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_RWTUN,
175 &zfs_write_to_degraded, 0, "Allow writing data to degraded vdevs");
176 TUNABLE_INT("vfs.zfs.write_to_degraded", &zfs_write_to_degraded);
179 * Max number of metaslabs per group to preload.
181 int metaslab_preload_limit = SPA_DVAS_PER_BP;
182 TUNABLE_INT("vfs.zfs.metaslab.preload_limit", &metaslab_preload_limit);
183 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_limit, CTLFLAG_RWTUN,
184 &metaslab_preload_limit, 0,
185 "Max number of metaslabs per group to preload");
188 * Enable/disable preloading of metaslab.
190 boolean_t metaslab_preload_enabled = B_TRUE;
191 TUNABLE_INT("vfs.zfs.metaslab.preload_enabled", &metaslab_preload_enabled);
192 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, preload_enabled, CTLFLAG_RWTUN,
193 &metaslab_preload_enabled, 0,
194 "Max number of metaslabs per group to preload");
197 * Enable/disable additional weight factor for each metaslab.
199 boolean_t metaslab_weight_factor_enable = B_FALSE;
200 TUNABLE_INT("vfs.zfs.metaslab.weight_factor_enable",
201 &metaslab_weight_factor_enable);
202 SYSCTL_INT(_vfs_zfs_metaslab, OID_AUTO, weight_factor_enable, CTLFLAG_RWTUN,
203 &metaslab_weight_factor_enable, 0,
204 "Enable additional weight factor for each metaslab");
208 * ==========================================================================
210 * ==========================================================================
213 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
215 metaslab_class_t *mc;
217 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
227 metaslab_class_destroy(metaslab_class_t *mc)
229 ASSERT(mc->mc_rotor == NULL);
230 ASSERT(mc->mc_alloc == 0);
231 ASSERT(mc->mc_deferred == 0);
232 ASSERT(mc->mc_space == 0);
233 ASSERT(mc->mc_dspace == 0);
235 kmem_free(mc, sizeof (metaslab_class_t));
239 metaslab_class_validate(metaslab_class_t *mc)
241 metaslab_group_t *mg;
245 * Must hold one of the spa_config locks.
247 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
248 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
250 if ((mg = mc->mc_rotor) == NULL)
255 ASSERT(vd->vdev_mg != NULL);
256 ASSERT3P(vd->vdev_top, ==, vd);
257 ASSERT3P(mg->mg_class, ==, mc);
258 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
259 } while ((mg = mg->mg_next) != mc->mc_rotor);
265 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
266 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
268 atomic_add_64(&mc->mc_alloc, alloc_delta);
269 atomic_add_64(&mc->mc_deferred, defer_delta);
270 atomic_add_64(&mc->mc_space, space_delta);
271 atomic_add_64(&mc->mc_dspace, dspace_delta);
275 metaslab_class_minblocksize_update(metaslab_class_t *mc)
277 metaslab_group_t *mg;
279 uint64_t minashift = UINT64_MAX;
281 if ((mg = mc->mc_rotor) == NULL) {
282 mc->mc_minblocksize = SPA_MINBLOCKSIZE;
288 if (vd->vdev_ashift < minashift)
289 minashift = vd->vdev_ashift;
290 } while ((mg = mg->mg_next) != mc->mc_rotor);
292 mc->mc_minblocksize = 1ULL << minashift;
296 metaslab_class_get_alloc(metaslab_class_t *mc)
298 return (mc->mc_alloc);
302 metaslab_class_get_deferred(metaslab_class_t *mc)
304 return (mc->mc_deferred);
308 metaslab_class_get_space(metaslab_class_t *mc)
310 return (mc->mc_space);
314 metaslab_class_get_dspace(metaslab_class_t *mc)
316 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
320 metaslab_class_get_minblocksize(metaslab_class_t *mc)
322 return (mc->mc_minblocksize);
326 * ==========================================================================
328 * ==========================================================================
331 metaslab_compare(const void *x1, const void *x2)
333 const metaslab_t *m1 = x1;
334 const metaslab_t *m2 = x2;
336 if (m1->ms_weight < m2->ms_weight)
338 if (m1->ms_weight > m2->ms_weight)
342 * If the weights are identical, use the offset to force uniqueness.
344 if (m1->ms_start < m2->ms_start)
346 if (m1->ms_start > m2->ms_start)
349 ASSERT3P(m1, ==, m2);
355 * Update the allocatable flag and the metaslab group's capacity.
356 * The allocatable flag is set to true if the capacity is below
357 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
358 * from allocatable to non-allocatable or vice versa then the metaslab
359 * group's class is updated to reflect the transition.
362 metaslab_group_alloc_update(metaslab_group_t *mg)
364 vdev_t *vd = mg->mg_vd;
365 metaslab_class_t *mc = mg->mg_class;
366 vdev_stat_t *vs = &vd->vdev_stat;
367 boolean_t was_allocatable;
369 ASSERT(vd == vd->vdev_top);
371 mutex_enter(&mg->mg_lock);
372 was_allocatable = mg->mg_allocatable;
374 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
377 mg->mg_allocatable = (mg->mg_free_capacity > zfs_mg_noalloc_threshold);
380 * The mc_alloc_groups maintains a count of the number of
381 * groups in this metaslab class that are still above the
382 * zfs_mg_noalloc_threshold. This is used by the allocating
383 * threads to determine if they should avoid allocations to
384 * a given group. The allocator will avoid allocations to a group
385 * if that group has reached or is below the zfs_mg_noalloc_threshold
386 * and there are still other groups that are above the threshold.
387 * When a group transitions from allocatable to non-allocatable or
388 * vice versa we update the metaslab class to reflect that change.
389 * When the mc_alloc_groups value drops to 0 that means that all
390 * groups have reached the zfs_mg_noalloc_threshold making all groups
391 * eligible for allocations. This effectively means that all devices
392 * are balanced again.
394 if (was_allocatable && !mg->mg_allocatable)
395 mc->mc_alloc_groups--;
396 else if (!was_allocatable && mg->mg_allocatable)
397 mc->mc_alloc_groups++;
398 mutex_exit(&mg->mg_lock);
402 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
404 metaslab_group_t *mg;
406 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
407 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
408 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
409 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
412 mg->mg_activation_count = 0;
414 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
415 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
421 metaslab_group_destroy(metaslab_group_t *mg)
423 ASSERT(mg->mg_prev == NULL);
424 ASSERT(mg->mg_next == NULL);
426 * We may have gone below zero with the activation count
427 * either because we never activated in the first place or
428 * because we're done, and possibly removing the vdev.
430 ASSERT(mg->mg_activation_count <= 0);
432 taskq_destroy(mg->mg_taskq);
433 avl_destroy(&mg->mg_metaslab_tree);
434 mutex_destroy(&mg->mg_lock);
435 kmem_free(mg, sizeof (metaslab_group_t));
439 metaslab_group_activate(metaslab_group_t *mg)
441 metaslab_class_t *mc = mg->mg_class;
442 metaslab_group_t *mgprev, *mgnext;
444 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
446 ASSERT(mc->mc_rotor != mg);
447 ASSERT(mg->mg_prev == NULL);
448 ASSERT(mg->mg_next == NULL);
449 ASSERT(mg->mg_activation_count <= 0);
451 if (++mg->mg_activation_count <= 0)
454 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
455 metaslab_group_alloc_update(mg);
457 if ((mgprev = mc->mc_rotor) == NULL) {
461 mgnext = mgprev->mg_next;
462 mg->mg_prev = mgprev;
463 mg->mg_next = mgnext;
464 mgprev->mg_next = mg;
465 mgnext->mg_prev = mg;
468 metaslab_class_minblocksize_update(mc);
472 metaslab_group_passivate(metaslab_group_t *mg)
474 metaslab_class_t *mc = mg->mg_class;
475 metaslab_group_t *mgprev, *mgnext;
477 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
479 if (--mg->mg_activation_count != 0) {
480 ASSERT(mc->mc_rotor != mg);
481 ASSERT(mg->mg_prev == NULL);
482 ASSERT(mg->mg_next == NULL);
483 ASSERT(mg->mg_activation_count < 0);
487 taskq_wait(mg->mg_taskq);
489 mgprev = mg->mg_prev;
490 mgnext = mg->mg_next;
495 mc->mc_rotor = mgnext;
496 mgprev->mg_next = mgnext;
497 mgnext->mg_prev = mgprev;
502 metaslab_class_minblocksize_update(mc);
506 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
508 mutex_enter(&mg->mg_lock);
509 ASSERT(msp->ms_group == NULL);
512 avl_add(&mg->mg_metaslab_tree, msp);
513 mutex_exit(&mg->mg_lock);
517 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
519 mutex_enter(&mg->mg_lock);
520 ASSERT(msp->ms_group == mg);
521 avl_remove(&mg->mg_metaslab_tree, msp);
522 msp->ms_group = NULL;
523 mutex_exit(&mg->mg_lock);
527 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
530 * Although in principle the weight can be any value, in
531 * practice we do not use values in the range [1, 510].
533 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
534 ASSERT(MUTEX_HELD(&msp->ms_lock));
536 mutex_enter(&mg->mg_lock);
537 ASSERT(msp->ms_group == mg);
538 avl_remove(&mg->mg_metaslab_tree, msp);
539 msp->ms_weight = weight;
540 avl_add(&mg->mg_metaslab_tree, msp);
541 mutex_exit(&mg->mg_lock);
545 * Determine if a given metaslab group should skip allocations. A metaslab
546 * group should avoid allocations if its used capacity has crossed the
547 * zfs_mg_noalloc_threshold and there is at least one metaslab group
548 * that can still handle allocations.
551 metaslab_group_allocatable(metaslab_group_t *mg)
553 vdev_t *vd = mg->mg_vd;
554 spa_t *spa = vd->vdev_spa;
555 metaslab_class_t *mc = mg->mg_class;
558 * A metaslab group is considered allocatable if its free capacity
559 * is greater than the set value of zfs_mg_noalloc_threshold, it's
560 * associated with a slog, or there are no other metaslab groups
561 * with free capacity greater than zfs_mg_noalloc_threshold.
563 return (mg->mg_free_capacity > zfs_mg_noalloc_threshold ||
564 mc != spa_normal_class(spa) || mc->mc_alloc_groups == 0);
568 * ==========================================================================
569 * Range tree callbacks
570 * ==========================================================================
574 * Comparison function for the private size-ordered tree. Tree is sorted
575 * by size, larger sizes at the end of the tree.
578 metaslab_rangesize_compare(const void *x1, const void *x2)
580 const range_seg_t *r1 = x1;
581 const range_seg_t *r2 = x2;
582 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
583 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
585 if (rs_size1 < rs_size2)
587 if (rs_size1 > rs_size2)
590 if (r1->rs_start < r2->rs_start)
593 if (r1->rs_start > r2->rs_start)
600 * Create any block allocator specific components. The current allocators
601 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
604 metaslab_rt_create(range_tree_t *rt, void *arg)
606 metaslab_t *msp = arg;
608 ASSERT3P(rt->rt_arg, ==, msp);
609 ASSERT(msp->ms_tree == NULL);
611 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
612 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
616 * Destroy the block allocator specific components.
619 metaslab_rt_destroy(range_tree_t *rt, void *arg)
621 metaslab_t *msp = arg;
623 ASSERT3P(rt->rt_arg, ==, msp);
624 ASSERT3P(msp->ms_tree, ==, rt);
625 ASSERT0(avl_numnodes(&msp->ms_size_tree));
627 avl_destroy(&msp->ms_size_tree);
631 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
633 metaslab_t *msp = arg;
635 ASSERT3P(rt->rt_arg, ==, msp);
636 ASSERT3P(msp->ms_tree, ==, rt);
637 VERIFY(!msp->ms_condensing);
638 avl_add(&msp->ms_size_tree, rs);
642 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
644 metaslab_t *msp = arg;
646 ASSERT3P(rt->rt_arg, ==, msp);
647 ASSERT3P(msp->ms_tree, ==, rt);
648 VERIFY(!msp->ms_condensing);
649 avl_remove(&msp->ms_size_tree, rs);
653 metaslab_rt_vacate(range_tree_t *rt, void *arg)
655 metaslab_t *msp = arg;
657 ASSERT3P(rt->rt_arg, ==, msp);
658 ASSERT3P(msp->ms_tree, ==, rt);
661 * Normally one would walk the tree freeing nodes along the way.
662 * Since the nodes are shared with the range trees we can avoid
663 * walking all nodes and just reinitialize the avl tree. The nodes
664 * will be freed by the range tree, so we don't want to free them here.
666 avl_create(&msp->ms_size_tree, metaslab_rangesize_compare,
667 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
670 static range_tree_ops_t metaslab_rt_ops = {
679 * ==========================================================================
680 * Metaslab block operations
681 * ==========================================================================
685 * Return the maximum contiguous segment within the metaslab.
688 metaslab_block_maxsize(metaslab_t *msp)
690 avl_tree_t *t = &msp->ms_size_tree;
693 if (t == NULL || (rs = avl_last(t)) == NULL)
696 return (rs->rs_end - rs->rs_start);
700 metaslab_block_alloc(metaslab_t *msp, uint64_t size)
703 range_tree_t *rt = msp->ms_tree;
705 VERIFY(!msp->ms_condensing);
707 start = msp->ms_ops->msop_alloc(msp, size);
708 if (start != -1ULL) {
709 vdev_t *vd = msp->ms_group->mg_vd;
711 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
712 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
713 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
714 range_tree_remove(rt, start, size);
720 * ==========================================================================
721 * Common allocator routines
722 * ==========================================================================
726 * This is a helper function that can be used by the allocator to find
727 * a suitable block to allocate. This will search the specified AVL
728 * tree looking for a block that matches the specified criteria.
731 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
734 range_seg_t *rs, rsearch;
737 rsearch.rs_start = *cursor;
738 rsearch.rs_end = *cursor + size;
740 rs = avl_find(t, &rsearch, &where);
742 rs = avl_nearest(t, where, AVL_AFTER);
745 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
747 if (offset + size <= rs->rs_end) {
748 *cursor = offset + size;
751 rs = AVL_NEXT(t, rs);
755 * If we know we've searched the whole map (*cursor == 0), give up.
756 * Otherwise, reset the cursor to the beginning and try again.
762 return (metaslab_block_picker(t, cursor, size, align));
766 * ==========================================================================
767 * The first-fit block allocator
768 * ==========================================================================
771 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
774 * Find the largest power of 2 block size that evenly divides the
775 * requested size. This is used to try to allocate blocks with similar
776 * alignment from the same area of the metaslab (i.e. same cursor
777 * bucket) but it does not guarantee that other allocations sizes
778 * may exist in the same region.
780 uint64_t align = size & -size;
781 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
782 avl_tree_t *t = &msp->ms_tree->rt_root;
784 return (metaslab_block_picker(t, cursor, size, align));
789 metaslab_ff_fragmented(metaslab_t *msp)
794 static metaslab_ops_t metaslab_ff_ops = {
796 metaslab_ff_fragmented
800 * ==========================================================================
801 * Dynamic block allocator -
802 * Uses the first fit allocation scheme until space get low and then
803 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
804 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
805 * ==========================================================================
808 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
811 * Find the largest power of 2 block size that evenly divides the
812 * requested size. This is used to try to allocate blocks with similar
813 * alignment from the same area of the metaslab (i.e. same cursor
814 * bucket) but it does not guarantee that other allocations sizes
815 * may exist in the same region.
817 uint64_t align = size & -size;
818 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
819 range_tree_t *rt = msp->ms_tree;
820 avl_tree_t *t = &rt->rt_root;
821 uint64_t max_size = metaslab_block_maxsize(msp);
822 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
824 ASSERT(MUTEX_HELD(&msp->ms_lock));
825 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
831 * If we're running low on space switch to using the size
832 * sorted AVL tree (best-fit).
834 if (max_size < metaslab_df_alloc_threshold ||
835 free_pct < metaslab_df_free_pct) {
836 t = &msp->ms_size_tree;
840 return (metaslab_block_picker(t, cursor, size, 1ULL));
844 metaslab_df_fragmented(metaslab_t *msp)
846 range_tree_t *rt = msp->ms_tree;
847 uint64_t max_size = metaslab_block_maxsize(msp);
848 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
850 if (max_size >= metaslab_df_alloc_threshold &&
851 free_pct >= metaslab_df_free_pct)
857 static metaslab_ops_t metaslab_df_ops = {
859 metaslab_df_fragmented
863 * ==========================================================================
864 * Cursor fit block allocator -
865 * Select the largest region in the metaslab, set the cursor to the beginning
866 * of the range and the cursor_end to the end of the range. As allocations
867 * are made advance the cursor. Continue allocating from the cursor until
868 * the range is exhausted and then find a new range.
869 * ==========================================================================
872 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
874 range_tree_t *rt = msp->ms_tree;
875 avl_tree_t *t = &msp->ms_size_tree;
876 uint64_t *cursor = &msp->ms_lbas[0];
877 uint64_t *cursor_end = &msp->ms_lbas[1];
880 ASSERT(MUTEX_HELD(&msp->ms_lock));
881 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
883 ASSERT3U(*cursor_end, >=, *cursor);
885 if ((*cursor + size) > *cursor_end) {
888 rs = avl_last(&msp->ms_size_tree);
889 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
892 *cursor = rs->rs_start;
893 *cursor_end = rs->rs_end;
903 metaslab_cf_fragmented(metaslab_t *msp)
905 return (metaslab_block_maxsize(msp) < metaslab_min_alloc_size);
908 static metaslab_ops_t metaslab_cf_ops = {
910 metaslab_cf_fragmented
914 * ==========================================================================
915 * New dynamic fit allocator -
916 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
917 * contiguous blocks. If no region is found then just use the largest segment
919 * ==========================================================================
923 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
924 * to request from the allocator.
926 uint64_t metaslab_ndf_clump_shift = 4;
929 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
931 avl_tree_t *t = &msp->ms_tree->rt_root;
933 range_seg_t *rs, rsearch;
934 uint64_t hbit = highbit64(size);
935 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
936 uint64_t max_size = metaslab_block_maxsize(msp);
938 ASSERT(MUTEX_HELD(&msp->ms_lock));
939 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
944 rsearch.rs_start = *cursor;
945 rsearch.rs_end = *cursor + size;
947 rs = avl_find(t, &rsearch, &where);
948 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
949 t = &msp->ms_size_tree;
951 rsearch.rs_start = 0;
952 rsearch.rs_end = MIN(max_size,
953 1ULL << (hbit + metaslab_ndf_clump_shift));
954 rs = avl_find(t, &rsearch, &where);
956 rs = avl_nearest(t, where, AVL_AFTER);
960 if ((rs->rs_end - rs->rs_start) >= size) {
961 *cursor = rs->rs_start + size;
962 return (rs->rs_start);
968 metaslab_ndf_fragmented(metaslab_t *msp)
970 return (metaslab_block_maxsize(msp) <=
971 (metaslab_min_alloc_size << metaslab_ndf_clump_shift));
974 static metaslab_ops_t metaslab_ndf_ops = {
976 metaslab_ndf_fragmented
979 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
982 * ==========================================================================
984 * ==========================================================================
988 * Wait for any in-progress metaslab loads to complete.
991 metaslab_load_wait(metaslab_t *msp)
993 ASSERT(MUTEX_HELD(&msp->ms_lock));
995 while (msp->ms_loading) {
996 ASSERT(!msp->ms_loaded);
997 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1002 metaslab_load(metaslab_t *msp)
1006 ASSERT(MUTEX_HELD(&msp->ms_lock));
1007 ASSERT(!msp->ms_loaded);
1008 ASSERT(!msp->ms_loading);
1010 msp->ms_loading = B_TRUE;
1013 * If the space map has not been allocated yet, then treat
1014 * all the space in the metaslab as free and add it to the
1017 if (msp->ms_sm != NULL)
1018 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1020 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1022 msp->ms_loaded = (error == 0);
1023 msp->ms_loading = B_FALSE;
1025 if (msp->ms_loaded) {
1026 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1027 range_tree_walk(msp->ms_defertree[t],
1028 range_tree_remove, msp->ms_tree);
1031 cv_broadcast(&msp->ms_load_cv);
1036 metaslab_unload(metaslab_t *msp)
1038 ASSERT(MUTEX_HELD(&msp->ms_lock));
1039 range_tree_vacate(msp->ms_tree, NULL, NULL);
1040 msp->ms_loaded = B_FALSE;
1041 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1045 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg)
1047 vdev_t *vd = mg->mg_vd;
1048 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1051 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1052 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1053 cv_init(&msp->ms_load_cv, NULL, CV_DEFAULT, NULL);
1055 msp->ms_start = id << vd->vdev_ms_shift;
1056 msp->ms_size = 1ULL << vd->vdev_ms_shift;
1059 * We only open space map objects that already exist. All others
1060 * will be opened when we finally allocate an object for it.
1063 VERIFY0(space_map_open(&msp->ms_sm, mos, object, msp->ms_start,
1064 msp->ms_size, vd->vdev_ashift, &msp->ms_lock));
1065 ASSERT(msp->ms_sm != NULL);
1069 * We create the main range tree here, but we don't create the
1070 * alloctree and freetree until metaslab_sync_done(). This serves
1071 * two purposes: it allows metaslab_sync_done() to detect the
1072 * addition of new space; and for debugging, it ensures that we'd
1073 * data fault on any attempt to use this metaslab before it's ready.
1075 msp->ms_tree = range_tree_create(&metaslab_rt_ops, msp, &msp->ms_lock);
1076 metaslab_group_add(mg, msp);
1078 msp->ms_ops = mg->mg_class->mc_ops;
1081 * If we're opening an existing pool (txg == 0) or creating
1082 * a new one (txg == TXG_INITIAL), all space is available now.
1083 * If we're adding space to an existing pool, the new space
1084 * does not become available until after this txg has synced.
1086 if (txg <= TXG_INITIAL)
1087 metaslab_sync_done(msp, 0);
1090 * If metaslab_debug_load is set and we're initializing a metaslab
1091 * that has an allocated space_map object then load the its space
1092 * map so that can verify frees.
1094 if (metaslab_debug_load && msp->ms_sm != NULL) {
1095 mutex_enter(&msp->ms_lock);
1096 VERIFY0(metaslab_load(msp));
1097 mutex_exit(&msp->ms_lock);
1101 vdev_dirty(vd, 0, NULL, txg);
1102 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1109 metaslab_fini(metaslab_t *msp)
1111 metaslab_group_t *mg = msp->ms_group;
1113 metaslab_group_remove(mg, msp);
1115 mutex_enter(&msp->ms_lock);
1117 VERIFY(msp->ms_group == NULL);
1118 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1120 space_map_close(msp->ms_sm);
1122 metaslab_unload(msp);
1123 range_tree_destroy(msp->ms_tree);
1125 for (int t = 0; t < TXG_SIZE; t++) {
1126 range_tree_destroy(msp->ms_alloctree[t]);
1127 range_tree_destroy(msp->ms_freetree[t]);
1130 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1131 range_tree_destroy(msp->ms_defertree[t]);
1134 ASSERT0(msp->ms_deferspace);
1136 mutex_exit(&msp->ms_lock);
1137 cv_destroy(&msp->ms_load_cv);
1138 mutex_destroy(&msp->ms_lock);
1140 kmem_free(msp, sizeof (metaslab_t));
1144 * Apply a weighting factor based on the histogram information for this
1145 * metaslab. The current weighting factor is somewhat arbitrary and requires
1146 * additional investigation. The implementation provides a measure of
1147 * "weighted" free space and gives a higher weighting for larger contiguous
1148 * regions. The weighting factor is determined by counting the number of
1149 * sm_shift sectors that exist in each region represented by the histogram.
1150 * That value is then multiplied by the power of 2 exponent and the sm_shift
1153 * For example, assume the 2^21 histogram bucket has 4 2MB regions and the
1154 * metaslab has an sm_shift value of 9 (512B):
1156 * 1) calculate the number of sm_shift sectors in the region:
1157 * 2^21 / 2^9 = 2^12 = 4096 * 4 (number of regions) = 16384
1158 * 2) multiply by the power of 2 exponent and the sm_shift value:
1159 * 16384 * 21 * 9 = 3096576
1160 * This value will be added to the weighting of the metaslab.
1163 metaslab_weight_factor(metaslab_t *msp)
1165 uint64_t factor = 0;
1170 * A null space map means that the entire metaslab is free,
1171 * calculate a weight factor that spans the entire size of the
1174 if (msp->ms_sm == NULL) {
1175 vdev_t *vd = msp->ms_group->mg_vd;
1177 i = highbit64(msp->ms_size) - 1;
1178 sectors = msp->ms_size >> vd->vdev_ashift;
1179 return (sectors * i * vd->vdev_ashift);
1182 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
1185 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE(msp->ms_sm); i++) {
1186 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1190 * Determine the number of sm_shift sectors in the region
1191 * indicated by the histogram. For example, given an
1192 * sm_shift value of 9 (512 bytes) and i = 4 then we know
1193 * that we're looking at an 8K region in the histogram
1194 * (i.e. 9 + 4 = 13, 2^13 = 8192). To figure out the
1195 * number of sm_shift sectors (512 bytes in this example),
1196 * we would take 8192 / 512 = 16. Since the histogram
1197 * is offset by sm_shift we can simply use the value of
1198 * of i to calculate this (i.e. 2^i = 16 where i = 4).
1200 sectors = msp->ms_sm->sm_phys->smp_histogram[i] << i;
1201 factor += (i + msp->ms_sm->sm_shift) * sectors;
1203 return (factor * msp->ms_sm->sm_shift);
1207 metaslab_weight(metaslab_t *msp)
1209 metaslab_group_t *mg = msp->ms_group;
1210 vdev_t *vd = mg->mg_vd;
1211 uint64_t weight, space;
1213 ASSERT(MUTEX_HELD(&msp->ms_lock));
1216 * This vdev is in the process of being removed so there is nothing
1217 * for us to do here.
1219 if (vd->vdev_removing) {
1220 ASSERT0(space_map_allocated(msp->ms_sm));
1221 ASSERT0(vd->vdev_ms_shift);
1226 * The baseline weight is the metaslab's free space.
1228 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1232 * Modern disks have uniform bit density and constant angular velocity.
1233 * Therefore, the outer recording zones are faster (higher bandwidth)
1234 * than the inner zones by the ratio of outer to inner track diameter,
1235 * which is typically around 2:1. We account for this by assigning
1236 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1237 * In effect, this means that we'll select the metaslab with the most
1238 * free bandwidth rather than simply the one with the most free space.
1240 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1241 ASSERT(weight >= space && weight <= 2 * space);
1243 msp->ms_factor = metaslab_weight_factor(msp);
1244 if (metaslab_weight_factor_enable)
1245 weight += msp->ms_factor;
1247 if (msp->ms_loaded && !msp->ms_ops->msop_fragmented(msp)) {
1249 * If this metaslab is one we're actively using, adjust its
1250 * weight to make it preferable to any inactive metaslab so
1251 * we'll polish it off.
1253 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1260 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
1262 ASSERT(MUTEX_HELD(&msp->ms_lock));
1264 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1265 metaslab_load_wait(msp);
1266 if (!msp->ms_loaded) {
1267 int error = metaslab_load(msp);
1269 metaslab_group_sort(msp->ms_group, msp, 0);
1274 metaslab_group_sort(msp->ms_group, msp,
1275 msp->ms_weight | activation_weight);
1277 ASSERT(msp->ms_loaded);
1278 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
1284 metaslab_passivate(metaslab_t *msp, uint64_t size)
1287 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1288 * this metaslab again. In that case, it had better be empty,
1289 * or we would be leaving space on the table.
1291 ASSERT(size >= SPA_MINBLOCKSIZE || range_tree_space(msp->ms_tree) == 0);
1292 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
1293 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
1297 metaslab_preload(void *arg)
1299 metaslab_t *msp = arg;
1300 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1302 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
1304 mutex_enter(&msp->ms_lock);
1305 metaslab_load_wait(msp);
1306 if (!msp->ms_loaded)
1307 (void) metaslab_load(msp);
1310 * Set the ms_access_txg value so that we don't unload it right away.
1312 msp->ms_access_txg = spa_syncing_txg(spa) + metaslab_unload_delay + 1;
1313 mutex_exit(&msp->ms_lock);
1317 metaslab_group_preload(metaslab_group_t *mg)
1319 spa_t *spa = mg->mg_vd->vdev_spa;
1321 avl_tree_t *t = &mg->mg_metaslab_tree;
1324 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
1325 taskq_wait(mg->mg_taskq);
1329 mutex_enter(&mg->mg_lock);
1331 * Load the next potential metaslabs
1334 while (msp != NULL) {
1335 metaslab_t *msp_next = AVL_NEXT(t, msp);
1337 /* If we have reached our preload limit then we're done */
1338 if (++m > metaslab_preload_limit)
1342 * We must drop the metaslab group lock here to preserve
1343 * lock ordering with the ms_lock (when grabbing both
1344 * the mg_lock and the ms_lock, the ms_lock must be taken
1345 * first). As a result, it is possible that the ordering
1346 * of the metaslabs within the avl tree may change before
1347 * we reacquire the lock. The metaslab cannot be removed from
1348 * the tree while we're in syncing context so it is safe to
1349 * drop the mg_lock here. If the metaslabs are reordered
1350 * nothing will break -- we just may end up loading a
1351 * less than optimal one.
1353 mutex_exit(&mg->mg_lock);
1354 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
1355 msp, TQ_SLEEP) != 0);
1356 mutex_enter(&mg->mg_lock);
1359 mutex_exit(&mg->mg_lock);
1363 * Determine if the space map's on-disk footprint is past our tolerance
1364 * for inefficiency. We would like to use the following criteria to make
1367 * 1. The size of the space map object should not dramatically increase as a
1368 * result of writing out the free space range tree.
1370 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1371 * times the size than the free space range tree representation
1372 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1374 * Checking the first condition is tricky since we don't want to walk
1375 * the entire AVL tree calculating the estimated on-disk size. Instead we
1376 * use the size-ordered range tree in the metaslab and calculate the
1377 * size required to write out the largest segment in our free tree. If the
1378 * size required to represent that segment on disk is larger than the space
1379 * map object then we avoid condensing this map.
1381 * To determine the second criterion we use a best-case estimate and assume
1382 * each segment can be represented on-disk as a single 64-bit entry. We refer
1383 * to this best-case estimate as the space map's minimal form.
1386 metaslab_should_condense(metaslab_t *msp)
1388 space_map_t *sm = msp->ms_sm;
1390 uint64_t size, entries, segsz;
1392 ASSERT(MUTEX_HELD(&msp->ms_lock));
1393 ASSERT(msp->ms_loaded);
1396 * Use the ms_size_tree range tree, which is ordered by size, to
1397 * obtain the largest segment in the free tree. If the tree is empty
1398 * then we should condense the map.
1400 rs = avl_last(&msp->ms_size_tree);
1405 * Calculate the number of 64-bit entries this segment would
1406 * require when written to disk. If this single segment would be
1407 * larger on-disk than the entire current on-disk structure, then
1408 * clearly condensing will increase the on-disk structure size.
1410 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
1411 entries = size / (MIN(size, SM_RUN_MAX));
1412 segsz = entries * sizeof (uint64_t);
1414 return (segsz <= space_map_length(msp->ms_sm) &&
1415 space_map_length(msp->ms_sm) >= (zfs_condense_pct *
1416 sizeof (uint64_t) * avl_numnodes(&msp->ms_tree->rt_root)) / 100);
1420 * Condense the on-disk space map representation to its minimized form.
1421 * The minimized form consists of a small number of allocations followed by
1422 * the entries of the free range tree.
1425 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1427 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1428 range_tree_t *freetree = msp->ms_freetree[txg & TXG_MASK];
1429 range_tree_t *condense_tree;
1430 space_map_t *sm = msp->ms_sm;
1432 ASSERT(MUTEX_HELD(&msp->ms_lock));
1433 ASSERT3U(spa_sync_pass(spa), ==, 1);
1434 ASSERT(msp->ms_loaded);
1436 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1437 "smp size %llu, segments %lu", txg, msp->ms_id, msp,
1438 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root));
1441 * Create an range tree that is 100% allocated. We remove segments
1442 * that have been freed in this txg, any deferred frees that exist,
1443 * and any allocation in the future. Removing segments should be
1444 * a relatively inexpensive operation since we expect these trees to
1445 * have a small number of nodes.
1447 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
1448 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
1451 * Remove what's been freed in this txg from the condense_tree.
1452 * Since we're in sync_pass 1, we know that all the frees from
1453 * this txg are in the freetree.
1455 range_tree_walk(freetree, range_tree_remove, condense_tree);
1457 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1458 range_tree_walk(msp->ms_defertree[t],
1459 range_tree_remove, condense_tree);
1462 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1463 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
1464 range_tree_remove, condense_tree);
1468 * We're about to drop the metaslab's lock thus allowing
1469 * other consumers to change it's content. Set the
1470 * metaslab's ms_condensing flag to ensure that
1471 * allocations on this metaslab do not occur while we're
1472 * in the middle of committing it to disk. This is only critical
1473 * for the ms_tree as all other range trees use per txg
1474 * views of their content.
1476 msp->ms_condensing = B_TRUE;
1478 mutex_exit(&msp->ms_lock);
1479 space_map_truncate(sm, tx);
1480 mutex_enter(&msp->ms_lock);
1483 * While we would ideally like to create a space_map representation
1484 * that consists only of allocation records, doing so can be
1485 * prohibitively expensive because the in-core free tree can be
1486 * large, and therefore computationally expensive to subtract
1487 * from the condense_tree. Instead we sync out two trees, a cheap
1488 * allocation only tree followed by the in-core free tree. While not
1489 * optimal, this is typically close to optimal, and much cheaper to
1492 space_map_write(sm, condense_tree, SM_ALLOC, tx);
1493 range_tree_vacate(condense_tree, NULL, NULL);
1494 range_tree_destroy(condense_tree);
1496 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
1497 msp->ms_condensing = B_FALSE;
1501 * Write a metaslab to disk in the context of the specified transaction group.
1504 metaslab_sync(metaslab_t *msp, uint64_t txg)
1506 metaslab_group_t *mg = msp->ms_group;
1507 vdev_t *vd = mg->mg_vd;
1508 spa_t *spa = vd->vdev_spa;
1509 objset_t *mos = spa_meta_objset(spa);
1510 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
1511 range_tree_t **freetree = &msp->ms_freetree[txg & TXG_MASK];
1512 range_tree_t **freed_tree =
1513 &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1515 uint64_t object = space_map_object(msp->ms_sm);
1517 ASSERT(!vd->vdev_ishole);
1520 * This metaslab has just been added so there's no work to do now.
1522 if (*freetree == NULL) {
1523 ASSERT3P(alloctree, ==, NULL);
1527 ASSERT3P(alloctree, !=, NULL);
1528 ASSERT3P(*freetree, !=, NULL);
1529 ASSERT3P(*freed_tree, !=, NULL);
1531 if (range_tree_space(alloctree) == 0 &&
1532 range_tree_space(*freetree) == 0)
1536 * The only state that can actually be changing concurrently with
1537 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1538 * be modifying this txg's alloctree, freetree, freed_tree, or
1539 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1540 * space_map ASSERTs. We drop it whenever we call into the DMU,
1541 * because the DMU can call down to us (e.g. via zio_free()) at
1545 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1547 if (msp->ms_sm == NULL) {
1548 uint64_t new_object;
1550 new_object = space_map_alloc(mos, tx);
1551 VERIFY3U(new_object, !=, 0);
1553 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
1554 msp->ms_start, msp->ms_size, vd->vdev_ashift,
1556 ASSERT(msp->ms_sm != NULL);
1559 mutex_enter(&msp->ms_lock);
1561 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
1562 metaslab_should_condense(msp)) {
1563 metaslab_condense(msp, txg, tx);
1565 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
1566 space_map_write(msp->ms_sm, *freetree, SM_FREE, tx);
1569 range_tree_vacate(alloctree, NULL, NULL);
1571 if (msp->ms_loaded) {
1573 * When the space map is loaded, we have an accruate
1574 * histogram in the range tree. This gives us an opportunity
1575 * to bring the space map's histogram up-to-date so we clear
1576 * it first before updating it.
1578 space_map_histogram_clear(msp->ms_sm);
1579 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
1582 * Since the space map is not loaded we simply update the
1583 * exisiting histogram with what was freed in this txg. This
1584 * means that the on-disk histogram may not have an accurate
1585 * view of the free space but it's close enough to allow
1586 * us to make allocation decisions.
1588 space_map_histogram_add(msp->ms_sm, *freetree, tx);
1592 * For sync pass 1, we avoid traversing this txg's free range tree
1593 * and instead will just swap the pointers for freetree and
1594 * freed_tree. We can safely do this since the freed_tree is
1595 * guaranteed to be empty on the initial pass.
1597 if (spa_sync_pass(spa) == 1) {
1598 range_tree_swap(freetree, freed_tree);
1600 range_tree_vacate(*freetree, range_tree_add, *freed_tree);
1603 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1604 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1606 mutex_exit(&msp->ms_lock);
1608 if (object != space_map_object(msp->ms_sm)) {
1609 object = space_map_object(msp->ms_sm);
1610 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1611 msp->ms_id, sizeof (uint64_t), &object, tx);
1617 * Called after a transaction group has completely synced to mark
1618 * all of the metaslab's free space as usable.
1621 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1623 metaslab_group_t *mg = msp->ms_group;
1624 vdev_t *vd = mg->mg_vd;
1625 range_tree_t **freed_tree;
1626 range_tree_t **defer_tree;
1627 int64_t alloc_delta, defer_delta;
1629 ASSERT(!vd->vdev_ishole);
1631 mutex_enter(&msp->ms_lock);
1634 * If this metaslab is just becoming available, initialize its
1635 * alloctrees, freetrees, and defertree and add its capacity to
1638 if (msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK] == NULL) {
1639 for (int t = 0; t < TXG_SIZE; t++) {
1640 ASSERT(msp->ms_alloctree[t] == NULL);
1641 ASSERT(msp->ms_freetree[t] == NULL);
1643 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
1645 msp->ms_freetree[t] = range_tree_create(NULL, msp,
1649 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1650 ASSERT(msp->ms_defertree[t] == NULL);
1652 msp->ms_defertree[t] = range_tree_create(NULL, msp,
1656 vdev_space_update(vd, 0, 0, msp->ms_size);
1659 freed_tree = &msp->ms_freetree[TXG_CLEAN(txg) & TXG_MASK];
1660 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
1662 alloc_delta = space_map_alloc_delta(msp->ms_sm);
1663 defer_delta = range_tree_space(*freed_tree) -
1664 range_tree_space(*defer_tree);
1666 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1668 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
1669 ASSERT0(range_tree_space(msp->ms_freetree[txg & TXG_MASK]));
1672 * If there's a metaslab_load() in progress, wait for it to complete
1673 * so that we have a consistent view of the in-core space map.
1675 metaslab_load_wait(msp);
1678 * Move the frees from the defer_tree back to the free
1679 * range tree (if it's loaded). Swap the freed_tree and the
1680 * defer_tree -- this is safe to do because we've just emptied out
1683 range_tree_vacate(*defer_tree,
1684 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
1685 range_tree_swap(freed_tree, defer_tree);
1687 space_map_update(msp->ms_sm);
1689 msp->ms_deferspace += defer_delta;
1690 ASSERT3S(msp->ms_deferspace, >=, 0);
1691 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
1692 if (msp->ms_deferspace != 0) {
1694 * Keep syncing this metaslab until all deferred frees
1695 * are back in circulation.
1697 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1700 if (msp->ms_loaded && msp->ms_access_txg < txg) {
1701 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
1702 VERIFY0(range_tree_space(
1703 msp->ms_alloctree[(txg + t) & TXG_MASK]));
1706 if (!metaslab_debug_unload)
1707 metaslab_unload(msp);
1710 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1711 mutex_exit(&msp->ms_lock);
1716 metaslab_sync_reassess(metaslab_group_t *mg)
1718 metaslab_group_alloc_update(mg);
1721 * Preload the next potential metaslabs
1723 metaslab_group_preload(mg);
1727 metaslab_distance(metaslab_t *msp, dva_t *dva)
1729 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1730 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1731 uint64_t start = msp->ms_id;
1733 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1734 return (1ULL << 63);
1737 return ((start - offset) << ms_shift);
1739 return ((offset - start) << ms_shift);
1744 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1745 uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
1747 spa_t *spa = mg->mg_vd->vdev_spa;
1748 metaslab_t *msp = NULL;
1749 uint64_t offset = -1ULL;
1750 avl_tree_t *t = &mg->mg_metaslab_tree;
1751 uint64_t activation_weight;
1752 uint64_t target_distance;
1755 activation_weight = METASLAB_WEIGHT_PRIMARY;
1756 for (i = 0; i < d; i++) {
1757 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1758 activation_weight = METASLAB_WEIGHT_SECONDARY;
1764 boolean_t was_active;
1766 mutex_enter(&mg->mg_lock);
1767 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1768 if (msp->ms_weight < asize) {
1769 spa_dbgmsg(spa, "%s: failed to meet weight "
1770 "requirement: vdev %llu, txg %llu, mg %p, "
1771 "msp %p, psize %llu, asize %llu, "
1772 "weight %llu", spa_name(spa),
1773 mg->mg_vd->vdev_id, txg,
1774 mg, msp, psize, asize, msp->ms_weight);
1775 mutex_exit(&mg->mg_lock);
1780 * If the selected metaslab is condensing, skip it.
1782 if (msp->ms_condensing)
1785 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1786 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1789 target_distance = min_distance +
1790 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
1793 for (i = 0; i < d; i++)
1794 if (metaslab_distance(msp, &dva[i]) <
1800 mutex_exit(&mg->mg_lock);
1804 mutex_enter(&msp->ms_lock);
1807 * Ensure that the metaslab we have selected is still
1808 * capable of handling our request. It's possible that
1809 * another thread may have changed the weight while we
1810 * were blocked on the metaslab lock.
1812 if (msp->ms_weight < asize || (was_active &&
1813 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1814 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1815 mutex_exit(&msp->ms_lock);
1819 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1820 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1821 metaslab_passivate(msp,
1822 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1823 mutex_exit(&msp->ms_lock);
1827 if (metaslab_activate(msp, activation_weight) != 0) {
1828 mutex_exit(&msp->ms_lock);
1833 * If this metaslab is currently condensing then pick again as
1834 * we can't manipulate this metaslab until it's committed
1837 if (msp->ms_condensing) {
1838 mutex_exit(&msp->ms_lock);
1842 if ((offset = metaslab_block_alloc(msp, asize)) != -1ULL)
1845 metaslab_passivate(msp, metaslab_block_maxsize(msp));
1846 mutex_exit(&msp->ms_lock);
1849 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
1850 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1852 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, asize);
1853 msp->ms_access_txg = txg + metaslab_unload_delay;
1855 mutex_exit(&msp->ms_lock);
1861 * Allocate a block for the specified i/o.
1864 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1865 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1867 metaslab_group_t *mg, *rotor;
1871 int zio_lock = B_FALSE;
1872 boolean_t allocatable;
1873 uint64_t offset = -1ULL;
1877 ASSERT(!DVA_IS_VALID(&dva[d]));
1880 * For testing, make some blocks above a certain size be gang blocks.
1882 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1883 return (SET_ERROR(ENOSPC));
1886 * Start at the rotor and loop through all mgs until we find something.
1887 * Note that there's no locking on mc_rotor or mc_aliquot because
1888 * nothing actually breaks if we miss a few updates -- we just won't
1889 * allocate quite as evenly. It all balances out over time.
1891 * If we are doing ditto or log blocks, try to spread them across
1892 * consecutive vdevs. If we're forced to reuse a vdev before we've
1893 * allocated all of our ditto blocks, then try and spread them out on
1894 * that vdev as much as possible. If it turns out to not be possible,
1895 * gradually lower our standards until anything becomes acceptable.
1896 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1897 * gives us hope of containing our fault domains to something we're
1898 * able to reason about. Otherwise, any two top-level vdev failures
1899 * will guarantee the loss of data. With consecutive allocation,
1900 * only two adjacent top-level vdev failures will result in data loss.
1902 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1903 * ourselves on the same vdev as our gang block header. That
1904 * way, we can hope for locality in vdev_cache, plus it makes our
1905 * fault domains something tractable.
1908 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1911 * It's possible the vdev we're using as the hint no
1912 * longer exists (i.e. removed). Consult the rotor when
1918 if (flags & METASLAB_HINTBP_AVOID &&
1919 mg->mg_next != NULL)
1924 } else if (d != 0) {
1925 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1926 mg = vd->vdev_mg->mg_next;
1932 * If the hint put us into the wrong metaslab class, or into a
1933 * metaslab group that has been passivated, just follow the rotor.
1935 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1942 ASSERT(mg->mg_activation_count == 1);
1947 * Don't allocate from faulted devices.
1950 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1951 allocatable = vdev_allocatable(vd);
1952 spa_config_exit(spa, SCL_ZIO, FTAG);
1954 allocatable = vdev_allocatable(vd);
1958 * Determine if the selected metaslab group is eligible
1959 * for allocations. If we're ganging or have requested
1960 * an allocation for the smallest gang block size
1961 * then we don't want to avoid allocating to the this
1962 * metaslab group. If we're in this condition we should
1963 * try to allocate from any device possible so that we
1964 * don't inadvertently return ENOSPC and suspend the pool
1965 * even though space is still available.
1967 if (allocatable && CAN_FASTGANG(flags) &&
1968 psize > SPA_GANGBLOCKSIZE)
1969 allocatable = metaslab_group_allocatable(mg);
1975 * Avoid writing single-copy data to a failing vdev
1976 * unless the user instructs us that it is okay.
1978 if ((vd->vdev_stat.vs_write_errors > 0 ||
1979 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1980 d == 0 && dshift == 3 &&
1981 !(zfs_write_to_degraded && vd->vdev_state ==
1982 VDEV_STATE_DEGRADED)) {
1987 ASSERT(mg->mg_class == mc);
1989 distance = vd->vdev_asize >> dshift;
1990 if (distance <= (1ULL << vd->vdev_ms_shift))
1995 asize = vdev_psize_to_asize(vd, psize);
1996 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1998 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
2000 if (offset != -1ULL) {
2002 * If we've just selected this metaslab group,
2003 * figure out whether the corresponding vdev is
2004 * over- or under-used relative to the pool,
2005 * and set an allocation bias to even it out.
2007 if (mc->mc_aliquot == 0) {
2008 vdev_stat_t *vs = &vd->vdev_stat;
2011 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
2012 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
2015 * Calculate how much more or less we should
2016 * try to allocate from this device during
2017 * this iteration around the rotor.
2018 * For example, if a device is 80% full
2019 * and the pool is 20% full then we should
2020 * reduce allocations by 60% on this device.
2022 * mg_bias = (20 - 80) * 512K / 100 = -307K
2024 * This reduces allocations by 307K for this
2027 mg->mg_bias = ((cu - vu) *
2028 (int64_t)mg->mg_aliquot) / 100;
2031 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
2032 mg->mg_aliquot + mg->mg_bias) {
2033 mc->mc_rotor = mg->mg_next;
2037 DVA_SET_VDEV(&dva[d], vd->vdev_id);
2038 DVA_SET_OFFSET(&dva[d], offset);
2039 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
2040 DVA_SET_ASIZE(&dva[d], asize);
2045 mc->mc_rotor = mg->mg_next;
2047 } while ((mg = mg->mg_next) != rotor);
2051 ASSERT(dshift < 64);
2055 if (!allocatable && !zio_lock) {
2061 bzero(&dva[d], sizeof (dva_t));
2063 return (SET_ERROR(ENOSPC));
2067 * Free the block represented by DVA in the context of the specified
2068 * transaction group.
2071 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
2073 uint64_t vdev = DVA_GET_VDEV(dva);
2074 uint64_t offset = DVA_GET_OFFSET(dva);
2075 uint64_t size = DVA_GET_ASIZE(dva);
2079 ASSERT(DVA_IS_VALID(dva));
2081 if (txg > spa_freeze_txg(spa))
2084 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2085 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
2086 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
2087 (u_longlong_t)vdev, (u_longlong_t)offset);
2092 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2094 if (DVA_GET_GANG(dva))
2095 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2097 mutex_enter(&msp->ms_lock);
2100 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
2103 VERIFY(!msp->ms_condensing);
2104 VERIFY3U(offset, >=, msp->ms_start);
2105 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
2106 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
2108 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2109 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2110 range_tree_add(msp->ms_tree, offset, size);
2112 if (range_tree_space(msp->ms_freetree[txg & TXG_MASK]) == 0)
2113 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2114 range_tree_add(msp->ms_freetree[txg & TXG_MASK],
2118 mutex_exit(&msp->ms_lock);
2122 * Intent log support: upon opening the pool after a crash, notify the SPA
2123 * of blocks that the intent log has allocated for immediate write, but
2124 * which are still considered free by the SPA because the last transaction
2125 * group didn't commit yet.
2128 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
2130 uint64_t vdev = DVA_GET_VDEV(dva);
2131 uint64_t offset = DVA_GET_OFFSET(dva);
2132 uint64_t size = DVA_GET_ASIZE(dva);
2137 ASSERT(DVA_IS_VALID(dva));
2139 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
2140 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
2141 return (SET_ERROR(ENXIO));
2143 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2145 if (DVA_GET_GANG(dva))
2146 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
2148 mutex_enter(&msp->ms_lock);
2150 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
2151 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
2153 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
2154 error = SET_ERROR(ENOENT);
2156 if (error || txg == 0) { /* txg == 0 indicates dry run */
2157 mutex_exit(&msp->ms_lock);
2161 VERIFY(!msp->ms_condensing);
2162 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
2163 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2164 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
2165 range_tree_remove(msp->ms_tree, offset, size);
2167 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
2168 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2169 vdev_dirty(vd, VDD_METASLAB, msp, txg);
2170 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
2173 mutex_exit(&msp->ms_lock);
2179 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
2180 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
2182 dva_t *dva = bp->blk_dva;
2183 dva_t *hintdva = hintbp->blk_dva;
2186 ASSERT(bp->blk_birth == 0);
2187 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
2189 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2191 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
2192 spa_config_exit(spa, SCL_ALLOC, FTAG);
2193 return (SET_ERROR(ENOSPC));
2196 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
2197 ASSERT(BP_GET_NDVAS(bp) == 0);
2198 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
2200 for (int d = 0; d < ndvas; d++) {
2201 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
2204 for (d--; d >= 0; d--) {
2205 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
2206 bzero(&dva[d], sizeof (dva_t));
2208 spa_config_exit(spa, SCL_ALLOC, FTAG);
2213 ASSERT(BP_GET_NDVAS(bp) == ndvas);
2215 spa_config_exit(spa, SCL_ALLOC, FTAG);
2217 BP_SET_BIRTH(bp, txg, txg);
2223 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
2225 const dva_t *dva = bp->blk_dva;
2226 int ndvas = BP_GET_NDVAS(bp);
2228 ASSERT(!BP_IS_HOLE(bp));
2229 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
2231 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
2233 for (int d = 0; d < ndvas; d++)
2234 metaslab_free_dva(spa, &dva[d], txg, now);
2236 spa_config_exit(spa, SCL_FREE, FTAG);
2240 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
2242 const dva_t *dva = bp->blk_dva;
2243 int ndvas = BP_GET_NDVAS(bp);
2246 ASSERT(!BP_IS_HOLE(bp));
2250 * First do a dry run to make sure all DVAs are claimable,
2251 * so we don't have to unwind from partial failures below.
2253 if ((error = metaslab_claim(spa, bp, 0)) != 0)
2257 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2259 for (int d = 0; d < ndvas; d++)
2260 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
2263 spa_config_exit(spa, SCL_ALLOC, FTAG);
2265 ASSERT(error == 0 || txg == 0);
2271 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
2273 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
2276 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
2277 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
2278 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
2279 vdev_t *vd = vdev_lookup_top(spa, vdev);
2280 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
2281 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
2282 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
2285 range_tree_verify(msp->ms_tree, offset, size);
2287 for (int j = 0; j < TXG_SIZE; j++)
2288 range_tree_verify(msp->ms_freetree[j], offset, size);
2289 for (int j = 0; j < TXG_DEFER_SIZE; j++)
2290 range_tree_verify(msp->ms_defertree[j], offset, size);
2292 spa_config_exit(spa, SCL_VDEV, FTAG);