Unleashed v1.4
[unleashed.git] / kernel / fs / zfs / metaslab.c
blob75a42283ef8bcd06f57025e61e24a971c9548cb3
1 /*
2 * CDDL HEADER START
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]
19 * CDDL HEADER END
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/zfs_context.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
37 #include <sys/vdev_indirect_mapping.h>
38 #include <sys/zap.h>
40 #define GANG_ALLOCATION(flags) \
41 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
43 uint64_t metaslab_aliquot = 512ULL << 10;
44 uint64_t metaslab_force_ganging = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
47 * Since we can touch multiple metaslabs (and their respective space maps)
48 * with each transaction group, we benefit from having a smaller space map
49 * block size since it allows us to issue more I/O operations scattered
50 * around the disk.
52 int zfs_metaslab_sm_blksz = (1 << 12);
55 * The in-core space map representation is more compact than its on-disk form.
56 * The zfs_condense_pct determines how much more compact the in-core
57 * space map representation must be before we compact it on-disk.
58 * Values should be greater than or equal to 100.
60 int zfs_condense_pct = 200;
63 * Condensing a metaslab is not guaranteed to actually reduce the amount of
64 * space used on disk. In particular, a space map uses data in increments of
65 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
66 * same number of blocks after condensing. Since the goal of condensing is to
67 * reduce the number of IOPs required to read the space map, we only want to
68 * condense when we can be sure we will reduce the number of blocks used by the
69 * space map. Unfortunately, we cannot precisely compute whether or not this is
70 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
71 * we apply the following heuristic: do not condense a spacemap unless the
72 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
73 * blocks.
75 int zfs_metaslab_condense_block_threshold = 4;
78 * The zfs_mg_noalloc_threshold defines which metaslab groups should
79 * be eligible for allocation. The value is defined as a percentage of
80 * free space. Metaslab groups that have more free space than
81 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
82 * a metaslab group's free space is less than or equal to the
83 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
84 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
85 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
86 * groups are allowed to accept allocations. Gang blocks are always
87 * eligible to allocate on any metaslab group. The default value of 0 means
88 * no metaslab group will be excluded based on this criterion.
90 int zfs_mg_noalloc_threshold = 0;
93 * Metaslab groups are considered eligible for allocations if their
94 * fragmenation metric (measured as a percentage) is less than or equal to
95 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
96 * then it will be skipped unless all metaslab groups within the metaslab
97 * class have also crossed this threshold.
99 int zfs_mg_fragmentation_threshold = 85;
102 * Allow metaslabs to keep their active state as long as their fragmentation
103 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
104 * active metaslab that exceeds this threshold will no longer keep its active
105 * status allowing better metaslabs to be selected.
107 int zfs_metaslab_fragmentation_threshold = 70;
110 * When set will load all metaslabs when pool is first opened.
112 int metaslab_debug_load = 0;
115 * When set will prevent metaslabs from being unloaded.
117 int metaslab_debug_unload = 0;
120 * Minimum size which forces the dynamic allocator to change
121 * it's allocation strategy. Once the space map cannot satisfy
122 * an allocation of this size then it switches to using more
123 * aggressive strategy (i.e search by size rather than offset).
125 uint64_t metaslab_df_alloc_threshold = SPA_OLD_MAXBLOCKSIZE;
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;
136 * A metaslab is considered "free" if it contains a contiguous
137 * segment which is greater than metaslab_min_alloc_size.
139 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
142 * Percentage of all cpus that can be used by the metaslab taskq.
144 int metaslab_load_pct = 50;
147 * Determines how many txgs a metaslab may remain loaded without having any
148 * allocations from it. As long as a metaslab continues to be used we will
149 * keep it loaded.
151 int metaslab_unload_delay = TXG_SIZE * 2;
154 * Max number of metaslabs per group to preload.
156 int metaslab_preload_limit = SPA_DVAS_PER_BP;
159 * Enable/disable preloading of metaslab.
161 boolean_t metaslab_preload_enabled = B_TRUE;
164 * Enable/disable fragmentation weighting on metaslabs.
166 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
169 * Enable/disable lba weighting (i.e. outer tracks are given preference).
171 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
174 * Enable/disable metaslab group biasing.
176 boolean_t metaslab_bias_enabled = B_TRUE;
179 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
181 boolean_t zfs_remap_blkptr_enable = B_TRUE;
184 * Enable/disable segment-based metaslab selection.
186 boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
189 * When using segment-based metaslab selection, we will continue
190 * allocating from the active metaslab until we have exhausted
191 * zfs_metaslab_switch_threshold of its buckets.
193 int zfs_metaslab_switch_threshold = 2;
196 * Internal switch to enable/disable the metaslab allocation tracing
197 * facility.
199 boolean_t metaslab_trace_enabled = B_TRUE;
202 * Maximum entries that the metaslab allocation tracing facility will keep
203 * in a given list when running in non-debug mode. We limit the number
204 * of entries in non-debug mode to prevent us from using up too much memory.
205 * The limit should be sufficiently large that we don't expect any allocation
206 * to every exceed this value. In debug mode, the system will panic if this
207 * limit is ever reached allowing for further investigation.
209 uint64_t metaslab_trace_max_entries = 5000;
211 static uint64_t metaslab_weight(metaslab_t *);
212 static void metaslab_set_fragmentation(metaslab_t *);
213 static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, boolean_t);
214 static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t);
215 static void metaslab_passivate(metaslab_t *msp, uint64_t weight);
216 static uint64_t metaslab_weight_from_range_tree(metaslab_t *msp);
218 kmem_cache_t *metaslab_alloc_trace_cache;
221 * ==========================================================================
222 * Metaslab classes
223 * ==========================================================================
225 metaslab_class_t *
226 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
228 metaslab_class_t *mc;
230 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
232 mc->mc_spa = spa;
233 mc->mc_rotor = NULL;
234 mc->mc_ops = ops;
235 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
236 mc->mc_alloc_slots = kmem_zalloc(spa->spa_alloc_count *
237 sizeof (refcount_t), KM_SLEEP);
238 mc->mc_alloc_max_slots = kmem_zalloc(spa->spa_alloc_count *
239 sizeof (uint64_t), KM_SLEEP);
240 for (int i = 0; i < spa->spa_alloc_count; i++)
241 refcount_create_tracked(&mc->mc_alloc_slots[i]);
243 return (mc);
246 void
247 metaslab_class_destroy(metaslab_class_t *mc)
249 ASSERT(mc->mc_rotor == NULL);
250 ASSERT(mc->mc_alloc == 0);
251 ASSERT(mc->mc_deferred == 0);
252 ASSERT(mc->mc_space == 0);
253 ASSERT(mc->mc_dspace == 0);
255 for (int i = 0; i < mc->mc_spa->spa_alloc_count; i++)
256 refcount_destroy(&mc->mc_alloc_slots[i]);
257 kmem_free(mc->mc_alloc_slots, mc->mc_spa->spa_alloc_count *
258 sizeof (refcount_t));
259 kmem_free(mc->mc_alloc_max_slots, mc->mc_spa->spa_alloc_count *
260 sizeof (uint64_t));
261 mutex_destroy(&mc->mc_lock);
262 kmem_free(mc, sizeof (metaslab_class_t));
266 metaslab_class_validate(metaslab_class_t *mc)
268 metaslab_group_t *mg;
269 vdev_t *vd;
272 * Must hold one of the spa_config locks.
274 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
275 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
277 if ((mg = mc->mc_rotor) == NULL)
278 return (0);
280 do {
281 vd = mg->mg_vd;
282 ASSERT(vd->vdev_mg != NULL);
283 ASSERT3P(vd->vdev_top, ==, vd);
284 ASSERT3P(mg->mg_class, ==, mc);
285 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
286 } while ((mg = mg->mg_next) != mc->mc_rotor);
288 return (0);
291 void
292 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
293 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
295 atomic_add_64(&mc->mc_alloc, alloc_delta);
296 atomic_add_64(&mc->mc_deferred, defer_delta);
297 atomic_add_64(&mc->mc_space, space_delta);
298 atomic_add_64(&mc->mc_dspace, dspace_delta);
301 uint64_t
302 metaslab_class_get_alloc(metaslab_class_t *mc)
304 return (mc->mc_alloc);
307 uint64_t
308 metaslab_class_get_deferred(metaslab_class_t *mc)
310 return (mc->mc_deferred);
313 uint64_t
314 metaslab_class_get_space(metaslab_class_t *mc)
316 return (mc->mc_space);
319 uint64_t
320 metaslab_class_get_dspace(metaslab_class_t *mc)
322 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
325 void
326 metaslab_class_histogram_verify(metaslab_class_t *mc)
328 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
329 uint64_t *mc_hist;
330 int i;
332 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
333 return;
335 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
336 KM_SLEEP);
338 for (int c = 0; c < rvd->vdev_children; c++) {
339 vdev_t *tvd = rvd->vdev_child[c];
340 metaslab_group_t *mg = tvd->vdev_mg;
343 * Skip any holes, uninitialized top-levels, or
344 * vdevs that are not in this metalab class.
346 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
347 mg->mg_class != mc) {
348 continue;
351 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
352 mc_hist[i] += mg->mg_histogram[i];
355 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
356 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
358 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
362 * Calculate the metaslab class's fragmentation metric. The metric
363 * is weighted based on the space contribution of each metaslab group.
364 * The return value will be a number between 0 and 100 (inclusive), or
365 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
366 * zfs_frag_table for more information about the metric.
368 uint64_t
369 metaslab_class_fragmentation(metaslab_class_t *mc)
371 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
372 uint64_t fragmentation = 0;
374 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
376 for (int c = 0; c < rvd->vdev_children; c++) {
377 vdev_t *tvd = rvd->vdev_child[c];
378 metaslab_group_t *mg = tvd->vdev_mg;
381 * Skip any holes, uninitialized top-levels,
382 * or vdevs that are not in this metalab class.
384 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
385 mg->mg_class != mc) {
386 continue;
390 * If a metaslab group does not contain a fragmentation
391 * metric then just bail out.
393 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
394 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
395 return (ZFS_FRAG_INVALID);
399 * Determine how much this metaslab_group is contributing
400 * to the overall pool fragmentation metric.
402 fragmentation += mg->mg_fragmentation *
403 metaslab_group_get_space(mg);
405 fragmentation /= metaslab_class_get_space(mc);
407 ASSERT3U(fragmentation, <=, 100);
408 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
409 return (fragmentation);
413 * Calculate the amount of expandable space that is available in
414 * this metaslab class. If a device is expanded then its expandable
415 * space will be the amount of allocatable space that is currently not
416 * part of this metaslab class.
418 uint64_t
419 metaslab_class_expandable_space(metaslab_class_t *mc)
421 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
422 uint64_t space = 0;
424 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
425 for (int c = 0; c < rvd->vdev_children; c++) {
426 uint64_t tspace;
427 vdev_t *tvd = rvd->vdev_child[c];
428 metaslab_group_t *mg = tvd->vdev_mg;
430 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
431 mg->mg_class != mc) {
432 continue;
436 * Calculate if we have enough space to add additional
437 * metaslabs. We report the expandable space in terms
438 * of the metaslab size since that's the unit of expansion.
439 * Adjust by efi system partition size.
441 tspace = tvd->vdev_max_asize - tvd->vdev_asize;
442 if (tspace > mc->mc_spa->spa_bootsize) {
443 tspace -= mc->mc_spa->spa_bootsize;
445 space += P2ALIGN(tspace, 1ULL << tvd->vdev_ms_shift);
447 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
448 return (space);
451 static int
452 metaslab_compare(const void *x1, const void *x2)
454 const metaslab_t *m1 = x1;
455 const metaslab_t *m2 = x2;
457 int sort1 = 0;
458 int sort2 = 0;
459 if (m1->ms_allocator != -1 && m1->ms_primary)
460 sort1 = 1;
461 else if (m1->ms_allocator != -1 && !m1->ms_primary)
462 sort1 = 2;
463 if (m2->ms_allocator != -1 && m2->ms_primary)
464 sort2 = 1;
465 else if (m2->ms_allocator != -1 && !m2->ms_primary)
466 sort2 = 2;
469 * Sort inactive metaslabs first, then primaries, then secondaries. When
470 * selecting a metaslab to allocate from, an allocator first tries its
471 * primary, then secondary active metaslab. If it doesn't have active
472 * metaslabs, or can't allocate from them, it searches for an inactive
473 * metaslab to activate. If it can't find a suitable one, it will steal
474 * a primary or secondary metaslab from another allocator.
476 if (sort1 < sort2)
477 return (-1);
478 if (sort1 > sort2)
479 return (1);
481 if (m1->ms_weight < m2->ms_weight)
482 return (1);
483 if (m1->ms_weight > m2->ms_weight)
484 return (-1);
487 * If the weights are identical, use the offset to force uniqueness.
489 if (m1->ms_start < m2->ms_start)
490 return (-1);
491 if (m1->ms_start > m2->ms_start)
492 return (1);
494 ASSERT3P(m1, ==, m2);
496 return (0);
500 * Verify that the space accounting on disk matches the in-core range_trees.
502 void
503 metaslab_verify_space(metaslab_t *msp, uint64_t txg)
505 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
506 uint64_t allocated = 0;
507 uint64_t sm_free_space, msp_free_space;
509 ASSERT(MUTEX_HELD(&msp->ms_lock));
511 if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
512 return;
515 * We can only verify the metaslab space when we're called
516 * from syncing context with a loaded metaslab that has an allocated
517 * space map. Calling this in non-syncing context does not
518 * provide a consistent view of the metaslab since we're performing
519 * allocations in the future.
521 if (txg != spa_syncing_txg(spa) || msp->ms_sm == NULL ||
522 !msp->ms_loaded)
523 return;
525 sm_free_space = msp->ms_size - space_map_allocated(msp->ms_sm) -
526 space_map_alloc_delta(msp->ms_sm);
529 * Account for future allocations since we would have already
530 * deducted that space from the ms_freetree.
532 for (int t = 0; t < TXG_CONCURRENT_STATES; t++) {
533 allocated +=
534 range_tree_space(msp->ms_allocating[(txg + t) & TXG_MASK]);
537 msp_free_space = range_tree_space(msp->ms_allocatable) + allocated +
538 msp->ms_deferspace + range_tree_space(msp->ms_freed);
540 VERIFY3U(sm_free_space, ==, msp_free_space);
544 * ==========================================================================
545 * Metaslab groups
546 * ==========================================================================
549 * Update the allocatable flag and the metaslab group's capacity.
550 * The allocatable flag is set to true if the capacity is below
551 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
552 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
553 * transitions from allocatable to non-allocatable or vice versa then the
554 * metaslab group's class is updated to reflect the transition.
556 static void
557 metaslab_group_alloc_update(metaslab_group_t *mg)
559 vdev_t *vd = mg->mg_vd;
560 metaslab_class_t *mc = mg->mg_class;
561 vdev_stat_t *vs = &vd->vdev_stat;
562 boolean_t was_allocatable;
563 boolean_t was_initialized;
565 ASSERT(vd == vd->vdev_top);
566 ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==,
567 SCL_ALLOC);
569 mutex_enter(&mg->mg_lock);
570 was_allocatable = mg->mg_allocatable;
571 was_initialized = mg->mg_initialized;
573 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
574 (vs->vs_space + 1);
576 mutex_enter(&mc->mc_lock);
579 * If the metaslab group was just added then it won't
580 * have any space until we finish syncing out this txg.
581 * At that point we will consider it initialized and available
582 * for allocations. We also don't consider non-activated
583 * metaslab groups (e.g. vdevs that are in the middle of being removed)
584 * to be initialized, because they can't be used for allocation.
586 mg->mg_initialized = metaslab_group_initialized(mg);
587 if (!was_initialized && mg->mg_initialized) {
588 mc->mc_groups++;
589 } else if (was_initialized && !mg->mg_initialized) {
590 ASSERT3U(mc->mc_groups, >, 0);
591 mc->mc_groups--;
593 if (mg->mg_initialized)
594 mg->mg_no_free_space = B_FALSE;
597 * A metaslab group is considered allocatable if it has plenty
598 * of free space or is not heavily fragmented. We only take
599 * fragmentation into account if the metaslab group has a valid
600 * fragmentation metric (i.e. a value between 0 and 100).
602 mg->mg_allocatable = (mg->mg_activation_count > 0 &&
603 mg->mg_free_capacity > zfs_mg_noalloc_threshold &&
604 (mg->mg_fragmentation == ZFS_FRAG_INVALID ||
605 mg->mg_fragmentation <= zfs_mg_fragmentation_threshold));
608 * The mc_alloc_groups maintains a count of the number of
609 * groups in this metaslab class that are still above the
610 * zfs_mg_noalloc_threshold. This is used by the allocating
611 * threads to determine if they should avoid allocations to
612 * a given group. The allocator will avoid allocations to a group
613 * if that group has reached or is below the zfs_mg_noalloc_threshold
614 * and there are still other groups that are above the threshold.
615 * When a group transitions from allocatable to non-allocatable or
616 * vice versa we update the metaslab class to reflect that change.
617 * When the mc_alloc_groups value drops to 0 that means that all
618 * groups have reached the zfs_mg_noalloc_threshold making all groups
619 * eligible for allocations. This effectively means that all devices
620 * are balanced again.
622 if (was_allocatable && !mg->mg_allocatable)
623 mc->mc_alloc_groups--;
624 else if (!was_allocatable && mg->mg_allocatable)
625 mc->mc_alloc_groups++;
626 mutex_exit(&mc->mc_lock);
628 mutex_exit(&mg->mg_lock);
631 metaslab_group_t *
632 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd, int allocators)
634 metaslab_group_t *mg;
636 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
637 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
638 mutex_init(&mg->mg_ms_initialize_lock, NULL, MUTEX_DEFAULT, NULL);
639 cv_init(&mg->mg_ms_initialize_cv, NULL, CV_DEFAULT, NULL);
640 mg->mg_primaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
641 KM_SLEEP);
642 mg->mg_secondaries = kmem_zalloc(allocators * sizeof (metaslab_t *),
643 KM_SLEEP);
644 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
645 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
646 mg->mg_vd = vd;
647 mg->mg_class = mc;
648 mg->mg_activation_count = 0;
649 mg->mg_initialized = B_FALSE;
650 mg->mg_no_free_space = B_TRUE;
651 mg->mg_allocators = allocators;
653 mg->mg_alloc_queue_depth = kmem_zalloc(allocators * sizeof (refcount_t),
654 KM_SLEEP);
655 mg->mg_cur_max_alloc_queue_depth = kmem_zalloc(allocators *
656 sizeof (uint64_t), KM_SLEEP);
657 for (int i = 0; i < allocators; i++) {
658 refcount_create_tracked(&mg->mg_alloc_queue_depth[i]);
659 mg->mg_cur_max_alloc_queue_depth[i] = 0;
662 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
663 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
665 return (mg);
668 void
669 metaslab_group_destroy(metaslab_group_t *mg)
671 ASSERT(mg->mg_prev == NULL);
672 ASSERT(mg->mg_next == NULL);
674 * We may have gone below zero with the activation count
675 * either because we never activated in the first place or
676 * because we're done, and possibly removing the vdev.
678 ASSERT(mg->mg_activation_count <= 0);
680 taskq_destroy(mg->mg_taskq);
681 avl_destroy(&mg->mg_metaslab_tree);
682 kmem_free(mg->mg_primaries, mg->mg_allocators * sizeof (metaslab_t *));
683 kmem_free(mg->mg_secondaries, mg->mg_allocators *
684 sizeof (metaslab_t *));
685 mutex_destroy(&mg->mg_lock);
686 mutex_destroy(&mg->mg_ms_initialize_lock);
687 cv_destroy(&mg->mg_ms_initialize_cv);
689 for (int i = 0; i < mg->mg_allocators; i++) {
690 refcount_destroy(&mg->mg_alloc_queue_depth[i]);
691 mg->mg_cur_max_alloc_queue_depth[i] = 0;
693 kmem_free(mg->mg_alloc_queue_depth, mg->mg_allocators *
694 sizeof (refcount_t));
695 kmem_free(mg->mg_cur_max_alloc_queue_depth, mg->mg_allocators *
696 sizeof (uint64_t));
698 kmem_free(mg, sizeof (metaslab_group_t));
701 void
702 metaslab_group_activate(metaslab_group_t *mg)
704 metaslab_class_t *mc = mg->mg_class;
705 metaslab_group_t *mgprev, *mgnext;
707 ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER), !=, 0);
709 ASSERT(mc->mc_rotor != mg);
710 ASSERT(mg->mg_prev == NULL);
711 ASSERT(mg->mg_next == NULL);
712 ASSERT(mg->mg_activation_count <= 0);
714 if (++mg->mg_activation_count <= 0)
715 return;
717 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
718 metaslab_group_alloc_update(mg);
720 if ((mgprev = mc->mc_rotor) == NULL) {
721 mg->mg_prev = mg;
722 mg->mg_next = mg;
723 } else {
724 mgnext = mgprev->mg_next;
725 mg->mg_prev = mgprev;
726 mg->mg_next = mgnext;
727 mgprev->mg_next = mg;
728 mgnext->mg_prev = mg;
730 mc->mc_rotor = mg;
734 * Passivate a metaslab group and remove it from the allocation rotor.
735 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
736 * a metaslab group. This function will momentarily drop spa_config_locks
737 * that are lower than the SCL_ALLOC lock (see comment below).
739 void
740 metaslab_group_passivate(metaslab_group_t *mg)
742 metaslab_class_t *mc = mg->mg_class;
743 spa_t *spa = mc->mc_spa;
744 metaslab_group_t *mgprev, *mgnext;
745 int locks = spa_config_held(spa, SCL_ALL, RW_WRITER);
747 ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==,
748 (SCL_ALLOC | SCL_ZIO));
750 if (--mg->mg_activation_count != 0) {
751 ASSERT(mc->mc_rotor != mg);
752 ASSERT(mg->mg_prev == NULL);
753 ASSERT(mg->mg_next == NULL);
754 ASSERT(mg->mg_activation_count < 0);
755 return;
759 * The spa_config_lock is an array of rwlocks, ordered as
760 * follows (from highest to lowest):
761 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
762 * SCL_ZIO > SCL_FREE > SCL_VDEV
763 * (For more information about the spa_config_lock see spa_misc.c)
764 * The higher the lock, the broader its coverage. When we passivate
765 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
766 * config locks. However, the metaslab group's taskq might be trying
767 * to preload metaslabs so we must drop the SCL_ZIO lock and any
768 * lower locks to allow the I/O to complete. At a minimum,
769 * we continue to hold the SCL_ALLOC lock, which prevents any future
770 * allocations from taking place and any changes to the vdev tree.
772 spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa);
773 taskq_wait(mg->mg_taskq);
774 spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
775 metaslab_group_alloc_update(mg);
776 for (int i = 0; i < mg->mg_allocators; i++) {
777 metaslab_t *msp = mg->mg_primaries[i];
778 if (msp != NULL) {
779 mutex_enter(&msp->ms_lock);
780 metaslab_passivate(msp,
781 metaslab_weight_from_range_tree(msp));
782 mutex_exit(&msp->ms_lock);
784 msp = mg->mg_secondaries[i];
785 if (msp != NULL) {
786 mutex_enter(&msp->ms_lock);
787 metaslab_passivate(msp,
788 metaslab_weight_from_range_tree(msp));
789 mutex_exit(&msp->ms_lock);
793 mgprev = mg->mg_prev;
794 mgnext = mg->mg_next;
796 if (mg == mgnext) {
797 mc->mc_rotor = NULL;
798 } else {
799 mc->mc_rotor = mgnext;
800 mgprev->mg_next = mgnext;
801 mgnext->mg_prev = mgprev;
804 mg->mg_prev = NULL;
805 mg->mg_next = NULL;
808 boolean_t
809 metaslab_group_initialized(metaslab_group_t *mg)
811 vdev_t *vd = mg->mg_vd;
812 vdev_stat_t *vs = &vd->vdev_stat;
814 return (vs->vs_space != 0 && mg->mg_activation_count > 0);
817 uint64_t
818 metaslab_group_get_space(metaslab_group_t *mg)
820 return ((1ULL << mg->mg_vd->vdev_ms_shift) * mg->mg_vd->vdev_ms_count);
823 void
824 metaslab_group_histogram_verify(metaslab_group_t *mg)
826 uint64_t *mg_hist;
827 vdev_t *vd = mg->mg_vd;
828 uint64_t ashift = vd->vdev_ashift;
829 int i;
831 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
832 return;
834 mg_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
835 KM_SLEEP);
837 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE, >=,
838 SPACE_MAP_HISTOGRAM_SIZE + ashift);
840 for (int m = 0; m < vd->vdev_ms_count; m++) {
841 metaslab_t *msp = vd->vdev_ms[m];
843 if (msp->ms_sm == NULL)
844 continue;
846 for (i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++)
847 mg_hist[i + ashift] +=
848 msp->ms_sm->sm_phys->smp_histogram[i];
851 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i ++)
852 VERIFY3U(mg_hist[i], ==, mg->mg_histogram[i]);
854 kmem_free(mg_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
857 static void
858 metaslab_group_histogram_add(metaslab_group_t *mg, metaslab_t *msp)
860 metaslab_class_t *mc = mg->mg_class;
861 uint64_t ashift = mg->mg_vd->vdev_ashift;
863 ASSERT(MUTEX_HELD(&msp->ms_lock));
864 if (msp->ms_sm == NULL)
865 return;
867 mutex_enter(&mg->mg_lock);
868 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
869 mg->mg_histogram[i + ashift] +=
870 msp->ms_sm->sm_phys->smp_histogram[i];
871 mc->mc_histogram[i + ashift] +=
872 msp->ms_sm->sm_phys->smp_histogram[i];
874 mutex_exit(&mg->mg_lock);
877 void
878 metaslab_group_histogram_remove(metaslab_group_t *mg, metaslab_t *msp)
880 metaslab_class_t *mc = mg->mg_class;
881 uint64_t ashift = mg->mg_vd->vdev_ashift;
883 ASSERT(MUTEX_HELD(&msp->ms_lock));
884 if (msp->ms_sm == NULL)
885 return;
887 mutex_enter(&mg->mg_lock);
888 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
889 ASSERT3U(mg->mg_histogram[i + ashift], >=,
890 msp->ms_sm->sm_phys->smp_histogram[i]);
891 ASSERT3U(mc->mc_histogram[i + ashift], >=,
892 msp->ms_sm->sm_phys->smp_histogram[i]);
894 mg->mg_histogram[i + ashift] -=
895 msp->ms_sm->sm_phys->smp_histogram[i];
896 mc->mc_histogram[i + ashift] -=
897 msp->ms_sm->sm_phys->smp_histogram[i];
899 mutex_exit(&mg->mg_lock);
902 static void
903 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
905 ASSERT(msp->ms_group == NULL);
906 mutex_enter(&mg->mg_lock);
907 msp->ms_group = mg;
908 msp->ms_weight = 0;
909 avl_add(&mg->mg_metaslab_tree, msp);
910 mutex_exit(&mg->mg_lock);
912 mutex_enter(&msp->ms_lock);
913 metaslab_group_histogram_add(mg, msp);
914 mutex_exit(&msp->ms_lock);
917 static void
918 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
920 mutex_enter(&msp->ms_lock);
921 metaslab_group_histogram_remove(mg, msp);
922 mutex_exit(&msp->ms_lock);
924 mutex_enter(&mg->mg_lock);
925 ASSERT(msp->ms_group == mg);
926 avl_remove(&mg->mg_metaslab_tree, msp);
927 msp->ms_group = NULL;
928 mutex_exit(&mg->mg_lock);
931 static void
932 metaslab_group_sort_impl(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
934 ASSERT(MUTEX_HELD(&mg->mg_lock));
935 ASSERT(msp->ms_group == mg);
936 avl_remove(&mg->mg_metaslab_tree, msp);
937 msp->ms_weight = weight;
938 avl_add(&mg->mg_metaslab_tree, msp);
942 static void
943 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
946 * Although in principle the weight can be any value, in
947 * practice we do not use values in the range [1, 511].
949 ASSERT(weight >= SPA_MINBLOCKSIZE || weight == 0);
950 ASSERT(MUTEX_HELD(&msp->ms_lock));
952 mutex_enter(&mg->mg_lock);
953 metaslab_group_sort_impl(mg, msp, weight);
954 mutex_exit(&mg->mg_lock);
958 * Calculate the fragmentation for a given metaslab group. We can use
959 * a simple average here since all metaslabs within the group must have
960 * the same size. The return value will be a value between 0 and 100
961 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
962 * group have a fragmentation metric.
964 uint64_t
965 metaslab_group_fragmentation(metaslab_group_t *mg)
967 vdev_t *vd = mg->mg_vd;
968 uint64_t fragmentation = 0;
969 uint64_t valid_ms = 0;
971 for (int m = 0; m < vd->vdev_ms_count; m++) {
972 metaslab_t *msp = vd->vdev_ms[m];
974 if (msp->ms_fragmentation == ZFS_FRAG_INVALID)
975 continue;
977 valid_ms++;
978 fragmentation += msp->ms_fragmentation;
981 if (valid_ms <= vd->vdev_ms_count / 2)
982 return (ZFS_FRAG_INVALID);
984 fragmentation /= valid_ms;
985 ASSERT3U(fragmentation, <=, 100);
986 return (fragmentation);
990 * Determine if a given metaslab group should skip allocations. A metaslab
991 * group should avoid allocations if its free capacity is less than the
992 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
993 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
994 * that can still handle allocations. If the allocation throttle is enabled
995 * then we skip allocations to devices that have reached their maximum
996 * allocation queue depth unless the selected metaslab group is the only
997 * eligible group remaining.
999 static boolean_t
1000 metaslab_group_allocatable(metaslab_group_t *mg, metaslab_group_t *rotor,
1001 uint64_t psize, int allocator)
1003 spa_t *spa = mg->mg_vd->vdev_spa;
1004 metaslab_class_t *mc = mg->mg_class;
1007 * We can only consider skipping this metaslab group if it's
1008 * in the normal metaslab class and there are other metaslab
1009 * groups to select from. Otherwise, we always consider it eligible
1010 * for allocations.
1012 if (mc != spa_normal_class(spa) || mc->mc_groups <= 1)
1013 return (B_TRUE);
1016 * If the metaslab group's mg_allocatable flag is set (see comments
1017 * in metaslab_group_alloc_update() for more information) and
1018 * the allocation throttle is disabled then allow allocations to this
1019 * device. However, if the allocation throttle is enabled then
1020 * check if we have reached our allocation limit (mg_alloc_queue_depth)
1021 * to determine if we should allow allocations to this metaslab group.
1022 * If all metaslab groups are no longer considered allocatable
1023 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1024 * gang block size then we allow allocations on this metaslab group
1025 * regardless of the mg_allocatable or throttle settings.
1027 if (mg->mg_allocatable) {
1028 metaslab_group_t *mgp;
1029 int64_t qdepth;
1030 uint64_t qmax = mg->mg_cur_max_alloc_queue_depth[allocator];
1032 if (!mc->mc_alloc_throttle_enabled)
1033 return (B_TRUE);
1036 * If this metaslab group does not have any free space, then
1037 * there is no point in looking further.
1039 if (mg->mg_no_free_space)
1040 return (B_FALSE);
1042 qdepth = refcount_count(&mg->mg_alloc_queue_depth[allocator]);
1045 * If this metaslab group is below its qmax or it's
1046 * the only allocatable metasable group, then attempt
1047 * to allocate from it.
1049 if (qdepth < qmax || mc->mc_alloc_groups == 1)
1050 return (B_TRUE);
1051 ASSERT3U(mc->mc_alloc_groups, >, 1);
1054 * Since this metaslab group is at or over its qmax, we
1055 * need to determine if there are metaslab groups after this
1056 * one that might be able to handle this allocation. This is
1057 * racy since we can't hold the locks for all metaslab
1058 * groups at the same time when we make this check.
1060 for (mgp = mg->mg_next; mgp != rotor; mgp = mgp->mg_next) {
1061 qmax = mgp->mg_cur_max_alloc_queue_depth[allocator];
1063 qdepth = refcount_count(
1064 &mgp->mg_alloc_queue_depth[allocator]);
1067 * If there is another metaslab group that
1068 * might be able to handle the allocation, then
1069 * we return false so that we skip this group.
1071 if (qdepth < qmax && !mgp->mg_no_free_space)
1072 return (B_FALSE);
1076 * We didn't find another group to handle the allocation
1077 * so we can't skip this metaslab group even though
1078 * we are at or over our qmax.
1080 return (B_TRUE);
1082 } else if (mc->mc_alloc_groups == 0 || psize == SPA_MINBLOCKSIZE) {
1083 return (B_TRUE);
1085 return (B_FALSE);
1089 * ==========================================================================
1090 * Range tree callbacks
1091 * ==========================================================================
1095 * Comparison function for the private size-ordered tree. Tree is sorted
1096 * by size, larger sizes at the end of the tree.
1098 static int
1099 metaslab_rangesize_compare(const void *x1, const void *x2)
1101 const range_seg_t *r1 = x1;
1102 const range_seg_t *r2 = x2;
1103 uint64_t rs_size1 = r1->rs_end - r1->rs_start;
1104 uint64_t rs_size2 = r2->rs_end - r2->rs_start;
1106 if (rs_size1 < rs_size2)
1107 return (-1);
1108 if (rs_size1 > rs_size2)
1109 return (1);
1111 if (r1->rs_start < r2->rs_start)
1112 return (-1);
1114 if (r1->rs_start > r2->rs_start)
1115 return (1);
1117 return (0);
1121 * Create any block allocator specific components. The current allocators
1122 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1124 static void
1125 metaslab_rt_create(range_tree_t *rt, void *arg)
1127 metaslab_t *msp = arg;
1129 ASSERT3P(rt->rt_arg, ==, msp);
1130 ASSERT(msp->ms_allocatable == NULL);
1132 avl_create(&msp->ms_allocatable_by_size, metaslab_rangesize_compare,
1133 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1137 * Destroy the block allocator specific components.
1139 static void
1140 metaslab_rt_destroy(range_tree_t *rt, void *arg)
1142 metaslab_t *msp = arg;
1144 ASSERT3P(rt->rt_arg, ==, msp);
1145 ASSERT3P(msp->ms_allocatable, ==, rt);
1146 ASSERT0(avl_numnodes(&msp->ms_allocatable_by_size));
1148 avl_destroy(&msp->ms_allocatable_by_size);
1151 static void
1152 metaslab_rt_add(range_tree_t *rt, range_seg_t *rs, void *arg)
1154 metaslab_t *msp = arg;
1156 ASSERT3P(rt->rt_arg, ==, msp);
1157 ASSERT3P(msp->ms_allocatable, ==, rt);
1158 VERIFY(!msp->ms_condensing);
1159 avl_add(&msp->ms_allocatable_by_size, rs);
1162 static void
1163 metaslab_rt_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
1165 metaslab_t *msp = arg;
1167 ASSERT3P(rt->rt_arg, ==, msp);
1168 ASSERT3P(msp->ms_allocatable, ==, rt);
1169 VERIFY(!msp->ms_condensing);
1170 avl_remove(&msp->ms_allocatable_by_size, rs);
1173 static void
1174 metaslab_rt_vacate(range_tree_t *rt, void *arg)
1176 metaslab_t *msp = arg;
1178 ASSERT3P(rt->rt_arg, ==, msp);
1179 ASSERT3P(msp->ms_allocatable, ==, rt);
1182 * Normally one would walk the tree freeing nodes along the way.
1183 * Since the nodes are shared with the range trees we can avoid
1184 * walking all nodes and just reinitialize the avl tree. The nodes
1185 * will be freed by the range tree, so we don't want to free them here.
1187 avl_create(&msp->ms_allocatable_by_size, metaslab_rangesize_compare,
1188 sizeof (range_seg_t), offsetof(range_seg_t, rs_pp_node));
1191 static range_tree_ops_t metaslab_rt_ops = {
1192 metaslab_rt_create,
1193 metaslab_rt_destroy,
1194 metaslab_rt_add,
1195 metaslab_rt_remove,
1196 metaslab_rt_vacate
1200 * ==========================================================================
1201 * Common allocator routines
1202 * ==========================================================================
1206 * Return the maximum contiguous segment within the metaslab.
1208 uint64_t
1209 metaslab_block_maxsize(metaslab_t *msp)
1211 avl_tree_t *t = &msp->ms_allocatable_by_size;
1212 range_seg_t *rs;
1214 if (t == NULL || (rs = avl_last(t)) == NULL)
1215 return (0ULL);
1217 return (rs->rs_end - rs->rs_start);
1220 static range_seg_t *
1221 metaslab_block_find(avl_tree_t *t, uint64_t start, uint64_t size)
1223 range_seg_t *rs, rsearch;
1224 avl_index_t where;
1226 rsearch.rs_start = start;
1227 rsearch.rs_end = start + size;
1229 rs = avl_find(t, &rsearch, &where);
1230 if (rs == NULL) {
1231 rs = avl_nearest(t, where, AVL_AFTER);
1234 return (rs);
1238 * This is a helper function that can be used by the allocator to find
1239 * a suitable block to allocate. This will search the specified AVL
1240 * tree looking for a block that matches the specified criteria.
1242 static uint64_t
1243 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1244 uint64_t align)
1246 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1248 while (rs != NULL) {
1249 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1251 if (offset + size <= rs->rs_end) {
1252 *cursor = offset + size;
1253 return (offset);
1255 rs = AVL_NEXT(t, rs);
1259 * If we know we've searched the whole map (*cursor == 0), give up.
1260 * Otherwise, reset the cursor to the beginning and try again.
1262 if (*cursor == 0)
1263 return (-1ULL);
1265 *cursor = 0;
1266 return (metaslab_block_picker(t, cursor, size, align));
1270 * ==========================================================================
1271 * The first-fit block allocator
1272 * ==========================================================================
1274 static uint64_t
1275 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1278 * Find the largest power of 2 block size that evenly divides the
1279 * requested size. This is used to try to allocate blocks with similar
1280 * alignment from the same area of the metaslab (i.e. same cursor
1281 * bucket) but it does not guarantee that other allocations sizes
1282 * may exist in the same region.
1284 uint64_t align = size & -size;
1285 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1286 avl_tree_t *t = &msp->ms_allocatable->rt_root;
1288 return (metaslab_block_picker(t, cursor, size, align));
1291 static metaslab_ops_t metaslab_ff_ops = {
1292 metaslab_ff_alloc
1296 * ==========================================================================
1297 * Dynamic block allocator -
1298 * Uses the first fit allocation scheme until space get low and then
1299 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1300 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1301 * ==========================================================================
1303 static uint64_t
1304 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1307 * Find the largest power of 2 block size that evenly divides the
1308 * requested size. This is used to try to allocate blocks with similar
1309 * alignment from the same area of the metaslab (i.e. same cursor
1310 * bucket) but it does not guarantee that other allocations sizes
1311 * may exist in the same region.
1313 uint64_t align = size & -size;
1314 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1315 range_tree_t *rt = msp->ms_allocatable;
1316 avl_tree_t *t = &rt->rt_root;
1317 uint64_t max_size = metaslab_block_maxsize(msp);
1318 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1320 ASSERT(MUTEX_HELD(&msp->ms_lock));
1321 ASSERT3U(avl_numnodes(t), ==,
1322 avl_numnodes(&msp->ms_allocatable_by_size));
1324 if (max_size < size)
1325 return (-1ULL);
1328 * If we're running low on space switch to using the size
1329 * sorted AVL tree (best-fit).
1331 if (max_size < metaslab_df_alloc_threshold ||
1332 free_pct < metaslab_df_free_pct) {
1333 t = &msp->ms_allocatable_by_size;
1334 *cursor = 0;
1337 return (metaslab_block_picker(t, cursor, size, 1ULL));
1340 static metaslab_ops_t metaslab_df_ops = {
1341 metaslab_df_alloc
1345 * ==========================================================================
1346 * Cursor fit block allocator -
1347 * Select the largest region in the metaslab, set the cursor to the beginning
1348 * of the range and the cursor_end to the end of the range. As allocations
1349 * are made advance the cursor. Continue allocating from the cursor until
1350 * the range is exhausted and then find a new range.
1351 * ==========================================================================
1353 static uint64_t
1354 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1356 range_tree_t *rt = msp->ms_allocatable;
1357 avl_tree_t *t = &msp->ms_allocatable_by_size;
1358 uint64_t *cursor = &msp->ms_lbas[0];
1359 uint64_t *cursor_end = &msp->ms_lbas[1];
1360 uint64_t offset = 0;
1362 ASSERT(MUTEX_HELD(&msp->ms_lock));
1363 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1365 ASSERT3U(*cursor_end, >=, *cursor);
1367 if ((*cursor + size) > *cursor_end) {
1368 range_seg_t *rs;
1370 rs = avl_last(&msp->ms_allocatable_by_size);
1371 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1372 return (-1ULL);
1374 *cursor = rs->rs_start;
1375 *cursor_end = rs->rs_end;
1378 offset = *cursor;
1379 *cursor += size;
1381 return (offset);
1384 static metaslab_ops_t metaslab_cf_ops = {
1385 metaslab_cf_alloc
1389 * ==========================================================================
1390 * New dynamic fit allocator -
1391 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1392 * contiguous blocks. If no region is found then just use the largest segment
1393 * that remains.
1394 * ==========================================================================
1398 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1399 * to request from the allocator.
1401 uint64_t metaslab_ndf_clump_shift = 4;
1403 static uint64_t
1404 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1406 avl_tree_t *t = &msp->ms_allocatable->rt_root;
1407 avl_index_t where;
1408 range_seg_t *rs, rsearch;
1409 uint64_t hbit = highbit64(size);
1410 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1411 uint64_t max_size = metaslab_block_maxsize(msp);
1413 ASSERT(MUTEX_HELD(&msp->ms_lock));
1414 ASSERT3U(avl_numnodes(t), ==,
1415 avl_numnodes(&msp->ms_allocatable_by_size));
1417 if (max_size < size)
1418 return (-1ULL);
1420 rsearch.rs_start = *cursor;
1421 rsearch.rs_end = *cursor + size;
1423 rs = avl_find(t, &rsearch, &where);
1424 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1425 t = &msp->ms_allocatable_by_size;
1427 rsearch.rs_start = 0;
1428 rsearch.rs_end = MIN(max_size,
1429 1ULL << (hbit + metaslab_ndf_clump_shift));
1430 rs = avl_find(t, &rsearch, &where);
1431 if (rs == NULL)
1432 rs = avl_nearest(t, where, AVL_AFTER);
1433 ASSERT(rs != NULL);
1436 if ((rs->rs_end - rs->rs_start) >= size) {
1437 *cursor = rs->rs_start + size;
1438 return (rs->rs_start);
1440 return (-1ULL);
1443 static metaslab_ops_t metaslab_ndf_ops = {
1444 metaslab_ndf_alloc
1447 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1450 * ==========================================================================
1451 * Metaslabs
1452 * ==========================================================================
1456 * Wait for any in-progress metaslab loads to complete.
1458 void
1459 metaslab_load_wait(metaslab_t *msp)
1461 ASSERT(MUTEX_HELD(&msp->ms_lock));
1463 while (msp->ms_loading) {
1464 ASSERT(!msp->ms_loaded);
1465 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1470 metaslab_load(metaslab_t *msp)
1472 int error = 0;
1473 boolean_t success = B_FALSE;
1475 ASSERT(MUTEX_HELD(&msp->ms_lock));
1476 ASSERT(!msp->ms_loaded);
1477 ASSERT(!msp->ms_loading);
1479 msp->ms_loading = B_TRUE;
1481 * Nobody else can manipulate a loading metaslab, so it's now safe
1482 * to drop the lock. This way we don't have to hold the lock while
1483 * reading the spacemap from disk.
1485 mutex_exit(&msp->ms_lock);
1488 * If the space map has not been allocated yet, then treat
1489 * all the space in the metaslab as free and add it to ms_allocatable.
1491 if (msp->ms_sm != NULL) {
1492 error = space_map_load(msp->ms_sm, msp->ms_allocatable,
1493 SM_FREE);
1494 } else {
1495 range_tree_add(msp->ms_allocatable,
1496 msp->ms_start, msp->ms_size);
1499 success = (error == 0);
1501 mutex_enter(&msp->ms_lock);
1502 msp->ms_loading = B_FALSE;
1504 if (success) {
1505 ASSERT3P(msp->ms_group, !=, NULL);
1506 msp->ms_loaded = B_TRUE;
1509 * If the metaslab already has a spacemap, then we need to
1510 * remove all segments from the defer tree; otherwise, the
1511 * metaslab is completely empty and we can skip this.
1513 if (msp->ms_sm != NULL) {
1514 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1515 range_tree_walk(msp->ms_defer[t],
1516 range_tree_remove, msp->ms_allocatable);
1519 msp->ms_max_size = metaslab_block_maxsize(msp);
1521 cv_broadcast(&msp->ms_load_cv);
1522 return (error);
1525 void
1526 metaslab_unload(metaslab_t *msp)
1528 ASSERT(MUTEX_HELD(&msp->ms_lock));
1529 range_tree_vacate(msp->ms_allocatable, NULL, NULL);
1530 msp->ms_loaded = B_FALSE;
1531 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1532 msp->ms_max_size = 0;
1536 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1537 metaslab_t **msp)
1539 vdev_t *vd = mg->mg_vd;
1540 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1541 metaslab_t *ms;
1542 int error;
1544 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1545 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1546 mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL);
1547 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1549 ms->ms_id = id;
1550 ms->ms_start = id << vd->vdev_ms_shift;
1551 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1552 ms->ms_allocator = -1;
1553 ms->ms_new = B_TRUE;
1556 * We only open space map objects that already exist. All others
1557 * will be opened when we finally allocate an object for it.
1559 if (object != 0) {
1560 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1561 ms->ms_size, vd->vdev_ashift);
1563 if (error != 0) {
1564 kmem_free(ms, sizeof (metaslab_t));
1565 return (error);
1568 ASSERT(ms->ms_sm != NULL);
1572 * We create the main range tree here, but we don't create the
1573 * other range trees until metaslab_sync_done(). This serves
1574 * two purposes: it allows metaslab_sync_done() to detect the
1575 * addition of new space; and for debugging, it ensures that we'd
1576 * data fault on any attempt to use this metaslab before it's ready.
1578 ms->ms_allocatable = range_tree_create(&metaslab_rt_ops, ms);
1579 metaslab_group_add(mg, ms);
1581 metaslab_set_fragmentation(ms);
1584 * If we're opening an existing pool (txg == 0) or creating
1585 * a new one (txg == TXG_INITIAL), all space is available now.
1586 * If we're adding space to an existing pool, the new space
1587 * does not become available until after this txg has synced.
1588 * The metaslab's weight will also be initialized when we sync
1589 * out this txg. This ensures that we don't attempt to allocate
1590 * from it before we have initialized it completely.
1592 if (txg <= TXG_INITIAL)
1593 metaslab_sync_done(ms, 0);
1596 * If metaslab_debug_load is set and we're initializing a metaslab
1597 * that has an allocated space map object then load the its space
1598 * map so that can verify frees.
1600 if (metaslab_debug_load && ms->ms_sm != NULL) {
1601 mutex_enter(&ms->ms_lock);
1602 VERIFY0(metaslab_load(ms));
1603 mutex_exit(&ms->ms_lock);
1606 if (txg != 0) {
1607 vdev_dirty(vd, 0, NULL, txg);
1608 vdev_dirty(vd, VDD_METASLAB, ms, txg);
1611 *msp = ms;
1613 return (0);
1616 void
1617 metaslab_fini(metaslab_t *msp)
1619 metaslab_group_t *mg = msp->ms_group;
1621 metaslab_group_remove(mg, msp);
1623 mutex_enter(&msp->ms_lock);
1624 VERIFY(msp->ms_group == NULL);
1625 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1626 0, -msp->ms_size);
1627 space_map_close(msp->ms_sm);
1629 metaslab_unload(msp);
1630 range_tree_destroy(msp->ms_allocatable);
1631 range_tree_destroy(msp->ms_freeing);
1632 range_tree_destroy(msp->ms_freed);
1634 for (int t = 0; t < TXG_SIZE; t++) {
1635 range_tree_destroy(msp->ms_allocating[t]);
1638 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1639 range_tree_destroy(msp->ms_defer[t]);
1641 ASSERT0(msp->ms_deferspace);
1643 range_tree_destroy(msp->ms_checkpointing);
1645 mutex_exit(&msp->ms_lock);
1646 cv_destroy(&msp->ms_load_cv);
1647 mutex_destroy(&msp->ms_lock);
1648 mutex_destroy(&msp->ms_sync_lock);
1649 ASSERT3U(msp->ms_allocator, ==, -1);
1651 kmem_free(msp, sizeof (metaslab_t));
1654 #define FRAGMENTATION_TABLE_SIZE 17
1657 * This table defines a segment size based fragmentation metric that will
1658 * allow each metaslab to derive its own fragmentation value. This is done
1659 * by calculating the space in each bucket of the spacemap histogram and
1660 * multiplying that by the fragmetation metric in this table. Doing
1661 * this for all buckets and dividing it by the total amount of free
1662 * space in this metaslab (i.e. the total free space in all buckets) gives
1663 * us the fragmentation metric. This means that a high fragmentation metric
1664 * equates to most of the free space being comprised of small segments.
1665 * Conversely, if the metric is low, then most of the free space is in
1666 * large segments. A 10% change in fragmentation equates to approximately
1667 * double the number of segments.
1669 * This table defines 0% fragmented space using 16MB segments. Testing has
1670 * shown that segments that are greater than or equal to 16MB do not suffer
1671 * from drastic performance problems. Using this value, we derive the rest
1672 * of the table. Since the fragmentation value is never stored on disk, it
1673 * is possible to change these calculations in the future.
1675 int zfs_frag_table[FRAGMENTATION_TABLE_SIZE] = {
1676 100, /* 512B */
1677 100, /* 1K */
1678 98, /* 2K */
1679 95, /* 4K */
1680 90, /* 8K */
1681 80, /* 16K */
1682 70, /* 32K */
1683 60, /* 64K */
1684 50, /* 128K */
1685 40, /* 256K */
1686 30, /* 512K */
1687 20, /* 1M */
1688 15, /* 2M */
1689 10, /* 4M */
1690 5, /* 8M */
1691 0 /* 16M */
1695 * Calclate the metaslab's fragmentation metric. A return value
1696 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1697 * not support this metric. Otherwise, the return value should be in the
1698 * range [0, 100].
1700 static void
1701 metaslab_set_fragmentation(metaslab_t *msp)
1703 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1704 uint64_t fragmentation = 0;
1705 uint64_t total = 0;
1706 boolean_t feature_enabled = spa_feature_is_enabled(spa,
1707 SPA_FEATURE_SPACEMAP_HISTOGRAM);
1709 if (!feature_enabled) {
1710 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1711 return;
1715 * A null space map means that the entire metaslab is free
1716 * and thus is not fragmented.
1718 if (msp->ms_sm == NULL) {
1719 msp->ms_fragmentation = 0;
1720 return;
1724 * If this metaslab's space map has not been upgraded, flag it
1725 * so that we upgrade next time we encounter it.
1727 if (msp->ms_sm->sm_dbuf->db_size != sizeof (space_map_phys_t)) {
1728 uint64_t txg = spa_syncing_txg(spa);
1729 vdev_t *vd = msp->ms_group->mg_vd;
1732 * If we've reached the final dirty txg, then we must
1733 * be shutting down the pool. We don't want to dirty
1734 * any data past this point so skip setting the condense
1735 * flag. We can retry this action the next time the pool
1736 * is imported.
1738 if (spa_writeable(spa) && txg < spa_final_dirty_txg(spa)) {
1739 msp->ms_condense_wanted = B_TRUE;
1740 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1741 zfs_dbgmsg("txg %llu, requesting force condense: "
1742 "ms_id %llu, vdev_id %llu", txg, msp->ms_id,
1743 vd->vdev_id);
1745 msp->ms_fragmentation = ZFS_FRAG_INVALID;
1746 return;
1749 for (int i = 0; i < SPACE_MAP_HISTOGRAM_SIZE; i++) {
1750 uint64_t space = 0;
1751 uint8_t shift = msp->ms_sm->sm_shift;
1753 int idx = MIN(shift - SPA_MINBLOCKSHIFT + i,
1754 FRAGMENTATION_TABLE_SIZE - 1);
1756 if (msp->ms_sm->sm_phys->smp_histogram[i] == 0)
1757 continue;
1759 space = msp->ms_sm->sm_phys->smp_histogram[i] << (i + shift);
1760 total += space;
1762 ASSERT3U(idx, <, FRAGMENTATION_TABLE_SIZE);
1763 fragmentation += space * zfs_frag_table[idx];
1766 if (total > 0)
1767 fragmentation /= total;
1768 ASSERT3U(fragmentation, <=, 100);
1770 msp->ms_fragmentation = fragmentation;
1774 * Compute a weight -- a selection preference value -- for the given metaslab.
1775 * This is based on the amount of free space, the level of fragmentation,
1776 * the LBA range, and whether the metaslab is loaded.
1778 static uint64_t
1779 metaslab_space_weight(metaslab_t *msp)
1781 metaslab_group_t *mg = msp->ms_group;
1782 vdev_t *vd = mg->mg_vd;
1783 uint64_t weight, space;
1785 ASSERT(MUTEX_HELD(&msp->ms_lock));
1786 ASSERT(!vd->vdev_removing);
1789 * The baseline weight is the metaslab's free space.
1791 space = msp->ms_size - space_map_allocated(msp->ms_sm);
1793 if (metaslab_fragmentation_factor_enabled &&
1794 msp->ms_fragmentation != ZFS_FRAG_INVALID) {
1796 * Use the fragmentation information to inversely scale
1797 * down the baseline weight. We need to ensure that we
1798 * don't exclude this metaslab completely when it's 100%
1799 * fragmented. To avoid this we reduce the fragmented value
1800 * by 1.
1802 space = (space * (100 - (msp->ms_fragmentation - 1))) / 100;
1805 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1806 * this metaslab again. The fragmentation metric may have
1807 * decreased the space to something smaller than
1808 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1809 * so that we can consume any remaining space.
1811 if (space > 0 && space < SPA_MINBLOCKSIZE)
1812 space = SPA_MINBLOCKSIZE;
1814 weight = space;
1817 * Modern disks have uniform bit density and constant angular velocity.
1818 * Therefore, the outer recording zones are faster (higher bandwidth)
1819 * than the inner zones by the ratio of outer to inner track diameter,
1820 * which is typically around 2:1. We account for this by assigning
1821 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1822 * In effect, this means that we'll select the metaslab with the most
1823 * free bandwidth rather than simply the one with the most free space.
1825 if (metaslab_lba_weighting_enabled) {
1826 weight = 2 * weight - (msp->ms_id * weight) / vd->vdev_ms_count;
1827 ASSERT(weight >= space && weight <= 2 * space);
1831 * If this metaslab is one we're actively using, adjust its
1832 * weight to make it preferable to any inactive metaslab so
1833 * we'll polish it off. If the fragmentation on this metaslab
1834 * has exceed our threshold, then don't mark it active.
1836 if (msp->ms_loaded && msp->ms_fragmentation != ZFS_FRAG_INVALID &&
1837 msp->ms_fragmentation <= zfs_metaslab_fragmentation_threshold) {
1838 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
1841 WEIGHT_SET_SPACEBASED(weight);
1842 return (weight);
1846 * Return the weight of the specified metaslab, according to the segment-based
1847 * weighting algorithm. The metaslab must be loaded. This function can
1848 * be called within a sync pass since it relies only on the metaslab's
1849 * range tree which is always accurate when the metaslab is loaded.
1851 static uint64_t
1852 metaslab_weight_from_range_tree(metaslab_t *msp)
1854 uint64_t weight = 0;
1855 uint32_t segments = 0;
1857 ASSERT(msp->ms_loaded);
1859 for (int i = RANGE_TREE_HISTOGRAM_SIZE - 1; i >= SPA_MINBLOCKSHIFT;
1860 i--) {
1861 uint8_t shift = msp->ms_group->mg_vd->vdev_ashift;
1862 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1864 segments <<= 1;
1865 segments += msp->ms_allocatable->rt_histogram[i];
1868 * The range tree provides more precision than the space map
1869 * and must be downgraded so that all values fit within the
1870 * space map's histogram. This allows us to compare loaded
1871 * vs. unloaded metaslabs to determine which metaslab is
1872 * considered "best".
1874 if (i > max_idx)
1875 continue;
1877 if (segments != 0) {
1878 WEIGHT_SET_COUNT(weight, segments);
1879 WEIGHT_SET_INDEX(weight, i);
1880 WEIGHT_SET_ACTIVE(weight, 0);
1881 break;
1884 return (weight);
1888 * Calculate the weight based on the on-disk histogram. This should only
1889 * be called after a sync pass has completely finished since the on-disk
1890 * information is updated in metaslab_sync().
1892 static uint64_t
1893 metaslab_weight_from_spacemap(metaslab_t *msp)
1895 uint64_t weight = 0;
1897 for (int i = SPACE_MAP_HISTOGRAM_SIZE - 1; i >= 0; i--) {
1898 if (msp->ms_sm->sm_phys->smp_histogram[i] != 0) {
1899 WEIGHT_SET_COUNT(weight,
1900 msp->ms_sm->sm_phys->smp_histogram[i]);
1901 WEIGHT_SET_INDEX(weight, i +
1902 msp->ms_sm->sm_shift);
1903 WEIGHT_SET_ACTIVE(weight, 0);
1904 break;
1907 return (weight);
1911 * Compute a segment-based weight for the specified metaslab. The weight
1912 * is determined by highest bucket in the histogram. The information
1913 * for the highest bucket is encoded into the weight value.
1915 static uint64_t
1916 metaslab_segment_weight(metaslab_t *msp)
1918 metaslab_group_t *mg = msp->ms_group;
1919 uint64_t weight = 0;
1920 uint8_t shift = mg->mg_vd->vdev_ashift;
1922 ASSERT(MUTEX_HELD(&msp->ms_lock));
1925 * The metaslab is completely free.
1927 if (space_map_allocated(msp->ms_sm) == 0) {
1928 int idx = highbit64(msp->ms_size) - 1;
1929 int max_idx = SPACE_MAP_HISTOGRAM_SIZE + shift - 1;
1931 if (idx < max_idx) {
1932 WEIGHT_SET_COUNT(weight, 1ULL);
1933 WEIGHT_SET_INDEX(weight, idx);
1934 } else {
1935 WEIGHT_SET_COUNT(weight, 1ULL << (idx - max_idx));
1936 WEIGHT_SET_INDEX(weight, max_idx);
1938 WEIGHT_SET_ACTIVE(weight, 0);
1939 ASSERT(!WEIGHT_IS_SPACEBASED(weight));
1941 return (weight);
1944 ASSERT3U(msp->ms_sm->sm_dbuf->db_size, ==, sizeof (space_map_phys_t));
1947 * If the metaslab is fully allocated then just make the weight 0.
1949 if (space_map_allocated(msp->ms_sm) == msp->ms_size)
1950 return (0);
1952 * If the metaslab is already loaded, then use the range tree to
1953 * determine the weight. Otherwise, we rely on the space map information
1954 * to generate the weight.
1956 if (msp->ms_loaded) {
1957 weight = metaslab_weight_from_range_tree(msp);
1958 } else {
1959 weight = metaslab_weight_from_spacemap(msp);
1963 * If the metaslab was active the last time we calculated its weight
1964 * then keep it active. We want to consume the entire region that
1965 * is associated with this weight.
1967 if (msp->ms_activation_weight != 0 && weight != 0)
1968 WEIGHT_SET_ACTIVE(weight, WEIGHT_GET_ACTIVE(msp->ms_weight));
1969 return (weight);
1973 * Determine if we should attempt to allocate from this metaslab. If the
1974 * metaslab has a maximum size then we can quickly determine if the desired
1975 * allocation size can be satisfied. Otherwise, if we're using segment-based
1976 * weighting then we can determine the maximum allocation that this metaslab
1977 * can accommodate based on the index encoded in the weight. If we're using
1978 * space-based weights then rely on the entire weight (excluding the weight
1979 * type bit).
1981 boolean_t
1982 metaslab_should_allocate(metaslab_t *msp, uint64_t asize)
1984 boolean_t should_allocate;
1986 if (msp->ms_max_size != 0)
1987 return (msp->ms_max_size >= asize);
1989 if (!WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
1991 * The metaslab segment weight indicates segments in the
1992 * range [2^i, 2^(i+1)), where i is the index in the weight.
1993 * Since the asize might be in the middle of the range, we
1994 * should attempt the allocation if asize < 2^(i+1).
1996 should_allocate = (asize <
1997 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1998 } else {
1999 should_allocate = (asize <=
2000 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
2002 return (should_allocate);
2005 static uint64_t
2006 metaslab_weight(metaslab_t *msp)
2008 vdev_t *vd = msp->ms_group->mg_vd;
2009 spa_t *spa = vd->vdev_spa;
2010 uint64_t weight;
2012 ASSERT(MUTEX_HELD(&msp->ms_lock));
2015 * If this vdev is in the process of being removed, there is nothing
2016 * for us to do here.
2018 if (vd->vdev_removing)
2019 return (0);
2021 metaslab_set_fragmentation(msp);
2024 * Update the maximum size if the metaslab is loaded. This will
2025 * ensure that we get an accurate maximum size if newly freed space
2026 * has been added back into the free tree.
2028 if (msp->ms_loaded)
2029 msp->ms_max_size = metaslab_block_maxsize(msp);
2032 * Segment-based weighting requires space map histogram support.
2034 if (zfs_metaslab_segment_weight_enabled &&
2035 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
2036 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
2037 sizeof (space_map_phys_t))) {
2038 weight = metaslab_segment_weight(msp);
2039 } else {
2040 weight = metaslab_space_weight(msp);
2042 return (weight);
2045 static int
2046 metaslab_activate_allocator(metaslab_group_t *mg, metaslab_t *msp,
2047 int allocator, uint64_t activation_weight)
2050 * If we're activating for the claim code, we don't want to actually
2051 * set the metaslab up for a specific allocator.
2053 if (activation_weight == METASLAB_WEIGHT_CLAIM)
2054 return (0);
2055 metaslab_t **arr = (activation_weight == METASLAB_WEIGHT_PRIMARY ?
2056 mg->mg_primaries : mg->mg_secondaries);
2058 ASSERT(MUTEX_HELD(&msp->ms_lock));
2059 mutex_enter(&mg->mg_lock);
2060 if (arr[allocator] != NULL) {
2061 mutex_exit(&mg->mg_lock);
2062 return (EEXIST);
2065 arr[allocator] = msp;
2066 ASSERT3S(msp->ms_allocator, ==, -1);
2067 msp->ms_allocator = allocator;
2068 msp->ms_primary = (activation_weight == METASLAB_WEIGHT_PRIMARY);
2069 mutex_exit(&mg->mg_lock);
2071 return (0);
2074 static int
2075 metaslab_activate(metaslab_t *msp, int allocator, uint64_t activation_weight)
2077 ASSERT(MUTEX_HELD(&msp->ms_lock));
2079 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
2080 int error = 0;
2081 metaslab_load_wait(msp);
2082 if (!msp->ms_loaded) {
2083 if ((error = metaslab_load(msp)) != 0) {
2084 metaslab_group_sort(msp->ms_group, msp, 0);
2085 return (error);
2088 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) != 0) {
2090 * The metaslab was activated for another allocator
2091 * while we were waiting, we should reselect.
2093 return (EBUSY);
2095 if ((error = metaslab_activate_allocator(msp->ms_group, msp,
2096 allocator, activation_weight)) != 0) {
2097 return (error);
2100 msp->ms_activation_weight = msp->ms_weight;
2101 metaslab_group_sort(msp->ms_group, msp,
2102 msp->ms_weight | activation_weight);
2104 ASSERT(msp->ms_loaded);
2105 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
2107 return (0);
2110 static void
2111 metaslab_passivate_allocator(metaslab_group_t *mg, metaslab_t *msp,
2112 uint64_t weight)
2114 ASSERT(MUTEX_HELD(&msp->ms_lock));
2115 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM) {
2116 metaslab_group_sort(mg, msp, weight);
2117 return;
2120 mutex_enter(&mg->mg_lock);
2121 ASSERT3P(msp->ms_group, ==, mg);
2122 if (msp->ms_primary) {
2123 ASSERT3U(0, <=, msp->ms_allocator);
2124 ASSERT3U(msp->ms_allocator, <, mg->mg_allocators);
2125 ASSERT3P(mg->mg_primaries[msp->ms_allocator], ==, msp);
2126 ASSERT(msp->ms_weight & METASLAB_WEIGHT_PRIMARY);
2127 mg->mg_primaries[msp->ms_allocator] = NULL;
2128 } else {
2129 ASSERT(msp->ms_weight & METASLAB_WEIGHT_SECONDARY);
2130 ASSERT3P(mg->mg_secondaries[msp->ms_allocator], ==, msp);
2131 mg->mg_secondaries[msp->ms_allocator] = NULL;
2133 msp->ms_allocator = -1;
2134 metaslab_group_sort_impl(mg, msp, weight);
2135 mutex_exit(&mg->mg_lock);
2138 static void
2139 metaslab_passivate(metaslab_t *msp, uint64_t weight)
2141 uint64_t size = weight & ~METASLAB_WEIGHT_TYPE;
2144 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
2145 * this metaslab again. In that case, it had better be empty,
2146 * or we would be leaving space on the table.
2148 ASSERT(size >= SPA_MINBLOCKSIZE ||
2149 range_tree_is_empty(msp->ms_allocatable));
2150 ASSERT0(weight & METASLAB_ACTIVE_MASK);
2152 msp->ms_activation_weight = 0;
2153 metaslab_passivate_allocator(msp->ms_group, msp, weight);
2154 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
2158 * Segment-based metaslabs are activated once and remain active until
2159 * we either fail an allocation attempt (similar to space-based metaslabs)
2160 * or have exhausted the free space in zfs_metaslab_switch_threshold
2161 * buckets since the metaslab was activated. This function checks to see
2162 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
2163 * metaslab and passivates it proactively. This will allow us to select a
2164 * metaslabs with larger contiguous region if any remaining within this
2165 * metaslab group. If we're in sync pass > 1, then we continue using this
2166 * metaslab so that we don't dirty more block and cause more sync passes.
2168 void
2169 metaslab_segment_may_passivate(metaslab_t *msp)
2171 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2173 if (WEIGHT_IS_SPACEBASED(msp->ms_weight) || spa_sync_pass(spa) > 1)
2174 return;
2177 * Since we are in the middle of a sync pass, the most accurate
2178 * information that is accessible to us is the in-core range tree
2179 * histogram; calculate the new weight based on that information.
2181 uint64_t weight = metaslab_weight_from_range_tree(msp);
2182 int activation_idx = WEIGHT_GET_INDEX(msp->ms_activation_weight);
2183 int current_idx = WEIGHT_GET_INDEX(weight);
2185 if (current_idx <= activation_idx - zfs_metaslab_switch_threshold)
2186 metaslab_passivate(msp, weight);
2189 static void
2190 metaslab_preload(void *arg)
2192 metaslab_t *msp = arg;
2193 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
2195 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
2197 mutex_enter(&msp->ms_lock);
2198 metaslab_load_wait(msp);
2199 if (!msp->ms_loaded)
2200 (void) metaslab_load(msp);
2201 msp->ms_selected_txg = spa_syncing_txg(spa);
2202 mutex_exit(&msp->ms_lock);
2205 static void
2206 metaslab_group_preload(metaslab_group_t *mg)
2208 spa_t *spa = mg->mg_vd->vdev_spa;
2209 metaslab_t *msp;
2210 avl_tree_t *t = &mg->mg_metaslab_tree;
2211 int m = 0;
2213 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
2214 taskq_wait(mg->mg_taskq);
2215 return;
2218 mutex_enter(&mg->mg_lock);
2221 * Load the next potential metaslabs
2223 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2224 ASSERT3P(msp->ms_group, ==, mg);
2227 * We preload only the maximum number of metaslabs specified
2228 * by metaslab_preload_limit. If a metaslab is being forced
2229 * to condense then we preload it too. This will ensure
2230 * that force condensing happens in the next txg.
2232 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2233 continue;
2236 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2237 msp, TQ_SLEEP) != (uintptr_t)NULL);
2239 mutex_exit(&mg->mg_lock);
2243 * Determine if the space map's on-disk footprint is past our tolerance
2244 * for inefficiency. We would like to use the following criteria to make
2245 * our decision:
2247 * 1. The size of the space map object should not dramatically increase as a
2248 * result of writing out the free space range tree.
2250 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2251 * times the size than the free space range tree representation
2252 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1MB).
2254 * 3. The on-disk size of the space map should actually decrease.
2256 * Unfortunately, we cannot compute the on-disk size of the space map in this
2257 * context because we cannot accurately compute the effects of compression, etc.
2258 * Instead, we apply the heuristic described in the block comment for
2259 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2260 * is greater than a threshold number of blocks.
2262 static boolean_t
2263 metaslab_should_condense(metaslab_t *msp)
2265 space_map_t *sm = msp->ms_sm;
2266 vdev_t *vd = msp->ms_group->mg_vd;
2267 uint64_t vdev_blocksize = 1 << vd->vdev_ashift;
2268 uint64_t current_txg = spa_syncing_txg(vd->vdev_spa);
2270 ASSERT(MUTEX_HELD(&msp->ms_lock));
2271 ASSERT(msp->ms_loaded);
2274 * Allocations and frees in early passes are generally more space
2275 * efficient (in terms of blocks described in space map entries)
2276 * than the ones in later passes (e.g. we don't compress after
2277 * sync pass 5) and condensing a metaslab multiple times in a txg
2278 * could degrade performance.
2280 * Thus we prefer condensing each metaslab at most once every txg at
2281 * the earliest sync pass possible. If a metaslab is eligible for
2282 * condensing again after being considered for condensing within the
2283 * same txg, it will hopefully be dirty in the next txg where it will
2284 * be condensed at an earlier pass.
2286 if (msp->ms_condense_checked_txg == current_txg)
2287 return (B_FALSE);
2288 msp->ms_condense_checked_txg = current_txg;
2291 * We always condense metaslabs that are empty and metaslabs for
2292 * which a condense request has been made.
2294 if (avl_is_empty(&msp->ms_allocatable_by_size) ||
2295 msp->ms_condense_wanted)
2296 return (B_TRUE);
2298 uint64_t object_size = space_map_length(msp->ms_sm);
2299 uint64_t optimal_size = space_map_estimate_optimal_size(sm,
2300 msp->ms_allocatable, SM_NO_VDEVID);
2302 dmu_object_info_t doi;
2303 dmu_object_info_from_db(sm->sm_dbuf, &doi);
2304 uint64_t record_size = MAX(doi.doi_data_block_size, vdev_blocksize);
2306 return (object_size >= (optimal_size * zfs_condense_pct / 100) &&
2307 object_size > zfs_metaslab_condense_block_threshold * record_size);
2311 * Condense the on-disk space map representation to its minimized form.
2312 * The minimized form consists of a small number of allocations followed by
2313 * the entries of the free range tree.
2315 static void
2316 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
2318 range_tree_t *condense_tree;
2319 space_map_t *sm = msp->ms_sm;
2321 ASSERT(MUTEX_HELD(&msp->ms_lock));
2322 ASSERT(msp->ms_loaded);
2324 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2325 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2326 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2327 msp->ms_group->mg_vd->vdev_spa->spa_name,
2328 space_map_length(msp->ms_sm),
2329 avl_numnodes(&msp->ms_allocatable->rt_root),
2330 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2332 msp->ms_condense_wanted = B_FALSE;
2335 * Create an range tree that is 100% allocated. We remove segments
2336 * that have been freed in this txg, any deferred frees that exist,
2337 * and any allocation in the future. Removing segments should be
2338 * a relatively inexpensive operation since we expect these trees to
2339 * have a small number of nodes.
2341 condense_tree = range_tree_create(NULL, NULL);
2342 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2344 range_tree_walk(msp->ms_freeing, range_tree_remove, condense_tree);
2345 range_tree_walk(msp->ms_freed, range_tree_remove, condense_tree);
2347 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2348 range_tree_walk(msp->ms_defer[t],
2349 range_tree_remove, condense_tree);
2352 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2353 range_tree_walk(msp->ms_allocating[(txg + t) & TXG_MASK],
2354 range_tree_remove, condense_tree);
2358 * We're about to drop the metaslab's lock thus allowing
2359 * other consumers to change it's content. Set the
2360 * metaslab's ms_condensing flag to ensure that
2361 * allocations on this metaslab do not occur while we're
2362 * in the middle of committing it to disk. This is only critical
2363 * for ms_allocatable as all other range trees use per txg
2364 * views of their content.
2366 msp->ms_condensing = B_TRUE;
2368 mutex_exit(&msp->ms_lock);
2369 space_map_truncate(sm, zfs_metaslab_sm_blksz, tx);
2372 * While we would ideally like to create a space map representation
2373 * that consists only of allocation records, doing so can be
2374 * prohibitively expensive because the in-core free tree can be
2375 * large, and therefore computationally expensive to subtract
2376 * from the condense_tree. Instead we sync out two trees, a cheap
2377 * allocation only tree followed by the in-core free tree. While not
2378 * optimal, this is typically close to optimal, and much cheaper to
2379 * compute.
2381 space_map_write(sm, condense_tree, SM_ALLOC, SM_NO_VDEVID, tx);
2382 range_tree_vacate(condense_tree, NULL, NULL);
2383 range_tree_destroy(condense_tree);
2385 space_map_write(sm, msp->ms_allocatable, SM_FREE, SM_NO_VDEVID, tx);
2386 mutex_enter(&msp->ms_lock);
2387 msp->ms_condensing = B_FALSE;
2391 * Write a metaslab to disk in the context of the specified transaction group.
2393 void
2394 metaslab_sync(metaslab_t *msp, uint64_t txg)
2396 metaslab_group_t *mg = msp->ms_group;
2397 vdev_t *vd = mg->mg_vd;
2398 spa_t *spa = vd->vdev_spa;
2399 objset_t *mos = spa_meta_objset(spa);
2400 range_tree_t *alloctree = msp->ms_allocating[txg & TXG_MASK];
2401 dmu_tx_t *tx;
2402 uint64_t object = space_map_object(msp->ms_sm);
2404 ASSERT(!vd->vdev_ishole);
2407 * This metaslab has just been added so there's no work to do now.
2409 if (msp->ms_freeing == NULL) {
2410 ASSERT3P(alloctree, ==, NULL);
2411 return;
2414 ASSERT3P(alloctree, !=, NULL);
2415 ASSERT3P(msp->ms_freeing, !=, NULL);
2416 ASSERT3P(msp->ms_freed, !=, NULL);
2417 ASSERT3P(msp->ms_checkpointing, !=, NULL);
2420 * Normally, we don't want to process a metaslab if there are no
2421 * allocations or frees to perform. However, if the metaslab is being
2422 * forced to condense and it's loaded, we need to let it through.
2424 if (range_tree_is_empty(alloctree) &&
2425 range_tree_is_empty(msp->ms_freeing) &&
2426 range_tree_is_empty(msp->ms_checkpointing) &&
2427 !(msp->ms_loaded && msp->ms_condense_wanted))
2428 return;
2431 VERIFY(txg <= spa_final_dirty_txg(spa));
2434 * The only state that can actually be changing concurrently with
2435 * metaslab_sync() is the metaslab's ms_allocatable. No other
2436 * thread can be modifying this txg's alloc, freeing,
2437 * freed, or space_map_phys_t. We drop ms_lock whenever we
2438 * could call into the DMU, because the DMU can call down to us
2439 * (e.g. via zio_free()) at any time.
2441 * The spa_vdev_remove_thread() can be reading metaslab state
2442 * concurrently, and it is locked out by the ms_sync_lock. Note
2443 * that the ms_lock is insufficient for this, because it is dropped
2444 * by space_map_write().
2446 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2448 if (msp->ms_sm == NULL) {
2449 uint64_t new_object;
2451 new_object = space_map_alloc(mos, zfs_metaslab_sm_blksz, tx);
2452 VERIFY3U(new_object, !=, 0);
2454 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2455 msp->ms_start, msp->ms_size, vd->vdev_ashift));
2456 ASSERT(msp->ms_sm != NULL);
2459 if (!range_tree_is_empty(msp->ms_checkpointing) &&
2460 vd->vdev_checkpoint_sm == NULL) {
2461 ASSERT(spa_has_checkpoint(spa));
2463 uint64_t new_object = space_map_alloc(mos,
2464 vdev_standard_sm_blksz, tx);
2465 VERIFY3U(new_object, !=, 0);
2467 VERIFY0(space_map_open(&vd->vdev_checkpoint_sm,
2468 mos, new_object, 0, vd->vdev_asize, vd->vdev_ashift));
2469 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
2472 * We save the space map object as an entry in vdev_top_zap
2473 * so it can be retrieved when the pool is reopened after an
2474 * export or through zdb.
2476 VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset,
2477 vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM,
2478 sizeof (new_object), 1, &new_object, tx));
2481 mutex_enter(&msp->ms_sync_lock);
2482 mutex_enter(&msp->ms_lock);
2485 * Note: metaslab_condense() clears the space map's histogram.
2486 * Therefore we must verify and remove this histogram before
2487 * condensing.
2489 metaslab_group_histogram_verify(mg);
2490 metaslab_class_histogram_verify(mg->mg_class);
2491 metaslab_group_histogram_remove(mg, msp);
2493 if (msp->ms_loaded && metaslab_should_condense(msp)) {
2494 metaslab_condense(msp, txg, tx);
2495 } else {
2496 mutex_exit(&msp->ms_lock);
2497 space_map_write(msp->ms_sm, alloctree, SM_ALLOC,
2498 SM_NO_VDEVID, tx);
2499 space_map_write(msp->ms_sm, msp->ms_freeing, SM_FREE,
2500 SM_NO_VDEVID, tx);
2501 mutex_enter(&msp->ms_lock);
2504 if (!range_tree_is_empty(msp->ms_checkpointing)) {
2505 ASSERT(spa_has_checkpoint(spa));
2506 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
2509 * Since we are doing writes to disk and the ms_checkpointing
2510 * tree won't be changing during that time, we drop the
2511 * ms_lock while writing to the checkpoint space map.
2513 mutex_exit(&msp->ms_lock);
2514 space_map_write(vd->vdev_checkpoint_sm,
2515 msp->ms_checkpointing, SM_FREE, SM_NO_VDEVID, tx);
2516 mutex_enter(&msp->ms_lock);
2517 space_map_update(vd->vdev_checkpoint_sm);
2519 spa->spa_checkpoint_info.sci_dspace +=
2520 range_tree_space(msp->ms_checkpointing);
2521 vd->vdev_stat.vs_checkpoint_space +=
2522 range_tree_space(msp->ms_checkpointing);
2523 ASSERT3U(vd->vdev_stat.vs_checkpoint_space, ==,
2524 -vd->vdev_checkpoint_sm->sm_alloc);
2526 range_tree_vacate(msp->ms_checkpointing, NULL, NULL);
2529 if (msp->ms_loaded) {
2531 * When the space map is loaded, we have an accurate
2532 * histogram in the range tree. This gives us an opportunity
2533 * to bring the space map's histogram up-to-date so we clear
2534 * it first before updating it.
2536 space_map_histogram_clear(msp->ms_sm);
2537 space_map_histogram_add(msp->ms_sm, msp->ms_allocatable, tx);
2540 * Since we've cleared the histogram we need to add back
2541 * any free space that has already been processed, plus
2542 * any deferred space. This allows the on-disk histogram
2543 * to accurately reflect all free space even if some space
2544 * is not yet available for allocation (i.e. deferred).
2546 space_map_histogram_add(msp->ms_sm, msp->ms_freed, tx);
2549 * Add back any deferred free space that has not been
2550 * added back into the in-core free tree yet. This will
2551 * ensure that we don't end up with a space map histogram
2552 * that is completely empty unless the metaslab is fully
2553 * allocated.
2555 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2556 space_map_histogram_add(msp->ms_sm,
2557 msp->ms_defer[t], tx);
2562 * Always add the free space from this sync pass to the space
2563 * map histogram. We want to make sure that the on-disk histogram
2564 * accounts for all free space. If the space map is not loaded,
2565 * then we will lose some accuracy but will correct it the next
2566 * time we load the space map.
2568 space_map_histogram_add(msp->ms_sm, msp->ms_freeing, tx);
2570 metaslab_group_histogram_add(mg, msp);
2571 metaslab_group_histogram_verify(mg);
2572 metaslab_class_histogram_verify(mg->mg_class);
2575 * For sync pass 1, we avoid traversing this txg's free range tree
2576 * and instead will just swap the pointers for freeing and
2577 * freed. We can safely do this since the freed_tree is
2578 * guaranteed to be empty on the initial pass.
2580 if (spa_sync_pass(spa) == 1) {
2581 range_tree_swap(&msp->ms_freeing, &msp->ms_freed);
2582 } else {
2583 range_tree_vacate(msp->ms_freeing,
2584 range_tree_add, msp->ms_freed);
2586 range_tree_vacate(alloctree, NULL, NULL);
2588 ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
2589 ASSERT0(range_tree_space(msp->ms_allocating[TXG_CLEAN(txg)
2590 & TXG_MASK]));
2591 ASSERT0(range_tree_space(msp->ms_freeing));
2592 ASSERT0(range_tree_space(msp->ms_checkpointing));
2594 mutex_exit(&msp->ms_lock);
2596 if (object != space_map_object(msp->ms_sm)) {
2597 object = space_map_object(msp->ms_sm);
2598 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2599 msp->ms_id, sizeof (uint64_t), &object, tx);
2601 mutex_exit(&msp->ms_sync_lock);
2602 dmu_tx_commit(tx);
2606 * Called after a transaction group has completely synced to mark
2607 * all of the metaslab's free space as usable.
2609 void
2610 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2612 metaslab_group_t *mg = msp->ms_group;
2613 vdev_t *vd = mg->mg_vd;
2614 spa_t *spa = vd->vdev_spa;
2615 range_tree_t **defer_tree;
2616 int64_t alloc_delta, defer_delta;
2617 boolean_t defer_allowed = B_TRUE;
2619 ASSERT(!vd->vdev_ishole);
2621 mutex_enter(&msp->ms_lock);
2624 * If this metaslab is just becoming available, initialize its
2625 * range trees and add its capacity to the vdev.
2627 if (msp->ms_freed == NULL) {
2628 for (int t = 0; t < TXG_SIZE; t++) {
2629 ASSERT(msp->ms_allocating[t] == NULL);
2631 msp->ms_allocating[t] = range_tree_create(NULL, NULL);
2634 ASSERT3P(msp->ms_freeing, ==, NULL);
2635 msp->ms_freeing = range_tree_create(NULL, NULL);
2637 ASSERT3P(msp->ms_freed, ==, NULL);
2638 msp->ms_freed = range_tree_create(NULL, NULL);
2640 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2641 ASSERT(msp->ms_defer[t] == NULL);
2643 msp->ms_defer[t] = range_tree_create(NULL, NULL);
2646 ASSERT3P(msp->ms_checkpointing, ==, NULL);
2647 msp->ms_checkpointing = range_tree_create(NULL, NULL);
2649 vdev_space_update(vd, 0, 0, msp->ms_size);
2651 ASSERT0(range_tree_space(msp->ms_freeing));
2652 ASSERT0(range_tree_space(msp->ms_checkpointing));
2654 defer_tree = &msp->ms_defer[txg % TXG_DEFER_SIZE];
2656 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2657 metaslab_class_get_alloc(spa_normal_class(spa));
2658 if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) {
2659 defer_allowed = B_FALSE;
2662 defer_delta = 0;
2663 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2664 if (defer_allowed) {
2665 defer_delta = range_tree_space(msp->ms_freed) -
2666 range_tree_space(*defer_tree);
2667 } else {
2668 defer_delta -= range_tree_space(*defer_tree);
2671 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2674 * If there's a metaslab_load() in progress, wait for it to complete
2675 * so that we have a consistent view of the in-core space map.
2677 metaslab_load_wait(msp);
2680 * Move the frees from the defer_tree back to the free
2681 * range tree (if it's loaded). Swap the freed_tree and
2682 * the defer_tree -- this is safe to do because we've
2683 * just emptied out the defer_tree.
2685 range_tree_vacate(*defer_tree,
2686 msp->ms_loaded ? range_tree_add : NULL, msp->ms_allocatable);
2687 if (defer_allowed) {
2688 range_tree_swap(&msp->ms_freed, defer_tree);
2689 } else {
2690 range_tree_vacate(msp->ms_freed,
2691 msp->ms_loaded ? range_tree_add : NULL,
2692 msp->ms_allocatable);
2694 space_map_update(msp->ms_sm);
2696 msp->ms_deferspace += defer_delta;
2697 ASSERT3S(msp->ms_deferspace, >=, 0);
2698 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2699 if (msp->ms_deferspace != 0) {
2701 * Keep syncing this metaslab until all deferred frees
2702 * are back in circulation.
2704 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2707 if (msp->ms_new) {
2708 msp->ms_new = B_FALSE;
2709 mutex_enter(&mg->mg_lock);
2710 mg->mg_ms_ready++;
2711 mutex_exit(&mg->mg_lock);
2714 * Calculate the new weights before unloading any metaslabs.
2715 * This will give us the most accurate weighting.
2717 metaslab_group_sort(mg, msp, metaslab_weight(msp) |
2718 (msp->ms_weight & METASLAB_ACTIVE_MASK));
2721 * If the metaslab is loaded and we've not tried to load or allocate
2722 * from it in 'metaslab_unload_delay' txgs, then unload it.
2724 if (msp->ms_loaded &&
2725 msp->ms_initializing == 0 &&
2726 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2727 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2728 VERIFY0(range_tree_space(
2729 msp->ms_allocating[(txg + t) & TXG_MASK]));
2731 if (msp->ms_allocator != -1) {
2732 metaslab_passivate(msp, msp->ms_weight &
2733 ~METASLAB_ACTIVE_MASK);
2736 if (!metaslab_debug_unload)
2737 metaslab_unload(msp);
2740 ASSERT0(range_tree_space(msp->ms_allocating[txg & TXG_MASK]));
2741 ASSERT0(range_tree_space(msp->ms_freeing));
2742 ASSERT0(range_tree_space(msp->ms_freed));
2743 ASSERT0(range_tree_space(msp->ms_checkpointing));
2745 mutex_exit(&msp->ms_lock);
2748 void
2749 metaslab_sync_reassess(metaslab_group_t *mg)
2751 spa_t *spa = mg->mg_class->mc_spa;
2753 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2754 metaslab_group_alloc_update(mg);
2755 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2758 * Preload the next potential metaslabs but only on active
2759 * metaslab groups. We can get into a state where the metaslab
2760 * is no longer active since we dirty metaslabs as we remove a
2761 * a device, thus potentially making the metaslab group eligible
2762 * for preloading.
2764 if (mg->mg_activation_count > 0) {
2765 metaslab_group_preload(mg);
2767 spa_config_exit(spa, SCL_ALLOC, FTAG);
2770 static uint64_t
2771 metaslab_distance(metaslab_t *msp, dva_t *dva)
2773 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2774 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2775 uint64_t start = msp->ms_id;
2777 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2778 return (1ULL << 63);
2780 if (offset < start)
2781 return ((start - offset) << ms_shift);
2782 if (offset > start)
2783 return ((offset - start) << ms_shift);
2784 return (0);
2788 * ==========================================================================
2789 * Metaslab allocation tracing facility
2790 * ==========================================================================
2792 kstat_t *metaslab_trace_ksp;
2793 kstat_named_t metaslab_trace_over_limit;
2795 void
2796 metaslab_alloc_trace_init(void)
2798 ASSERT(metaslab_alloc_trace_cache == NULL);
2799 metaslab_alloc_trace_cache = kmem_cache_create(
2800 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t),
2801 0, NULL, NULL, NULL, NULL, NULL, 0);
2802 metaslab_trace_ksp = kstat_create("zfs", 0, "metaslab_trace_stats",
2803 "misc", KSTAT_TYPE_NAMED, 1, KSTAT_FLAG_VIRTUAL);
2804 if (metaslab_trace_ksp != NULL) {
2805 metaslab_trace_ksp->ks_data = &metaslab_trace_over_limit;
2806 kstat_named_init(&metaslab_trace_over_limit,
2807 "metaslab_trace_over_limit", KSTAT_DATA_UINT64);
2808 kstat_install(metaslab_trace_ksp);
2812 void
2813 metaslab_alloc_trace_fini(void)
2815 if (metaslab_trace_ksp != NULL) {
2816 kstat_delete(metaslab_trace_ksp);
2817 metaslab_trace_ksp = NULL;
2819 kmem_cache_destroy(metaslab_alloc_trace_cache);
2820 metaslab_alloc_trace_cache = NULL;
2824 * Add an allocation trace element to the allocation tracing list.
2826 static void
2827 metaslab_trace_add(zio_alloc_list_t *zal, metaslab_group_t *mg,
2828 metaslab_t *msp, uint64_t psize, uint32_t dva_id, uint64_t offset,
2829 int allocator)
2831 if (!metaslab_trace_enabled)
2832 return;
2835 * When the tracing list reaches its maximum we remove
2836 * the second element in the list before adding a new one.
2837 * By removing the second element we preserve the original
2838 * entry as a clue to what allocations steps have already been
2839 * performed.
2841 if (zal->zal_size == metaslab_trace_max_entries) {
2842 metaslab_alloc_trace_t *mat_next;
2843 #ifdef DEBUG
2844 panic("too many entries in allocation list");
2845 #endif
2846 atomic_inc_64(&metaslab_trace_over_limit.value.ui64);
2847 zal->zal_size--;
2848 mat_next = list_next(&zal->zal_list, list_head(&zal->zal_list));
2849 list_remove(&zal->zal_list, mat_next);
2850 kmem_cache_free(metaslab_alloc_trace_cache, mat_next);
2853 metaslab_alloc_trace_t *mat =
2854 kmem_cache_alloc(metaslab_alloc_trace_cache, KM_SLEEP);
2855 list_link_init(&mat->mat_list_node);
2856 mat->mat_mg = mg;
2857 mat->mat_msp = msp;
2858 mat->mat_size = psize;
2859 mat->mat_dva_id = dva_id;
2860 mat->mat_offset = offset;
2861 mat->mat_weight = 0;
2862 mat->mat_allocator = allocator;
2864 if (msp != NULL)
2865 mat->mat_weight = msp->ms_weight;
2868 * The list is part of the zio so locking is not required. Only
2869 * a single thread will perform allocations for a given zio.
2871 list_insert_tail(&zal->zal_list, mat);
2872 zal->zal_size++;
2874 ASSERT3U(zal->zal_size, <=, metaslab_trace_max_entries);
2877 void
2878 metaslab_trace_init(zio_alloc_list_t *zal)
2880 list_create(&zal->zal_list, sizeof (metaslab_alloc_trace_t),
2881 offsetof(metaslab_alloc_trace_t, mat_list_node));
2882 zal->zal_size = 0;
2885 void
2886 metaslab_trace_fini(zio_alloc_list_t *zal)
2888 metaslab_alloc_trace_t *mat;
2890 while ((mat = list_remove_head(&zal->zal_list)) != NULL)
2891 kmem_cache_free(metaslab_alloc_trace_cache, mat);
2892 list_destroy(&zal->zal_list);
2893 zal->zal_size = 0;
2897 * ==========================================================================
2898 * Metaslab block operations
2899 * ==========================================================================
2902 static void
2903 metaslab_group_alloc_increment(spa_t *spa, uint64_t vdev, void *tag, int flags,
2904 int allocator)
2906 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2907 (flags & METASLAB_DONT_THROTTLE))
2908 return;
2910 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2911 if (!mg->mg_class->mc_alloc_throttle_enabled)
2912 return;
2914 (void) refcount_add(&mg->mg_alloc_queue_depth[allocator], tag);
2917 static void
2918 metaslab_group_increment_qdepth(metaslab_group_t *mg, int allocator)
2920 uint64_t max = mg->mg_max_alloc_queue_depth;
2921 uint64_t cur = mg->mg_cur_max_alloc_queue_depth[allocator];
2922 while (cur < max) {
2923 if (atomic_cas_64(&mg->mg_cur_max_alloc_queue_depth[allocator],
2924 cur, cur + 1) == cur) {
2925 atomic_inc_64(
2926 &mg->mg_class->mc_alloc_max_slots[allocator]);
2927 return;
2929 cur = mg->mg_cur_max_alloc_queue_depth[allocator];
2933 void
2934 metaslab_group_alloc_decrement(spa_t *spa, uint64_t vdev, void *tag, int flags,
2935 int allocator, boolean_t io_complete)
2937 if (!(flags & METASLAB_ASYNC_ALLOC) ||
2938 (flags & METASLAB_DONT_THROTTLE))
2939 return;
2941 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2942 if (!mg->mg_class->mc_alloc_throttle_enabled)
2943 return;
2945 (void) refcount_remove(&mg->mg_alloc_queue_depth[allocator], tag);
2946 if (io_complete)
2947 metaslab_group_increment_qdepth(mg, allocator);
2950 void
2951 metaslab_group_alloc_verify(spa_t *spa, const blkptr_t *bp, void *tag,
2952 int allocator)
2954 #ifdef ZFS_DEBUG
2955 const dva_t *dva = bp->blk_dva;
2956 int ndvas = BP_GET_NDVAS(bp);
2958 for (int d = 0; d < ndvas; d++) {
2959 uint64_t vdev = DVA_GET_VDEV(&dva[d]);
2960 metaslab_group_t *mg = vdev_lookup_top(spa, vdev)->vdev_mg;
2961 VERIFY(refcount_not_held(&mg->mg_alloc_queue_depth[allocator],
2962 tag));
2964 #endif
2967 static uint64_t
2968 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2970 uint64_t start;
2971 range_tree_t *rt = msp->ms_allocatable;
2972 metaslab_class_t *mc = msp->ms_group->mg_class;
2974 VERIFY(!msp->ms_condensing);
2975 VERIFY0(msp->ms_initializing);
2977 start = mc->mc_ops->msop_alloc(msp, size);
2978 if (start != -1ULL) {
2979 metaslab_group_t *mg = msp->ms_group;
2980 vdev_t *vd = mg->mg_vd;
2982 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2983 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2984 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2985 range_tree_remove(rt, start, size);
2987 if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
2988 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2990 range_tree_add(msp->ms_allocating[txg & TXG_MASK], start, size);
2992 /* Track the last successful allocation */
2993 msp->ms_alloc_txg = txg;
2994 metaslab_verify_space(msp, txg);
2998 * Now that we've attempted the allocation we need to update the
2999 * metaslab's maximum block size since it may have changed.
3001 msp->ms_max_size = metaslab_block_maxsize(msp);
3002 return (start);
3006 * Find the metaslab with the highest weight that is less than what we've
3007 * already tried. In the common case, this means that we will examine each
3008 * metaslab at most once. Note that concurrent callers could reorder metaslabs
3009 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
3010 * activated by another thread, and we fail to allocate from the metaslab we
3011 * have selected, we may not try the newly-activated metaslab, and instead
3012 * activate another metaslab. This is not optimal, but generally does not cause
3013 * any problems (a possible exception being if every metaslab is completely full
3014 * except for the the newly-activated metaslab which we fail to examine).
3016 static metaslab_t *
3017 find_valid_metaslab(metaslab_group_t *mg, uint64_t activation_weight,
3018 dva_t *dva, int d, uint64_t min_distance, uint64_t asize, int allocator,
3019 zio_alloc_list_t *zal, metaslab_t *search, boolean_t *was_active)
3021 avl_index_t idx;
3022 avl_tree_t *t = &mg->mg_metaslab_tree;
3023 metaslab_t *msp = avl_find(t, search, &idx);
3024 if (msp == NULL)
3025 msp = avl_nearest(t, idx, AVL_AFTER);
3027 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
3028 int i;
3029 if (!metaslab_should_allocate(msp, asize)) {
3030 metaslab_trace_add(zal, mg, msp, asize, d,
3031 TRACE_TOO_SMALL, allocator);
3032 continue;
3036 * If the selected metaslab is condensing or being
3037 * initialized, skip it.
3039 if (msp->ms_condensing || msp->ms_initializing > 0)
3040 continue;
3042 *was_active = msp->ms_allocator != -1;
3044 * If we're activating as primary, this is our first allocation
3045 * from this disk, so we don't need to check how close we are.
3046 * If the metaslab under consideration was already active,
3047 * we're getting desperate enough to steal another allocator's
3048 * metaslab, so we still don't care about distances.
3050 if (activation_weight == METASLAB_WEIGHT_PRIMARY || *was_active)
3051 break;
3053 uint64_t target_distance = min_distance
3054 + (space_map_allocated(msp->ms_sm) != 0 ? 0 :
3055 min_distance >> 1);
3057 for (i = 0; i < d; i++) {
3058 if (metaslab_distance(msp, &dva[i]) < target_distance)
3059 break;
3061 if (i == d)
3062 break;
3065 if (msp != NULL) {
3066 search->ms_weight = msp->ms_weight;
3067 search->ms_start = msp->ms_start + 1;
3068 search->ms_allocator = msp->ms_allocator;
3069 search->ms_primary = msp->ms_primary;
3071 return (msp);
3074 /* ARGSUSED */
3075 static uint64_t
3076 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
3077 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d,
3078 int allocator)
3080 metaslab_t *msp = NULL;
3081 uint64_t offset = -1ULL;
3082 uint64_t activation_weight;
3084 activation_weight = METASLAB_WEIGHT_PRIMARY;
3085 for (int i = 0; i < d; i++) {
3086 if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
3087 DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
3088 activation_weight = METASLAB_WEIGHT_SECONDARY;
3089 } else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
3090 DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
3091 activation_weight = METASLAB_WEIGHT_CLAIM;
3092 break;
3097 * If we don't have enough metaslabs active to fill the entire array, we
3098 * just use the 0th slot.
3100 if (mg->mg_ms_ready < mg->mg_allocators * 3)
3101 allocator = 0;
3103 ASSERT3U(mg->mg_vd->vdev_ms_count, >=, 2);
3105 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
3106 search->ms_weight = UINT64_MAX;
3107 search->ms_start = 0;
3109 * At the end of the metaslab tree are the already-active metaslabs,
3110 * first the primaries, then the secondaries. When we resume searching
3111 * through the tree, we need to consider ms_allocator and ms_primary so
3112 * we start in the location right after where we left off, and don't
3113 * accidentally loop forever considering the same metaslabs.
3115 search->ms_allocator = -1;
3116 search->ms_primary = B_TRUE;
3117 for (;;) {
3118 boolean_t was_active = B_FALSE;
3120 mutex_enter(&mg->mg_lock);
3122 if (activation_weight == METASLAB_WEIGHT_PRIMARY &&
3123 mg->mg_primaries[allocator] != NULL) {
3124 msp = mg->mg_primaries[allocator];
3125 was_active = B_TRUE;
3126 } else if (activation_weight == METASLAB_WEIGHT_SECONDARY &&
3127 mg->mg_secondaries[allocator] != NULL) {
3128 msp = mg->mg_secondaries[allocator];
3129 was_active = B_TRUE;
3130 } else {
3131 msp = find_valid_metaslab(mg, activation_weight, dva, d,
3132 min_distance, asize, allocator, zal, search,
3133 &was_active);
3136 mutex_exit(&mg->mg_lock);
3137 if (msp == NULL) {
3138 kmem_free(search, sizeof (*search));
3139 return (-1ULL);
3142 mutex_enter(&msp->ms_lock);
3144 * Ensure that the metaslab we have selected is still
3145 * capable of handling our request. It's possible that
3146 * another thread may have changed the weight while we
3147 * were blocked on the metaslab lock. We check the
3148 * active status first to see if we need to reselect
3149 * a new metaslab.
3151 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
3152 mutex_exit(&msp->ms_lock);
3153 continue;
3157 * If the metaslab is freshly activated for an allocator that
3158 * isn't the one we're allocating from, or if it's a primary and
3159 * we're seeking a secondary (or vice versa), we go back and
3160 * select a new metaslab.
3162 if (!was_active && (msp->ms_weight & METASLAB_ACTIVE_MASK) &&
3163 (msp->ms_allocator != -1) &&
3164 (msp->ms_allocator != allocator || ((activation_weight ==
3165 METASLAB_WEIGHT_PRIMARY) != msp->ms_primary))) {
3166 mutex_exit(&msp->ms_lock);
3167 continue;
3170 if (msp->ms_weight & METASLAB_WEIGHT_CLAIM &&
3171 activation_weight != METASLAB_WEIGHT_CLAIM) {
3172 metaslab_passivate(msp, msp->ms_weight &
3173 ~METASLAB_WEIGHT_CLAIM);
3174 mutex_exit(&msp->ms_lock);
3175 continue;
3178 if (metaslab_activate(msp, allocator, activation_weight) != 0) {
3179 mutex_exit(&msp->ms_lock);
3180 continue;
3183 msp->ms_selected_txg = txg;
3186 * Now that we have the lock, recheck to see if we should
3187 * continue to use this metaslab for this allocation. The
3188 * the metaslab is now loaded so metaslab_should_allocate() can
3189 * accurately determine if the allocation attempt should
3190 * proceed.
3192 if (!metaslab_should_allocate(msp, asize)) {
3193 /* Passivate this metaslab and select a new one. */
3194 metaslab_trace_add(zal, mg, msp, asize, d,
3195 TRACE_TOO_SMALL, allocator);
3196 goto next;
3200 * If this metaslab is currently condensing then pick again as
3201 * we can't manipulate this metaslab until it's committed
3202 * to disk. If this metaslab is being initialized, we shouldn't
3203 * allocate from it since the allocated region might be
3204 * overwritten after allocation.
3206 if (msp->ms_condensing) {
3207 metaslab_trace_add(zal, mg, msp, asize, d,
3208 TRACE_CONDENSING, allocator);
3209 metaslab_passivate(msp, msp->ms_weight &
3210 ~METASLAB_ACTIVE_MASK);
3211 mutex_exit(&msp->ms_lock);
3212 continue;
3213 } else if (msp->ms_initializing > 0) {
3214 metaslab_trace_add(zal, mg, msp, asize, d,
3215 TRACE_INITIALIZING, allocator);
3216 metaslab_passivate(msp, msp->ms_weight &
3217 ~METASLAB_ACTIVE_MASK);
3218 mutex_exit(&msp->ms_lock);
3219 continue;
3222 offset = metaslab_block_alloc(msp, asize, txg);
3223 metaslab_trace_add(zal, mg, msp, asize, d, offset, allocator);
3225 if (offset != -1ULL) {
3226 /* Proactively passivate the metaslab, if needed */
3227 metaslab_segment_may_passivate(msp);
3228 break;
3230 next:
3231 ASSERT(msp->ms_loaded);
3234 * We were unable to allocate from this metaslab so determine
3235 * a new weight for this metaslab. Now that we have loaded
3236 * the metaslab we can provide a better hint to the metaslab
3237 * selector.
3239 * For space-based metaslabs, we use the maximum block size.
3240 * This information is only available when the metaslab
3241 * is loaded and is more accurate than the generic free
3242 * space weight that was calculated by metaslab_weight().
3243 * This information allows us to quickly compare the maximum
3244 * available allocation in the metaslab to the allocation
3245 * size being requested.
3247 * For segment-based metaslabs, determine the new weight
3248 * based on the highest bucket in the range tree. We
3249 * explicitly use the loaded segment weight (i.e. the range
3250 * tree histogram) since it contains the space that is
3251 * currently available for allocation and is accurate
3252 * even within a sync pass.
3254 if (WEIGHT_IS_SPACEBASED(msp->ms_weight)) {
3255 uint64_t weight = metaslab_block_maxsize(msp);
3256 WEIGHT_SET_SPACEBASED(weight);
3257 metaslab_passivate(msp, weight);
3258 } else {
3259 metaslab_passivate(msp,
3260 metaslab_weight_from_range_tree(msp));
3264 * We have just failed an allocation attempt, check
3265 * that metaslab_should_allocate() agrees. Otherwise,
3266 * we may end up in an infinite loop retrying the same
3267 * metaslab.
3269 ASSERT(!metaslab_should_allocate(msp, asize));
3270 mutex_exit(&msp->ms_lock);
3272 mutex_exit(&msp->ms_lock);
3273 kmem_free(search, sizeof (*search));
3274 return (offset);
3277 static uint64_t
3278 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
3279 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d,
3280 int allocator)
3282 uint64_t offset;
3283 ASSERT(mg->mg_initialized);
3285 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
3286 min_distance, dva, d, allocator);
3288 mutex_enter(&mg->mg_lock);
3289 if (offset == -1ULL) {
3290 mg->mg_failed_allocations++;
3291 metaslab_trace_add(zal, mg, NULL, asize, d,
3292 TRACE_GROUP_FAILURE, allocator);
3293 if (asize == SPA_GANGBLOCKSIZE) {
3295 * This metaslab group was unable to allocate
3296 * the minimum gang block size so it must be out of
3297 * space. We must notify the allocation throttle
3298 * to start skipping allocation attempts to this
3299 * metaslab group until more space becomes available.
3300 * Note: this failure cannot be caused by the
3301 * allocation throttle since the allocation throttle
3302 * is only responsible for skipping devices and
3303 * not failing block allocations.
3305 mg->mg_no_free_space = B_TRUE;
3308 mg->mg_allocations++;
3309 mutex_exit(&mg->mg_lock);
3310 return (offset);
3314 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
3315 * on the same vdev as an existing DVA of this BP, then try to allocate it
3316 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
3317 * existing DVAs.
3319 int ditto_same_vdev_distance_shift = 3;
3322 * Allocate a block for the specified i/o.
3325 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
3326 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
3327 zio_alloc_list_t *zal, int allocator)
3329 metaslab_group_t *mg, *rotor;
3330 vdev_t *vd;
3331 boolean_t try_hard = B_FALSE;
3333 ASSERT(!DVA_IS_VALID(&dva[d]));
3336 * For testing, make some blocks above a certain size be gang blocks.
3338 if (psize >= metaslab_force_ganging && (ddi_get_lbolt() & 3) == 0) {
3339 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG,
3340 allocator);
3341 return (SET_ERROR(ENOSPC));
3345 * Start at the rotor and loop through all mgs until we find something.
3346 * Note that there's no locking on mc_rotor or mc_aliquot because
3347 * nothing actually breaks if we miss a few updates -- we just won't
3348 * allocate quite as evenly. It all balances out over time.
3350 * If we are doing ditto or log blocks, try to spread them across
3351 * consecutive vdevs. If we're forced to reuse a vdev before we've
3352 * allocated all of our ditto blocks, then try and spread them out on
3353 * that vdev as much as possible. If it turns out to not be possible,
3354 * gradually lower our standards until anything becomes acceptable.
3355 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3356 * gives us hope of containing our fault domains to something we're
3357 * able to reason about. Otherwise, any two top-level vdev failures
3358 * will guarantee the loss of data. With consecutive allocation,
3359 * only two adjacent top-level vdev failures will result in data loss.
3361 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3362 * ourselves on the same vdev as our gang block header. That
3363 * way, we can hope for locality in vdev_cache, plus it makes our
3364 * fault domains something tractable.
3366 if (hintdva) {
3367 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3370 * It's possible the vdev we're using as the hint no
3371 * longer exists or its mg has been closed (e.g. by
3372 * device removal). Consult the rotor when
3373 * all else fails.
3375 if (vd != NULL && vd->vdev_mg != NULL) {
3376 mg = vd->vdev_mg;
3378 if (flags & METASLAB_HINTBP_AVOID &&
3379 mg->mg_next != NULL)
3380 mg = mg->mg_next;
3381 } else {
3382 mg = mc->mc_rotor;
3384 } else if (d != 0) {
3385 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3386 mg = vd->vdev_mg->mg_next;
3387 } else {
3388 mg = mc->mc_rotor;
3392 * If the hint put us into the wrong metaslab class, or into a
3393 * metaslab group that has been passivated, just follow the rotor.
3395 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3396 mg = mc->mc_rotor;
3398 rotor = mg;
3399 top:
3400 do {
3401 boolean_t allocatable;
3403 ASSERT(mg->mg_activation_count == 1);
3404 vd = mg->mg_vd;
3407 * Don't allocate from faulted devices.
3409 if (try_hard) {
3410 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
3411 allocatable = vdev_allocatable(vd);
3412 spa_config_exit(spa, SCL_ZIO, FTAG);
3413 } else {
3414 allocatable = vdev_allocatable(vd);
3418 * Determine if the selected metaslab group is eligible
3419 * for allocations. If we're ganging then don't allow
3420 * this metaslab group to skip allocations since that would
3421 * inadvertently return ENOSPC and suspend the pool
3422 * even though space is still available.
3424 if (allocatable && !GANG_ALLOCATION(flags) && !try_hard) {
3425 allocatable = metaslab_group_allocatable(mg, rotor,
3426 psize, allocator);
3429 if (!allocatable) {
3430 metaslab_trace_add(zal, mg, NULL, psize, d,
3431 TRACE_NOT_ALLOCATABLE, allocator);
3432 goto next;
3435 ASSERT(mg->mg_initialized);
3438 * Avoid writing single-copy data to a failing,
3439 * non-redundant vdev, unless we've already tried all
3440 * other vdevs.
3442 if ((vd->vdev_stat.vs_write_errors > 0 ||
3443 vd->vdev_state < VDEV_STATE_HEALTHY) &&
3444 d == 0 && !try_hard && vd->vdev_children == 0) {
3445 metaslab_trace_add(zal, mg, NULL, psize, d,
3446 TRACE_VDEV_ERROR, allocator);
3447 goto next;
3450 ASSERT(mg->mg_class == mc);
3453 * If we don't need to try hard, then require that the
3454 * block be 1/8th of the device away from any other DVAs
3455 * in this BP. If we are trying hard, allow any offset
3456 * to be used (distance=0).
3458 uint64_t distance = 0;
3459 if (!try_hard) {
3460 distance = vd->vdev_asize >>
3461 ditto_same_vdev_distance_shift;
3462 if (distance <= (1ULL << vd->vdev_ms_shift))
3463 distance = 0;
3466 uint64_t asize = vdev_psize_to_asize(vd, psize);
3467 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3469 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
3470 distance, dva, d, allocator);
3472 if (offset != -1ULL) {
3474 * If we've just selected this metaslab group,
3475 * figure out whether the corresponding vdev is
3476 * over- or under-used relative to the pool,
3477 * and set an allocation bias to even it out.
3479 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3480 vdev_stat_t *vs = &vd->vdev_stat;
3481 int64_t vu, cu;
3483 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
3484 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
3487 * Calculate how much more or less we should
3488 * try to allocate from this device during
3489 * this iteration around the rotor.
3490 * For example, if a device is 80% full
3491 * and the pool is 20% full then we should
3492 * reduce allocations by 60% on this device.
3494 * mg_bias = (20 - 80) * 512K / 100 = -307K
3496 * This reduces allocations by 307K for this
3497 * iteration.
3499 mg->mg_bias = ((cu - vu) *
3500 (int64_t)mg->mg_aliquot) / 100;
3501 } else if (!metaslab_bias_enabled) {
3502 mg->mg_bias = 0;
3505 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3506 mg->mg_aliquot + mg->mg_bias) {
3507 mc->mc_rotor = mg->mg_next;
3508 mc->mc_aliquot = 0;
3511 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3512 DVA_SET_OFFSET(&dva[d], offset);
3513 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
3514 DVA_SET_ASIZE(&dva[d], asize);
3516 return (0);
3518 next:
3519 mc->mc_rotor = mg->mg_next;
3520 mc->mc_aliquot = 0;
3521 } while ((mg = mg->mg_next) != rotor);
3524 * If we haven't tried hard, do so now.
3526 if (!try_hard) {
3527 try_hard = B_TRUE;
3528 goto top;
3531 bzero(&dva[d], sizeof (dva_t));
3533 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC, allocator);
3534 return (SET_ERROR(ENOSPC));
3537 void
3538 metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize,
3539 boolean_t checkpoint)
3541 metaslab_t *msp;
3542 spa_t *spa = vd->vdev_spa;
3544 ASSERT(vdev_is_concrete(vd));
3545 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3546 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
3548 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3550 VERIFY(!msp->ms_condensing);
3551 VERIFY3U(offset, >=, msp->ms_start);
3552 VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size);
3553 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3554 VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift));
3556 metaslab_check_free_impl(vd, offset, asize);
3558 mutex_enter(&msp->ms_lock);
3559 if (range_tree_is_empty(msp->ms_freeing) &&
3560 range_tree_is_empty(msp->ms_checkpointing)) {
3561 vdev_dirty(vd, VDD_METASLAB, msp, spa_syncing_txg(spa));
3564 if (checkpoint) {
3565 ASSERT(spa_has_checkpoint(spa));
3566 range_tree_add(msp->ms_checkpointing, offset, asize);
3567 } else {
3568 range_tree_add(msp->ms_freeing, offset, asize);
3570 mutex_exit(&msp->ms_lock);
3573 /* ARGSUSED */
3574 void
3575 metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3576 uint64_t size, void *arg)
3578 boolean_t *checkpoint = arg;
3580 ASSERT3P(checkpoint, !=, NULL);
3582 if (vd->vdev_ops->vdev_op_remap != NULL)
3583 vdev_indirect_mark_obsolete(vd, offset, size);
3584 else
3585 metaslab_free_impl(vd, offset, size, *checkpoint);
3588 static void
3589 metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
3590 boolean_t checkpoint)
3592 spa_t *spa = vd->vdev_spa;
3594 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3596 if (spa_syncing_txg(spa) > spa_freeze_txg(spa))
3597 return;
3599 if (spa->spa_vdev_removal != NULL &&
3600 spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id &&
3601 vdev_is_concrete(vd)) {
3603 * Note: we check if the vdev is concrete because when
3604 * we complete the removal, we first change the vdev to be
3605 * an indirect vdev (in open context), and then (in syncing
3606 * context) clear spa_vdev_removal.
3608 free_from_removing_vdev(vd, offset, size);
3609 } else if (vd->vdev_ops->vdev_op_remap != NULL) {
3610 vdev_indirect_mark_obsolete(vd, offset, size);
3611 vd->vdev_ops->vdev_op_remap(vd, offset, size,
3612 metaslab_free_impl_cb, &checkpoint);
3613 } else {
3614 metaslab_free_concrete(vd, offset, size, checkpoint);
3618 typedef struct remap_blkptr_cb_arg {
3619 blkptr_t *rbca_bp;
3620 spa_remap_cb_t rbca_cb;
3621 vdev_t *rbca_remap_vd;
3622 uint64_t rbca_remap_offset;
3623 void *rbca_cb_arg;
3624 } remap_blkptr_cb_arg_t;
3626 void
3627 remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3628 uint64_t size, void *arg)
3630 remap_blkptr_cb_arg_t *rbca = arg;
3631 blkptr_t *bp = rbca->rbca_bp;
3633 /* We can not remap split blocks. */
3634 if (size != DVA_GET_ASIZE(&bp->blk_dva[0]))
3635 return;
3636 ASSERT0(inner_offset);
3638 if (rbca->rbca_cb != NULL) {
3640 * At this point we know that we are not handling split
3641 * blocks and we invoke the callback on the previous
3642 * vdev which must be indirect.
3644 ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops);
3646 rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id,
3647 rbca->rbca_remap_offset, size, rbca->rbca_cb_arg);
3649 /* set up remap_blkptr_cb_arg for the next call */
3650 rbca->rbca_remap_vd = vd;
3651 rbca->rbca_remap_offset = offset;
3655 * The phys birth time is that of dva[0]. This ensures that we know
3656 * when each dva was written, so that resilver can determine which
3657 * blocks need to be scrubbed (i.e. those written during the time
3658 * the vdev was offline). It also ensures that the key used in
3659 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
3660 * we didn't change the phys_birth, a lookup in the ARC for a
3661 * remapped BP could find the data that was previously stored at
3662 * this vdev + offset.
3664 vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa,
3665 DVA_GET_VDEV(&bp->blk_dva[0]));
3666 vdev_indirect_births_t *vib = oldvd->vdev_indirect_births;
3667 bp->blk_phys_birth = vdev_indirect_births_physbirth(vib,
3668 DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0]));
3670 DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id);
3671 DVA_SET_OFFSET(&bp->blk_dva[0], offset);
3675 * If the block pointer contains any indirect DVAs, modify them to refer to
3676 * concrete DVAs. Note that this will sometimes not be possible, leaving
3677 * the indirect DVA in place. This happens if the indirect DVA spans multiple
3678 * segments in the mapping (i.e. it is a "split block").
3680 * If the BP was remapped, calls the callback on the original dva (note the
3681 * callback can be called multiple times if the original indirect DVA refers
3682 * to another indirect DVA, etc).
3684 * Returns TRUE if the BP was remapped.
3686 boolean_t
3687 spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg)
3689 remap_blkptr_cb_arg_t rbca;
3691 if (!zfs_remap_blkptr_enable)
3692 return (B_FALSE);
3694 if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS))
3695 return (B_FALSE);
3698 * Dedup BP's can not be remapped, because ddt_phys_select() depends
3699 * on DVA[0] being the same in the BP as in the DDT (dedup table).
3701 if (BP_GET_DEDUP(bp))
3702 return (B_FALSE);
3705 * Gang blocks can not be remapped, because
3706 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
3707 * the BP used to read the gang block header (GBH) being the same
3708 * as the DVA[0] that we allocated for the GBH.
3710 if (BP_IS_GANG(bp))
3711 return (B_FALSE);
3714 * Embedded BP's have no DVA to remap.
3716 if (BP_GET_NDVAS(bp) < 1)
3717 return (B_FALSE);
3720 * Note: we only remap dva[0]. If we remapped other dvas, we
3721 * would no longer know what their phys birth txg is.
3723 dva_t *dva = &bp->blk_dva[0];
3725 uint64_t offset = DVA_GET_OFFSET(dva);
3726 uint64_t size = DVA_GET_ASIZE(dva);
3727 vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3729 if (vd->vdev_ops->vdev_op_remap == NULL)
3730 return (B_FALSE);
3732 rbca.rbca_bp = bp;
3733 rbca.rbca_cb = callback;
3734 rbca.rbca_remap_vd = vd;
3735 rbca.rbca_remap_offset = offset;
3736 rbca.rbca_cb_arg = arg;
3739 * remap_blkptr_cb() will be called in order for each level of
3740 * indirection, until a concrete vdev is reached or a split block is
3741 * encountered. old_vd and old_offset are updated within the callback
3742 * as we go from the one indirect vdev to the next one (either concrete
3743 * or indirect again) in that order.
3745 vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca);
3747 /* Check if the DVA wasn't remapped because it is a split block */
3748 if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id)
3749 return (B_FALSE);
3751 return (B_TRUE);
3755 * Undo the allocation of a DVA which happened in the given transaction group.
3757 void
3758 metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3760 metaslab_t *msp;
3761 vdev_t *vd;
3762 uint64_t vdev = DVA_GET_VDEV(dva);
3763 uint64_t offset = DVA_GET_OFFSET(dva);
3764 uint64_t size = DVA_GET_ASIZE(dva);
3766 ASSERT(DVA_IS_VALID(dva));
3767 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3769 if (txg > spa_freeze_txg(spa))
3770 return;
3772 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3773 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3774 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
3775 (u_longlong_t)vdev, (u_longlong_t)offset);
3776 ASSERT(0);
3777 return;
3780 ASSERT(!vd->vdev_removing);
3781 ASSERT(vdev_is_concrete(vd));
3782 ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
3783 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
3785 if (DVA_GET_GANG(dva))
3786 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3788 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3790 mutex_enter(&msp->ms_lock);
3791 range_tree_remove(msp->ms_allocating[txg & TXG_MASK],
3792 offset, size);
3794 VERIFY(!msp->ms_condensing);
3795 VERIFY3U(offset, >=, msp->ms_start);
3796 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3797 VERIFY3U(range_tree_space(msp->ms_allocatable) + size, <=,
3798 msp->ms_size);
3799 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3800 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3801 range_tree_add(msp->ms_allocatable, offset, size);
3802 mutex_exit(&msp->ms_lock);
3806 * Free the block represented by the given DVA.
3808 void
3809 metaslab_free_dva(spa_t *spa, const dva_t *dva, boolean_t checkpoint)
3811 uint64_t vdev = DVA_GET_VDEV(dva);
3812 uint64_t offset = DVA_GET_OFFSET(dva);
3813 uint64_t size = DVA_GET_ASIZE(dva);
3814 vdev_t *vd = vdev_lookup_top(spa, vdev);
3816 ASSERT(DVA_IS_VALID(dva));
3817 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3819 if (DVA_GET_GANG(dva)) {
3820 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3823 metaslab_free_impl(vd, offset, size, checkpoint);
3827 * Reserve some allocation slots. The reservation system must be called
3828 * before we call into the allocator. If there aren't any available slots
3829 * then the I/O will be throttled until an I/O completes and its slots are
3830 * freed up. The function returns true if it was successful in placing
3831 * the reservation.
3833 boolean_t
3834 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, int allocator,
3835 zio_t *zio, int flags)
3837 uint64_t available_slots = 0;
3838 boolean_t slot_reserved = B_FALSE;
3839 uint64_t max = mc->mc_alloc_max_slots[allocator];
3841 ASSERT(mc->mc_alloc_throttle_enabled);
3842 mutex_enter(&mc->mc_lock);
3844 uint64_t reserved_slots =
3845 refcount_count(&mc->mc_alloc_slots[allocator]);
3846 if (reserved_slots < max)
3847 available_slots = max - reserved_slots;
3849 if (slots <= available_slots || GANG_ALLOCATION(flags)) {
3851 * We reserve the slots individually so that we can unreserve
3852 * them individually when an I/O completes.
3854 for (int d = 0; d < slots; d++) {
3855 reserved_slots =
3856 refcount_add(&mc->mc_alloc_slots[allocator],
3857 zio);
3859 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3860 slot_reserved = B_TRUE;
3863 mutex_exit(&mc->mc_lock);
3864 return (slot_reserved);
3867 void
3868 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots,
3869 int allocator, zio_t *zio)
3871 ASSERT(mc->mc_alloc_throttle_enabled);
3872 mutex_enter(&mc->mc_lock);
3873 for (int d = 0; d < slots; d++) {
3874 (void) refcount_remove(&mc->mc_alloc_slots[allocator],
3875 zio);
3877 mutex_exit(&mc->mc_lock);
3880 static int
3881 metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
3882 uint64_t txg)
3884 metaslab_t *msp;
3885 spa_t *spa = vd->vdev_spa;
3886 int error = 0;
3888 if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count)
3889 return (ENXIO);
3891 ASSERT3P(vd->vdev_ms, !=, NULL);
3892 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3894 mutex_enter(&msp->ms_lock);
3896 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3897 error = metaslab_activate(msp, 0, METASLAB_WEIGHT_CLAIM);
3899 * No need to fail in that case; someone else has activated the
3900 * metaslab, but that doesn't preclude us from using it.
3902 if (error == EBUSY)
3903 error = 0;
3905 if (error == 0 &&
3906 !range_tree_contains(msp->ms_allocatable, offset, size))
3907 error = SET_ERROR(ENOENT);
3909 if (error || txg == 0) { /* txg == 0 indicates dry run */
3910 mutex_exit(&msp->ms_lock);
3911 return (error);
3914 VERIFY(!msp->ms_condensing);
3915 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3916 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3917 VERIFY3U(range_tree_space(msp->ms_allocatable) - size, <=,
3918 msp->ms_size);
3919 range_tree_remove(msp->ms_allocatable, offset, size);
3921 if (spa_writeable(spa)) { /* don't dirty if we're zdb(8) */
3922 if (range_tree_is_empty(msp->ms_allocating[txg & TXG_MASK]))
3923 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3924 range_tree_add(msp->ms_allocating[txg & TXG_MASK],
3925 offset, size);
3928 mutex_exit(&msp->ms_lock);
3930 return (0);
3933 typedef struct metaslab_claim_cb_arg_t {
3934 uint64_t mcca_txg;
3935 int mcca_error;
3936 } metaslab_claim_cb_arg_t;
3938 /* ARGSUSED */
3939 static void
3940 metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3941 uint64_t size, void *arg)
3943 metaslab_claim_cb_arg_t *mcca_arg = arg;
3945 if (mcca_arg->mcca_error == 0) {
3946 mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset,
3947 size, mcca_arg->mcca_txg);
3952 metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
3954 if (vd->vdev_ops->vdev_op_remap != NULL) {
3955 metaslab_claim_cb_arg_t arg;
3958 * Only zdb(8) can claim on indirect vdevs. This is used
3959 * to detect leaks of mapped space (that are not accounted
3960 * for in the obsolete counts, spacemap, or bpobj).
3962 ASSERT(!spa_writeable(vd->vdev_spa));
3963 arg.mcca_error = 0;
3964 arg.mcca_txg = txg;
3966 vd->vdev_ops->vdev_op_remap(vd, offset, size,
3967 metaslab_claim_impl_cb, &arg);
3969 if (arg.mcca_error == 0) {
3970 arg.mcca_error = metaslab_claim_concrete(vd,
3971 offset, size, txg);
3973 return (arg.mcca_error);
3974 } else {
3975 return (metaslab_claim_concrete(vd, offset, size, txg));
3980 * Intent log support: upon opening the pool after a crash, notify the SPA
3981 * of blocks that the intent log has allocated for immediate write, but
3982 * which are still considered free by the SPA because the last transaction
3983 * group didn't commit yet.
3985 static int
3986 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3988 uint64_t vdev = DVA_GET_VDEV(dva);
3989 uint64_t offset = DVA_GET_OFFSET(dva);
3990 uint64_t size = DVA_GET_ASIZE(dva);
3991 vdev_t *vd;
3993 if ((vd = vdev_lookup_top(spa, vdev)) == NULL) {
3994 return (SET_ERROR(ENXIO));
3997 ASSERT(DVA_IS_VALID(dva));
3999 if (DVA_GET_GANG(dva))
4000 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
4002 return (metaslab_claim_impl(vd, offset, size, txg));
4006 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
4007 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
4008 zio_alloc_list_t *zal, zio_t *zio, int allocator)
4010 dva_t *dva = bp->blk_dva;
4011 dva_t *hintdva = hintbp->blk_dva;
4012 int error = 0;
4014 ASSERT(bp->blk_birth == 0);
4015 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
4017 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4019 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
4020 spa_config_exit(spa, SCL_ALLOC, FTAG);
4021 return (SET_ERROR(ENOSPC));
4024 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
4025 ASSERT(BP_GET_NDVAS(bp) == 0);
4026 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
4027 ASSERT3P(zal, !=, NULL);
4029 for (int d = 0; d < ndvas; d++) {
4030 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
4031 txg, flags, zal, allocator);
4032 if (error != 0) {
4033 for (d--; d >= 0; d--) {
4034 metaslab_unalloc_dva(spa, &dva[d], txg);
4035 metaslab_group_alloc_decrement(spa,
4036 DVA_GET_VDEV(&dva[d]), zio, flags,
4037 allocator, B_FALSE);
4038 bzero(&dva[d], sizeof (dva_t));
4040 spa_config_exit(spa, SCL_ALLOC, FTAG);
4041 return (error);
4042 } else {
4044 * Update the metaslab group's queue depth
4045 * based on the newly allocated dva.
4047 metaslab_group_alloc_increment(spa,
4048 DVA_GET_VDEV(&dva[d]), zio, flags, allocator);
4052 ASSERT(error == 0);
4053 ASSERT(BP_GET_NDVAS(bp) == ndvas);
4055 spa_config_exit(spa, SCL_ALLOC, FTAG);
4057 BP_SET_BIRTH(bp, txg, txg);
4059 return (0);
4062 void
4063 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
4065 const dva_t *dva = bp->blk_dva;
4066 int ndvas = BP_GET_NDVAS(bp);
4068 ASSERT(!BP_IS_HOLE(bp));
4069 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
4072 * If we have a checkpoint for the pool we need to make sure that
4073 * the blocks that we free that are part of the checkpoint won't be
4074 * reused until the checkpoint is discarded or we revert to it.
4076 * The checkpoint flag is passed down the metaslab_free code path
4077 * and is set whenever we want to add a block to the checkpoint's
4078 * accounting. That is, we "checkpoint" blocks that existed at the
4079 * time the checkpoint was created and are therefore referenced by
4080 * the checkpointed uberblock.
4082 * Note that, we don't checkpoint any blocks if the current
4083 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
4084 * normally as they will be referenced by the checkpointed uberblock.
4086 boolean_t checkpoint = B_FALSE;
4087 if (bp->blk_birth <= spa->spa_checkpoint_txg &&
4088 spa_syncing_txg(spa) > spa->spa_checkpoint_txg) {
4090 * At this point, if the block is part of the checkpoint
4091 * there is no way it was created in the current txg.
4093 ASSERT(!now);
4094 ASSERT3U(spa_syncing_txg(spa), ==, txg);
4095 checkpoint = B_TRUE;
4098 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
4100 for (int d = 0; d < ndvas; d++) {
4101 if (now) {
4102 metaslab_unalloc_dva(spa, &dva[d], txg);
4103 } else {
4104 ASSERT3U(txg, ==, spa_syncing_txg(spa));
4105 metaslab_free_dva(spa, &dva[d], checkpoint);
4109 spa_config_exit(spa, SCL_FREE, FTAG);
4113 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
4115 const dva_t *dva = bp->blk_dva;
4116 int ndvas = BP_GET_NDVAS(bp);
4117 int error = 0;
4119 ASSERT(!BP_IS_HOLE(bp));
4121 if (txg != 0) {
4123 * First do a dry run to make sure all DVAs are claimable,
4124 * so we don't have to unwind from partial failures below.
4126 if ((error = metaslab_claim(spa, bp, 0)) != 0)
4127 return (error);
4130 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
4132 for (int d = 0; d < ndvas; d++)
4133 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
4134 break;
4136 spa_config_exit(spa, SCL_ALLOC, FTAG);
4138 ASSERT(error == 0 || txg == 0);
4140 return (error);
4143 /* ARGSUSED */
4144 static void
4145 metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset,
4146 uint64_t size, void *arg)
4148 if (vd->vdev_ops == &vdev_indirect_ops)
4149 return;
4151 metaslab_check_free_impl(vd, offset, size);
4154 static void
4155 metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size)
4157 metaslab_t *msp;
4158 spa_t *spa = vd->vdev_spa;
4160 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
4161 return;
4163 if (vd->vdev_ops->vdev_op_remap != NULL) {
4164 vd->vdev_ops->vdev_op_remap(vd, offset, size,
4165 metaslab_check_free_impl_cb, NULL);
4166 return;
4169 ASSERT(vdev_is_concrete(vd));
4170 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
4171 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
4173 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
4175 mutex_enter(&msp->ms_lock);
4176 if (msp->ms_loaded)
4177 range_tree_verify(msp->ms_allocatable, offset, size);
4179 range_tree_verify(msp->ms_freeing, offset, size);
4180 range_tree_verify(msp->ms_checkpointing, offset, size);
4181 range_tree_verify(msp->ms_freed, offset, size);
4182 for (int j = 0; j < TXG_DEFER_SIZE; j++)
4183 range_tree_verify(msp->ms_defer[j], offset, size);
4184 mutex_exit(&msp->ms_lock);
4187 void
4188 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
4190 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
4191 return;
4193 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4194 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
4195 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
4196 vdev_t *vd = vdev_lookup_top(spa, vdev);
4197 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
4198 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
4200 if (DVA_GET_GANG(&bp->blk_dva[i]))
4201 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
4203 ASSERT3P(vd, !=, NULL);
4205 metaslab_check_free_impl(vd, offset, size);
4207 spa_config_exit(spa, SCL_VDEV, FTAG);