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>
36 SYSCTL_DECL(_vfs_zfs);
37 SYSCTL_NODE(_vfs_zfs, OID_AUTO, metaslab, CTLFLAG_RW, 0, "ZFS metaslab");
40 * Allow allocations to switch to gang blocks quickly. We do this to
41 * avoid having to load lots of space_maps in a given txg. There are,
42 * however, some cases where we want to avoid "fast" ganging and instead
43 * we want to do an exhaustive search of all metaslabs on this device.
44 * Currently we don't allow any gang, slog, or dump device related allocations
47 #define CAN_FASTGANG(flags) \
48 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
49 METASLAB_GANG_AVOID)))
51 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
52 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
53 #define METASLAB_ACTIVE_MASK \
54 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
56 uint64_t metaslab_aliquot = 512ULL << 10;
57 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
58 TUNABLE_QUAD("vfs.zfs.metaslab.gang_bang", &metaslab_gang_bang);
59 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, gang_bang, CTLFLAG_RWTUN,
60 &metaslab_gang_bang, 0,
61 "Force gang block allocation for blocks larger than or equal to this value");
64 * The in-core space map representation is more compact than its on-disk form.
65 * The zfs_condense_pct determines how much more compact the in-core
66 * space_map representation must be before we compact it on-disk.
67 * Values should be greater than or equal to 100.
69 int zfs_condense_pct = 200;
70 TUNABLE_INT("vfs.zfs.condense_pct", &zfs_condense_pct);
71 SYSCTL_INT(_vfs_zfs, OID_AUTO, condense_pct, CTLFLAG_RWTUN,
73 "Condense on-disk spacemap when it is more than this many percents"
74 " of in-memory counterpart");
77 * The zfs_mg_noalloc_threshold defines which metaslab groups should
78 * be eligible for allocation. The value is defined as a percentage of
79 * a free space. Metaslab groups that have more free space than
80 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
81 * a metaslab group's free space is less than or equal to the
82 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
83 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
84 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
85 * groups are allowed to accept allocations. Gang blocks are always
86 * eligible to allocate on any metaslab group. The default value of 0 means
87 * no metaslab group will be excluded based on this criterion.
89 int zfs_mg_noalloc_threshold = 0;
90 TUNABLE_INT("vfs.zfs.mg_noalloc_threshold", &zfs_mg_noalloc_threshold);
91 SYSCTL_INT(_vfs_zfs, OID_AUTO, mg_noalloc_threshold, CTLFLAG_RWTUN,
92 &zfs_mg_noalloc_threshold, 0,
93 "Percentage of metaslab group size that should be free"
94 " to make it eligible for allocation");
97 * When set will load all metaslabs when pool is first opened.
99 int metaslab_debug_load = 0;
100 TUNABLE_INT("vfs.zfs.metaslab.debug_load", &metaslab_debug_load);
101 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_load, CTLFLAG_RWTUN,
102 &metaslab_debug_load, 0,
103 "Load all metaslabs when pool is first opened");
106 * When set will prevent metaslabs from being unloaded.
108 int metaslab_debug_unload = 0;
109 TUNABLE_INT("vfs.zfs.metaslab.debug_unload", &metaslab_debug_unload);
110 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, debug_unload, CTLFLAG_RWTUN,
111 &metaslab_debug_unload, 0,
112 "Prevent metaslabs from being unloaded");
115 * Minimum size which forces the dynamic allocator to change
116 * it's allocation strategy. Once the space map cannot satisfy
117 * an allocation of this size then it switches to using more
118 * aggressive strategy (i.e search by size rather than offset).
120 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
121 TUNABLE_QUAD("vfs.zfs.metaslab.df_alloc_threshold",
122 &metaslab_df_alloc_threshold);
123 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, df_alloc_threshold, CTLFLAG_RWTUN,
124 &metaslab_df_alloc_threshold, 0,
125 "Minimum size which forces the dynamic allocator to change it's allocation strategy");
128 * The minimum free space, in percent, which must be available
129 * in a space map to continue allocations in a first-fit fashion.
130 * Once the space_map's free space drops below this level we dynamically
131 * switch to using best-fit allocations.
133 int metaslab_df_free_pct = 4;
134 TUNABLE_INT("vfs.zfs.metaslab.df_free_pct", &metaslab_df_free_pct);
135 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, df_free_pct, CTLFLAG_RWTUN,
136 &metaslab_df_free_pct, 0,
137 "The minimum free space, in percent, which must be available in a space map to continue allocations in a first-fit fashion");
140 * A metaslab is considered "free" if it contains a contiguous
141 * segment which is greater than metaslab_min_alloc_size.
143 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
144 TUNABLE_QUAD("vfs.zfs.metaslab.min_alloc_size",
145 &metaslab_min_alloc_size);
146 SYSCTL_QUAD(_vfs_zfs_metaslab, OID_AUTO, min_alloc_size, CTLFLAG_RWTUN,
147 &metaslab_min_alloc_size, 0,
148 "A metaslab is considered \"free\" if it contains a contiguous segment which is greater than vfs.zfs.metaslab.min_alloc_size");
151 * Percentage of all cpus that can be used by the metaslab taskq.
153 int metaslab_load_pct = 50;
154 TUNABLE_INT("vfs.zfs.metaslab.load_pct", &metaslab_load_pct);
155 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, load_pct, CTLFLAG_RWTUN,
156 &metaslab_load_pct, 0,
157 "Percentage of cpus that can be used by the metaslab taskq");
160 * Determines how many txgs a metaslab may remain loaded without having any
161 * allocations from it. As long as a metaslab continues to be used we will
164 int metaslab_unload_delay = TXG_SIZE * 2;
165 TUNABLE_INT("vfs.zfs.metaslab.unload_delay", &metaslab_unload_delay);
166 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, unload_delay, CTLFLAG_RWTUN,
167 &metaslab_unload_delay, 0,
168 "Number of TXGs that an unused metaslab can be kept in memory");
171 * Should we be willing to write data to degraded vdevs?
173 boolean_t zfs_write_to_degraded = B_FALSE;
174 SYSCTL_INT(_vfs_zfs, OID_AUTO, write_to_degraded, CTLFLAG_RW,
175 &zfs_write_to_degraded, 0,
176 "Allow writing data to degraded vdevs");
177 TUNABLE_INT("vfs.zfs.write_to_degraded", &zfs_write_to_degraded);
180 * Max number of metaslabs per group to preload.
182 int metaslab_preload_limit = SPA_DVAS_PER_BP;
183 TUNABLE_INT("vfs.zfs.metaslab.preload_limit", &metaslab_preload_limit);
184 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
185 &metaslab_preload_limit, 0,
186 "Max number of metaslabs per group to preload");
189 * Enable/disable preloading of metaslab.
191 boolean_t metaslab_preload_enabled = B_TRUE;
192 TUNABLE_INT("vfs.zfs.metaslab.preload_enabled", &metaslab_preload_enabled);
193 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
194 &metaslab_preload_enabled, 0,
195 "Max number of metaslabs per group to preload");
198 * Enable/disable additional weight factor for each metaslab.
200 boolean_t metaslab_weight_factor_enable = B_FALSE;
201 TUNABLE_INT("vfs.zfs.metaslab.weight_factor_enable",
202 &metaslab_weight_factor_enable);
203 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, weight_factor_enable, CTLFLAG_RWTUN,
204 &metaslab_weight_factor_enable, 0,
205 "Enable additional weight factor for each metaslab");
209 * ==========================================================================
211 * ==========================================================================
214 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
216 metaslab_class_t *mc;
218 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
228 metaslab_class_destroy(metaslab_class_t *mc)
230 ASSERT(mc->mc_rotor == NULL);
231 ASSERT(mc->mc_alloc == 0);
232 ASSERT(mc->mc_deferred == 0);
233 ASSERT(mc->mc_space == 0);
234 ASSERT(mc->mc_dspace == 0);
236 kmem_free(mc, sizeof (metaslab_class_t));
240 metaslab_class_validate(metaslab_class_t *mc)
242 metaslab_group_t *mg;
246 * Must hold one of the spa_config locks.
248 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
249 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
251 if ((mg = mc->mc_rotor) == NULL)
256 ASSERT(vd->vdev_mg != NULL);
257 ASSERT3P(vd->vdev_top, ==, vd);
258 ASSERT3P(mg->mg_class, ==, mc);
259 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
260 } while ((mg = mg->mg_next) != mc->mc_rotor);
266 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
267 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
269 atomic_add_64(&mc->mc_alloc, alloc_delta);
270 atomic_add_64(&mc->mc_deferred, defer_delta);
271 atomic_add_64(&mc->mc_space, space_delta);
272 atomic_add_64(&mc->mc_dspace, dspace_delta);
276 metaslab_class_minblocksize_update(metaslab_class_t *mc)
278 metaslab_group_t *mg;
280 uint64_t minashift = UINT64_MAX;
282 if ((mg = mc->mc_rotor) == NULL) {
283 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
289 if (vd->vdev_ashift < minashift)
290 minashift = vd->vdev_ashift;
291 } while ((mg = mg->mg_next) != mc->mc_rotor);
293 mc->mc_minblocksize = 1ULL << minashift;
297 metaslab_class_get_alloc(metaslab_class_t *mc)
299 return (mc->mc_alloc);
303 metaslab_class_get_deferred(metaslab_class_t *mc)
305 return (mc->mc_deferred);
309 metaslab_class_get_space(metaslab_class_t *mc)
311 return (mc->mc_space);
315 metaslab_class_get_dspace(metaslab_class_t *mc)
317 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
321 metaslab_class_get_minblocksize(metaslab_class_t *mc)
323 return (mc->mc_minblocksize);
327 * ==========================================================================
329 * ==========================================================================
332 metaslab_compare(const void *x1, const void *x2)
334 const metaslab_t *m1 = x1;
335 const metaslab_t *m2 = x2;
337 if (m1->ms_weight < m2->ms_weight)
339 if (m1->ms_weight > m2->ms_weight)
343 * If the weights are identical, use the offset to force uniqueness.
345 if (m1->ms_start < m2->ms_start)
347 if (m1->ms_start > m2->ms_start)
350 ASSERT3P(m1, ==, m2);
356 * Update the allocatable flag and the metaslab group's capacity.
357 * The allocatable flag is set to true if the capacity is below
358 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
359 * from allocatable to non-allocatable or vice versa then the metaslab
360 * group's class is updated to reflect the transition.
363 metaslab_group_alloc_update(metaslab_group_t *mg)
365 vdev_t *vd = mg->mg_vd;
366 metaslab_class_t *mc = mg->mg_class;
367 vdev_stat_t *vs = &vd->vdev_stat;
368 boolean_t was_allocatable;
370 ASSERT(vd == vd->vdev_top);
372 mutex_enter(&mg->mg_lock);
373 was_allocatable = mg->mg_allocatable;
375 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
378 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold);
381 * The mc_alloc_groups maintains a count of the number of
382 * groups in this metaslab class that are still above the
383 * zfs_mg_noalloc_threshold. This is used by the allocating
384 * threads to determine if they should avoid allocations to
385 * a given group. The allocator will avoid allocations to a group
386 * if that group has reached or is below the zfs_mg_noalloc_threshold
387 * and there are still other groups that are above the threshold.
388 * When a group transitions from allocatable to non-allocatable or
389 * vice versa we update the metaslab class to reflect that change.
390 * When the mc_alloc_groups value drops to 0 that means that all
391 * groups have reached the zfs_mg_noalloc_threshold making all groups
392 * eligible for allocations. This effectively means that all devices
393 * are balanced again.
395 if (was_allocatable && !mg->mg_allocatable)
396 mc->mc_alloc_groups--;
397 else if (!was_allocatable && mg->mg_allocatable)
398 mc->mc_alloc_groups++;
399 mutex_exit(&mg->mg_lock);
403 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
405 metaslab_group_t *mg;
407 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
408 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
409 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
410 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
413 mg->mg_activation_count = 0;
415 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
416 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
422 metaslab_group_destroy(metaslab_group_t *mg)
424 ASSERT(mg->mg_prev == NULL);
425 ASSERT(mg->mg_next == NULL);
427 * We may have gone below zero with the activation count
428 * either because we never activated in the first place or
429 * because we're done, and possibly removing the vdev.
431 ASSERT(mg->mg_activation_count <= 0);
433 taskq_destroy(mg->mg_taskq);
434 avl_destroy(&mg->mg_metaslab_tree);
435 mutex_destroy(&mg->mg_lock);
436 kmem_free(mg, sizeof (metaslab_group_t));
440 metaslab_group_activate(metaslab_group_t *mg)
442 metaslab_class_t *mc = mg->mg_class;
443 metaslab_group_t *mgprev, *mgnext;
445 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
447 ASSERT(mc->mc_rotor != mg);
448 ASSERT(mg->mg_prev == NULL);
449 ASSERT(mg->mg_next == NULL);
450 ASSERT(mg->mg_activation_count <= 0);
452 if (++mg->mg_activation_count <= 0)
455 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
456 metaslab_group_alloc_update(mg);
458 if ((mgprev = mc->mc_rotor) == NULL) {
462 mgnext = mgprev->mg_next;
463 mg->mg_prev = mgprev;
464 mg->mg_next = mgnext;
465 mgprev->mg_next = mg;
466 mgnext->mg_prev = mg;
469 metaslab_class_minblocksize_update(mc);
473 metaslab_group_passivate(metaslab_group_t *mg)
475 metaslab_class_t *mc = mg->mg_class;
476 metaslab_group_t *mgprev, *mgnext;
478 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
480 if (--mg->mg_activation_count != 0) {
481 ASSERT(mc->mc_rotor != mg);
482 ASSERT(mg->mg_prev == NULL);
483 ASSERT(mg->mg_next == NULL);
484 ASSERT(mg->mg_activation_count < 0);
488 taskq_wait(mg->mg_taskq);
490 mgprev = mg->mg_prev;
491 mgnext = mg->mg_next;
496 mc->mc_rotor = mgnext;
497 mgprev->mg_next = mgnext;
498 mgnext->mg_prev = mgprev;
503 metaslab_class_minblocksize_update(mc);
507 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
509 mutex_enter(&mg->mg_lock);
510 ASSERT(msp->ms_group == NULL);
513 avl_add(&mg->mg_metaslab_tree, msp);
514 mutex_exit(&mg->mg_lock);
518 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
520 mutex_enter(&mg->mg_lock);
521 ASSERT(msp->ms_group == mg);
522 avl_remove(&mg->mg_metaslab_tree, msp);
523 msp->ms_group = NULL;
524 mutex_exit(&mg->mg_lock);
528 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
531 * Although in principle the weight can be any value, in
532 * practice we do not use values in the range [1, 510].
534 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
535 ASSERT(MUTEX_HELD(&msp->ms_lock));
537 mutex_enter(&mg->mg_lock);
538 ASSERT(msp->ms_group == mg);
539 avl_remove(&mg->mg_metaslab_tree, msp);
540 msp->ms_weight = weight;
541 avl_add(&mg->mg_metaslab_tree, msp);
542 mutex_exit(&mg->mg_lock);
546 * Determine if a given metaslab group should skip allocations. A metaslab
547 * group should avoid allocations if its used capacity has crossed the
548 * zfs_mg_noalloc_threshold and there is at least one metaslab group
549 * that can still handle allocations.
552 metaslab_group_allocatable(metaslab_group_t *mg)
554 vdev_t *vd = mg->mg_vd;
555 spa_t *spa = vd->vdev_spa;
556 metaslab_class_t *mc = mg->mg_class;
559 * A metaslab group is considered allocatable if its free capacity
560 * is greater than the set value of zfs_mg_noalloc_threshold, it's
561 * associated with a slog, or there are no other metaslab groups
562 * with free capacity greater than zfs_mg_noalloc_threshold.
564 return (mg->mg_free_capacity > zfs_mg_noalloc_threshold ||
565 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
569 * ==========================================================================
570 * Range tree callbacks
571 * ==========================================================================
575 * Comparison function for the private size-ordered tree. Tree is sorted
576 * by size, larger sizes at the end of the tree.
579 metaslab_rangesize_compare(const void *x1, const void *x2)
581 const range_seg_t *r1 = x1;
582 const range_seg_t *r2 = x2;
583 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
584 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
586 if (rs_size1 < rs_size2)
588 if (rs_size1 > rs_size2)
591 if (r1->rs_start < r2->rs_start)
594 if (r1->rs_start > r2->rs_start)
601 * Create any block allocator specific components. The current allocators
602 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
605 metaslab_rt_create(range_tree_t *rt, void *arg)
607 metaslab_t *msp = arg;
609 ASSERT3P(rt->rt_arg, ==, msp);
610 ASSERT(msp->ms_tree == NULL);
612 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
613 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
617 * Destroy the block allocator specific components.
620 metaslab_rt_destroy(range_tree_t *rt, void *arg)
622 metaslab_t *msp = arg;
624 ASSERT3P(rt->rt_arg, ==, msp);
625 ASSERT3P(msp->ms_tree, ==, rt);
626 ASSERT0(avl_numnodes(&msp->ms_size_tree));
628 avl_destroy(&msp->ms_size_tree);
632 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
634 metaslab_t *msp = arg;
636 ASSERT3P(rt->rt_arg, ==, msp);
637 ASSERT3P(msp->ms_tree, ==, rt);
638 VERIFY(!msp->ms_condensing);
639 avl_add(&msp->ms_size_tree, rs);
643 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
645 metaslab_t *msp = arg;
647 ASSERT3P(rt->rt_arg, ==, msp);
648 ASSERT3P(msp->ms_tree, ==, rt);
649 VERIFY(!msp->ms_condensing);
650 avl_remove(&msp->ms_size_tree, rs);
654 metaslab_rt_vacate(range_tree_t *rt, void *arg)
656 metaslab_t *msp = arg;
658 ASSERT3P(rt->rt_arg, ==, msp);
659 ASSERT3P(msp->ms_tree, ==, rt);
662 * Normally one would walk the tree freeing nodes along the way.
663 * Since the nodes are shared with the range trees we can avoid
664 * walking all nodes and just reinitialize the avl tree. The nodes
665 * will be freed by the range tree, so we don't want to free them here.
667 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
668 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
671 static range_tree_ops_t metaslab_rt_ops = {
680 * ==========================================================================
681 * Metaslab block operations
682 * ==========================================================================
686 * Return the maximum contiguous segment within the metaslab.
689 metaslab_block_maxsize(metaslab_t *msp)
691 avl_tree_t *t = &msp->ms_size_tree;
694 if (t == NULL || (rs = avl_last(t)) == NULL)
697 return (rs->rs_end - rs->rs_start);
701 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
704 range_tree_t *rt = msp->ms_tree;
706 VERIFY(!msp->ms_condensing);
708 start = msp->ms_ops->msop_alloc(msp, size);
709 if (start != -1ULL) {
710 vdev_t *vd = msp->ms_group->mg_vd;
712 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
713 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
714 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
715 range_tree_remove(rt, start, size);
721 * ==========================================================================
722 * Common allocator routines
723 * ==========================================================================
727 * This is a helper function that can be used by the allocator to find
728 * a suitable block to allocate. This will search the specified AVL
729 * tree looking for a block that matches the specified criteria.
732 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
735 range_seg_t *rs, rsearch;
738 rsearch.rs_start = *cursor;
739 rsearch.rs_end = *cursor + size;
741 rs = avl_find(t, &rsearch, &where);
743 rs = avl_nearest(t, where, AVL_AFTER);
746 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
748 if (offset + size <= rs->rs_end) {
749 *cursor = offset + size;
752 rs = AVL_NEXT(t, rs);
756 * If we know we've searched the whole map (*cursor == 0), give up.
757 * Otherwise, reset the cursor to the beginning and try again.
763 return (metaslab_block_picker(t, cursor, size, align));
767 * ==========================================================================
768 * The first-fit block allocator
769 * ==========================================================================
772 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
775 * Find the largest power of 2 block size that evenly divides the
776 * requested size. This is used to try to allocate blocks with similar
777 * alignment from the same area of the metaslab (i.e. same cursor
778 * bucket) but it does not guarantee that other allocations sizes
779 * may exist in the same region.
781 uint64_t align = size & -size;
782 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
783 avl_tree_t *t = &msp->ms_tree->rt_root;
785 return (metaslab_block_picker(t, cursor, size, align));
790 metaslab_ff_fragmented(metaslab_t *msp)
795 static metaslab_ops_t metaslab_ff_ops = {
797 metaslab_ff_fragmented
801 * ==========================================================================
802 * Dynamic block allocator -
803 * Uses the first fit allocation scheme until space get low and then
804 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
805 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
806 * ==========================================================================
809 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
812 * Find the largest power of 2 block size that evenly divides the
813 * requested size. This is used to try to allocate blocks with similar
814 * alignment from the same area of the metaslab (i.e. same cursor
815 * bucket) but it does not guarantee that other allocations sizes
816 * may exist in the same region.
818 uint64_t align = size & -size;
819 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
820 range_tree_t *rt = msp->ms_tree;
821 avl_tree_t *t = &rt->rt_root;
822 uint64_t max_size = metaslab_block_maxsize(msp);
823 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
825 ASSERT(MUTEX_HELD(&msp->ms_lock));
826 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
832 * If we're running low on space switch to using the size
833 * sorted AVL tree (best-fit).
835 if (max_size < metaslab_df_alloc_threshold ||
836 free_pct < metaslab_df_free_pct) {
837 t = &msp->ms_size_tree;
841 return (metaslab_block_picker(t, cursor, size, 1ULL));
845 metaslab_df_fragmented(metaslab_t *msp)
847 range_tree_t *rt = msp->ms_tree;
848 uint64_t max_size = metaslab_block_maxsize(msp);
849 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
851 if (max_size >= metaslab_df_alloc_threshold &&
852 free_pct >= metaslab_df_free_pct)
858 static metaslab_ops_t metaslab_df_ops = {
860 metaslab_df_fragmented
864 * ==========================================================================
865 * Cursor fit block allocator -
866 * Select the largest region in the metaslab, set the cursor to the beginning
867 * of the range and the cursor_end to the end of the range. As allocations
868 * are made advance the cursor. Continue allocating from the cursor until
869 * the range is exhausted and then find a new range.
870 * ==========================================================================
873 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
875 range_tree_t *rt = msp->ms_tree;
876 avl_tree_t *t = &msp->ms_size_tree;
877 uint64_t *cursor = &msp->ms_lbas[0];
878 uint64_t *cursor_end = &msp->ms_lbas[1];
881 ASSERT(MUTEX_HELD(&msp->ms_lock));
882 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
884 ASSERT3U(*cursor_end, >=, *cursor);
886 if ((*cursor + size) > *cursor_end) {
889 rs = avl_last(&msp->ms_size_tree);
890 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
893 *cursor = rs->rs_start;
894 *cursor_end = rs->rs_end;
904 metaslab_cf_fragmented(metaslab_t *msp)
906 return (metaslab_block_maxsize(msp) < metaslab_min_alloc_size);
909 static metaslab_ops_t metaslab_cf_ops = {
911 metaslab_cf_fragmented
915 * ==========================================================================
916 * New dynamic fit allocator -
917 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
918 * contiguous blocks. If no region is found then just use the largest segment
920 * ==========================================================================
924 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
925 * to request from the allocator.
927 uint64_t metaslab_ndf_clump_shift = 4;
930 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
932 avl_tree_t *t = &msp->ms_tree->rt_root;
934 range_seg_t *rs, rsearch;
935 uint64_t hbit = highbit64(size);
936 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
937 uint64_t max_size = metaslab_block_maxsize(msp);
939 ASSERT(MUTEX_HELD(&msp->ms_lock));
940 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
945 rsearch.rs_start = *cursor;
946 rsearch.rs_end = *cursor + size;
948 rs = avl_find(t, &rsearch, &where);
949 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
950 t = &msp->ms_size_tree;
952 rsearch.rs_start = 0;
953 rsearch.rs_end = MIN(max_size,
954 1ULL << (hbit + metaslab_ndf_clump_shift));
955 rs = avl_find(t, &rsearch, &where);
957 rs = avl_nearest(t, where, AVL_AFTER);
961 if ((rs->rs_end - rs->rs_start) >= size) {
962 *cursor = rs->rs_start + size;
963 return (rs->rs_start);
969 metaslab_ndf_fragmented(metaslab_t *msp)
971 return (metaslab_block_maxsize(msp) <=
972 (metaslab_min_alloc_size << metaslab_ndf_clump_shift));
975 static metaslab_ops_t metaslab_ndf_ops = {
977 metaslab_ndf_fragmented
980 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
983 * ==========================================================================
985 * ==========================================================================
989 * Wait for any in-progress metaslab loads to complete.
992 metaslab_load_wait(metaslab_t *msp)
994 ASSERT(MUTEX_HELD(&msp->ms_lock));
996 while (msp->ms_loading) {
997 ASSERT(!msp->ms_loaded);
998 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1003 metaslab_load(metaslab_t *msp)
1007 ASSERT(MUTEX_HELD(&msp->ms_lock));
1008 ASSERT(!msp->ms_loaded);
1009 ASSERT(!msp->ms_loading);
1011 msp->ms_loading = B_TRUE;
1014 * If the space map has not been allocated yet, then treat
1015 * all the space in the metaslab as free and add it to the
1018 if (msp->ms_sm != NULL)
1019 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1021 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1023 msp->ms_loaded = (error == 0);
1024 msp->ms_loading = B_FALSE;
1026 if (msp->ms_loaded) {
1027 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1028 range_tree_walk(msp->ms_defertree[t],
1029 range_tree_remove, msp->ms_tree);
1032 cv_broadcast(&msp->ms_load_cv);
1037 metaslab_unload(metaslab_t *msp)
1039 ASSERT(MUTEX_HELD(&msp->ms_lock));
1040 range_tree_vacate(msp->ms_tree, NULL, NULL);
1041 msp->ms_loaded = B_FALSE;
1042 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1046 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg)
1048 vdev_t *vd = mg->mg_vd;
1049 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1052 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1053 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1054 cv_init(&msp->ms_load_cv, NULL, CV_DEFAULT, NULL);
1056 msp->ms_start = id << vd->vdev_ms_shift;
1057 msp->ms_size = 1ULL << vd->vdev_ms_shift;
1060 * We only open space map objects that already exist. All others
1061 * will be opened when we finally allocate an object for it.
1064 VERIFY0(space_map_open(&msp->ms_sm, mos, object, msp->ms_start,
1065 msp->ms_size, vd->vdev_ashift, &msp->ms_lock));
1066 ASSERT(msp->ms_sm != NULL);
1070 * We create the main range tree here, but we don't create the
1071 * alloctree and freetree until metaslab_sync_done(). This serves
1072 * two purposes: it allows metaslab_sync_done() to detect the
1073 * addition of new space; and for debugging, it ensures that we'd
1074 * data fault on any attempt to use this metaslab before it's ready.
1076 msp->ms_tree = range_tree_create(&metaslab_rt_ops, msp, &msp->ms_lock);
1077 metaslab_group_add(mg, msp);
1079 msp->ms_ops = mg->mg_class->mc_ops;
1082 * If we're opening an existing pool (txg == 0) or creating
1083 * a new one (txg == TXG_INITIAL), all space is available now.
1084 * If we're adding space to an existing pool, the new space
1085 * does not become available until after this txg has synced.
1087 if (txg <= TXG_INITIAL)
1088 metaslab_sync_done(msp, 0);
1091 * If metaslab_debug_load is set and we're initializing a metaslab
1092 * that has an allocated space_map object then load the its space
1093 * map so that can verify frees.
1095 if (metaslab_debug_load && msp->ms_sm != NULL) {
1096 mutex_enter(&msp->ms_lock);
1097 VERIFY0(metaslab_load(msp));
1098 mutex_exit(&msp->ms_lock);
1102 vdev_dirty(vd, 0, NULL, txg);
1103 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1110 metaslab_fini(metaslab_t *msp)
1112 metaslab_group_t *mg = msp->ms_group;
1114 metaslab_group_remove(mg, msp);
1116 mutex_enter(&msp->ms_lock);
1118 VERIFY(msp->ms_group == NULL);
1119 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1121 space_map_close(msp->ms_sm);
1123 metaslab_unload(msp);
1124 range_tree_destroy(msp->ms_tree);
1126 for (int t = 0; t < TXG_SIZE; t++) {
1127 range_tree_destroy(msp->ms_alloctree[t]);
1128 range_tree_destroy(msp->ms_freetree[t]);
1131 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1132 range_tree_destroy(msp->ms_defertree[t]);
1135 ASSERT0(msp->ms_deferspace);
1137 mutex_exit(&msp->ms_lock);
1138 cv_destroy(&msp->ms_load_cv);
1139 mutex_destroy(&msp->ms_lock);
1141 kmem_free(msp, sizeof (metaslab_t));
1145 * Apply a weighting factor based on the histogram information for this
1146 * metaslab. The current weighting factor is somewhat arbitrary and requires
1147 * additional investigation. The implementation provides a measure of
1148 * "weighted" free space and gives a higher weighting for larger contiguous
1149 * regions. The weighting factor is determined by counting the number of
1150 * sm_shift sectors that exist in each region represented by the histogram.
1151 * That value is then multiplied by the power of 2 exponent and the sm_shift
1154 * For example, assume the 2^21 histogram bucket has 4 2MB regions and the
1155 * metaslab has an sm_shift value of 9 (512B):
1157 * 1) calculate the number of sm_shift sectors in the region:
1158 * 2^21 / 2^9 = 2^12 = 4096 * 4 (number of regions) = 16384
1159 * 2) multiply by the power of 2 exponent and the sm_shift value:
1160 * 16384 * 21 * 9 = 3096576
1161 * This value will be added to the weighting of the metaslab.
1164 metaslab_weight_factor(metaslab_t *msp)
1166 uint64_t factor = 0;
1171 * A null space map means that the entire metaslab is free,
1172 * calculate a weight factor that spans the entire size of the
1175 if (msp->ms_sm == NULL) {
1176 vdev_t *vd = msp->ms_group->mg_vd;
1178 i = highbit64(msp->ms_size) - 1;
1179 sectors = msp->ms_size >> vd->vdev_ashift;
1180 return (sectors * i * vd->vdev_ashift);
1183 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
1186 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE(msp->ms_sm); i++) {
1187 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1191 * Determine the number of sm_shift sectors in the region
1192 * indicated by the histogram. For example, given an
1193 * sm_shift value of 9 (512 bytes) and i = 4 then we know
1194 * that we're looking at an 8K region in the histogram
1195 * (i.e. 9 + 4 = 13, 2^13 = 8192). To figure out the
1196 * number of sm_shift sectors (512 bytes in this example),
1197 * we would take 8192 / 512 = 16. Since the histogram
1198 * is offset by sm_shift we can simply use the value of
1199 * of i to calculate this (i.e. 2^i = 16 where i = 4).
1201 sectors = msp->ms_sm->sm_phys->smp_histogram[i] << i;
1202 factor += (i + msp->ms_sm->sm_shift) * sectors;
1204 return (factor * msp->ms_sm->sm_shift);
1208 metaslab_weight(metaslab_t *msp)
1210 metaslab_group_t *mg = msp->ms_group;
1211 vdev_t *vd = mg->mg_vd;
1212 uint64_t weight, space;
1214 ASSERT(MUTEX_HELD(&msp->ms_lock));
1217 * This vdev is in the process of being removed so there is nothing
1218 * for us to do here.
1220 if (vd->vdev_removing) {
1221 ASSERT0(space_map_allocated(msp->ms_sm));
1222 ASSERT0(vd->vdev_ms_shift);
1227 * The baseline weight is the metaslab's free space.
1229 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1233 * Modern disks have uniform bit density and constant angular velocity.
1234 * Therefore, the outer recording zones are faster (higher bandwidth)
1235 * than the inner zones by the ratio of outer to inner track diameter,
1236 * which is typically around 2:1. We account for this by assigning
1237 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1238 * In effect, this means that we'll select the metaslab with the most
1239 * free bandwidth rather than simply the one with the most free space.
1241 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1242 ASSERT(weight >= space && weight <= 2 * space);
1244 msp->ms_factor = metaslab_weight_factor(msp);
1245 if (metaslab_weight_factor_enable)
1246 weight += msp->ms_factor;
1248 if (msp->ms_loaded && !msp->ms_ops->msop_fragmented(msp)) {
1250 * If this metaslab is one we're actively using, adjust its
1251 * weight to make it preferable to any inactive metaslab so
1252 * we'll polish it off.
1254 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1261 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1263 ASSERT(MUTEX_HELD(&msp->ms_lock));
1265 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1266 metaslab_load_wait(msp);
1267 if (!msp->ms_loaded) {
1268 int error = metaslab_load(msp);
1270 metaslab_group_sort(msp->ms_group, msp, 0);
1275 metaslab_group_sort(msp->ms_group, msp,
1276 msp->ms_weight | activation_weight);
1278 ASSERT(msp->ms_loaded);
1279 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1285 metaslab_passivate(metaslab_t *msp, uint64_t size)
1288 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1289 * this metaslab again. In that case, it had better be empty,
1290 * or we would be leaving space on the table.
1292 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1293 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1294 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1298 metaslab_preload(void *arg)
1300 metaslab_t *msp = arg;
1301 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1303 mutex_enter(&msp->ms_lock);
1304 metaslab_load_wait(msp);
1305 if (!msp->ms_loaded)
1306 (void) metaslab_load(msp);
1309 * Set the ms_access_txg value so that we don't unload it right away.
1311 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1312 mutex_exit(&msp->ms_lock);
1316 metaslab_group_preload(metaslab_group_t *mg)
1318 spa_t *spa = mg->mg_vd->vdev_spa;
1320 avl_tree_t *t = &mg->mg_metaslab_tree;
1323 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1324 taskq_wait(mg->mg_taskq);
1327 mutex_enter(&mg->mg_lock);
1330 * Prefetch the next potential metaslabs
1332 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
1334 /* If we have reached our preload limit then we're done */
1335 if (++m > metaslab_preload_limit)
1338 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1339 msp, TQ_SLEEP) != 0);
1341 mutex_exit(&mg->mg_lock);
1345 * Determine if the space map's on-disk footprint is past our tolerance
1346 * for inefficiency. We would like to use the following criteria to make
1349 * 1. The size of the space map object should not dramatically increase as a
1350 * result of writing out the free space range tree.
1352 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1353 * times the size than the free space range tree representation
1354 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1356 * Checking the first condition is tricky since we don't want to walk
1357 * the entire AVL tree calculating the estimated on-disk size. Instead we
1358 * use the size-ordered range tree in the metaslab and calculate the
1359 * size required to write out the largest segment in our free tree. If the
1360 * size required to represent that segment on disk is larger than the space
1361 * map object then we avoid condensing this map.
1363 * To determine the second criterion we use a best-case estimate and assume
1364 * each segment can be represented on-disk as a single 64-bit entry. We refer
1365 * to this best-case estimate as the space map's minimal form.
1368 metaslab_should_condense(metaslab_t *msp)
1370 space_map_t *sm = msp->ms_sm;
1372 uint64_t size, entries, segsz;
1374 ASSERT(MUTEX_HELD(&msp->ms_lock));
1375 ASSERT(msp->ms_loaded);
1378 * Use the ms_size_tree range tree, which is ordered by size, to
1379 * obtain the largest segment in the free tree. If the tree is empty
1380 * then we should condense the map.
1382 rs = avl_last(&msp->ms_size_tree);
1387 * Calculate the number of 64-bit entries this segment would
1388 * require when written to disk. If this single segment would be
1389 * larger on-disk than the entire current on-disk structure, then
1390 * clearly condensing will increase the on-disk structure size.
1392 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1393 entries = size / (MIN(size, SM_RUN_MAX));
1394 segsz = entries * sizeof (uint64_t);
1396 return (segsz <= space_map_length(msp->ms_sm) &&
1397 space_map_length(msp->ms_sm) >= (zfs_condense_pct *
1398 sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root)) / 100);
1402 * Condense the on-disk space map representation to its minimized form.
1403 * The minimized form consists of a small number of allocations followed by
1404 * the entries of the free range tree.
1407 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1409 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1410 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1411 range_tree_t *condense_tree;
1412 space_map_t *sm = msp->ms_sm;
1414 ASSERT(MUTEX_HELD(&msp->ms_lock));
1415 ASSERT3U(spa_sync_pass(spa), ==, 1);
1416 ASSERT(msp->ms_loaded);
1418 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1419 "smp size %llu, segments %lu", txg, msp->ms_id, msp,
1420 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root));
1423 * Create an range tree that is 100% allocated. We remove segments
1424 * that have been freed in this txg, any deferred frees that exist,
1425 * and any allocation in the future. Removing segments should be
1426 * a relatively inexpensive operation since we expect these trees to
1427 * have a small number of nodes.
1429 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1430 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1433 * Remove what's been freed in this txg from the condense_tree.
1434 * Since we're in sync_pass 1, we know that all the frees from
1435 * this txg are in the freetree.
1437 range_tree_walk(freetree, range_tree_remove, condense_tree);
1439 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1440 range_tree_walk(msp->ms_defertree[t],
1441 range_tree_remove, condense_tree);
1444 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1445 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1446 range_tree_remove, condense_tree);
1450 * We're about to drop the metaslab's lock thus allowing
1451 * other consumers to change it's content. Set the
1452 * metaslab's ms_condensing flag to ensure that
1453 * allocations on this metaslab do not occur while we're
1454 * in the middle of committing it to disk. This is only critical
1455 * for the ms_tree as all other range trees use per txg
1456 * views of their content.
1458 msp->ms_condensing = B_TRUE;
1460 mutex_exit(&msp->ms_lock);
1461 space_map_truncate(sm, tx);
1462 mutex_enter(&msp->ms_lock);
1465 * While we would ideally like to create a space_map representation
1466 * that consists only of allocation records, doing so can be
1467 * prohibitively expensive because the in-core free tree can be
1468 * large, and therefore computationally expensive to subtract
1469 * from the condense_tree. Instead we sync out two trees, a cheap
1470 * allocation only tree followed by the in-core free tree. While not
1471 * optimal, this is typically close to optimal, and much cheaper to
1474 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1475 range_tree_vacate(condense_tree, NULL, NULL);
1476 range_tree_destroy(condense_tree);
1478 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1479 msp->ms_condensing = B_FALSE;
1483 * Write a metaslab to disk in the context of the specified transaction group.
1486 metaslab_sync(metaslab_t *msp, uint64_t txg)
1488 metaslab_group_t *mg = msp->ms_group;
1489 vdev_t *vd = mg->mg_vd;
1490 spa_t *spa = vd->vdev_spa;
1491 objset_t *mos = spa_meta_objset(spa);
1492 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1493 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1494 range_tree_t **freed_tree =
1495 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1497 uint64_t object = space_map_object(msp->ms_sm);
1499 ASSERT(!vd->vdev_ishole);
1502 * This metaslab has just been added so there's no work to do now.
1504 if (*freetree == NULL) {
1505 ASSERT3P(alloctree, ==, NULL);
1509 ASSERT3P(alloctree, !=, NULL);
1510 ASSERT3P(*freetree, !=, NULL);
1511 ASSERT3P(*freed_tree, !=, NULL);
1513 if (range_tree_space(alloctree) == 0 &&
1514 range_tree_space(*freetree) == 0)
1518 * The only state that can actually be changing concurrently with
1519 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1520 * be modifying this txg's alloctree, freetree, freed_tree, or
1521 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1522 * space_map ASSERTs. We drop it whenever we call into the DMU,
1523 * because the DMU can call down to us (e.g. via zio_free()) at
1527 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1529 if (msp->ms_sm == NULL) {
1530 uint64_t new_object;
1532 new_object = space_map_alloc(mos, tx);
1533 VERIFY3U(new_object, !=, 0);
1535 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1536 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1538 ASSERT(msp->ms_sm != NULL);
1541 mutex_enter(&msp->ms_lock);
1543 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1544 metaslab_should_condense(msp)) {
1545 metaslab_condense(msp, txg, tx);
1547 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1548 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1551 range_tree_vacate(alloctree, NULL, NULL);
1553 if (msp->ms_loaded) {
1555 * When the space map is loaded, we have an accruate
1556 * histogram in the range tree. This gives us an opportunity
1557 * to bring the space map's histogram up-to-date so we clear
1558 * it first before updating it.
1560 space_map_histogram_clear(msp->ms_sm);
1561 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1564 * Since the space map is not loaded we simply update the
1565 * exisiting histogram with what was freed in this txg. This
1566 * means that the on-disk histogram may not have an accurate
1567 * view of the free space but it's close enough to allow
1568 * us to make allocation decisions.
1570 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1574 * For sync pass 1, we avoid traversing this txg's free range tree
1575 * and instead will just swap the pointers for freetree and
1576 * freed_tree. We can safely do this since the freed_tree is
1577 * guaranteed to be empty on the initial pass.
1579 if (spa_sync_pass(spa) == 1) {
1580 range_tree_swap(freetree, freed_tree);
1582 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1585 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1586 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1588 mutex_exit(&msp->ms_lock);
1590 if (object != space_map_object(msp->ms_sm)) {
1591 object = space_map_object(msp->ms_sm);
1592 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1593 msp->ms_id, sizeof (uint64_t), &object, tx);
1599 * Called after a transaction group has completely synced to mark
1600 * all of the metaslab's free space as usable.
1603 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1605 metaslab_group_t *mg = msp->ms_group;
1606 vdev_t *vd = mg->mg_vd;
1607 range_tree_t **freed_tree;
1608 range_tree_t **defer_tree;
1609 int64_t alloc_delta, defer_delta;
1611 ASSERT(!vd->vdev_ishole);
1613 mutex_enter(&msp->ms_lock);
1616 * If this metaslab is just becoming available, initialize its
1617 * alloctrees, freetrees, and defertree and add its capacity to
1620 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
1621 for (int t = 0; t < TXG_SIZE; t++) {
1622 ASSERT(msp->ms_alloctree[t] == NULL);
1623 ASSERT(msp->ms_freetree[t] == NULL);
1625 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
1627 msp->ms_freetree[t] = range_tree_create(NULL, msp,
1631 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1632 ASSERT(msp->ms_defertree[t] == NULL);
1634 msp->ms_defertree[t] = range_tree_create(NULL, msp,
1638 vdev_space_update(vd, 0, 0, msp->ms_size);
1641 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1642 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
1644 alloc_delta = space_map_alloc_delta(msp->ms_sm);
1645 defer_delta = range_tree_space(*freed_tree) -
1646 range_tree_space(*defer_tree);
1648 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1650 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1651 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1654 * If there's a metaslab_load() in progress, wait for it to complete
1655 * so that we have a consistent view of the in-core space map.
1657 metaslab_load_wait(msp);
1660 * Move the frees from the defer_tree back to the free
1661 * range tree (if it's loaded). Swap the freed_tree and the
1662 * defer_tree -- this is safe to do because we've just emptied out
1665 range_tree_vacate(*defer_tree,
1666 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
1667 range_tree_swap(freed_tree, defer_tree);
1669 space_map_update(msp->ms_sm);
1671 msp->ms_deferspace += defer_delta;
1672 ASSERT3S(msp->ms_deferspace, >=, 0);
1673 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
1674 if (msp->ms_deferspace != 0) {
1676 * Keep syncing this metaslab until all deferred frees
1677 * are back in circulation.
1679 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1682 if (msp->ms_loaded && msp->ms_access_txg < txg) {
1683 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1684 VERIFY0(range_tree_space(
1685 msp->ms_alloctree[(txg + t) & TXG_MASK]));
1688 if (!metaslab_debug_unload)
1689 metaslab_unload(msp);
1692 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1693 mutex_exit(&msp->ms_lock);
1698 metaslab_sync_reassess(metaslab_group_t *mg)
1700 metaslab_group_alloc_update(mg);
1703 * Preload the next potential metaslabs
1705 metaslab_group_preload(mg);
1709 metaslab_distance(metaslab_t *msp, dva_t *dva)
1711 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1712 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1713 uint64_t start = msp->ms_id;
1715 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1716 return (1ULL << 63);
1719 return ((start - offset) << ms_shift);
1721 return ((offset - start) << ms_shift);
1726 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1727 uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
1729 spa_t *spa = mg->mg_vd->vdev_spa;
1730 metaslab_t *msp = NULL;
1731 uint64_t offset = -1ULL;
1732 avl_tree_t *t = &mg->mg_metaslab_tree;
1733 uint64_t activation_weight;
1734 uint64_t target_distance;
1737 activation_weight = METASLAB_WEIGHT_PRIMARY;
1738 for (i = 0; i < d; i++) {
1739 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1740 activation_weight = METASLAB_WEIGHT_SECONDARY;
1746 boolean_t was_active;
1748 mutex_enter(&mg->mg_lock);
1749 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1750 if (msp->ms_weight < asize) {
1751 spa_dbgmsg(spa, "%s: failed to meet weight "
1752 "requirement: vdev %llu, txg %llu, mg %p, "
1753 "msp %p, psize %llu, asize %llu, "
1754 "weight %llu", spa_name(spa),
1755 mg->mg_vd->vdev_id, txg,
1756 mg, msp, psize, asize, msp->ms_weight);
1757 mutex_exit(&mg->mg_lock);
1762 * If the selected metaslab is condensing, skip it.
1764 if (msp->ms_condensing)
1767 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1768 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1771 target_distance = min_distance +
1772 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
1775 for (i = 0; i < d; i++)
1776 if (metaslab_distance(msp, &dva[i]) <
1782 mutex_exit(&mg->mg_lock);
1786 mutex_enter(&msp->ms_lock);
1789 * Ensure that the metaslab we have selected is still
1790 * capable of handling our request. It's possible that
1791 * another thread may have changed the weight while we
1792 * were blocked on the metaslab lock.
1794 if (msp->ms_weight < asize || (was_active &&
1795 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1796 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1797 mutex_exit(&msp->ms_lock);
1801 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1802 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1803 metaslab_passivate(msp,
1804 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1805 mutex_exit(&msp->ms_lock);
1809 if (metaslab_activate(msp, activation_weight) != 0) {
1810 mutex_exit(&msp->ms_lock);
1815 * If this metaslab is currently condensing then pick again as
1816 * we can't manipulate this metaslab until it's committed
1819 if (msp->ms_condensing) {
1820 mutex_exit(&msp->ms_lock);
1824 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
1827 metaslab_passivate(msp, metaslab_block_maxsize(msp));
1828 mutex_exit(&msp->ms_lock);
1831 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
1832 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1834 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
1835 msp->ms_access_txg = txg + metaslab_unload_delay;
1837 mutex_exit(&msp->ms_lock);
1843 * Allocate a block for the specified i/o.
1846 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1847 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1849 metaslab_group_t *mg, *rotor;
1853 int zio_lock = B_FALSE;
1854 boolean_t allocatable;
1855 uint64_t offset = -1ULL;
1859 ASSERT(!DVA_IS_VALID(&dva[d]));
1862 * For testing, make some blocks above a certain size be gang blocks.
1864 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1865 return (SET_ERROR(ENOSPC));
1868 * Start at the rotor and loop through all mgs until we find something.
1869 * Note that there's no locking on mc_rotor or mc_aliquot because
1870 * nothing actually breaks if we miss a few updates -- we just won't
1871 * allocate quite as evenly. It all balances out over time.
1873 * If we are doing ditto or log blocks, try to spread them across
1874 * consecutive vdevs. If we're forced to reuse a vdev before we've
1875 * allocated all of our ditto blocks, then try and spread them out on
1876 * that vdev as much as possible. If it turns out to not be possible,
1877 * gradually lower our standards until anything becomes acceptable.
1878 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1879 * gives us hope of containing our fault domains to something we're
1880 * able to reason about. Otherwise, any two top-level vdev failures
1881 * will guarantee the loss of data. With consecutive allocation,
1882 * only two adjacent top-level vdev failures will result in data loss.
1884 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1885 * ourselves on the same vdev as our gang block header. That
1886 * way, we can hope for locality in vdev_cache, plus it makes our
1887 * fault domains something tractable.
1890 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1893 * It's possible the vdev we're using as the hint no
1894 * longer exists (i.e. removed). Consult the rotor when
1900 if (flags & METASLAB_HINTBP_AVOID &&
1901 mg->mg_next != NULL)
1906 } else if (d != 0) {
1907 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1908 mg = vd->vdev_mg->mg_next;
1914 * If the hint put us into the wrong metaslab class, or into a
1915 * metaslab group that has been passivated, just follow the rotor.
1917 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1924 ASSERT(mg->mg_activation_count == 1);
1929 * Don't allocate from faulted devices.
1932 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1933 allocatable = vdev_allocatable(vd);
1934 spa_config_exit(spa, SCL_ZIO, FTAG);
1936 allocatable = vdev_allocatable(vd);
1940 * Determine if the selected metaslab group is eligible
1941 * for allocations. If we're ganging or have requested
1942 * an allocation for the smallest gang block size
1943 * then we don't want to avoid allocating to the this
1944 * metaslab group. If we're in this condition we should
1945 * try to allocate from any device possible so that we
1946 * don't inadvertently return ENOSPC and suspend the pool
1947 * even though space is still available.
1949 if (allocatable && CAN_FASTGANG(flags) &&
1950 psize > SPA_GANGBLOCKSIZE)
1951 allocatable = metaslab_group_allocatable(mg);
1957 * Avoid writing single-copy data to a failing vdev
1958 * unless the user instructs us that it is okay.
1960 if ((vd->vdev_stat.vs_write_errors > 0 ||
1961 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1962 d == 0 && dshift == 3 &&
1963 !(zfs_write_to_degraded && vd->vdev_state ==
1964 VDEV_STATE_DEGRADED)) {
1969 ASSERT(mg->mg_class == mc);
1971 distance = vd->vdev_asize >> dshift;
1972 if (distance <= (1ULL << vd->vdev_ms_shift))
1977 asize = vdev_psize_to_asize(vd, psize);
1978 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1980 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1982 if (offset != -1ULL) {
1984 * If we've just selected this metaslab group,
1985 * figure out whether the corresponding vdev is
1986 * over- or under-used relative to the pool,
1987 * and set an allocation bias to even it out.
1989 if (mc->mc_aliquot == 0) {
1990 vdev_stat_t *vs = &vd->vdev_stat;
1993 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1994 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1997 * Calculate how much more or less we should
1998 * try to allocate from this device during
1999 * this iteration around the rotor.
2000 * For example, if a device is 80% full
2001 * and the pool is 20% full then we should
2002 * reduce allocations by 60% on this device.
2004 * mg_bias = (20 - 80) * 512K / 100 = -307K
2006 * This reduces allocations by 307K for this
2009 mg->mg_bias = ((cu - vu) *
2010 (int64_t)mg->mg_aliquot) / 100;
2013 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2014 mg->mg_aliquot + mg->mg_bias) {
2015 mc->mc_rotor = mg->mg_next;
2019 DVA_SET_VDEV(&dva[d], vd->vdev_id);
2020 DVA_SET_OFFSET(&dva[d], offset);
2021 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2022 DVA_SET_ASIZE(&dva[d], asize);
2027 mc->mc_rotor = mg->mg_next;
2029 } while ((mg = mg->mg_next) != rotor);
2033 ASSERT(dshift < 64);
2037 if (!allocatable && !zio_lock) {
2043 bzero(&dva[d], sizeof (dva_t));
2045 return (SET_ERROR(ENOSPC));
2049 * Free the block represented by DVA in the context of the specified
2050 * transaction group.
2053 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2055 uint64_t vdev = DVA_GET_VDEV(dva);
2056 uint64_t offset = DVA_GET_OFFSET(dva);
2057 uint64_t size = DVA_GET_ASIZE(dva);
2061 ASSERT(DVA_IS_VALID(dva));
2063 if (txg > spa_freeze_txg(spa))
2066 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2067 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2068 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2069 (u_longlong_t)vdev, (u_longlong_t)offset);
2074 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2076 if (DVA_GET_GANG(dva))
2077 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2079 mutex_enter(&msp->ms_lock);
2082 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2085 VERIFY(!msp->ms_condensing);
2086 VERIFY3U(offset, >=, msp->ms_start);
2087 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2088 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2090 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2091 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2092 range_tree_add(msp->ms_tree, offset, size);
2094 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2095 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2096 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2100 mutex_exit(&msp->ms_lock);
2104 * Intent log support: upon opening the pool after a crash, notify the SPA
2105 * of blocks that the intent log has allocated for immediate write, but
2106 * which are still considered free by the SPA because the last transaction
2107 * group didn't commit yet.
2110 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2112 uint64_t vdev = DVA_GET_VDEV(dva);
2113 uint64_t offset = DVA_GET_OFFSET(dva);
2114 uint64_t size = DVA_GET_ASIZE(dva);
2119 ASSERT(DVA_IS_VALID(dva));
2121 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2122 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2123 return (SET_ERROR(ENXIO));
2125 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2127 if (DVA_GET_GANG(dva))
2128 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2130 mutex_enter(&msp->ms_lock);
2132 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2133 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2135 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2136 error = SET_ERROR(ENOENT);
2138 if (error || txg == 0) { /* txg == 0 indicates dry run */
2139 mutex_exit(&msp->ms_lock);
2143 VERIFY(!msp->ms_condensing);
2144 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2145 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2146 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2147 range_tree_remove(msp->ms_tree, offset, size);
2149 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2150 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2151 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2152 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2155 mutex_exit(&msp->ms_lock);
2161 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2162 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2164 dva_t *dva = bp->blk_dva;
2165 dva_t *hintdva = hintbp->blk_dva;
2168 ASSERT(bp->blk_birth == 0);
2169 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2171 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2173 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2174 spa_config_exit(spa, SCL_ALLOC, FTAG);
2175 return (SET_ERROR(ENOSPC));
2178 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2179 ASSERT(BP_GET_NDVAS(bp) == 0);
2180 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2182 for (int d = 0; d < ndvas; d++) {
2183 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2186 for (d--; d >= 0; d--) {
2187 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2188 bzero(&dva[d], sizeof (dva_t));
2190 spa_config_exit(spa, SCL_ALLOC, FTAG);
2195 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2197 spa_config_exit(spa, SCL_ALLOC, FTAG);
2199 BP_SET_BIRTH(bp, txg, txg);
2205 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2207 const dva_t *dva = bp->blk_dva;
2208 int ndvas = BP_GET_NDVAS(bp);
2210 ASSERT(!BP_IS_HOLE(bp));
2211 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2213 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2215 for (int d = 0; d < ndvas; d++)
2216 metaslab_free_dva(spa, &dva[d], txg, now);
2218 spa_config_exit(spa, SCL_FREE, FTAG);
2222 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2224 const dva_t *dva = bp->blk_dva;
2225 int ndvas = BP_GET_NDVAS(bp);
2228 ASSERT(!BP_IS_HOLE(bp));
2232 * First do a dry run to make sure all DVAs are claimable,
2233 * so we don't have to unwind from partial failures below.
2235 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2239 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2241 for (int d = 0; d < ndvas; d++)
2242 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2245 spa_config_exit(spa, SCL_ALLOC, FTAG);
2247 ASSERT(error == 0 || txg == 0);
2253 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2255 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2258 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2259 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2260 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2261 vdev_t *vd = vdev_lookup_top(spa, vdev);
2262 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2263 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2264 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2267 range_tree_verify(msp->ms_tree, offset, size);
2269 for (int j = 0; j < TXG_SIZE; j++)
2270 range_tree_verify(msp->ms_freetree[j], offset, size);
2271 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2272 range_tree_verify(msp->ms_defertree[j], offset, size);
2274 spa_config_exit(spa, SCL_VDEV, FTAG);