4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/space_map.h>
30 #include <sys/metaslab_impl.h>
31 #include <sys/vdev_impl.h>
34 #define WITH_DF_BLOCK_ALLOCATOR
37 * Allow allocations to switch to gang blocks quickly. We do this to
38 * avoid having to load lots of space_maps in a given txg. There are,
39 * however, some cases where we want to avoid "fast" ganging and instead
40 * we want to do an exhaustive search of all metaslabs on this device.
41 * Currently we don't allow any gang, zil, or dump device related allocations
44 #define CAN_FASTGANG(flags) \
45 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
46 METASLAB_GANG_AVOID)))
48 uint64_t metaslab_aliquot = 512ULL << 10;
49 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
52 * The in-core space map representation is more compact than its on-disk form.
53 * The zfs_condense_pct determines how much more compact the in-core
54 * space_map representation must be before we compact it on-disk.
55 * Values should be greater than or equal to 100.
57 int zfs_condense_pct = 200;
60 * This value defines the number of allowed allocation failures per vdev.
61 * If a device reaches this threshold in a given txg then we consider skipping
62 * allocations on that device.
64 int zfs_mg_alloc_failures;
67 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
69 int metaslab_debug = 0;
72 * Minimum size which forces the dynamic allocator to change
73 * it's allocation strategy. Once the space map cannot satisfy
74 * an allocation of this size then it switches to using more
75 * aggressive strategy (i.e search by size rather than offset).
77 uint64_t metaslab_df_alloc_threshold = SPA_MAXBLOCKSIZE;
80 * The minimum free space, in percent, which must be available
81 * in a space map to continue allocations in a first-fit fashion.
82 * Once the space_map's free space drops below this level we dynamically
83 * switch to using best-fit allocations.
85 int metaslab_df_free_pct = 4;
88 * A metaslab is considered "free" if it contains a contiguous
89 * segment which is greater than metaslab_min_alloc_size.
91 uint64_t metaslab_min_alloc_size = DMU_MAX_ACCESS;
94 * Max number of space_maps to prefetch.
96 int metaslab_prefetch_limit = SPA_DVAS_PER_BP;
99 * Percentage bonus multiplier for metaslabs that are in the bonus area.
101 int metaslab_smo_bonus_pct = 150;
104 * ==========================================================================
106 * ==========================================================================
109 metaslab_class_create(spa_t *spa, space_map_ops_t *ops)
111 metaslab_class_t *mc;
113 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_PUSHPAGE);
118 mutex_init(&mc->mc_fastwrite_lock, NULL, MUTEX_DEFAULT, NULL);
124 metaslab_class_destroy(metaslab_class_t *mc)
126 ASSERT(mc->mc_rotor == NULL);
127 ASSERT(mc->mc_alloc == 0);
128 ASSERT(mc->mc_deferred == 0);
129 ASSERT(mc->mc_space == 0);
130 ASSERT(mc->mc_dspace == 0);
132 mutex_destroy(&mc->mc_fastwrite_lock);
133 kmem_free(mc, sizeof (metaslab_class_t));
137 metaslab_class_validate(metaslab_class_t *mc)
139 metaslab_group_t *mg;
143 * Must hold one of the spa_config locks.
145 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
146 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
148 if ((mg = mc->mc_rotor) == NULL)
153 ASSERT(vd->vdev_mg != NULL);
154 ASSERT3P(vd->vdev_top, ==, vd);
155 ASSERT3P(mg->mg_class, ==, mc);
156 ASSERT3P(vd->vdev_ops, !=, &vdev_hole_ops);
157 } while ((mg = mg->mg_next) != mc->mc_rotor);
163 metaslab_class_space_update(metaslab_class_t *mc, int64_t alloc_delta,
164 int64_t defer_delta, int64_t space_delta, int64_t dspace_delta)
166 atomic_add_64(&mc->mc_alloc, alloc_delta);
167 atomic_add_64(&mc->mc_deferred, defer_delta);
168 atomic_add_64(&mc->mc_space, space_delta);
169 atomic_add_64(&mc->mc_dspace, dspace_delta);
173 metaslab_class_get_alloc(metaslab_class_t *mc)
175 return (mc->mc_alloc);
179 metaslab_class_get_deferred(metaslab_class_t *mc)
181 return (mc->mc_deferred);
185 metaslab_class_get_space(metaslab_class_t *mc)
187 return (mc->mc_space);
191 metaslab_class_get_dspace(metaslab_class_t *mc)
193 return (spa_deflate(mc->mc_spa) ? mc->mc_dspace : mc->mc_space);
197 * ==========================================================================
199 * ==========================================================================
202 metaslab_compare(const void *x1, const void *x2)
204 const metaslab_t *m1 = x1;
205 const metaslab_t *m2 = x2;
207 if (m1->ms_weight < m2->ms_weight)
209 if (m1->ms_weight > m2->ms_weight)
213 * If the weights are identical, use the offset to force uniqueness.
215 if (m1->ms_map->sm_start < m2->ms_map->sm_start)
217 if (m1->ms_map->sm_start > m2->ms_map->sm_start)
220 ASSERT3P(m1, ==, m2);
226 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
228 metaslab_group_t *mg;
230 mg = kmem_zalloc(sizeof (metaslab_group_t), KM_PUSHPAGE);
231 mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
232 avl_create(&mg->mg_metaslab_tree, metaslab_compare,
233 sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
236 mg->mg_activation_count = 0;
242 metaslab_group_destroy(metaslab_group_t *mg)
244 ASSERT(mg->mg_prev == NULL);
245 ASSERT(mg->mg_next == NULL);
247 * We may have gone below zero with the activation count
248 * either because we never activated in the first place or
249 * because we're done, and possibly removing the vdev.
251 ASSERT(mg->mg_activation_count <= 0);
253 avl_destroy(&mg->mg_metaslab_tree);
254 mutex_destroy(&mg->mg_lock);
255 kmem_free(mg, sizeof (metaslab_group_t));
259 metaslab_group_activate(metaslab_group_t *mg)
261 metaslab_class_t *mc = mg->mg_class;
262 metaslab_group_t *mgprev, *mgnext;
264 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
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);
271 if (++mg->mg_activation_count <= 0)
274 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
276 if ((mgprev = mc->mc_rotor) == NULL) {
280 mgnext = mgprev->mg_next;
281 mg->mg_prev = mgprev;
282 mg->mg_next = mgnext;
283 mgprev->mg_next = mg;
284 mgnext->mg_prev = mg;
290 metaslab_group_passivate(metaslab_group_t *mg)
292 metaslab_class_t *mc = mg->mg_class;
293 metaslab_group_t *mgprev, *mgnext;
295 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
297 if (--mg->mg_activation_count != 0) {
298 ASSERT(mc->mc_rotor != mg);
299 ASSERT(mg->mg_prev == NULL);
300 ASSERT(mg->mg_next == NULL);
301 ASSERT(mg->mg_activation_count < 0);
305 mgprev = mg->mg_prev;
306 mgnext = mg->mg_next;
311 mc->mc_rotor = mgnext;
312 mgprev->mg_next = mgnext;
313 mgnext->mg_prev = mgprev;
321 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
323 mutex_enter(&mg->mg_lock);
324 ASSERT(msp->ms_group == NULL);
327 avl_add(&mg->mg_metaslab_tree, msp);
328 mutex_exit(&mg->mg_lock);
332 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
334 mutex_enter(&mg->mg_lock);
335 ASSERT(msp->ms_group == mg);
336 avl_remove(&mg->mg_metaslab_tree, msp);
337 msp->ms_group = NULL;
338 mutex_exit(&mg->mg_lock);
342 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
345 * Although in principle the weight can be any value, in
346 * practice we do not use values in the range [1, 510].
348 ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
349 ASSERT(MUTEX_HELD(&msp->ms_lock));
351 mutex_enter(&mg->mg_lock);
352 ASSERT(msp->ms_group == mg);
353 avl_remove(&mg->mg_metaslab_tree, msp);
354 msp->ms_weight = weight;
355 avl_add(&mg->mg_metaslab_tree, msp);
356 mutex_exit(&mg->mg_lock);
360 * ==========================================================================
361 * Common allocator routines
362 * ==========================================================================
365 metaslab_segsize_compare(const void *x1, const void *x2)
367 const space_seg_t *s1 = x1;
368 const space_seg_t *s2 = x2;
369 uint64_t ss_size1 = s1->ss_end - s1->ss_start;
370 uint64_t ss_size2 = s2->ss_end - s2->ss_start;
372 if (ss_size1 < ss_size2)
374 if (ss_size1 > ss_size2)
377 if (s1->ss_start < s2->ss_start)
379 if (s1->ss_start > s2->ss_start)
385 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
386 defined(WITH_DF_BLOCK_ALLOCATOR) || \
387 defined(WITH_CDF_BLOCK_ALLOCATOR)
389 * This is a helper function that can be used by the allocator to find
390 * a suitable block to allocate. This will search the specified AVL
391 * tree looking for a block that matches the specified criteria.
394 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
397 space_seg_t *ss, ssearch;
400 ssearch.ss_start = *cursor;
401 ssearch.ss_end = *cursor + size;
403 ss = avl_find(t, &ssearch, &where);
405 ss = avl_nearest(t, where, AVL_AFTER);
408 uint64_t offset = P2ROUNDUP(ss->ss_start, align);
410 if (offset + size <= ss->ss_end) {
411 *cursor = offset + size;
414 ss = AVL_NEXT(t, ss);
418 * If we know we've searched the whole map (*cursor == 0), give up.
419 * Otherwise, reset the cursor to the beginning and try again.
425 return (metaslab_block_picker(t, cursor, size, align));
427 #endif /* WITH_FF/DF/CDF_BLOCK_ALLOCATOR */
430 metaslab_pp_load(space_map_t *sm)
434 ASSERT(sm->sm_ppd == NULL);
435 sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_PUSHPAGE);
437 sm->sm_pp_root = kmem_alloc(sizeof (avl_tree_t), KM_PUSHPAGE);
438 avl_create(sm->sm_pp_root, metaslab_segsize_compare,
439 sizeof (space_seg_t), offsetof(struct space_seg, ss_pp_node));
441 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
442 avl_add(sm->sm_pp_root, ss);
446 metaslab_pp_unload(space_map_t *sm)
450 kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
453 while (avl_destroy_nodes(sm->sm_pp_root, &cookie) != NULL) {
454 /* tear down the tree */
457 avl_destroy(sm->sm_pp_root);
458 kmem_free(sm->sm_pp_root, sizeof (avl_tree_t));
459 sm->sm_pp_root = NULL;
464 metaslab_pp_claim(space_map_t *sm, uint64_t start, uint64_t size)
466 /* No need to update cursor */
471 metaslab_pp_free(space_map_t *sm, uint64_t start, uint64_t size)
473 /* No need to update cursor */
477 * Return the maximum contiguous segment within the metaslab.
480 metaslab_pp_maxsize(space_map_t *sm)
482 avl_tree_t *t = sm->sm_pp_root;
485 if (t == NULL || (ss = avl_last(t)) == NULL)
488 return (ss->ss_end - ss->ss_start);
491 #if defined(WITH_FF_BLOCK_ALLOCATOR)
493 * ==========================================================================
494 * The first-fit block allocator
495 * ==========================================================================
498 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
500 avl_tree_t *t = &sm->sm_root;
501 uint64_t align = size & -size;
502 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
504 return (metaslab_block_picker(t, cursor, size, align));
509 metaslab_ff_fragmented(space_map_t *sm)
514 static space_map_ops_t metaslab_ff_ops = {
521 metaslab_ff_fragmented
524 space_map_ops_t *zfs_metaslab_ops = &metaslab_ff_ops;
525 #endif /* WITH_FF_BLOCK_ALLOCATOR */
527 #if defined(WITH_DF_BLOCK_ALLOCATOR)
529 * ==========================================================================
530 * Dynamic block allocator -
531 * Uses the first fit allocation scheme until space get low and then
532 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
533 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
534 * ==========================================================================
537 metaslab_df_alloc(space_map_t *sm, uint64_t size)
539 avl_tree_t *t = &sm->sm_root;
540 uint64_t align = size & -size;
541 uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
542 uint64_t max_size = metaslab_pp_maxsize(sm);
543 int free_pct = sm->sm_space * 100 / sm->sm_size;
545 ASSERT(MUTEX_HELD(sm->sm_lock));
546 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
552 * If we're running low on space switch to using the size
553 * sorted AVL tree (best-fit).
555 if (max_size < metaslab_df_alloc_threshold ||
556 free_pct < metaslab_df_free_pct) {
561 return (metaslab_block_picker(t, cursor, size, 1ULL));
565 metaslab_df_fragmented(space_map_t *sm)
567 uint64_t max_size = metaslab_pp_maxsize(sm);
568 int free_pct = sm->sm_space * 100 / sm->sm_size;
570 if (max_size >= metaslab_df_alloc_threshold &&
571 free_pct >= metaslab_df_free_pct)
577 static space_map_ops_t metaslab_df_ops = {
584 metaslab_df_fragmented
587 space_map_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
588 #endif /* WITH_DF_BLOCK_ALLOCATOR */
591 * ==========================================================================
592 * Other experimental allocators
593 * ==========================================================================
595 #if defined(WITH_CDF_BLOCK_ALLOCATOR)
597 metaslab_cdf_alloc(space_map_t *sm, uint64_t size)
599 avl_tree_t *t = &sm->sm_root;
600 uint64_t *cursor = (uint64_t *)sm->sm_ppd;
601 uint64_t *extent_end = (uint64_t *)sm->sm_ppd + 1;
602 uint64_t max_size = metaslab_pp_maxsize(sm);
603 uint64_t rsize = size;
606 ASSERT(MUTEX_HELD(sm->sm_lock));
607 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
612 ASSERT3U(*extent_end, >=, *cursor);
615 * If we're running low on space switch to using the size
616 * sorted AVL tree (best-fit).
618 if ((*cursor + size) > *extent_end) {
621 *cursor = *extent_end = 0;
623 if (max_size > 2 * SPA_MAXBLOCKSIZE)
624 rsize = MIN(metaslab_min_alloc_size, max_size);
625 offset = metaslab_block_picker(t, extent_end, rsize, 1ULL);
627 *cursor = offset + size;
629 offset = metaslab_block_picker(t, cursor, rsize, 1ULL);
631 ASSERT3U(*cursor, <=, *extent_end);
636 metaslab_cdf_fragmented(space_map_t *sm)
638 uint64_t max_size = metaslab_pp_maxsize(sm);
640 if (max_size > (metaslab_min_alloc_size * 10))
645 static space_map_ops_t metaslab_cdf_ops = {
652 metaslab_cdf_fragmented
655 space_map_ops_t *zfs_metaslab_ops = &metaslab_cdf_ops;
656 #endif /* WITH_CDF_BLOCK_ALLOCATOR */
658 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
659 uint64_t metaslab_ndf_clump_shift = 4;
662 metaslab_ndf_alloc(space_map_t *sm, uint64_t size)
664 avl_tree_t *t = &sm->sm_root;
666 space_seg_t *ss, ssearch;
667 uint64_t hbit = highbit(size);
668 uint64_t *cursor = (uint64_t *)sm->sm_ppd + hbit - 1;
669 uint64_t max_size = metaslab_pp_maxsize(sm);
671 ASSERT(MUTEX_HELD(sm->sm_lock));
672 ASSERT3U(avl_numnodes(&sm->sm_root), ==, avl_numnodes(sm->sm_pp_root));
677 ssearch.ss_start = *cursor;
678 ssearch.ss_end = *cursor + size;
680 ss = avl_find(t, &ssearch, &where);
681 if (ss == NULL || (ss->ss_start + size > ss->ss_end)) {
684 ssearch.ss_start = 0;
685 ssearch.ss_end = MIN(max_size,
686 1ULL << (hbit + metaslab_ndf_clump_shift));
687 ss = avl_find(t, &ssearch, &where);
689 ss = avl_nearest(t, where, AVL_AFTER);
694 if (ss->ss_start + size <= ss->ss_end) {
695 *cursor = ss->ss_start + size;
696 return (ss->ss_start);
703 metaslab_ndf_fragmented(space_map_t *sm)
705 uint64_t max_size = metaslab_pp_maxsize(sm);
707 if (max_size > (metaslab_min_alloc_size << metaslab_ndf_clump_shift))
713 static space_map_ops_t metaslab_ndf_ops = {
720 metaslab_ndf_fragmented
723 space_map_ops_t *zfs_metaslab_ops = &metaslab_ndf_ops;
724 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
727 * ==========================================================================
729 * ==========================================================================
732 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
733 uint64_t start, uint64_t size, uint64_t txg)
735 vdev_t *vd = mg->mg_vd;
738 msp = kmem_zalloc(sizeof (metaslab_t), KM_PUSHPAGE);
739 mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
741 msp->ms_smo_syncing = *smo;
744 * We create the main space map here, but we don't create the
745 * allocmaps and freemaps until metaslab_sync_done(). This serves
746 * two purposes: it allows metaslab_sync_done() to detect the
747 * addition of new space; and for debugging, it ensures that we'd
748 * data fault on any attempt to use this metaslab before it's ready.
750 msp->ms_map = kmem_zalloc(sizeof (space_map_t), KM_PUSHPAGE);
751 space_map_create(msp->ms_map, start, size,
752 vd->vdev_ashift, &msp->ms_lock);
754 metaslab_group_add(mg, msp);
756 if (metaslab_debug && smo->smo_object != 0) {
757 mutex_enter(&msp->ms_lock);
758 VERIFY(space_map_load(msp->ms_map, mg->mg_class->mc_ops,
759 SM_FREE, smo, spa_meta_objset(vd->vdev_spa)) == 0);
760 mutex_exit(&msp->ms_lock);
764 * If we're opening an existing pool (txg == 0) or creating
765 * a new one (txg == TXG_INITIAL), all space is available now.
766 * If we're adding space to an existing pool, the new space
767 * does not become available until after this txg has synced.
769 if (txg <= TXG_INITIAL)
770 metaslab_sync_done(msp, 0);
773 vdev_dirty(vd, 0, NULL, txg);
774 vdev_dirty(vd, VDD_METASLAB, msp, txg);
781 metaslab_fini(metaslab_t *msp)
783 metaslab_group_t *mg = msp->ms_group;
786 vdev_space_update(mg->mg_vd,
787 -msp->ms_smo.smo_alloc, 0, -msp->ms_map->sm_size);
789 metaslab_group_remove(mg, msp);
791 mutex_enter(&msp->ms_lock);
793 space_map_unload(msp->ms_map);
794 space_map_destroy(msp->ms_map);
795 kmem_free(msp->ms_map, sizeof (*msp->ms_map));
797 for (t = 0; t < TXG_SIZE; t++) {
798 space_map_destroy(msp->ms_allocmap[t]);
799 space_map_destroy(msp->ms_freemap[t]);
800 kmem_free(msp->ms_allocmap[t], sizeof (*msp->ms_allocmap[t]));
801 kmem_free(msp->ms_freemap[t], sizeof (*msp->ms_freemap[t]));
804 for (t = 0; t < TXG_DEFER_SIZE; t++) {
805 space_map_destroy(msp->ms_defermap[t]);
806 kmem_free(msp->ms_defermap[t], sizeof (*msp->ms_defermap[t]));
809 ASSERT0(msp->ms_deferspace);
811 mutex_exit(&msp->ms_lock);
812 mutex_destroy(&msp->ms_lock);
814 kmem_free(msp, sizeof (metaslab_t));
817 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
818 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
819 #define METASLAB_ACTIVE_MASK \
820 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
823 metaslab_weight(metaslab_t *msp)
825 metaslab_group_t *mg = msp->ms_group;
826 space_map_t *sm = msp->ms_map;
827 space_map_obj_t *smo = &msp->ms_smo;
828 vdev_t *vd = mg->mg_vd;
829 uint64_t weight, space;
831 ASSERT(MUTEX_HELD(&msp->ms_lock));
834 * The baseline weight is the metaslab's free space.
836 space = sm->sm_size - smo->smo_alloc;
840 * Modern disks have uniform bit density and constant angular velocity.
841 * Therefore, the outer recording zones are faster (higher bandwidth)
842 * than the inner zones by the ratio of outer to inner track diameter,
843 * which is typically around 2:1. We account for this by assigning
844 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
845 * In effect, this means that we'll select the metaslab with the most
846 * free bandwidth rather than simply the one with the most free space.
848 weight = 2 * weight -
849 ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
850 ASSERT(weight >= space && weight <= 2 * space);
853 * For locality, assign higher weight to metaslabs which have
854 * a lower offset than what we've already activated.
856 if (sm->sm_start <= mg->mg_bonus_area)
857 weight *= (metaslab_smo_bonus_pct / 100);
858 ASSERT(weight >= space &&
859 weight <= 2 * (metaslab_smo_bonus_pct / 100) * space);
861 if (sm->sm_loaded && !sm->sm_ops->smop_fragmented(sm)) {
863 * If this metaslab is one we're actively using, adjust its
864 * weight to make it preferable to any inactive metaslab so
865 * we'll polish it off.
867 weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
873 metaslab_prefetch(metaslab_group_t *mg)
875 spa_t *spa = mg->mg_vd->vdev_spa;
877 avl_tree_t *t = &mg->mg_metaslab_tree;
880 mutex_enter(&mg->mg_lock);
883 * Prefetch the next potential metaslabs
885 for (msp = avl_first(t), m = 0; msp; msp = AVL_NEXT(t, msp), m++) {
886 space_map_t *sm = msp->ms_map;
887 space_map_obj_t *smo = &msp->ms_smo;
889 /* If we have reached our prefetch limit then we're done */
890 if (m >= metaslab_prefetch_limit)
893 if (!sm->sm_loaded && smo->smo_object != 0) {
894 mutex_exit(&mg->mg_lock);
895 dmu_prefetch(spa_meta_objset(spa), smo->smo_object,
896 0ULL, smo->smo_objsize);
897 mutex_enter(&mg->mg_lock);
900 mutex_exit(&mg->mg_lock);
904 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
906 metaslab_group_t *mg = msp->ms_group;
907 space_map_t *sm = msp->ms_map;
908 space_map_ops_t *sm_ops = msp->ms_group->mg_class->mc_ops;
911 ASSERT(MUTEX_HELD(&msp->ms_lock));
913 if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
914 space_map_load_wait(sm);
915 if (!sm->sm_loaded) {
916 space_map_obj_t *smo = &msp->ms_smo;
918 int error = space_map_load(sm, sm_ops, SM_FREE, smo,
919 spa_meta_objset(msp->ms_group->mg_vd->vdev_spa));
921 metaslab_group_sort(msp->ms_group, msp, 0);
924 for (t = 0; t < TXG_DEFER_SIZE; t++)
925 space_map_walk(msp->ms_defermap[t],
926 space_map_claim, sm);
931 * Track the bonus area as we activate new metaslabs.
933 if (sm->sm_start > mg->mg_bonus_area) {
934 mutex_enter(&mg->mg_lock);
935 mg->mg_bonus_area = sm->sm_start;
936 mutex_exit(&mg->mg_lock);
939 metaslab_group_sort(msp->ms_group, msp,
940 msp->ms_weight | activation_weight);
942 ASSERT(sm->sm_loaded);
943 ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
949 metaslab_passivate(metaslab_t *msp, uint64_t size)
952 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
953 * this metaslab again. In that case, it had better be empty,
954 * or we would be leaving space on the table.
956 ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map->sm_space == 0);
957 metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
958 ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
962 * Determine if the in-core space map representation can be condensed on-disk.
963 * We would like to use the following criteria to make our decision:
965 * 1. The size of the space map object should not dramatically increase as a
966 * result of writing out our in-core free map.
968 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
969 * times the size than the in-core representation (i.e. zfs_condense_pct = 110
970 * and in-core = 1MB, minimal = 1.1.MB).
972 * Checking the first condition is tricky since we don't want to walk
973 * the entire AVL tree calculating the estimated on-disk size. Instead we
974 * use the size-ordered AVL tree in the space map and calculate the
975 * size required for the largest segment in our in-core free map. If the
976 * size required to represent that segment on disk is larger than the space
977 * map object then we avoid condensing this map.
979 * To determine the second criterion we use a best-case estimate and assume
980 * each segment can be represented on-disk as a single 64-bit entry. We refer
981 * to this best-case estimate as the space map's minimal form.
984 metaslab_should_condense(metaslab_t *msp)
986 space_map_t *sm = msp->ms_map;
987 space_map_obj_t *smo = &msp->ms_smo_syncing;
989 uint64_t size, entries, segsz;
991 ASSERT(MUTEX_HELD(&msp->ms_lock));
992 ASSERT(sm->sm_loaded);
995 * Use the sm_pp_root AVL tree, which is ordered by size, to obtain
996 * the largest segment in the in-core free map. If the tree is
997 * empty then we should condense the map.
999 ss = avl_last(sm->sm_pp_root);
1004 * Calculate the number of 64-bit entries this segment would
1005 * require when written to disk. If this single segment would be
1006 * larger on-disk than the entire current on-disk structure, then
1007 * clearly condensing will increase the on-disk structure size.
1009 size = (ss->ss_end - ss->ss_start) >> sm->sm_shift;
1010 entries = size / (MIN(size, SM_RUN_MAX));
1011 segsz = entries * sizeof (uint64_t);
1013 return (segsz <= smo->smo_objsize &&
1014 smo->smo_objsize >= (zfs_condense_pct *
1015 sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) / 100);
1019 * Condense the on-disk space map representation to its minimized form.
1020 * The minimized form consists of a small number of allocations followed by
1021 * the in-core free map.
1024 metaslab_condense(metaslab_t *msp, uint64_t txg, dmu_tx_t *tx)
1026 spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
1027 space_map_t *freemap = msp->ms_freemap[txg & TXG_MASK];
1028 space_map_t condense_map;
1029 space_map_t *sm = msp->ms_map;
1030 objset_t *mos = spa_meta_objset(spa);
1031 space_map_obj_t *smo = &msp->ms_smo_syncing;
1034 ASSERT(MUTEX_HELD(&msp->ms_lock));
1035 ASSERT3U(spa_sync_pass(spa), ==, 1);
1036 ASSERT(sm->sm_loaded);
1038 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, "
1039 "smo size %llu, segments %lu", txg,
1040 (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
1041 smo->smo_objsize, avl_numnodes(&sm->sm_root));
1044 * Create an map that is a 100% allocated map. We remove segments
1045 * that have been freed in this txg, any deferred frees that exist,
1046 * and any allocation in the future. Removing segments should be
1047 * a relatively inexpensive operation since we expect these maps to
1048 * a small number of nodes.
1050 space_map_create(&condense_map, sm->sm_start, sm->sm_size,
1051 sm->sm_shift, sm->sm_lock);
1052 space_map_add(&condense_map, condense_map.sm_start,
1053 condense_map.sm_size);
1056 * Remove what's been freed in this txg from the condense_map.
1057 * Since we're in sync_pass 1, we know that all the frees from
1058 * this txg are in the freemap.
1060 space_map_walk(freemap, space_map_remove, &condense_map);
1062 for (t = 0; t < TXG_DEFER_SIZE; t++)
1063 space_map_walk(msp->ms_defermap[t],
1064 space_map_remove, &condense_map);
1066 for (t = 1; t < TXG_CONCURRENT_STATES; t++)
1067 space_map_walk(msp->ms_allocmap[(txg + t) & TXG_MASK],
1068 space_map_remove, &condense_map);
1071 * We're about to drop the metaslab's lock thus allowing
1072 * other consumers to change it's content. Set the
1073 * space_map's sm_condensing flag to ensure that
1074 * allocations on this metaslab do not occur while we're
1075 * in the middle of committing it to disk. This is only critical
1076 * for the ms_map as all other space_maps use per txg
1077 * views of their content.
1079 sm->sm_condensing = B_TRUE;
1081 mutex_exit(&msp->ms_lock);
1082 space_map_truncate(smo, mos, tx);
1083 mutex_enter(&msp->ms_lock);
1086 * While we would ideally like to create a space_map representation
1087 * that consists only of allocation records, doing so can be
1088 * prohibitively expensive because the in-core free map can be
1089 * large, and therefore computationally expensive to subtract
1090 * from the condense_map. Instead we sync out two maps, a cheap
1091 * allocation only map followed by the in-core free map. While not
1092 * optimal, this is typically close to optimal, and much cheaper to
1095 space_map_sync(&condense_map, SM_ALLOC, smo, mos, tx);
1096 space_map_vacate(&condense_map, NULL, NULL);
1097 space_map_destroy(&condense_map);
1099 space_map_sync(sm, SM_FREE, smo, mos, tx);
1100 sm->sm_condensing = B_FALSE;
1102 spa_dbgmsg(spa, "condensed: txg %llu, msp[%llu] %p, "
1103 "smo size %llu", txg,
1104 (msp->ms_map->sm_start / msp->ms_map->sm_size), msp,
1109 * Write a metaslab to disk in the context of the specified transaction group.
1112 metaslab_sync(metaslab_t *msp, uint64_t txg)
1114 vdev_t *vd = msp->ms_group->mg_vd;
1115 spa_t *spa = vd->vdev_spa;
1116 objset_t *mos = spa_meta_objset(spa);
1117 space_map_t *allocmap = msp->ms_allocmap[txg & TXG_MASK];
1118 space_map_t **freemap = &msp->ms_freemap[txg & TXG_MASK];
1119 space_map_t **freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1120 space_map_t *sm = msp->ms_map;
1121 space_map_obj_t *smo = &msp->ms_smo_syncing;
1125 ASSERT(!vd->vdev_ishole);
1128 * This metaslab has just been added so there's no work to do now.
1130 if (*freemap == NULL) {
1131 ASSERT3P(allocmap, ==, NULL);
1135 ASSERT3P(allocmap, !=, NULL);
1136 ASSERT3P(*freemap, !=, NULL);
1137 ASSERT3P(*freed_map, !=, NULL);
1139 if (allocmap->sm_space == 0 && (*freemap)->sm_space == 0)
1143 * The only state that can actually be changing concurrently with
1144 * metaslab_sync() is the metaslab's ms_map. No other thread can
1145 * be modifying this txg's allocmap, freemap, freed_map, or smo.
1146 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
1147 * We drop it whenever we call into the DMU, because the DMU
1148 * can call down to us (e.g. via zio_free()) at any time.
1151 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
1153 if (smo->smo_object == 0) {
1154 ASSERT(smo->smo_objsize == 0);
1155 ASSERT(smo->smo_alloc == 0);
1156 smo->smo_object = dmu_object_alloc(mos,
1157 DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1158 DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1159 ASSERT(smo->smo_object != 0);
1160 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
1161 (sm->sm_start >> vd->vdev_ms_shift),
1162 sizeof (uint64_t), &smo->smo_object, tx);
1165 mutex_enter(&msp->ms_lock);
1167 if (sm->sm_loaded && spa_sync_pass(spa) == 1 &&
1168 metaslab_should_condense(msp)) {
1169 metaslab_condense(msp, txg, tx);
1171 space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
1172 space_map_sync(*freemap, SM_FREE, smo, mos, tx);
1175 space_map_vacate(allocmap, NULL, NULL);
1178 * For sync pass 1, we avoid walking the entire space map and
1179 * instead will just swap the pointers for freemap and
1180 * freed_map. We can safely do this since the freed_map is
1181 * guaranteed to be empty on the initial pass.
1183 if (spa_sync_pass(spa) == 1) {
1184 ASSERT0((*freed_map)->sm_space);
1185 ASSERT0(avl_numnodes(&(*freed_map)->sm_root));
1186 space_map_swap(freemap, freed_map);
1188 space_map_vacate(*freemap, space_map_add, *freed_map);
1191 ASSERT0(msp->ms_allocmap[txg & TXG_MASK]->sm_space);
1192 ASSERT0(msp->ms_freemap[txg & TXG_MASK]->sm_space);
1194 mutex_exit(&msp->ms_lock);
1196 VERIFY0(dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1197 dmu_buf_will_dirty(db, tx);
1198 ASSERT3U(db->db_size, >=, sizeof (*smo));
1199 bcopy(smo, db->db_data, sizeof (*smo));
1200 dmu_buf_rele(db, FTAG);
1206 * Called after a transaction group has completely synced to mark
1207 * all of the metaslab's free space as usable.
1210 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
1212 space_map_obj_t *smo = &msp->ms_smo;
1213 space_map_obj_t *smosync = &msp->ms_smo_syncing;
1214 space_map_t *sm = msp->ms_map;
1215 space_map_t *freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1216 space_map_t *defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE];
1217 metaslab_group_t *mg = msp->ms_group;
1218 vdev_t *vd = mg->mg_vd;
1219 int64_t alloc_delta, defer_delta;
1222 ASSERT(!vd->vdev_ishole);
1224 mutex_enter(&msp->ms_lock);
1227 * If this metaslab is just becoming available, initialize its
1228 * allocmaps, freemaps, and defermap and add its capacity to the vdev.
1230 if (freed_map == NULL) {
1231 ASSERT(defer_map == NULL);
1232 for (t = 0; t < TXG_SIZE; t++) {
1233 msp->ms_allocmap[t] = kmem_zalloc(sizeof (space_map_t),
1235 space_map_create(msp->ms_allocmap[t], sm->sm_start,
1236 sm->sm_size, sm->sm_shift, sm->sm_lock);
1237 msp->ms_freemap[t] = kmem_zalloc(sizeof (space_map_t),
1239 space_map_create(msp->ms_freemap[t], sm->sm_start,
1240 sm->sm_size, sm->sm_shift, sm->sm_lock);
1243 for (t = 0; t < TXG_DEFER_SIZE; t++) {
1244 msp->ms_defermap[t] = kmem_zalloc(sizeof (space_map_t),
1246 space_map_create(msp->ms_defermap[t], sm->sm_start,
1247 sm->sm_size, sm->sm_shift, sm->sm_lock);
1250 freed_map = msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
1251 defer_map = msp->ms_defermap[txg % TXG_DEFER_SIZE];
1253 vdev_space_update(vd, 0, 0, sm->sm_size);
1256 alloc_delta = smosync->smo_alloc - smo->smo_alloc;
1257 defer_delta = freed_map->sm_space - defer_map->sm_space;
1259 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
1261 ASSERT(msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0);
1262 ASSERT(msp->ms_freemap[txg & TXG_MASK]->sm_space == 0);
1265 * If there's a space_map_load() in progress, wait for it to complete
1266 * so that we have a consistent view of the in-core space map.
1267 * Then, add defer_map (oldest deferred frees) to this map and
1268 * transfer freed_map (this txg's frees) to defer_map.
1270 space_map_load_wait(sm);
1271 space_map_vacate(defer_map, sm->sm_loaded ? space_map_free : NULL, sm);
1272 space_map_vacate(freed_map, space_map_add, defer_map);
1276 msp->ms_deferspace += defer_delta;
1277 ASSERT3S(msp->ms_deferspace, >=, 0);
1278 ASSERT3S(msp->ms_deferspace, <=, sm->sm_size);
1279 if (msp->ms_deferspace != 0) {
1281 * Keep syncing this metaslab until all deferred frees
1282 * are back in circulation.
1284 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
1288 * If the map is loaded but no longer active, evict it as soon as all
1289 * future allocations have synced. (If we unloaded it now and then
1290 * loaded a moment later, the map wouldn't reflect those allocations.)
1292 if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
1295 for (t = 1; t < TXG_CONCURRENT_STATES; t++)
1296 if (msp->ms_allocmap[(txg + t) & TXG_MASK]->sm_space)
1299 if (evictable && !metaslab_debug)
1300 space_map_unload(sm);
1303 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1305 mutex_exit(&msp->ms_lock);
1309 metaslab_sync_reassess(metaslab_group_t *mg)
1311 vdev_t *vd = mg->mg_vd;
1312 int64_t failures = mg->mg_alloc_failures;
1316 * Re-evaluate all metaslabs which have lower offsets than the
1319 for (m = 0; m < vd->vdev_ms_count; m++) {
1320 metaslab_t *msp = vd->vdev_ms[m];
1322 if (msp->ms_map->sm_start > mg->mg_bonus_area)
1325 mutex_enter(&msp->ms_lock);
1326 metaslab_group_sort(mg, msp, metaslab_weight(msp));
1327 mutex_exit(&msp->ms_lock);
1330 atomic_add_64(&mg->mg_alloc_failures, -failures);
1333 * Prefetch the next potential metaslabs
1335 metaslab_prefetch(mg);
1339 metaslab_distance(metaslab_t *msp, dva_t *dva)
1341 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
1342 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
1343 uint64_t start = msp->ms_map->sm_start >> ms_shift;
1345 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
1346 return (1ULL << 63);
1349 return ((start - offset) << ms_shift);
1351 return ((offset - start) << ms_shift);
1356 metaslab_group_alloc(metaslab_group_t *mg, uint64_t psize, uint64_t asize,
1357 uint64_t txg, uint64_t min_distance, dva_t *dva, int d, int flags)
1359 spa_t *spa = mg->mg_vd->vdev_spa;
1360 metaslab_t *msp = NULL;
1361 uint64_t offset = -1ULL;
1362 avl_tree_t *t = &mg->mg_metaslab_tree;
1363 uint64_t activation_weight;
1364 uint64_t target_distance;
1367 activation_weight = METASLAB_WEIGHT_PRIMARY;
1368 for (i = 0; i < d; i++) {
1369 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
1370 activation_weight = METASLAB_WEIGHT_SECONDARY;
1376 boolean_t was_active;
1378 mutex_enter(&mg->mg_lock);
1379 for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
1380 if (msp->ms_weight < asize) {
1381 spa_dbgmsg(spa, "%s: failed to meet weight "
1382 "requirement: vdev %llu, txg %llu, mg %p, "
1383 "msp %p, psize %llu, asize %llu, "
1384 "failures %llu, weight %llu",
1385 spa_name(spa), mg->mg_vd->vdev_id, txg,
1386 mg, msp, psize, asize,
1387 mg->mg_alloc_failures, msp->ms_weight);
1388 mutex_exit(&mg->mg_lock);
1391 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
1392 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
1395 target_distance = min_distance +
1396 (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
1398 for (i = 0; i < d; i++)
1399 if (metaslab_distance(msp, &dva[i]) <
1405 mutex_exit(&mg->mg_lock);
1410 * If we've already reached the allowable number of failed
1411 * allocation attempts on this metaslab group then we
1412 * consider skipping it. We skip it only if we're allowed
1413 * to "fast" gang, the physical size is larger than
1414 * a gang block, and we're attempting to allocate from
1415 * the primary metaslab.
1417 if (mg->mg_alloc_failures > zfs_mg_alloc_failures &&
1418 CAN_FASTGANG(flags) && psize > SPA_GANGBLOCKSIZE &&
1419 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1420 spa_dbgmsg(spa, "%s: skipping metaslab group: "
1421 "vdev %llu, txg %llu, mg %p, psize %llu, "
1422 "asize %llu, failures %llu", spa_name(spa),
1423 mg->mg_vd->vdev_id, txg, mg, psize, asize,
1424 mg->mg_alloc_failures);
1428 mutex_enter(&msp->ms_lock);
1431 * If this metaslab is currently condensing then pick again as
1432 * we can't manipulate this metaslab until it's committed
1435 if (msp->ms_map->sm_condensing) {
1436 mutex_exit(&msp->ms_lock);
1441 * Ensure that the metaslab we have selected is still
1442 * capable of handling our request. It's possible that
1443 * another thread may have changed the weight while we
1444 * were blocked on the metaslab lock.
1446 if (msp->ms_weight < asize || (was_active &&
1447 !(msp->ms_weight & METASLAB_ACTIVE_MASK) &&
1448 activation_weight == METASLAB_WEIGHT_PRIMARY)) {
1449 mutex_exit(&msp->ms_lock);
1453 if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
1454 activation_weight == METASLAB_WEIGHT_PRIMARY) {
1455 metaslab_passivate(msp,
1456 msp->ms_weight & ~METASLAB_ACTIVE_MASK);
1457 mutex_exit(&msp->ms_lock);
1461 if (metaslab_activate(msp, activation_weight) != 0) {
1462 mutex_exit(&msp->ms_lock);
1466 if ((offset = space_map_alloc(msp->ms_map, asize)) != -1ULL)
1469 atomic_inc_64(&mg->mg_alloc_failures);
1471 metaslab_passivate(msp, space_map_maxsize(msp->ms_map));
1473 mutex_exit(&msp->ms_lock);
1476 if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
1477 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
1479 space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, asize);
1481 mutex_exit(&msp->ms_lock);
1487 * Allocate a block for the specified i/o.
1490 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
1491 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags)
1493 metaslab_group_t *mg, *fast_mg, *rotor;
1497 int zio_lock = B_FALSE;
1498 boolean_t allocatable;
1499 uint64_t offset = -1ULL;
1503 ASSERT(!DVA_IS_VALID(&dva[d]));
1506 * For testing, make some blocks above a certain size be gang blocks.
1508 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0)
1511 if (flags & METASLAB_FASTWRITE)
1512 mutex_enter(&mc->mc_fastwrite_lock);
1515 * Start at the rotor and loop through all mgs until we find something.
1516 * Note that there's no locking on mc_rotor or mc_aliquot because
1517 * nothing actually breaks if we miss a few updates -- we just won't
1518 * allocate quite as evenly. It all balances out over time.
1520 * If we are doing ditto or log blocks, try to spread them across
1521 * consecutive vdevs. If we're forced to reuse a vdev before we've
1522 * allocated all of our ditto blocks, then try and spread them out on
1523 * that vdev as much as possible. If it turns out to not be possible,
1524 * gradually lower our standards until anything becomes acceptable.
1525 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1526 * gives us hope of containing our fault domains to something we're
1527 * able to reason about. Otherwise, any two top-level vdev failures
1528 * will guarantee the loss of data. With consecutive allocation,
1529 * only two adjacent top-level vdev failures will result in data loss.
1531 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1532 * ourselves on the same vdev as our gang block header. That
1533 * way, we can hope for locality in vdev_cache, plus it makes our
1534 * fault domains something tractable.
1537 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
1540 * It's possible the vdev we're using as the hint no
1541 * longer exists (i.e. removed). Consult the rotor when
1547 if (flags & METASLAB_HINTBP_AVOID &&
1548 mg->mg_next != NULL)
1553 } else if (d != 0) {
1554 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
1555 mg = vd->vdev_mg->mg_next;
1556 } else if (flags & METASLAB_FASTWRITE) {
1557 mg = fast_mg = mc->mc_rotor;
1560 if (fast_mg->mg_vd->vdev_pending_fastwrite <
1561 mg->mg_vd->vdev_pending_fastwrite)
1563 } while ((fast_mg = fast_mg->mg_next) != mc->mc_rotor);
1570 * If the hint put us into the wrong metaslab class, or into a
1571 * metaslab group that has been passivated, just follow the rotor.
1573 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
1580 ASSERT(mg->mg_activation_count == 1);
1585 * Don't allocate from faulted devices.
1588 spa_config_enter(spa, SCL_ZIO, FTAG, RW_READER);
1589 allocatable = vdev_allocatable(vd);
1590 spa_config_exit(spa, SCL_ZIO, FTAG);
1592 allocatable = vdev_allocatable(vd);
1598 * Avoid writing single-copy data to a failing vdev
1600 if ((vd->vdev_stat.vs_write_errors > 0 ||
1601 vd->vdev_state < VDEV_STATE_HEALTHY) &&
1602 d == 0 && dshift == 3) {
1607 ASSERT(mg->mg_class == mc);
1609 distance = vd->vdev_asize >> dshift;
1610 if (distance <= (1ULL << vd->vdev_ms_shift))
1615 asize = vdev_psize_to_asize(vd, psize);
1616 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
1618 offset = metaslab_group_alloc(mg, psize, asize, txg, distance,
1620 if (offset != -1ULL) {
1622 * If we've just selected this metaslab group,
1623 * figure out whether the corresponding vdev is
1624 * over- or under-used relative to the pool,
1625 * and set an allocation bias to even it out.
1627 if (mc->mc_aliquot == 0) {
1628 vdev_stat_t *vs = &vd->vdev_stat;
1631 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
1632 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
1635 * Calculate how much more or less we should
1636 * try to allocate from this device during
1637 * this iteration around the rotor.
1638 * For example, if a device is 80% full
1639 * and the pool is 20% full then we should
1640 * reduce allocations by 60% on this device.
1642 * mg_bias = (20 - 80) * 512K / 100 = -307K
1644 * This reduces allocations by 307K for this
1647 mg->mg_bias = ((cu - vu) *
1648 (int64_t)mg->mg_aliquot) / 100;
1651 if ((flags & METASLAB_FASTWRITE) ||
1652 atomic_add_64_nv(&mc->mc_aliquot, asize) >=
1653 mg->mg_aliquot + mg->mg_bias) {
1654 mc->mc_rotor = mg->mg_next;
1658 DVA_SET_VDEV(&dva[d], vd->vdev_id);
1659 DVA_SET_OFFSET(&dva[d], offset);
1660 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
1661 DVA_SET_ASIZE(&dva[d], asize);
1663 if (flags & METASLAB_FASTWRITE) {
1664 atomic_add_64(&vd->vdev_pending_fastwrite,
1666 mutex_exit(&mc->mc_fastwrite_lock);
1672 mc->mc_rotor = mg->mg_next;
1674 } while ((mg = mg->mg_next) != rotor);
1678 ASSERT(dshift < 64);
1682 if (!allocatable && !zio_lock) {
1688 bzero(&dva[d], sizeof (dva_t));
1690 if (flags & METASLAB_FASTWRITE)
1691 mutex_exit(&mc->mc_fastwrite_lock);
1696 * Free the block represented by DVA in the context of the specified
1697 * transaction group.
1700 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
1702 uint64_t vdev = DVA_GET_VDEV(dva);
1703 uint64_t offset = DVA_GET_OFFSET(dva);
1704 uint64_t size = DVA_GET_ASIZE(dva);
1708 ASSERT(DVA_IS_VALID(dva));
1710 if (txg > spa_freeze_txg(spa))
1713 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1714 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
1715 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
1716 (u_longlong_t)vdev, (u_longlong_t)offset);
1721 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1723 if (DVA_GET_GANG(dva))
1724 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1726 mutex_enter(&msp->ms_lock);
1729 space_map_remove(msp->ms_allocmap[txg & TXG_MASK],
1731 space_map_free(msp->ms_map, offset, size);
1733 if (msp->ms_freemap[txg & TXG_MASK]->sm_space == 0)
1734 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1735 space_map_add(msp->ms_freemap[txg & TXG_MASK], offset, size);
1738 mutex_exit(&msp->ms_lock);
1742 * Intent log support: upon opening the pool after a crash, notify the SPA
1743 * of blocks that the intent log has allocated for immediate write, but
1744 * which are still considered free by the SPA because the last transaction
1745 * group didn't commit yet.
1748 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
1750 uint64_t vdev = DVA_GET_VDEV(dva);
1751 uint64_t offset = DVA_GET_OFFSET(dva);
1752 uint64_t size = DVA_GET_ASIZE(dva);
1757 ASSERT(DVA_IS_VALID(dva));
1759 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
1760 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
1763 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1765 if (DVA_GET_GANG(dva))
1766 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
1768 mutex_enter(&msp->ms_lock);
1770 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_map->sm_loaded)
1771 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
1773 if (error == 0 && !space_map_contains(msp->ms_map, offset, size))
1776 if (error || txg == 0) { /* txg == 0 indicates dry run */
1777 mutex_exit(&msp->ms_lock);
1781 space_map_claim(msp->ms_map, offset, size);
1783 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
1784 if (msp->ms_allocmap[txg & TXG_MASK]->sm_space == 0)
1785 vdev_dirty(vd, VDD_METASLAB, msp, txg);
1786 space_map_add(msp->ms_allocmap[txg & TXG_MASK], offset, size);
1789 mutex_exit(&msp->ms_lock);
1795 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
1796 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags)
1798 dva_t *dva = bp->blk_dva;
1799 dva_t *hintdva = hintbp->blk_dva;
1802 ASSERT(bp->blk_birth == 0);
1803 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
1805 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1807 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
1808 spa_config_exit(spa, SCL_ALLOC, FTAG);
1812 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1813 ASSERT(BP_GET_NDVAS(bp) == 0);
1814 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1816 for (d = 0; d < ndvas; d++) {
1817 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1820 for (d--; d >= 0; d--) {
1821 metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1822 bzero(&dva[d], sizeof (dva_t));
1824 spa_config_exit(spa, SCL_ALLOC, FTAG);
1829 ASSERT(BP_GET_NDVAS(bp) == ndvas);
1831 spa_config_exit(spa, SCL_ALLOC, FTAG);
1833 BP_SET_BIRTH(bp, txg, txg);
1839 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1841 const dva_t *dva = bp->blk_dva;
1842 int d, ndvas = BP_GET_NDVAS(bp);
1844 ASSERT(!BP_IS_HOLE(bp));
1845 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
1847 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
1849 for (d = 0; d < ndvas; d++)
1850 metaslab_free_dva(spa, &dva[d], txg, now);
1852 spa_config_exit(spa, SCL_FREE, FTAG);
1856 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1858 const dva_t *dva = bp->blk_dva;
1859 int ndvas = BP_GET_NDVAS(bp);
1862 ASSERT(!BP_IS_HOLE(bp));
1866 * First do a dry run to make sure all DVAs are claimable,
1867 * so we don't have to unwind from partial failures below.
1869 if ((error = metaslab_claim(spa, bp, 0)) != 0)
1873 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
1875 for (d = 0; d < ndvas; d++)
1876 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1879 spa_config_exit(spa, SCL_ALLOC, FTAG);
1881 ASSERT(error == 0 || txg == 0);
1886 void metaslab_fastwrite_mark(spa_t *spa, const blkptr_t *bp)
1888 const dva_t *dva = bp->blk_dva;
1889 int ndvas = BP_GET_NDVAS(bp);
1890 uint64_t psize = BP_GET_PSIZE(bp);
1894 ASSERT(!BP_IS_HOLE(bp));
1897 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
1899 for (d = 0; d < ndvas; d++) {
1900 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
1902 atomic_add_64(&vd->vdev_pending_fastwrite, psize);
1905 spa_config_exit(spa, SCL_VDEV, FTAG);
1908 void metaslab_fastwrite_unmark(spa_t *spa, const blkptr_t *bp)
1910 const dva_t *dva = bp->blk_dva;
1911 int ndvas = BP_GET_NDVAS(bp);
1912 uint64_t psize = BP_GET_PSIZE(bp);
1916 ASSERT(!BP_IS_HOLE(bp));
1919 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
1921 for (d = 0; d < ndvas; d++) {
1922 if ((vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d]))) == NULL)
1924 ASSERT3U(vd->vdev_pending_fastwrite, >=, psize);
1925 atomic_sub_64(&vd->vdev_pending_fastwrite, psize);
1928 spa_config_exit(spa, SCL_VDEV, FTAG);
1931 #if defined(_KERNEL) && defined(HAVE_SPL)
1932 module_param(metaslab_debug, int, 0644);
1933 MODULE_PARM_DESC(metaslab_debug, "keep space maps in core to verify frees");
1934 #endif /* _KERNEL && HAVE_SPL */