4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
25 #include <sys/zfs_context.h>
27 #include <sys/dmu_tx.h>
28 #include <sys/space_map.h>
29 #include <sys/metaslab_impl.h>
30 #include <sys/vdev_impl.h>
33 uint64_t metaslab_aliquot = 512ULL << 10;
34 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
37 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
39 static int metaslab_debug = 0;
42 * Minimum size which forces the dynamic allocator to change
43 * it's allocation strategy. Once the space map cannot satisfy
44 * an allocation of this size then it switches to using more
45 * aggressive strategy (i.e search by size rather than offset).
47 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
50 * The minimum free space, in percent, which must be available
51 * in a space map to continue allocations in a first-fit fashion.
52 * Once the space_map's free space drops below this level we dynamically
53 * switch to using best-fit allocations.
55 int metaslab_df_free_pct = 4;
58 * A metaslab is considered "free" if it contains a contiguous
59 * segment which is greater than metaslab_min_alloc_size.
61 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
64 * Max number of space_maps to prefetch.
66 int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
69 * Percentage bonus multiplier for metaslabs that are in the bonus area.
71 int metaslab_smo_bonus_pct = 150;
74 * ==========================================================================
76 * ==========================================================================
79 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
83 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
93 metaslab_class_destroy(metaslab_class_t *mc)
95 ASSERT(mc->mc_rotor == NULL);
96 ASSERT(mc->mc_alloc == 0);
97 ASSERT(mc->mc_deferred == 0);
98 ASSERT(mc->mc_space == 0);
99 ASSERT(mc->mc_dspace == 0);
101 kmem_free(mc, sizeof (metaslab_class_t));
105 metaslab_class_validate(metaslab_class_t *mc)
107 metaslab_group_t *mg;
111 * Must hold one of the spa_config locks.
113 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
114 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
116 if ((mg = mc->mc_rotor) == NULL)
121 ASSERT(vd->vdev_mg != NULL);
122 ASSERT3P(vd->vdev_top, ==, vd);
123 ASSERT3P(mg->mg_class, ==, mc);
124 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
125 } while ((mg = mg->mg_next) != mc->mc_rotor);
131 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
132 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
134 atomic_add_64(&mc->mc_alloc, alloc_delta);
135 atomic_add_64(&mc->mc_deferred, defer_delta);
136 atomic_add_64(&mc->mc_space, space_delta);
137 atomic_add_64(&mc->mc_dspace, dspace_delta);
141 metaslab_class_get_alloc(metaslab_class_t *mc)
143 return (mc->mc_alloc);
147 metaslab_class_get_deferred(metaslab_class_t *mc)
149 return (mc->mc_deferred);
153 metaslab_class_get_space(metaslab_class_t *mc)
155 return (mc->mc_space);
159 metaslab_class_get_dspace(metaslab_class_t *mc)
161 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
165 * ==========================================================================
167 * ==========================================================================
170 metaslab_compare(const void *x1, const void *x2)
172 const metaslab_t *m1 = x1;
173 const metaslab_t *m2 = x2;
175 if (m1->ms_weight < m2->ms_weight)
177 if (m1->ms_weight > m2->ms_weight)
181 * If the weights are identical, use the offset to force uniqueness.
183 if (m1->ms_map.sm_start < m2->ms_map.sm_start)
185 if (m1->ms_map.sm_start > m2->ms_map.sm_start)
188 ASSERT3P(m1, ==, m2);
194 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
196 metaslab_group_t *mg;
198 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
199 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
200 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
201 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
204 mg->mg_activation_count = 0;
210 metaslab_group_destroy(metaslab_group_t *mg)
212 ASSERT(mg->mg_prev == NULL);
213 ASSERT(mg->mg_next == NULL);
215 * We may have gone below zero with the activation count
216 * either because we never activated in the first place or
217 * because we're done, and possibly removing the vdev.
219 ASSERT(mg->mg_activation_count <= 0);
221 avl_destroy(&mg->mg_metaslab_tree);
222 mutex_destroy(&mg->mg_lock);
223 kmem_free(mg, sizeof (metaslab_group_t));
227 metaslab_group_activate(metaslab_group_t *mg)
229 metaslab_class_t *mc = mg->mg_class;
230 metaslab_group_t *mgprev, *mgnext;
232 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
234 ASSERT(mc->mc_rotor != mg);
235 ASSERT(mg->mg_prev == NULL);
236 ASSERT(mg->mg_next == NULL);
237 ASSERT(mg->mg_activation_count <= 0);
239 if (++mg->mg_activation_count <= 0)
242 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
244 if ((mgprev = mc->mc_rotor) == NULL) {
248 mgnext = mgprev->mg_next;
249 mg->mg_prev = mgprev;
250 mg->mg_next = mgnext;
251 mgprev->mg_next = mg;
252 mgnext->mg_prev = mg;
258 metaslab_group_passivate(metaslab_group_t *mg)
260 metaslab_class_t *mc = mg->mg_class;
261 metaslab_group_t *mgprev, *mgnext;
263 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
265 if (--mg->mg_activation_count != 0) {
266 ASSERT(mc->mc_rotor != mg);
267 ASSERT(mg->mg_prev == NULL);
268 ASSERT(mg->mg_next == NULL);
269 ASSERT(mg->mg_activation_count < 0);
273 mgprev = mg->mg_prev;
274 mgnext = mg->mg_next;
279 mc->mc_rotor = mgnext;
280 mgprev->mg_next = mgnext;
281 mgnext->mg_prev = mgprev;
289 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
291 mutex_enter(&mg->mg_lock);
292 ASSERT(msp->ms_group == NULL);
295 avl_add(&mg->mg_metaslab_tree, msp);
296 mutex_exit(&mg->mg_lock);
300 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
302 mutex_enter(&mg->mg_lock);
303 ASSERT(msp->ms_group == mg);
304 avl_remove(&mg->mg_metaslab_tree, msp);
305 msp->ms_group = NULL;
306 mutex_exit(&mg->mg_lock);
310 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
313 * Although in principle the weight can be any value, in
314 * practice we do not use values in the range [1, 510].
316 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
317 ASSERT(MUTEX_HELD(&msp->ms_lock));
319 mutex_enter(&mg->mg_lock);
320 ASSERT(msp->ms_group == mg);
321 avl_remove(&mg->mg_metaslab_tree, msp);
322 msp->ms_weight = weight;
323 avl_add(&mg->mg_metaslab_tree, msp);
324 mutex_exit(&mg->mg_lock);
328 * ==========================================================================
329 * Common allocator routines
330 * ==========================================================================
333 metaslab_segsize_compare(const void *x1, const void *x2)
335 const space_seg_t *s1 = x1;
336 const space_seg_t *s2 = x2;
337 uint64_t ss_size1 = s1->ss_end - s1->ss_start;
338 uint64_t ss_size2 = s2->ss_end - s2->ss_start;
340 if (ss_size1 < ss_size2)
342 if (ss_size1 > ss_size2)
345 if (s1->ss_start < s2->ss_start)
347 if (s1->ss_start > s2->ss_start)
354 * This is a helper function that can be used by the allocator to find
355 * a suitable block to allocate. This will search the specified AVL
356 * tree looking for a block that matches the specified criteria.
359 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
362 space_seg_t *ss, ssearch;
365 ssearch.ss_start = *cursor;
366 ssearch.ss_end = *cursor + size;
368 ss = avl_find(t, &ssearch, &where);
370 ss = avl_nearest(t, where, AVL_AFTER);
373 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
375 if (offset + size <= ss->ss_end) {
376 *cursor = offset + size;
379 ss = AVL_NEXT(t, ss);
383 * If we know we've searched the whole map (*cursor == 0), give up.
384 * Otherwise, reset the cursor to the beginning and try again.
390 return (metaslab_block_picker(t, cursor, size, align));
394 metaslab_pp_load(space_map_t *sm)
398 ASSERT(sm->sm_ppd == NULL);
399 sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
401 sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
402 avl_create(sm->sm_pp_root, metaslab_segsize_compare,
403 sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
405 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
406 avl_add(sm->sm_pp_root, ss);
410 metaslab_pp_unload(space_map_t *sm)
414 kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
417 while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
418 /* tear down the tree */
421 avl_destroy(sm->sm_pp_root);
422 kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
423 sm->sm_pp_root = NULL;
428 metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
430 /* No need to update cursor */
435 metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
437 /* No need to update cursor */
441 * Return the maximum contiguous segment within the metaslab.
444 metaslab_pp_maxsize(space_map_t *sm)
446 avl_tree_t *t = sm->sm_pp_root;
449 if (t == NULL || (ss = avl_last(t)) == NULL)
452 return (ss->ss_end - ss->ss_start);
456 * ==========================================================================
457 * The first-fit block allocator
458 * ==========================================================================
461 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
463 avl_tree_t *t = &sm->sm_root;
464 uint64_t align = size & -size;
465 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
467 return (metaslab_block_picker(t, cursor, size, align));
472 metaslab_ff_fragmented(space_map_t *sm)
477 static space_map_ops_t metaslab_ff_ops = {
484 metaslab_ff_fragmented
488 * ==========================================================================
489 * Dynamic block allocator -
490 * Uses the first fit allocation scheme until space get low and then
491 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
492 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
493 * ==========================================================================
496 metaslab_df_alloc(space_map_t *sm, uint64_t size)
498 avl_tree_t *t = &sm->sm_root;
499 uint64_t align = size & -size;
500 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
501 uint64_t max_size = metaslab_pp_maxsize(sm);
502 int free_pct = sm->sm_space * 100 / sm->sm_size;
504 ASSERT(MUTEX_HELD(sm->sm_lock));
505 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
511 * If we're running low on space switch to using the size
512 * sorted AVL tree (best-fit).
514 if (max_size < metaslab_df_alloc_threshold ||
515 free_pct < metaslab_df_free_pct) {
520 return (metaslab_block_picker(t, cursor, size, 1ULL));
524 metaslab_df_fragmented(space_map_t *sm)
526 uint64_t max_size = metaslab_pp_maxsize(sm);
527 int free_pct = sm->sm_space * 100 / sm->sm_size;
529 if (max_size >= metaslab_df_alloc_threshold &&
530 free_pct >= metaslab_df_free_pct)
536 static space_map_ops_t metaslab_df_ops = {
543 metaslab_df_fragmented
547 * ==========================================================================
548 * Other experimental allocators
549 * ==========================================================================
552 metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
554 avl_tree_t *t = &sm->sm_root;
555 uint64_t *cursor = (uint64_t *)sm->sm_ppd;
556 uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
557 uint64_t max_size = metaslab_pp_maxsize(sm);
558 uint64_t rsize = size;
561 ASSERT(MUTEX_HELD(sm->sm_lock));
562 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
567 ASSERT3U(*extent_end, >=, *cursor);
570 * If we're running low on space switch to using the size
571 * sorted AVL tree (best-fit).
573 if ((*cursor + size) > *extent_end) {
576 *cursor = *extent_end = 0;
578 if (max_size > 2 * SPA_MAXBLOCKSIZE)
579 rsize = MIN(metaslab_min_alloc_size, max_size);
580 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
582 *cursor = offset + size;
584 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
586 ASSERT3U(*cursor, <=, *extent_end);
591 metaslab_cdf_fragmented(space_map_t *sm)
593 uint64_t max_size = metaslab_pp_maxsize(sm);
595 if (max_size > (metaslab_min_alloc_size * 10))
600 static space_map_ops_t metaslab_cdf_ops = {
607 metaslab_cdf_fragmented
610 uint64_t metaslab_ndf_clump_shift = 4;
613 metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
615 avl_tree_t *t = &sm->sm_root;
617 space_seg_t *ss, ssearch;
618 uint64_t hbit = highbit(size);
619 uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
620 uint64_t max_size = metaslab_pp_maxsize(sm);
622 ASSERT(MUTEX_HELD(sm->sm_lock));
623 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
628 ssearch.ss_start = *cursor;
629 ssearch.ss_end = *cursor + size;
631 ss = avl_find(t, &ssearch, &where);
632 if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
635 ssearch.ss_start = 0;
636 ssearch.ss_end = MIN(max_size,
637 1ULL << (hbit + metaslab_ndf_clump_shift));
638 ss = avl_find(t, &ssearch, &where);
640 ss = avl_nearest(t, where, AVL_AFTER);
645 if (ss->ss_start + size <= ss->ss_end) {
646 *cursor = ss->ss_start + size;
647 return (ss->ss_start);
654 metaslab_ndf_fragmented(space_map_t *sm)
656 uint64_t max_size = metaslab_pp_maxsize(sm);
658 if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
664 static space_map_ops_t metaslab_ndf_ops = {
671 metaslab_ndf_fragmented
674 space_map_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops;
677 * ==========================================================================
679 * ==========================================================================
682 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
683 uint64_t start, uint64_t size, uint64_t txg)
685 vdev_t *vd = mg->mg_vd;
688 msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
689 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
691 msp->ms_smo_syncing = *smo;
694 * We create the main space map here, but we don't create the
695 * allocmaps and freemaps until metaslab_sync_done(). This serves
696 * two purposes: it allows metaslab_sync_done() to detect the
697 * addition of new space; and for debugging, it ensures that we'd
698 * data fault on any attempt to use this metaslab before it's ready.
700 space_map_create(&msp->ms_map, start, size,
701 vd->vdev_ashift, &msp->ms_lock);
703 metaslab_group_add(mg, msp);
705 if (metaslab_debug && smo->smo_object != 0) {
706 mutex_enter(&msp->ms_lock);
707 VERIFY(space_map_load(&msp->ms_map, mg->mg_class->mc_ops,
708 SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
709 mutex_exit(&msp->ms_lock);
713 * If we're opening an existing pool (txg == 0) or creating
714 * a new one (txg == TXG_INITIAL), all space is available now.
715 * If we're adding space to an existing pool, the new space
716 * does not become available until after this txg has synced.
718 if (txg <= TXG_INITIAL)
719 metaslab_sync_done(msp, 0);
722 vdev_dirty(vd, 0, NULL, txg);
723 vdev_dirty(vd, VDD_METASLAB, msp, txg);
730 metaslab_fini(metaslab_t *msp)
732 metaslab_group_t *mg = msp->ms_group;
735 vdev_space_update(mg->mg_vd,
736 -msp->ms_smo.smo_alloc, 0, -msp->ms_map.sm_size);
738 metaslab_group_remove(mg, msp);
740 mutex_enter(&msp->ms_lock);
742 space_map_unload(&msp->ms_map);
743 space_map_destroy(&msp->ms_map);
745 for (t = 0; t < TXG_SIZE; t++) {
746 space_map_destroy(&msp->ms_allocmap[t]);
747 space_map_destroy(&msp->ms_freemap[t]);
750 for (t = 0; t < TXG_DEFER_SIZE; t++)
751 space_map_destroy(&msp->ms_defermap[t]);
753 ASSERT3S(msp->ms_deferspace, ==, 0);
755 mutex_exit(&msp->ms_lock);
756 mutex_destroy(&msp->ms_lock);
758 kmem_free(msp, sizeof (metaslab_t));
761 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
762 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
763 #define METASLAB_ACTIVE_MASK \
764 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
767 metaslab_weight(metaslab_t *msp)
769 metaslab_group_t *mg = msp->ms_group;
770 space_map_t *sm = &msp->ms_map;
771 space_map_obj_t *smo = &msp->ms_smo;
772 vdev_t *vd = mg->mg_vd;
773 uint64_t weight, space;
775 ASSERT(MUTEX_HELD(&msp->ms_lock));
778 * The baseline weight is the metaslab's free space.
780 space = sm->sm_size - smo->smo_alloc;
784 * Modern disks have uniform bit density and constant angular velocity.
785 * Therefore, the outer recording zones are faster (higher bandwidth)
786 * than the inner zones by the ratio of outer to inner track diameter,
787 * which is typically around 2:1. We account for this by assigning
788 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
789 * In effect, this means that we'll select the metaslab with the most
790 * free bandwidth rather than simply the one with the most free space.
792 weight = 2 * weight -
793 ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
794 ASSERT(weight >= space && weight <= 2 * space);
797 * For locality, assign higher weight to metaslabs which have
798 * a lower offset than what we've already activated.
800 if (sm->sm_start <= mg->mg_bonus_area)
801 weight *= (metaslab_smo_bonus_pct / 100);
802 ASSERT(weight >= space &&
803 weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
805 if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
807 * If this metaslab is one we're actively using, adjust its
808 * weight to make it preferable to any inactive metaslab so
809 * we'll polish it off.
811 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
817 metaslab_prefetch(metaslab_group_t *mg)
819 spa_t *spa = mg->mg_vd->vdev_spa;
821 avl_tree_t *t = &mg->mg_metaslab_tree;
824 mutex_enter(&mg->mg_lock);
827 * Prefetch the next potential metaslabs
829 for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
830 space_map_t *sm = &msp->ms_map;
831 space_map_obj_t *smo = &msp->ms_smo;
833 /* If we have reached our prefetch limit then we're done */
834 if (m >= metaslab_prefetch_limit)
837 if (!sm->sm_loaded && smo->smo_object != 0) {
838 mutex_exit(&mg->mg_lock);
839 dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
840 0ULL, smo->smo_objsize);
841 mutex_enter(&mg->mg_lock);
844 mutex_exit(&mg->mg_lock);
848 metaslab_activate(metaslab_t *msp, uint64_t activation_weight, uint64_t size)
850 metaslab_group_t *mg = msp->ms_group;
851 space_map_t *sm = &msp->ms_map;
852 space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
855 ASSERT(MUTEX_HELD(&msp->ms_lock));
857 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
858 space_map_load_wait(sm);
859 if (!sm->sm_loaded) {
860 int error = space_map_load(sm, sm_ops, SM_FREE,
862 spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
864 metaslab_group_sort(msp->ms_group, msp, 0);
867 for (t = 0; t < TXG_DEFER_SIZE; t++)
868 space_map_walk(&msp->ms_defermap[t],
869 space_map_claim, sm);
874 * Track the bonus area as we activate new metaslabs.
876 if (sm->sm_start > mg->mg_bonus_area) {
877 mutex_enter(&mg->mg_lock);
878 mg->mg_bonus_area = sm->sm_start;
879 mutex_exit(&mg->mg_lock);
883 * If we were able to load the map then make sure
884 * that this map is still able to satisfy our request.
886 if (msp->ms_weight < size)
889 metaslab_group_sort(msp->ms_group, msp,
890 msp->ms_weight | activation_weight);
892 ASSERT(sm->sm_loaded);
893 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
899 metaslab_passivate(metaslab_t *msp, uint64_t size)
902 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
903 * this metaslab again. In that case, it had better be empty,
904 * or we would be leaving space on the table.
906 ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
907 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
908 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
912 * Write a metaslab to disk in the context of the specified transaction group.
915 metaslab_sync(metaslab_t *msp, uint64_t txg)
917 vdev_t *vd = msp->ms_group->mg_vd;
918 spa_t *spa = vd->vdev_spa;
919 objset_t *mos = spa_meta_objset(spa);
920 space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
921 space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
922 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
923 space_map_t *sm = &msp->ms_map;
924 space_map_obj_t *smo = &msp->ms_smo_syncing;
929 ASSERT(!vd->vdev_ishole);
931 if (allocmap->sm_space == 0 && freemap->sm_space == 0)
935 * The only state that can actually be changing concurrently with
936 * metaslab_sync() is the metaslab's ms_map. No other thread can
937 * be modifying this txg's allocmap, freemap, freed_map, or smo.
938 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
939 * We drop it whenever we call into the DMU, because the DMU
940 * can call down to us (e.g. via zio_free()) at any time.
943 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
945 if (smo->smo_object == 0) {
946 ASSERT(smo->smo_objsize == 0);
947 ASSERT(smo->smo_alloc == 0);
948 smo->smo_object = dmu_object_alloc(mos,
949 DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
950 DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
951 ASSERT(smo->smo_object != 0);
952 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
953 (sm->sm_start >> vd->vdev_ms_shift),
954 sizeof (uint64_t), &smo->smo_object, tx);
957 mutex_enter(&msp->ms_lock);
959 space_map_walk(freemap, space_map_add, freed_map);
961 if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
962 2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
964 * The in-core space map representation is twice as compact
965 * as the on-disk one, so it's time to condense the latter
966 * by generating a pure allocmap from first principles.
968 * This metaslab is 100% allocated,
969 * minus the content of the in-core map (sm),
970 * minus what's been freed this txg (freed_map),
971 * minus deferred frees (ms_defermap[]),
972 * minus allocations from txgs in the future
973 * (because they haven't been committed yet).
975 space_map_vacate(allocmap, NULL, NULL);
976 space_map_vacate(freemap, NULL, NULL);
978 space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
980 space_map_walk(sm, space_map_remove, allocmap);
981 space_map_walk(freed_map, space_map_remove, allocmap);
983 for (t = 0; t < TXG_DEFER_SIZE; t++)
984 space_map_walk(&msp->ms_defermap[t],
985 space_map_remove, allocmap);
987 for (t = 1; t < TXG_CONCURRENT_STATES; t++)
988 space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
989 space_map_remove, allocmap);
991 mutex_exit(&msp->ms_lock);
992 space_map_truncate(smo, mos, tx);
993 mutex_enter(&msp->ms_lock);
996 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
997 space_map_sync(freemap, SM_FREE, smo, mos, tx);
999 mutex_exit(&msp->ms_lock);
1001 VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1002 dmu_buf_will_dirty(db, tx);
1003 ASSERT3U(db->db_size, >=, sizeof (*smo));
1004 bcopy(smo, db->db_data, sizeof (*smo));
1005 dmu_buf_rele(db, FTAG);
1011 * Called after a transaction group has completely synced to mark
1012 * all of the metaslab's free space as usable.
1015 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1017 space_map_obj_t *smo = &msp->ms_smo;
1018 space_map_obj_t *smosync = &msp->ms_smo_syncing;
1019 space_map_t *sm = &msp->ms_map;
1020 space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1021 space_map_t *defer_map = &msp->ms_defermap[txg % TXG_DEFER_SIZE];
1022 metaslab_group_t *mg = msp->ms_group;
1023 vdev_t *vd = mg->mg_vd;
1024 int64_t alloc_delta, defer_delta;
1027 ASSERT(!vd->vdev_ishole);
1029 mutex_enter(&msp->ms_lock);
1032 * If this metaslab is just becoming available, initialize its
1033 * allocmaps and freemaps and add its capacity to the vdev.
1035 if (freed_map->sm_size == 0) {
1036 for (t = 0; t < TXG_SIZE; t++) {
1037 space_map_create(&msp->ms_allocmap[t], sm->sm_start,
1038 sm->sm_size, sm->sm_shift, sm->sm_lock);
1039 space_map_create(&msp->ms_freemap[t], sm->sm_start,
1040 sm->sm_size, sm->sm_shift, sm->sm_lock);
1043 for (t = 0; t < TXG_DEFER_SIZE; t++)
1044 space_map_create(&msp->ms_defermap[t], sm->sm_start,
1045 sm->sm_size, sm->sm_shift, sm->sm_lock);
1047 vdev_space_update(vd, 0, 0, sm->sm_size);
1050 alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1051 defer_delta = freed_map->sm_space - defer_map->sm_space;
1053 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1055 ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
1056 ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
1059 * If there's a space_map_load() in progress, wait for it to complete
1060 * so that we have a consistent view of the in-core space map.
1061 * Then, add defer_map (oldest deferred frees) to this map and
1062 * transfer freed_map (this txg's frees) to defer_map.
1064 space_map_load_wait(sm);
1065 space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1066 space_map_vacate(freed_map, space_map_add, defer_map);
1070 msp->ms_deferspace += defer_delta;
1071 ASSERT3S(msp->ms_deferspace, >=, 0);
1072 ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1073 if (msp->ms_deferspace != 0) {
1075 * Keep syncing this metaslab until all deferred frees
1076 * are back in circulation.
1078 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1082 * If the map is loaded but no longer active, evict it as soon as all
1083 * future allocations have synced. (If we unloaded it now and then
1084 * loaded a moment later, the map wouldn't reflect those allocations.)
1086 if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1089 for (t = 1; t < TXG_CONCURRENT_STATES; t++)
1090 if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
1093 if (evictable && !metaslab_debug)
1094 space_map_unload(sm);
1097 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1099 mutex_exit(&msp->ms_lock);
1103 metaslab_sync_reassess(metaslab_group_t *mg)
1105 vdev_t *vd = mg->mg_vd;
1109 * Re-evaluate all metaslabs which have lower offsets than the
1112 for (m = 0; m < vd->vdev_ms_count; m++) {
1113 metaslab_t *msp = vd->vdev_ms[m];
1115 if (msp->ms_map.sm_start > mg->mg_bonus_area)
1118 mutex_enter(&msp->ms_lock);
1119 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1120 mutex_exit(&msp->ms_lock);
1124 * Prefetch the next potential metaslabs
1126 metaslab_prefetch(mg);
1130 metaslab_distance(metaslab_t *msp, dva_t *dva)
1132 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1133 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1134 uint64_t start = msp->ms_map.sm_start >> ms_shift;
1136 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1137 return (1ULL << 63);
1140 return ((start - offset) << ms_shift);
1142 return ((offset - start) << ms_shift);
1147 metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
1148 uint64_t min_distance, dva_t *dva, int d)
1150 metaslab_t *msp = NULL;
1151 uint64_t offset = -1ULL;
1152 avl_tree_t *t = &mg->mg_metaslab_tree;
1153 uint64_t activation_weight;
1154 uint64_t target_distance;
1157 activation_weight = METASLAB_WEIGHT_PRIMARY;
1158 for (i = 0; i < d; i++) {
1159 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1160 activation_weight = METASLAB_WEIGHT_SECONDARY;
1166 boolean_t was_active;
1168 mutex_enter(&mg->mg_lock);
1169 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1170 if (msp->ms_weight < size) {
1171 mutex_exit(&mg->mg_lock);
1175 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1176 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1179 target_distance = min_distance +
1180 (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1182 for (i = 0; i < d; i++)
1183 if (metaslab_distance(msp, &dva[i]) <
1189 mutex_exit(&mg->mg_lock);
1193 mutex_enter(&msp->ms_lock);
1196 * Ensure that the metaslab we have selected is still
1197 * capable of handling our request. It's possible that
1198 * another thread may have changed the weight while we
1199 * were blocked on the metaslab lock.
1201 if (msp->ms_weight < size || (was_active &&
1202 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1203 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1204 mutex_exit(&msp->ms_lock);
1208 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1209 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1210 metaslab_passivate(msp,
1211 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1212 mutex_exit(&msp->ms_lock);
1216 if (metaslab_activate(msp, activation_weight, size) != 0) {
1217 mutex_exit(&msp->ms_lock);
1221 if ((offset = space_map_alloc(&msp->ms_map, size)) != -1ULL)
1224 metaslab_passivate(msp, space_map_maxsize(&msp->ms_map));
1226 mutex_exit(&msp->ms_lock);
1229 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1230 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1232 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1234 mutex_exit(&msp->ms_lock);
1240 * Allocate a block for the specified i/o.
1243 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1244 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1246 metaslab_group_t *mg, *rotor;
1250 int zio_lock = B_FALSE;
1251 boolean_t allocatable;
1252 uint64_t offset = -1ULL;
1256 ASSERT(!DVA_IS_VALID(&dva[d]));
1259 * For testing, make some blocks above a certain size be gang blocks.
1261 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1265 * Start at the rotor and loop through all mgs until we find something.
1266 * Note that there's no locking on mc_rotor or mc_aliquot because
1267 * nothing actually breaks if we miss a few updates -- we just won't
1268 * allocate quite as evenly. It all balances out over time.
1270 * If we are doing ditto or log blocks, try to spread them across
1271 * consecutive vdevs. If we're forced to reuse a vdev before we've
1272 * allocated all of our ditto blocks, then try and spread them out on
1273 * that vdev as much as possible. If it turns out to not be possible,
1274 * gradually lower our standards until anything becomes acceptable.
1275 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1276 * gives us hope of containing our fault domains to something we're
1277 * able to reason about. Otherwise, any two top-level vdev failures
1278 * will guarantee the loss of data. With consecutive allocation,
1279 * only two adjacent top-level vdev failures will result in data loss.
1281 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1282 * ourselves on the same vdev as our gang block header. That
1283 * way, we can hope for locality in vdev_cache, plus it makes our
1284 * fault domains something tractable.
1287 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1290 * It's possible the vdev we're using as the hint no
1291 * longer exists (i.e. removed). Consult the rotor when
1297 if (flags & METASLAB_HINTBP_AVOID &&
1298 mg->mg_next != NULL)
1303 } else if (d != 0) {
1304 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1305 mg = vd->vdev_mg->mg_next;
1311 * If the hint put us into the wrong metaslab class, or into a
1312 * metaslab group that has been passivated, just follow the rotor.
1314 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1321 ASSERT(mg->mg_activation_count == 1);
1326 * Don't allocate from faulted devices.
1329 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1330 allocatable = vdev_allocatable(vd);
1331 spa_config_exit(spa, SCL_ZIO, FTAG);
1333 allocatable = vdev_allocatable(vd);
1339 * Avoid writing single-copy data to a failing vdev
1341 if ((vd->vdev_stat.vs_write_errors > 0 ||
1342 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1343 d == 0 && dshift == 3) {
1348 ASSERT(mg->mg_class == mc);
1350 distance = vd->vdev_asize >> dshift;
1351 if (distance <= (1ULL << vd->vdev_ms_shift))
1356 asize = vdev_psize_to_asize(vd, psize);
1357 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1359 offset = metaslab_group_alloc(mg, asize, txg, distance, dva, d);
1360 if (offset != -1ULL) {
1362 * If we've just selected this metaslab group,
1363 * figure out whether the corresponding vdev is
1364 * over- or under-used relative to the pool,
1365 * and set an allocation bias to even it out.
1367 if (mc->mc_aliquot == 0) {
1368 vdev_stat_t *vs = &vd->vdev_stat;
1372 * Determine percent used in units of 0..1024.
1373 * (This is just to avoid floating point.)
1375 vu = (vs->vs_alloc << 10) / (vs->vs_space + 1);
1376 cu = (mc->mc_alloc << 10) / (mc->mc_space + 1);
1379 * Bias by at most +/- 25% of the aliquot.
1381 mg->mg_bias = ((cu - vu) *
1382 (int64_t)mg->mg_aliquot) / (1024 * 4);
1385 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1386 mg->mg_aliquot + mg->mg_bias) {
1387 mc->mc_rotor = mg->mg_next;
1391 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1392 DVA_SET_OFFSET(&dva[d], offset);
1393 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1394 DVA_SET_ASIZE(&dva[d], asize);
1399 mc->mc_rotor = mg->mg_next;
1401 } while ((mg = mg->mg_next) != rotor);
1405 ASSERT(dshift < 64);
1409 if (!allocatable && !zio_lock) {
1415 bzero(&dva[d], sizeof (dva_t));
1421 * Free the block represented by DVA in the context of the specified
1422 * transaction group.
1425 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1427 uint64_t vdev = DVA_GET_VDEV(dva);
1428 uint64_t offset = DVA_GET_OFFSET(dva);
1429 uint64_t size = DVA_GET_ASIZE(dva);
1433 ASSERT(DVA_IS_VALID(dva));
1435 if (txg > spa_freeze_txg(spa))
1438 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1439 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1440 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1441 (u_longlong_t)vdev, (u_longlong_t)offset);
1446 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1448 if (DVA_GET_GANG(dva))
1449 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1451 mutex_enter(&msp->ms_lock);
1454 space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
1456 space_map_free(&msp->ms_map, offset, size);
1458 if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
1459 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1460 space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
1463 mutex_exit(&msp->ms_lock);
1467 * Intent log support: upon opening the pool after a crash, notify the SPA
1468 * of blocks that the intent log has allocated for immediate write, but
1469 * which are still considered free by the SPA because the last transaction
1470 * group didn't commit yet.
1473 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1475 uint64_t vdev = DVA_GET_VDEV(dva);
1476 uint64_t offset = DVA_GET_OFFSET(dva);
1477 uint64_t size = DVA_GET_ASIZE(dva);
1482 ASSERT(DVA_IS_VALID(dva));
1484 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1485 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1488 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1490 if (DVA_GET_GANG(dva))
1491 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1493 mutex_enter(&msp->ms_lock);
1495 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map.sm_loaded)
1496 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY, 0);
1498 if (error == 0 && !space_map_contains(&msp->ms_map, offset, size))
1501 if (error || txg == 0) { /* txg == 0 indicates dry run */
1502 mutex_exit(&msp->ms_lock);
1506 space_map_claim(&msp->ms_map, offset, size);
1508 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
1509 if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
1510 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1511 space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
1514 mutex_exit(&msp->ms_lock);
1520 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1521 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1523 dva_t *dva = bp->blk_dva;
1524 dva_t *hintdva = hintbp->blk_dva;
1527 ASSERT(bp->blk_birth == 0);
1528 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1530 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1532 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
1533 spa_config_exit(spa, SCL_ALLOC, FTAG);
1537 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1538 ASSERT(BP_GET_NDVAS(bp) == 0);
1539 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1541 for (d = 0; d < ndvas; d++) {
1542 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1545 for (d--; d >= 0; d--) {
1546 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1547 bzero(&dva[d], sizeof (dva_t));
1549 spa_config_exit(spa, SCL_ALLOC, FTAG);
1554 ASSERT(BP_GET_NDVAS(bp) == ndvas);
1556 spa_config_exit(spa, SCL_ALLOC, FTAG);
1558 BP_SET_BIRTH(bp, txg, txg);
1564 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1566 const dva_t *dva = bp->blk_dva;
1567 int d, ndvas = BP_GET_NDVAS(bp);
1569 ASSERT(!BP_IS_HOLE(bp));
1570 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1572 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1574 for (d = 0; d < ndvas; d++)
1575 metaslab_free_dva(spa, &dva[d], txg, now);
1577 spa_config_exit(spa, SCL_FREE, FTAG);
1581 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1583 const dva_t *dva = bp->blk_dva;
1584 int ndvas = BP_GET_NDVAS(bp);
1587 ASSERT(!BP_IS_HOLE(bp));
1591 * First do a dry run to make sure all DVAs are claimable,
1592 * so we don't have to unwind from partial failures below.
1594 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1598 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1600 for (d = 0; d < ndvas; d++)
1601 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1604 spa_config_exit(spa, SCL_ALLOC, FTAG);
1606 ASSERT(error == 0 || txg == 0);