
236 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
237 TUNABLE_INT("vfs.zfs.metaslab_fragmentation_factor_enabled",
238 &metaslab_fragmentation_factor_enabled);
239 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, fragmentation_factor_enabled, CTLFLAG_RWTUN,
240 &metaslab_fragmentation_factor_enabled, 0,
241 "Enable fragmentation weighting on metaslabs");
244 * Enable/disable lba weighting (i.e. outer tracks are given preference).
246 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
247 TUNABLE_INT("vfs.zfs.metaslab.lba_weighting_enabled",
248 &metaslab_lba_weighting_enabled);
249 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, lba_weighting_enabled, CTLFLAG_RWTUN,
250 &metaslab_lba_weighting_enabled, 0,
251 "Enable LBA weighting (i.e. outer tracks are given preference)");
254 * Enable/disable metaslab group biasing.
256 boolean_t metaslab_bias_enabled = B_TRUE;
257 TUNABLE_INT("vfs.zfs.metaslab.bias_enabled",
258 &metaslab_bias_enabled);
259 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, bias_enabled, CTLFLAG_RWTUN,
260 &metaslab_bias_enabled, 0,
261 "Enable metaslab group biasing");
263 static uint64_t metaslab_fragmentation(metaslab_t *);
266 * ==========================================================================
268 * ==========================================================================
271 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
273 metaslab_class_t *mc;
275 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
285 metaslab_class_destroy(metaslab_class_t *mc)
287 ASSERT(mc->mc_rotor == NULL);
288 ASSERT(mc->mc_alloc == 0);
289 ASSERT(mc->mc_deferred == 0);
290 ASSERT(mc->mc_space == 0);
291 ASSERT(mc->mc_dspace == 0);
293 kmem_free(mc, sizeof (metaslab_class_t));
297 metaslab_class_validate(metaslab_class_t *mc)
299 metaslab_group_t *mg;
303 * Must hold one of the spa_config locks.
305 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
306 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
308 if ((mg = mc->mc_rotor) == NULL)
313 ASSERT(vd->vdev_mg != NULL);
314 ASSERT3P(vd->vdev_top, ==, vd);
315 ASSERT3P(mg->mg_class, ==, mc);
316 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
317 } while ((mg = mg->mg_next) != mc->mc_rotor);
323 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
324 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
326 atomic_add_64(&mc->mc_alloc, alloc_delta);
327 atomic_add_64(&mc->mc_deferred, defer_delta);
328 atomic_add_64(&mc->mc_space, space_delta);
329 atomic_add_64(&mc->mc_dspace, dspace_delta);
333 metaslab_class_minblocksize_update(metaslab_class_t *mc)
335 metaslab_group_t *mg;
337 uint64_t minashift = UINT64_MAX;
339 if ((mg = mc->mc_rotor) == NULL) {
340 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
346 if (vd->vdev_ashift < minashift)
347 minashift = vd->vdev_ashift;
348 } while ((mg = mg->mg_next) != mc->mc_rotor);
350 mc->mc_minblocksize = 1ULL << minashift;
354 metaslab_class_get_alloc(metaslab_class_t *mc)
356 return (mc->mc_alloc);
360 metaslab_class_get_deferred(metaslab_class_t *mc)
362 return (mc->mc_deferred);
366 metaslab_class_get_space(metaslab_class_t *mc)
368 return (mc->mc_space);
372 metaslab_class_get_dspace(metaslab_class_t *mc)
374 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
378 metaslab_class_get_minblocksize(metaslab_class_t *mc)
380 return (mc->mc_minblocksize);
384 metaslab_class_histogram_verify(metaslab_class_t *mc)
386 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
390 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
393 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
396 for (int c = 0; c < rvd->vdev_children; c++) {
397 vdev_t *tvd = rvd->vdev_child[c];
398 metaslab_group_t *mg = tvd->vdev_mg;
401 * Skip any holes, uninitialized top-levels, or
402 * vdevs that are not in this metalab class.
404 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
405 mg->mg_class != mc) {
409 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
410 mc_hist[i] += mg->mg_histogram[i];
413 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
414 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
416 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
420 * Calculate the metaslab class's fragmentation metric. The metric
421 * is weighted based on the space contribution of each metaslab group.
422 * The return value will be a number between 0 and 100 (inclusive), or
423 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
424 * zfs_frag_table for more information about the metric.
427 metaslab_class_fragmentation(metaslab_class_t *mc)
429 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
430 uint64_t fragmentation = 0;
432 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
434 for (int c = 0; c < rvd->vdev_children; c++) {
435 vdev_t *tvd = rvd->vdev_child[c];
436 metaslab_group_t *mg = tvd->vdev_mg;
439 * Skip any holes, uninitialized top-levels, or
440 * vdevs that are not in this metalab class.
442 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
443 mg->mg_class != mc) {
448 * If a metaslab group does not contain a fragmentation
449 * metric then just bail out.
451 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
452 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
453 return (ZFS_FRAG_INVALID);
457 * Determine how much this metaslab_group is contributing
458 * to the overall pool fragmentation metric.
460 fragmentation += mg->mg_fragmentation *
461 metaslab_group_get_space(mg);
463 fragmentation /= metaslab_class_get_space(mc);
465 ASSERT3U(fragmentation, <=, 100);
466 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
467 return (fragmentation);
471 * Calculate the amount of expandable space that is available in
472 * this metaslab class. If a device is expanded then its expandable
473 * space will be the amount of allocatable space that is currently not
474 * part of this metaslab class.
477 metaslab_class_expandable_space(metaslab_class_t *mc)
479 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
482 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
483 for (int c = 0; c < rvd->vdev_children; c++) {
484 vdev_t *tvd = rvd->vdev_child[c];
485 metaslab_group_t *mg = tvd->vdev_mg;
487 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
488 mg->mg_class != mc) {
492 space += tvd->vdev_max_asize - tvd->vdev_asize;
494 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
499 * ==========================================================================
501 * ==========================================================================
504 metaslab_compare(const void *x1, const void *x2)
506 const metaslab_t *m1 = x1;
507 const metaslab_t *m2 = x2;
509 if (m1->ms_weight < m2->ms_weight)
511 if (m1->ms_weight > m2->ms_weight)
515 * If the weights are identical, use the offset to force uniqueness.
517 if (m1->ms_start < m2->ms_start)
519 if (m1->ms_start > m2->ms_start)
522 ASSERT3P(m1, ==, m2);
528 * Update the allocatable flag and the metaslab group's capacity.
529 * The allocatable flag is set to true if the capacity is below
530 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
531 * from allocatable to non-allocatable or vice versa then the metaslab
532 * group's class is updated to reflect the transition.
535 metaslab_group_alloc_update(metaslab_group_t *mg)
537 vdev_t *vd = mg->mg_vd;
538 metaslab_class_t *mc = mg->mg_class;
539 vdev_stat_t *vs = &vd->vdev_stat;
540 boolean_t was_allocatable;
542 ASSERT(vd == vd->vdev_top);
544 mutex_enter(&mg->mg_lock);
545 was_allocatable = mg->mg_allocatable;
547 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
551 * A metaslab group is considered allocatable if it has plenty
552 * of free space or is not heavily fragmented. We only take
553 * fragmentation into account if the metaslab group has a valid
554 * fragmentation metric (i.e. a value between 0 and 100).
556 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
557 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
558 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
561 * The mc_alloc_groups maintains a count of the number of
562 * groups in this metaslab class that are still above the
563 * zfs_mg_noalloc_threshold. This is used by the allocating
564 * threads to determine if they should avoid allocations to
565 * a given group. The allocator will avoid allocations to a group
566 * if that group has reached or is below the zfs_mg_noalloc_threshold
567 * and there are still other groups that are above the threshold.
568 * When a group transitions from allocatable to non-allocatable or
569 * vice versa we update the metaslab class to reflect that change.
570 * When the mc_alloc_groups value drops to 0 that means that all
571 * groups have reached the zfs_mg_noalloc_threshold making all groups
572 * eligible for allocations. This effectively means that all devices
573 * are balanced again.
575 if (was_allocatable && !mg->mg_allocatable)
576 mc->mc_alloc_groups--;
577 else if (!was_allocatable && mg->mg_allocatable)
578 mc->mc_alloc_groups++;
580 mutex_exit(&mg->mg_lock);
584 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
586 metaslab_group_t *mg;
588 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
589 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
590 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
591 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
594 mg->mg_activation_count = 0;
596 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
597 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
603 metaslab_group_destroy(metaslab_group_t *mg)
605 ASSERT(mg->mg_prev == NULL);
606 ASSERT(mg->mg_next == NULL);
608 * We may have gone below zero with the activation count
609 * either because we never activated in the first place or
610 * because we're done, and possibly removing the vdev.
612 ASSERT(mg->mg_activation_count <= 0);
614 taskq_destroy(mg->mg_taskq);
615 avl_destroy(&mg->mg_metaslab_tree);
616 mutex_destroy(&mg->mg_lock);
617 kmem_free(mg, sizeof (metaslab_group_t));
621 metaslab_group_activate(metaslab_group_t *mg)
623 metaslab_class_t *mc = mg->mg_class;
624 metaslab_group_t *mgprev, *mgnext;
626 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
628 ASSERT(mc->mc_rotor != mg);
629 ASSERT(mg->mg_prev == NULL);
630 ASSERT(mg->mg_next == NULL);
631 ASSERT(mg->mg_activation_count <= 0);
633 if (++mg->mg_activation_count <= 0)
636 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
637 metaslab_group_alloc_update(mg);
639 if ((mgprev = mc->mc_rotor) == NULL) {
643 mgnext = mgprev->mg_next;
644 mg->mg_prev = mgprev;
645 mg->mg_next = mgnext;
646 mgprev->mg_next = mg;
647 mgnext->mg_prev = mg;
650 metaslab_class_minblocksize_update(mc);
654 metaslab_group_passivate(metaslab_group_t *mg)
656 metaslab_class_t *mc = mg->mg_class;
657 metaslab_group_t *mgprev, *mgnext;
659 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
661 if (--mg->mg_activation_count != 0) {
662 ASSERT(mc->mc_rotor != mg);
663 ASSERT(mg->mg_prev == NULL);
664 ASSERT(mg->mg_next == NULL);
665 ASSERT(mg->mg_activation_count < 0);
669 taskq_wait(mg->mg_taskq);
670 metaslab_group_alloc_update(mg);
672 mgprev = mg->mg_prev;
673 mgnext = mg->mg_next;
678 mc->mc_rotor = mgnext;
679 mgprev->mg_next = mgnext;
680 mgnext->mg_prev = mgprev;
685 metaslab_class_minblocksize_update(mc);
689 metaslab_group_get_space(metaslab_group_t *mg)
691 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
695 metaslab_group_histogram_verify(metaslab_group_t *mg)
698 vdev_t *vd = mg->mg_vd;
699 uint64_t ashift = vd->vdev_ashift;
702 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
705 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
708 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
709 SPACE_MAP_HISTOGRAM_SIZE + ashift);
711 for (int m = 0; m < vd->vdev_ms_count; m++) {
712 metaslab_t *msp = vd->vdev_ms[m];
714 if (msp->ms_sm == NULL)
717 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
718 mg_hist[i + ashift] +=
719 msp->ms_sm->sm_phys->smp_histogram[i];
722 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
723 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
725 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
729 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
731 metaslab_class_t *mc = mg->mg_class;
732 uint64_t ashift = mg->mg_vd->vdev_ashift;
734 ASSERT(MUTEX_HELD(&msp->ms_lock));
735 if (msp->ms_sm == NULL)
738 mutex_enter(&mg->mg_lock);
739 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
740 mg->mg_histogram[i + ashift] +=
741 msp->ms_sm->sm_phys->smp_histogram[i];
742 mc->mc_histogram[i + ashift] +=
743 msp->ms_sm->sm_phys->smp_histogram[i];
745 mutex_exit(&mg->mg_lock);
749 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
751 metaslab_class_t *mc = mg->mg_class;
752 uint64_t ashift = mg->mg_vd->vdev_ashift;
754 ASSERT(MUTEX_HELD(&msp->ms_lock));
755 if (msp->ms_sm == NULL)
758 mutex_enter(&mg->mg_lock);
759 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
760 ASSERT3U(mg->mg_histogram[i + ashift], >=,
761 msp->ms_sm->sm_phys->smp_histogram[i]);
762 ASSERT3U(mc->mc_histogram[i + ashift], >=,
763 msp->ms_sm->sm_phys->smp_histogram[i]);
765 mg->mg_histogram[i + ashift] -=
766 msp->ms_sm->sm_phys->smp_histogram[i];
767 mc->mc_histogram[i + ashift] -=
768 msp->ms_sm->sm_phys->smp_histogram[i];
770 mutex_exit(&mg->mg_lock);
774 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
776 ASSERT(msp->ms_group == NULL);
777 mutex_enter(&mg->mg_lock);
780 avl_add(&mg->mg_metaslab_tree, msp);
781 mutex_exit(&mg->mg_lock);
783 mutex_enter(&msp->ms_lock);
784 metaslab_group_histogram_add(mg, msp);
785 mutex_exit(&msp->ms_lock);
789 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
791 mutex_enter(&msp->ms_lock);
792 metaslab_group_histogram_remove(mg, msp);
793 mutex_exit(&msp->ms_lock);
795 mutex_enter(&mg->mg_lock);
796 ASSERT(msp->ms_group == mg);
797 avl_remove(&mg->mg_metaslab_tree, msp);
798 msp->ms_group = NULL;
799 mutex_exit(&mg->mg_lock);
803 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
806 * Although in principle the weight can be any value, in
807 * practice we do not use values in the range [1, 511].
809 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
810 ASSERT(MUTEX_HELD(&msp->ms_lock));
812 mutex_enter(&mg->mg_lock);
813 ASSERT(msp->ms_group == mg);
814 avl_remove(&mg->mg_metaslab_tree, msp);
815 msp->ms_weight = weight;
816 avl_add(&mg->mg_metaslab_tree, msp);
817 mutex_exit(&mg->mg_lock);
821 * Calculate the fragmentation for a given metaslab group. We can use
822 * a simple average here since all metaslabs within the group must have
823 * the same size. The return value will be a value between 0 and 100
824 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
825 * group have a fragmentation metric.
828 metaslab_group_fragmentation(metaslab_group_t *mg)
830 vdev_t *vd = mg->mg_vd;
831 uint64_t fragmentation = 0;
832 uint64_t valid_ms = 0;
834 for (int m = 0; m < vd->vdev_ms_count; m++) {
835 metaslab_t *msp = vd->vdev_ms[m];
837 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
841 fragmentation += msp->ms_fragmentation;
844 if (valid_ms <= vd->vdev_ms_count / 2)
845 return (ZFS_FRAG_INVALID);
847 fragmentation /= valid_ms;
848 ASSERT3U(fragmentation, <=, 100);
849 return (fragmentation);
853 * Determine if a given metaslab group should skip allocations. A metaslab
854 * group should avoid allocations if its free capacity is less than the
855 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
856 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
857 * that can still handle allocations.
860 metaslab_group_allocatable(metaslab_group_t *mg)
862 vdev_t *vd = mg->mg_vd;
863 spa_t *spa = vd->vdev_spa;
864 metaslab_class_t *mc = mg->mg_class;
867 * We use two key metrics to determine if a metaslab group is
868 * considered allocatable -- free space and fragmentation. If
869 * the free space is greater than the free space threshold and
870 * the fragmentation is less than the fragmentation threshold then
871 * consider the group allocatable. There are two case when we will
872 * not consider these key metrics. The first is if the group is
873 * associated with a slog device and the second is if all groups
874 * in this metaslab class have already been consider ineligible
877 return ((mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
878 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
879 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold)) ||
880 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
884 * ==========================================================================
885 * Range tree callbacks
886 * ==========================================================================
890 * Comparison function for the private size-ordered tree. Tree is sorted
891 * by size, larger sizes at the end of the tree.
894 metaslab_rangesize_compare(const void *x1, const void *x2)
896 const range_seg_t *r1 = x1;
897 const range_seg_t *r2 = x2;
898 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
899 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
901 if (rs_size1 < rs_size2)
903 if (rs_size1 > rs_size2)
906 if (r1->rs_start < r2->rs_start)
909 if (r1->rs_start > r2->rs_start)
916 * Create any block allocator specific components. The current allocators
917 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
920 metaslab_rt_create(range_tree_t *rt, void *arg)
922 metaslab_t *msp = arg;
924 ASSERT3P(rt->rt_arg, ==, msp);
925 ASSERT(msp->ms_tree == NULL);
927 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
928 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
932 * Destroy the block allocator specific components.
935 metaslab_rt_destroy(range_tree_t *rt, void *arg)
937 metaslab_t *msp = arg;
939 ASSERT3P(rt->rt_arg, ==, msp);
940 ASSERT3P(msp->ms_tree, ==, rt);
941 ASSERT0(avl_numnodes(&msp->ms_size_tree));
943 avl_destroy(&msp->ms_size_tree);
947 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
949 metaslab_t *msp = arg;
951 ASSERT3P(rt->rt_arg, ==, msp);
952 ASSERT3P(msp->ms_tree, ==, rt);
953 VERIFY(!msp->ms_condensing);
954 avl_add(&msp->ms_size_tree, rs);
958 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
960 metaslab_t *msp = arg;
962 ASSERT3P(rt->rt_arg, ==, msp);
963 ASSERT3P(msp->ms_tree, ==, rt);
964 VERIFY(!msp->ms_condensing);
965 avl_remove(&msp->ms_size_tree, rs);
969 metaslab_rt_vacate(range_tree_t *rt, void *arg)
971 metaslab_t *msp = arg;
973 ASSERT3P(rt->rt_arg, ==, msp);
974 ASSERT3P(msp->ms_tree, ==, rt);
977 * Normally one would walk the tree freeing nodes along the way.
978 * Since the nodes are shared with the range trees we can avoid
979 * walking all nodes and just reinitialize the avl tree. The nodes
980 * will be freed by the range tree, so we don't want to free them here.
982 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
983 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
986 static range_tree_ops_t metaslab_rt_ops = {
995 * ==========================================================================
996 * Metaslab block operations
997 * ==========================================================================
1001 * Return the maximum contiguous segment within the metaslab.
1004 metaslab_block_maxsize(metaslab_t *msp)
1006 avl_tree_t *t = &msp->ms_size_tree;
1009 if (t == NULL || (rs = avl_last(t)) == NULL)
1012 return (rs->rs_end - rs->rs_start);
1016 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
1019 range_tree_t *rt = msp->ms_tree;
1021 VERIFY(!msp->ms_condensing);
1023 start = msp->ms_ops->msop_alloc(msp, size);
1024 if (start != -1ULL) {
1025 vdev_t *vd = msp->ms_group->mg_vd;
1027 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
1028 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
1029 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
1030 range_tree_remove(rt, start, size);
1036 * ==========================================================================
1037 * Common allocator routines
1038 * ==========================================================================
1042 * This is a helper function that can be used by the allocator to find
1043 * a suitable block to allocate. This will search the specified AVL
1044 * tree looking for a block that matches the specified criteria.
1047 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1050 range_seg_t *rs, rsearch;
1053 rsearch.rs_start = *cursor;
1054 rsearch.rs_end = *cursor + size;
1056 rs = avl_find(t, &rsearch, &where);
1058 rs = avl_nearest(t, where, AVL_AFTER);
1060 while (rs != NULL) {
1061 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1063 if (offset + size <= rs->rs_end) {
1064 *cursor = offset + size;
1067 rs = AVL_NEXT(t, rs);
1071 * If we know we've searched the whole map (*cursor == 0), give up.
1072 * Otherwise, reset the cursor to the beginning and try again.
1078 return (metaslab_block_picker(t, cursor, size, align));
1082 * ==========================================================================
1083 * The first-fit block allocator
1084 * ==========================================================================
1087 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1090 * Find the largest power of 2 block size that evenly divides the
1091 * requested size. This is used to try to allocate blocks with similar
1092 * alignment from the same area of the metaslab (i.e. same cursor
1093 * bucket) but it does not guarantee that other allocations sizes
1094 * may exist in the same region.
1096 uint64_t align = size & -size;
1097 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1098 avl_tree_t *t = &msp->ms_tree->rt_root;
1100 return (metaslab_block_picker(t, cursor, size, align));
1103 static metaslab_ops_t metaslab_ff_ops = {
1108 * ==========================================================================
1109 * Dynamic block allocator -
1110 * Uses the first fit allocation scheme until space get low and then
1111 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1112 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1113 * ==========================================================================
1116 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1119 * Find the largest power of 2 block size that evenly divides the
1120 * requested size. This is used to try to allocate blocks with similar
1121 * alignment from the same area of the metaslab (i.e. same cursor
1122 * bucket) but it does not guarantee that other allocations sizes
1123 * may exist in the same region.
1125 uint64_t align = size & -size;
1126 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1127 range_tree_t *rt = msp->ms_tree;
1128 avl_tree_t *t = &rt->rt_root;
1129 uint64_t max_size = metaslab_block_maxsize(msp);
1130 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1132 ASSERT(MUTEX_HELD(&msp->ms_lock));
1133 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1135 if (max_size < size)
1139 * If we're running low on space switch to using the size
1140 * sorted AVL tree (best-fit).
1142 if (max_size < metaslab_df_alloc_threshold ||
1143 free_pct < metaslab_df_free_pct) {
1144 t = &msp->ms_size_tree;
1148 return (metaslab_block_picker(t, cursor, size, 1ULL));
1151 static metaslab_ops_t metaslab_df_ops = {
1156 * ==========================================================================
1157 * Cursor fit block allocator -
1158 * Select the largest region in the metaslab, set the cursor to the beginning
1159 * of the range and the cursor_end to the end of the range. As allocations
1160 * are made advance the cursor. Continue allocating from the cursor until
1161 * the range is exhausted and then find a new range.
1162 * ==========================================================================
1165 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1167 range_tree_t *rt = msp->ms_tree;
1168 avl_tree_t *t = &msp->ms_size_tree;
1169 uint64_t *cursor = &msp->ms_lbas[0];
1170 uint64_t *cursor_end = &msp->ms_lbas[1];
1171 uint64_t offset = 0;
1173 ASSERT(MUTEX_HELD(&msp->ms_lock));
1174 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1176 ASSERT3U(*cursor_end, >=, *cursor);
1178 if ((*cursor + size) > *cursor_end) {
1181 rs = avl_last(&msp->ms_size_tree);
1182 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1185 *cursor = rs->rs_start;
1186 *cursor_end = rs->rs_end;
1195 static metaslab_ops_t metaslab_cf_ops = {
1200 * ==========================================================================
1201 * New dynamic fit allocator -
1202 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1203 * contiguous blocks. If no region is found then just use the largest segment
1205 * ==========================================================================
1209 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1210 * to request from the allocator.
1212 uint64_t metaslab_ndf_clump_shift = 4;
1215 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1217 avl_tree_t *t = &msp->ms_tree->rt_root;
1219 range_seg_t *rs, rsearch;
1220 uint64_t hbit = highbit64(size);
1221 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1222 uint64_t max_size = metaslab_block_maxsize(msp);
1224 ASSERT(MUTEX_HELD(&msp->ms_lock));
1225 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1227 if (max_size < size)
1230 rsearch.rs_start = *cursor;
1231 rsearch.rs_end = *cursor + size;
1233 rs = avl_find(t, &rsearch, &where);
1234 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1235 t = &msp->ms_size_tree;
1237 rsearch.rs_start = 0;
1238 rsearch.rs_end = MIN(max_size,
1239 1ULL << (hbit + metaslab_ndf_clump_shift));
1240 rs = avl_find(t, &rsearch, &where);
1242 rs = avl_nearest(t, where, AVL_AFTER);
1246 if ((rs->rs_end - rs->rs_start) >= size) {
1247 *cursor = rs->rs_start + size;
1248 return (rs->rs_start);
1253 static metaslab_ops_t metaslab_ndf_ops = {
1257 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1260 * ==========================================================================
1262 * ==========================================================================
1266 * Wait for any in-progress metaslab loads to complete.
1269 metaslab_load_wait(metaslab_t *msp)
1271 ASSERT(MUTEX_HELD(&msp->ms_lock));
1273 while (msp->ms_loading) {
1274 ASSERT(!msp->ms_loaded);
1275 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1280 metaslab_load(metaslab_t *msp)
1284 ASSERT(MUTEX_HELD(&msp->ms_lock));
1285 ASSERT(!msp->ms_loaded);
1286 ASSERT(!msp->ms_loading);
1288 msp->ms_loading = B_TRUE;
1291 * If the space map has not been allocated yet, then treat
1292 * all the space in the metaslab as free and add it to the
1295 if (msp->ms_sm != NULL)
1296 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1298 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1300 msp->ms_loaded = (error == 0);
1301 msp->ms_loading = B_FALSE;
1303 if (msp->ms_loaded) {
1304 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1305 range_tree_walk(msp->ms_defertree[t],
1306 range_tree_remove, msp->ms_tree);
1309 cv_broadcast(&msp->ms_load_cv);
1314 metaslab_unload(metaslab_t *msp)
1316 ASSERT(MUTEX_HELD(&msp->ms_lock));
1317 range_tree_vacate(msp->ms_tree, NULL, NULL);
1318 msp->ms_loaded = B_FALSE;
1319 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1323 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1326 vdev_t *vd = mg->mg_vd;
1327 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1331 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1332 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1333 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1335 ms->ms_start = id << vd->vdev_ms_shift;
1336 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1339 * We only open space map objects that already exist. All others
1340 * will be opened when we finally allocate an object for it.
1343 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1344 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1347 kmem_free(ms, sizeof (metaslab_t));
1351 ASSERT(ms->ms_sm != NULL);
1355 * We create the main range tree here, but we don't create the
1356 * alloctree and freetree until metaslab_sync_done(). This serves
1357 * two purposes: it allows metaslab_sync_done() to detect the
1358 * addition of new space; and for debugging, it ensures that we'd
1359 * data fault on any attempt to use this metaslab before it's ready.
1361 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1362 metaslab_group_add(mg, ms);
1364 ms->ms_fragmentation = metaslab_fragmentation(ms);
1365 ms->ms_ops = mg->mg_class->mc_ops;
1368 * If we're opening an existing pool (txg == 0) or creating
1369 * a new one (txg == TXG_INITIAL), all space is available now.
1370 * If we're adding space to an existing pool, the new space
1371 * does not become available until after this txg has synced.
1373 if (txg <= TXG_INITIAL)
1374 metaslab_sync_done(ms, 0);
1377 * If metaslab_debug_load is set and we're initializing a metaslab
1378 * that has an allocated space_map object then load the its space
1379 * map so that can verify frees.
1381 if (metaslab_debug_load && ms->ms_sm != NULL) {
1382 mutex_enter(&ms->ms_lock);
1383 VERIFY0(metaslab_load(ms));
1384 mutex_exit(&ms->ms_lock);
1388 vdev_dirty(vd, 0, NULL, txg);
1389 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1398 metaslab_fini(metaslab_t *msp)
1400 metaslab_group_t *mg = msp->ms_group;
1402 metaslab_group_remove(mg, msp);
1404 mutex_enter(&msp->ms_lock);
1406 VERIFY(msp->ms_group == NULL);
1407 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1409 space_map_close(msp->ms_sm);
1411 metaslab_unload(msp);
1412 range_tree_destroy(msp->ms_tree);
1414 for (int t = 0; t < TXG_SIZE; t++) {
1415 range_tree_destroy(msp->ms_alloctree[t]);
1416 range_tree_destroy(msp->ms_freetree[t]);
1419 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1420 range_tree_destroy(msp->ms_defertree[t]);
1423 ASSERT0(msp->ms_deferspace);
1425 mutex_exit(&msp->ms_lock);
1426 cv_destroy(&msp->ms_load_cv);
1427 mutex_destroy(&msp->ms_lock);
1429 kmem_free(msp, sizeof (metaslab_t));
1432 #define FRAGMENTATION_TABLE_SIZE 17
1435 * This table defines a segment size based fragmentation metric that will
1436 * allow each metaslab to derive its own fragmentation value. This is done
1437 * by calculating the space in each bucket of the spacemap histogram and
1438 * multiplying that by the fragmetation metric in this table. Doing
1439 * this for all buckets and dividing it by the total amount of free
1440 * space in this metaslab (i.e. the total free space in all buckets) gives
1441 * us the fragmentation metric. This means that a high fragmentation metric
1442 * equates to most of the free space being comprised of small segments.
1443 * Conversely, if the metric is low, then most of the free space is in
1444 * large segments. A 10% change in fragmentation equates to approximately
1445 * double the number of segments.
1447 * This table defines 0% fragmented space using 16MB segments. Testing has
1448 * shown that segments that are greater than or equal to 16MB do not suffer
1449 * from drastic performance problems. Using this value, we derive the rest
1450 * of the table. Since the fragmentation value is never stored on disk, it
1451 * is possible to change these calculations in the future.
1453 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1473 * Calclate the metaslab's fragmentation metric. A return value
1474 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1475 * not support this metric. Otherwise, the return value should be in the
1479 metaslab_fragmentation(metaslab_t *msp)
1481 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1482 uint64_t fragmentation = 0;
1484 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1485 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1487 if (!feature_enabled)
1488 return (ZFS_FRAG_INVALID);
1491 * A null space map means that the entire metaslab is free
1492 * and thus is not fragmented.
1494 if (msp->ms_sm == NULL)
1498 * If this metaslab's space_map has not been upgraded, flag it
1499 * so that we upgrade next time we encounter it.
1501 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1502 uint64_t txg = spa_syncing_txg(spa);
1503 vdev_t *vd = msp->ms_group->mg_vd;
1505 if (spa_writeable(spa)) {
1506 msp->ms_condense_wanted = B_TRUE;
1507 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1508 spa_dbgmsg(spa, "txg %llu, requesting force condense: "
1509 "msp %p, vd %p", txg, msp, vd);
1511 return (ZFS_FRAG_INVALID);
1514 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1516 uint8_t shift = msp->ms_sm->sm_shift;
1517 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1518 FRAGMENTATION_TABLE_SIZE - 1);
1520 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1523 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1526 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1527 fragmentation += space * zfs_frag_table[idx];
1531 fragmentation /= total;
1532 ASSERT3U(fragmentation, <=, 100);
1533 return (fragmentation);
1537 * Compute a weight -- a selection preference value -- for the given metaslab.
1538 * This is based on the amount of free space, the level of fragmentation,
1539 * the LBA range, and whether the metaslab is loaded.
1542 metaslab_weight(metaslab_t *msp)
1544 metaslab_group_t *mg = msp->ms_group;
1545 vdev_t *vd = mg->mg_vd;
1546 uint64_t weight, space;
1548 ASSERT(MUTEX_HELD(&msp->ms_lock));
1551 * This vdev is in the process of being removed so there is nothing
1552 * for us to do here.
1554 if (vd->vdev_removing) {
1555 ASSERT0(space_map_allocated(msp->ms_sm));
1556 ASSERT0(vd->vdev_ms_shift);
1561 * The baseline weight is the metaslab's free space.
1563 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1565 msp->ms_fragmentation = metaslab_fragmentation(msp);
1566 if (metaslab_fragmentation_factor_enabled &&
1567 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1569 * Use the fragmentation information to inversely scale
1570 * down the baseline weight. We need to ensure that we
1571 * don't exclude this metaslab completely when it's 100%
1572 * fragmented. To avoid this we reduce the fragmented value
1575 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1578 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1579 * this metaslab again. The fragmentation metric may have
1580 * decreased the space to something smaller than
1581 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1582 * so that we can consume any remaining space.
1584 if (space > 0 && space < SPA_MINBLOCKSIZE)
1585 space = SPA_MINBLOCKSIZE;
1590 * Modern disks have uniform bit density and constant angular velocity.
1591 * Therefore, the outer recording zones are faster (higher bandwidth)
1592 * than the inner zones by the ratio of outer to inner track diameter,
1593 * which is typically around 2:1. We account for this by assigning
1594 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1595 * In effect, this means that we'll select the metaslab with the most
1596 * free bandwidth rather than simply the one with the most free space.
1598 if (metaslab_lba_weighting_enabled) {
1599 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1600 ASSERT(weight >= space && weight <= 2 * space);
1604 * If this metaslab is one we're actively using, adjust its
1605 * weight to make it preferable to any inactive metaslab so
1606 * we'll polish it off. If the fragmentation on this metaslab
1607 * has exceed our threshold, then don't mark it active.
1609 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1610 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1611 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1618 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1620 ASSERT(MUTEX_HELD(&msp->ms_lock));
1622 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1623 metaslab_load_wait(msp);
1624 if (!msp->ms_loaded) {
1625 int error = metaslab_load(msp);
1627 metaslab_group_sort(msp->ms_group, msp, 0);
1632 metaslab_group_sort(msp->ms_group, msp,
1633 msp->ms_weight | activation_weight);
1635 ASSERT(msp->ms_loaded);
1636 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1642 metaslab_passivate(metaslab_t *msp, uint64_t size)
1645 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1646 * this metaslab again. In that case, it had better be empty,
1647 * or we would be leaving space on the table.
1649 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1650 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1651 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1655 metaslab_preload(void *arg)
1657 metaslab_t *msp = arg;
1658 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1660 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1662 mutex_enter(&msp->ms_lock);
1663 metaslab_load_wait(msp);
1664 if (!msp->ms_loaded)
1665 (void) metaslab_load(msp);
1668 * Set the ms_access_txg value so that we don't unload it right away.
1670 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1671 mutex_exit(&msp->ms_lock);
1675 metaslab_group_preload(metaslab_group_t *mg)
1677 spa_t *spa = mg->mg_vd->vdev_spa;
1679 avl_tree_t *t = &mg->mg_metaslab_tree;
1682 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1683 taskq_wait(mg->mg_taskq);
1687 mutex_enter(&mg->mg_lock);
1689 * Load the next potential metaslabs
1692 while (msp != NULL) {
1693 metaslab_t *msp_next = AVL_NEXT(t, msp);
1696 * We preload only the maximum number of metaslabs specified
1697 * by metaslab_preload_limit. If a metaslab is being forced
1698 * to condense then we preload it too. This will ensure
1699 * that force condensing happens in the next txg.
1701 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
1707 * We must drop the metaslab group lock here to preserve
1708 * lock ordering with the ms_lock (when grabbing both
1709 * the mg_lock and the ms_lock, the ms_lock must be taken
1710 * first). As a result, it is possible that the ordering
1711 * of the metaslabs within the avl tree may change before
1712 * we reacquire the lock. The metaslab cannot be removed from
1713 * the tree while we're in syncing context so it is safe to
1714 * drop the mg_lock here. If the metaslabs are reordered
1715 * nothing will break -- we just may end up loading a
1716 * less than optimal one.
1718 mutex_exit(&mg->mg_lock);
1719 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1720 msp, TQ_SLEEP) != 0);
1721 mutex_enter(&mg->mg_lock);
1724 mutex_exit(&mg->mg_lock);
1728 * Determine if the space map's on-disk footprint is past our tolerance
1729 * for inefficiency. We would like to use the following criteria to make
1732 * 1. The size of the space map object should not dramatically increase as a
1733 * result of writing out the free space range tree.
1735 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1736 * times the size than the free space range tree representation
1737 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1739 * 3. The on-disk size of the space map should actually decrease.
1741 * Checking the first condition is tricky since we don't want to walk
1742 * the entire AVL tree calculating the estimated on-disk size. Instead we
1743 * use the size-ordered range tree in the metaslab and calculate the
1744 * size required to write out the largest segment in our free tree. If the
1745 * size required to represent that segment on disk is larger than the space
1746 * map object then we avoid condensing this map.
1748 * To determine the second criterion we use a best-case estimate and assume
1749 * each segment can be represented on-disk as a single 64-bit entry. We refer
1750 * to this best-case estimate as the space map's minimal form.
1752 * Unfortunately, we cannot compute the on-disk size of the space map in this
1753 * context because we cannot accurately compute the effects of compression, etc.
1754 * Instead, we apply the heuristic described in the block comment for
1755 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1756 * is greater than a threshold number of blocks.
1759 metaslab_should_condense(metaslab_t *msp)
1761 space_map_t *sm = msp->ms_sm;
1763 uint64_t size, entries, segsz, object_size, optimal_size, record_size;
1764 dmu_object_info_t doi;
1765 uint64_t vdev_blocksize = 1 << msp->ms_group->mg_vd->vdev_ashift;
1767 ASSERT(MUTEX_HELD(&msp->ms_lock));
1768 ASSERT(msp->ms_loaded);
1771 * Use the ms_size_tree range tree, which is ordered by size, to
1772 * obtain the largest segment in the free tree. We always condense
1773 * metaslabs that are empty and metaslabs for which a condense
1774 * request has been made.
1776 rs = avl_last(&msp->ms_size_tree);
1777 if (rs == NULL || msp->ms_condense_wanted)
1781 * Calculate the number of 64-bit entries this segment would
1782 * require when written to disk. If this single segment would be
1783 * larger on-disk than the entire current on-disk structure, then
1784 * clearly condensing will increase the on-disk structure size.
1786 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1787 entries = size / (MIN(size, SM_RUN_MAX));
1788 segsz = entries * sizeof (uint64_t);
1790 optimal_size = sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root);
1791 object_size = space_map_length(msp->ms_sm);
1793 dmu_object_info_from_db(sm->sm_dbuf, &doi);
1794 record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
1796 return (segsz <= object_size &&
1797 object_size >= (optimal_size * zfs_condense_pct / 100) &&
1798 object_size > zfs_metaslab_condense_block_threshold * record_size);
1802 * Condense the on-disk space map representation to its minimized form.
1803 * The minimized form consists of a small number of allocations followed by
1804 * the entries of the free range tree.
1807 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1809 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1810 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1811 range_tree_t *condense_tree;
1812 space_map_t *sm = msp->ms_sm;
1814 ASSERT(MUTEX_HELD(&msp->ms_lock));
1815 ASSERT3U(spa_sync_pass(spa), ==, 1);
1816 ASSERT(msp->ms_loaded);
1819 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
1820 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
1821 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
1822 msp->ms_group->mg_vd->vdev_spa->spa_name,
1823 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
1824 msp->ms_condense_wanted ? "TRUE" : "FALSE");
1826 msp->ms_condense_wanted = B_FALSE;
1829 * Create an range tree that is 100% allocated. We remove segments
1830 * that have been freed in this txg, any deferred frees that exist,
1831 * and any allocation in the future. Removing segments should be
1832 * a relatively inexpensive operation since we expect these trees to
1833 * have a small number of nodes.
1835 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1836 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1839 * Remove what's been freed in this txg from the condense_tree.
1840 * Since we're in sync_pass 1, we know that all the frees from
1841 * this txg are in the freetree.
1843 range_tree_walk(freetree, range_tree_remove, condense_tree);
1845 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1846 range_tree_walk(msp->ms_defertree[t],
1847 range_tree_remove, condense_tree);
1850 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1851 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1852 range_tree_remove, condense_tree);
1856 * We're about to drop the metaslab's lock thus allowing
1857 * other consumers to change it's content. Set the
1858 * metaslab's ms_condensing flag to ensure that
1859 * allocations on this metaslab do not occur while we're
1860 * in the middle of committing it to disk. This is only critical
1861 * for the ms_tree as all other range trees use per txg
1862 * views of their content.
1864 msp->ms_condensing = B_TRUE;
1866 mutex_exit(&msp->ms_lock);
1867 space_map_truncate(sm, tx);
1868 mutex_enter(&msp->ms_lock);
1871 * While we would ideally like to create a space_map representation
1872 * that consists only of allocation records, doing so can be
1873 * prohibitively expensive because the in-core free tree can be
1874 * large, and therefore computationally expensive to subtract
1875 * from the condense_tree. Instead we sync out two trees, a cheap
1876 * allocation only tree followed by the in-core free tree. While not
1877 * optimal, this is typically close to optimal, and much cheaper to
1880 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1881 range_tree_vacate(condense_tree, NULL, NULL);
1882 range_tree_destroy(condense_tree);
1884 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1885 msp->ms_condensing = B_FALSE;
1889 * Write a metaslab to disk in the context of the specified transaction group.
1892 metaslab_sync(metaslab_t *msp, uint64_t txg)
1894 metaslab_group_t *mg = msp->ms_group;
1895 vdev_t *vd = mg->mg_vd;
1896 spa_t *spa = vd->vdev_spa;
1897 objset_t *mos = spa_meta_objset(spa);
1898 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1899 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1900 range_tree_t **freed_tree =
1901 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1903 uint64_t object = space_map_object(msp->ms_sm);
1905 ASSERT(!vd->vdev_ishole);
1908 * This metaslab has just been added so there's no work to do now.
1910 if (*freetree == NULL) {
1911 ASSERT3P(alloctree, ==, NULL);
1915 ASSERT3P(alloctree, !=, NULL);
1916 ASSERT3P(*freetree, !=, NULL);
1917 ASSERT3P(*freed_tree, !=, NULL);
1920 * Normally, we don't want to process a metaslab if there
1921 * are no allocations or frees to perform. However, if the metaslab
1922 * is being forced to condense we need to let it through.
1924 if (range_tree_space(alloctree) == 0 &&
1925 range_tree_space(*freetree) == 0 &&
1926 !msp->ms_condense_wanted)
1930 * The only state that can actually be changing concurrently with
1931 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1932 * be modifying this txg's alloctree, freetree, freed_tree, or
1933 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1934 * space_map ASSERTs. We drop it whenever we call into the DMU,
1935 * because the DMU can call down to us (e.g. via zio_free()) at
1939 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1941 if (msp->ms_sm == NULL) {
1942 uint64_t new_object;
1944 new_object = space_map_alloc(mos, tx);
1945 VERIFY3U(new_object, !=, 0);
1947 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1948 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1950 ASSERT(msp->ms_sm != NULL);
1953 mutex_enter(&msp->ms_lock);
1956 * Note: metaslab_condense() clears the space_map's histogram.
1957 * Therefore we must verify and remove this histogram before
1960 metaslab_group_histogram_verify(mg);
1961 metaslab_class_histogram_verify(mg->mg_class);
1962 metaslab_group_histogram_remove(mg, msp);
1964 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1965 metaslab_should_condense(msp)) {
1966 metaslab_condense(msp, txg, tx);
1968 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1969 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1972 if (msp->ms_loaded) {
1974 * When the space map is loaded, we have an accruate
1975 * histogram in the range tree. This gives us an opportunity
1976 * to bring the space map's histogram up-to-date so we clear
1977 * it first before updating it.
1979 space_map_histogram_clear(msp->ms_sm);
1980 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1983 * Since the space map is not loaded we simply update the
1984 * exisiting histogram with what was freed in this txg. This
1985 * means that the on-disk histogram may not have an accurate
1986 * view of the free space but it's close enough to allow
1987 * us to make allocation decisions.
1989 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1991 metaslab_group_histogram_add(mg, msp);
1992 metaslab_group_histogram_verify(mg);
1993 metaslab_class_histogram_verify(mg->mg_class);
1996 * For sync pass 1, we avoid traversing this txg's free range tree
1997 * and instead will just swap the pointers for freetree and
1998 * freed_tree. We can safely do this since the freed_tree is
1999 * guaranteed to be empty on the initial pass.
2001 if (spa_sync_pass(spa) == 1) {
2002 range_tree_swap(freetree, freed_tree);
2004 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
2006 range_tree_vacate(alloctree, NULL, NULL);
2008 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2009 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2011 mutex_exit(&msp->ms_lock);
2013 if (object != space_map_object(msp->ms_sm)) {
2014 object = space_map_object(msp->ms_sm);
2015 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2016 msp->ms_id, sizeof (uint64_t), &object, tx);
2022 * Called after a transaction group has completely synced to mark
2023 * all of the metaslab's free space as usable.
2026 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2028 metaslab_group_t *mg = msp->ms_group;
2029 vdev_t *vd = mg->mg_vd;
2030 range_tree_t **freed_tree;
2031 range_tree_t **defer_tree;
2032 int64_t alloc_delta, defer_delta;
2034 ASSERT(!vd->vdev_ishole);
2036 mutex_enter(&msp->ms_lock);
2039 * If this metaslab is just becoming available, initialize its
2040 * alloctrees, freetrees, and defertree and add its capacity to
2043 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
2044 for (int t = 0; t < TXG_SIZE; t++) {
2045 ASSERT(msp->ms_alloctree[t] == NULL);
2046 ASSERT(msp->ms_freetree[t] == NULL);
2048 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2050 msp->ms_freetree[t] = range_tree_create(NULL, msp,
2054 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2055 ASSERT(msp->ms_defertree[t] == NULL);
2057 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2061 vdev_space_update(vd, 0, 0, msp->ms_size);
2064 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
2065 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2067 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2068 defer_delta = range_tree_space(*freed_tree) -
2069 range_tree_space(*defer_tree);
2071 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2073 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2074 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
2077 * If there's a metaslab_load() in progress, wait for it to complete
2078 * so that we have a consistent view of the in-core space map.
2080 metaslab_load_wait(msp);
2083 * Move the frees from the defer_tree back to the free
2084 * range tree (if it's loaded). Swap the freed_tree and the
2085 * defer_tree -- this is safe to do because we've just emptied out
2088 range_tree_vacate(*defer_tree,
2089 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2090 range_tree_swap(freed_tree, defer_tree);
2092 space_map_update(msp->ms_sm);
2094 msp->ms_deferspace += defer_delta;
2095 ASSERT3S(msp->ms_deferspace, >=, 0);
2096 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2097 if (msp->ms_deferspace != 0) {
2099 * Keep syncing this metaslab until all deferred frees
2100 * are back in circulation.
2102 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2105 if (msp->ms_loaded && msp->ms_access_txg < txg) {
2106 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2107 VERIFY0(range_tree_space(
2108 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2111 if (!metaslab_debug_unload)
2112 metaslab_unload(msp);
2115 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2116 mutex_exit(&msp->ms_lock);
2120 metaslab_sync_reassess(metaslab_group_t *mg)
2122 metaslab_group_alloc_update(mg);
2123 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2126 * Preload the next potential metaslabs
2128 metaslab_group_preload(mg);
2132 metaslab_distance(metaslab_t *msp, dva_t *dva)
2134 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2135 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2136 uint64_t start = msp->ms_id;
2138 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2139 return (1ULL << 63);
2142 return ((start - offset) << ms_shift);
2144 return ((offset - start) << ms_shift);
2149 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
2150 uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2152 spa_t *spa = mg->mg_vd->vdev_spa;
2153 metaslab_t *msp = NULL;
2154 uint64_t offset = -1ULL;
2155 avl_tree_t *t = &mg->mg_metaslab_tree;
2156 uint64_t activation_weight;
2157 uint64_t target_distance;
2160 activation_weight = METASLAB_WEIGHT_PRIMARY;
2161 for (i = 0; i < d; i++) {
2162 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2163 activation_weight = METASLAB_WEIGHT_SECONDARY;
2169 boolean_t was_active;
2171 mutex_enter(&mg->mg_lock);
2172 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
2173 if (msp->ms_weight < asize) {
2174 spa_dbgmsg(spa, "%s: failed to meet weight "
2175 "requirement: vdev %llu, txg %llu, mg %p, "
2176 "msp %p, psize %llu, asize %llu, "
2177 "weight %llu", spa_name(spa),
2178 mg->mg_vd->vdev_id, txg,
2179 mg, msp, psize, asize, msp->ms_weight);
2180 mutex_exit(&mg->mg_lock);
2185 * If the selected metaslab is condensing, skip it.
2187 if (msp->ms_condensing)
2190 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2191 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2194 target_distance = min_distance +
2195 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2198 for (i = 0; i < d; i++)
2199 if (metaslab_distance(msp, &dva[i]) <
2205 mutex_exit(&mg->mg_lock);
2209 mutex_enter(&msp->ms_lock);
2212 * Ensure that the metaslab we have selected is still
2213 * capable of handling our request. It's possible that
2214 * another thread may have changed the weight while we
2215 * were blocked on the metaslab lock.
2217 if (msp->ms_weight < asize || (was_active &&
2218 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
2219 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
2220 mutex_exit(&msp->ms_lock);
2224 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
2225 activation_weight == METASLAB_WEIGHT_PRIMARY) {
2226 metaslab_passivate(msp,
2227 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
2228 mutex_exit(&msp->ms_lock);
2232 if (metaslab_activate(msp, activation_weight) != 0) {
2233 mutex_exit(&msp->ms_lock);
2238 * If this metaslab is currently condensing then pick again as
2239 * we can't manipulate this metaslab until it's committed
2242 if (msp->ms_condensing) {
2243 mutex_exit(&msp->ms_lock);
2247 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
2250 metaslab_passivate(msp, metaslab_block_maxsize(msp));
2251 mutex_exit(&msp->ms_lock);
2254 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2255 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2257 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
2258 msp->ms_access_txg = txg + metaslab_unload_delay;
2260 mutex_exit(&msp->ms_lock);
2266 * Allocate a block for the specified i/o.
2269 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2270 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
2272 metaslab_group_t *mg, *rotor;
2276 int zio_lock = B_FALSE;
2277 boolean_t allocatable;
2278 uint64_t offset = -1ULL;
2282 ASSERT(!DVA_IS_VALID(&dva[d]));
2285 * For testing, make some blocks above a certain size be gang blocks.
2287 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
2288 return (SET_ERROR(ENOSPC));
2291 * Start at the rotor and loop through all mgs until we find something.
2292 * Note that there's no locking on mc_rotor or mc_aliquot because
2293 * nothing actually breaks if we miss a few updates -- we just won't
2294 * allocate quite as evenly. It all balances out over time.
2296 * If we are doing ditto or log blocks, try to spread them across
2297 * consecutive vdevs. If we're forced to reuse a vdev before we've
2298 * allocated all of our ditto blocks, then try and spread them out on
2299 * that vdev as much as possible. If it turns out to not be possible,
2300 * gradually lower our standards until anything becomes acceptable.
2301 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2302 * gives us hope of containing our fault domains to something we're
2303 * able to reason about. Otherwise, any two top-level vdev failures
2304 * will guarantee the loss of data. With consecutive allocation,
2305 * only two adjacent top-level vdev failures will result in data loss.
2307 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2308 * ourselves on the same vdev as our gang block header. That
2309 * way, we can hope for locality in vdev_cache, plus it makes our
2310 * fault domains something tractable.
2313 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
2316 * It's possible the vdev we're using as the hint no
2317 * longer exists (i.e. removed). Consult the rotor when
2323 if (flags & METASLAB_HINTBP_AVOID &&
2324 mg->mg_next != NULL)
2329 } else if (d != 0) {
2330 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
2331 mg = vd->vdev_mg->mg_next;
2337 * If the hint put us into the wrong metaslab class, or into a
2338 * metaslab group that has been passivated, just follow the rotor.
2340 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
2347 ASSERT(mg->mg_activation_count == 1);
2352 * Don't allocate from faulted devices.
2355 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
2356 allocatable = vdev_allocatable(vd);
2357 spa_config_exit(spa, SCL_ZIO, FTAG);
2359 allocatable = vdev_allocatable(vd);
2363 * Determine if the selected metaslab group is eligible
2364 * for allocations. If we're ganging or have requested
2365 * an allocation for the smallest gang block size
2366 * then we don't want to avoid allocating to the this
2367 * metaslab group. If we're in this condition we should
2368 * try to allocate from any device possible so that we
2369 * don't inadvertently return ENOSPC and suspend the pool
2370 * even though space is still available.
2372 if (allocatable && CAN_FASTGANG(flags) &&
2373 psize > SPA_GANGBLOCKSIZE)
2374 allocatable = metaslab_group_allocatable(mg);
2380 * Avoid writing single-copy data to a failing vdev
2381 * unless the user instructs us that it is okay.
2383 if ((vd->vdev_stat.vs_write_errors > 0 ||
2384 vd->vdev_state < VDEV_STATE_HEALTHY) &&
2385 d == 0 && dshift == 3 && vd->vdev_children == 0) {
2390 ASSERT(mg->mg_class == mc);
2392 distance = vd->vdev_asize >> dshift;
2393 if (distance <= (1ULL << vd->vdev_ms_shift))
2398 asize = vdev_psize_to_asize(vd, psize);
2399 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
2401 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
2403 if (offset != -1ULL) {
2405 * If we've just selected this metaslab group,
2406 * figure out whether the corresponding vdev is
2407 * over- or under-used relative to the pool,
2408 * and set an allocation bias to even it out.
2410 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
2411 vdev_stat_t *vs = &vd->vdev_stat;
2414 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
2415 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2418 * Calculate how much more or less we should
2419 * try to allocate from this device during
2420 * this iteration around the rotor.
2421 * For example, if a device is 80% full
2422 * and the pool is 20% full then we should
2423 * reduce allocations by 60% on this device.
2425 * mg_bias = (20 - 80) * 512K / 100 = -307K
2427 * This reduces allocations by 307K for this
2430 mg->mg_bias = ((cu - vu) *
2431 (int64_t)mg->mg_aliquot) / 100;
2432 } else if (!metaslab_bias_enabled) {
2436 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2437 mg->mg_aliquot + mg->mg_bias) {
2438 mc->mc_rotor = mg->mg_next;
2442 DVA_SET_VDEV(&dva[d], vd->vdev_id);
2443 DVA_SET_OFFSET(&dva[d], offset);
2444 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2445 DVA_SET_ASIZE(&dva[d], asize);
2450 mc->mc_rotor = mg->mg_next;
2452 } while ((mg = mg->mg_next) != rotor);
2456 ASSERT(dshift < 64);
2460 if (!allocatable && !zio_lock) {
2466 bzero(&dva[d], sizeof (dva_t));
2468 return (SET_ERROR(ENOSPC));
2472 * Free the block represented by DVA in the context of the specified
2473 * transaction group.
2476 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2478 uint64_t vdev = DVA_GET_VDEV(dva);
2479 uint64_t offset = DVA_GET_OFFSET(dva);
2480 uint64_t size = DVA_GET_ASIZE(dva);
2484 ASSERT(DVA_IS_VALID(dva));
2486 if (txg > spa_freeze_txg(spa))
2489 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2490 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2491 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2492 (u_longlong_t)vdev, (u_longlong_t)offset);
2497 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2499 if (DVA_GET_GANG(dva))
2500 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2502 mutex_enter(&msp->ms_lock);
2505 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2508 VERIFY(!msp->ms_condensing);
2509 VERIFY3U(offset, >=, msp->ms_start);
2510 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2511 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2513 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2514 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2515 range_tree_add(msp->ms_tree, offset, size);
2517 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2518 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2519 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2523 mutex_exit(&msp->ms_lock);
2527 * Intent log support: upon opening the pool after a crash, notify the SPA
2528 * of blocks that the intent log has allocated for immediate write, but
2529 * which are still considered free by the SPA because the last transaction
2530 * group didn't commit yet.
2533 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2535 uint64_t vdev = DVA_GET_VDEV(dva);
2536 uint64_t offset = DVA_GET_OFFSET(dva);
2537 uint64_t size = DVA_GET_ASIZE(dva);
2542 ASSERT(DVA_IS_VALID(dva));
2544 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2545 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2546 return (SET_ERROR(ENXIO));
2548 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2550 if (DVA_GET_GANG(dva))
2551 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2553 mutex_enter(&msp->ms_lock);
2555 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2556 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2558 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2559 error = SET_ERROR(ENOENT);
2561 if (error || txg == 0) { /* txg == 0 indicates dry run */
2562 mutex_exit(&msp->ms_lock);
2566 VERIFY(!msp->ms_condensing);
2567 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2568 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2569 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2570 range_tree_remove(msp->ms_tree, offset, size);
2572 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2573 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2574 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2575 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2578 mutex_exit(&msp->ms_lock);
2584 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2585 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2587 dva_t *dva = bp->blk_dva;
2588 dva_t *hintdva = hintbp->blk_dva;
2591 ASSERT(bp->blk_birth == 0);
2592 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2594 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2596 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2597 spa_config_exit(spa, SCL_ALLOC, FTAG);
2598 return (SET_ERROR(ENOSPC));
2601 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2602 ASSERT(BP_GET_NDVAS(bp) == 0);
2603 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2605 for (int d = 0; d < ndvas; d++) {
2606 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2609 for (d--; d >= 0; d--) {
2610 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2611 bzero(&dva[d], sizeof (dva_t));
2613 spa_config_exit(spa, SCL_ALLOC, FTAG);
2618 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2620 spa_config_exit(spa, SCL_ALLOC, FTAG);
2622 BP_SET_BIRTH(bp, txg, txg);
2628 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2630 const dva_t *dva = bp->blk_dva;
2631 int ndvas = BP_GET_NDVAS(bp);
2633 ASSERT(!BP_IS_HOLE(bp));
2634 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2636 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2638 for (int d = 0; d < ndvas; d++)
2639 metaslab_free_dva(spa, &dva[d], txg, now);
2641 spa_config_exit(spa, SCL_FREE, FTAG);
2645 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2647 const dva_t *dva = bp->blk_dva;
2648 int ndvas = BP_GET_NDVAS(bp);
2651 ASSERT(!BP_IS_HOLE(bp));
2655 * First do a dry run to make sure all DVAs are claimable,
2656 * so we don't have to unwind from partial failures below.
2658 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2662 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2664 for (int d = 0; d < ndvas; d++)
2665 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2668 spa_config_exit(spa, SCL_ALLOC, FTAG);
2670 ASSERT(error == 0 || txg == 0);
2676 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2678 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2681 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2682 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2683 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2684 vdev_t *vd = vdev_lookup_top(spa, vdev);
2685 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2686 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2687 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2690 range_tree_verify(msp->ms_tree, offset, size);
2692 for (int j = 0; j < TXG_SIZE; j++)
2693 range_tree_verify(msp->ms_freetree[j], offset, size);
2694 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2695 range_tree_verify(msp->ms_defertree[j], offset, size);
2697 spa_config_exit(spa, SCL_VDEV, FTAG);