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>
35 * Allow allocations to switch to gang blocks quickly. We do this to
36 * avoid having to load lots of space_maps in a given txg. There are,
37 * however, some cases where we want to avoid "fast" ganging and instead
38 * we want to do an exhaustive search of all metaslabs on this device.
39 * Currently we don't allow any gang, zil, or dump device related allocations
42 #define CAN_FASTGANG(flags) \
43 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
44 METASLAB_GANG_AVOID)))
46 uint64_t metaslab_aliquot
= 512ULL << 10;
47 uint64_t metaslab_gang_bang
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
50 * This value defines the number of allowed allocation failures per vdev.
51 * If a device reaches this threshold in a given txg then we consider skipping
52 * allocations on that device.
54 int zfs_mg_alloc_failures
;
57 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
59 static int metaslab_debug
= 0;
62 * Minimum size which forces the dynamic allocator to change
63 * it's allocation strategy. Once the space map cannot satisfy
64 * an allocation of this size then it switches to using more
65 * aggressive strategy (i.e search by size rather than offset).
67 uint64_t metaslab_df_alloc_threshold
= SPA_MAXBLOCKSIZE
;
70 * The minimum free space, in percent, which must be available
71 * in a space map to continue allocations in a first-fit fashion.
72 * Once the space_map's free space drops below this level we dynamically
73 * switch to using best-fit allocations.
75 int metaslab_df_free_pct
= 4;
78 * A metaslab is considered "free" if it contains a contiguous
79 * segment which is greater than metaslab_min_alloc_size.
81 uint64_t metaslab_min_alloc_size
= DMU_MAX_ACCESS
;
84 * Max number of space_maps to prefetch.
86 int metaslab_prefetch_limit
= SPA_DVAS_PER_BP
;
89 * Percentage bonus multiplier for metaslabs that are in the bonus area.
91 int metaslab_smo_bonus_pct
= 150;
94 * ==========================================================================
96 * ==========================================================================
99 metaslab_class_create(spa_t
*spa
, space_map_ops_t
*ops
)
101 metaslab_class_t
*mc
;
103 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
113 metaslab_class_destroy(metaslab_class_t
*mc
)
115 ASSERT(mc
->mc_rotor
== NULL
);
116 ASSERT(mc
->mc_alloc
== 0);
117 ASSERT(mc
->mc_deferred
== 0);
118 ASSERT(mc
->mc_space
== 0);
119 ASSERT(mc
->mc_dspace
== 0);
121 kmem_free(mc
, sizeof (metaslab_class_t
));
125 metaslab_class_validate(metaslab_class_t
*mc
)
127 metaslab_group_t
*mg
;
131 * Must hold one of the spa_config locks.
133 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
134 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
136 if ((mg
= mc
->mc_rotor
) == NULL
)
141 ASSERT(vd
->vdev_mg
!= NULL
);
142 ASSERT3P(vd
->vdev_top
, ==, vd
);
143 ASSERT3P(mg
->mg_class
, ==, mc
);
144 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
145 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
151 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
152 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
154 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
155 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
156 atomic_add_64(&mc
->mc_space
, space_delta
);
157 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
161 metaslab_class_get_alloc(metaslab_class_t
*mc
)
163 return (mc
->mc_alloc
);
167 metaslab_class_get_deferred(metaslab_class_t
*mc
)
169 return (mc
->mc_deferred
);
173 metaslab_class_get_space(metaslab_class_t
*mc
)
175 return (mc
->mc_space
);
179 metaslab_class_get_dspace(metaslab_class_t
*mc
)
181 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
185 * ==========================================================================
187 * ==========================================================================
190 metaslab_compare(const void *x1
, const void *x2
)
192 const metaslab_t
*m1
= x1
;
193 const metaslab_t
*m2
= x2
;
195 if (m1
->ms_weight
< m2
->ms_weight
)
197 if (m1
->ms_weight
> m2
->ms_weight
)
201 * If the weights are identical, use the offset to force uniqueness.
203 if (m1
->ms_map
.sm_start
< m2
->ms_map
.sm_start
)
205 if (m1
->ms_map
.sm_start
> m2
->ms_map
.sm_start
)
208 ASSERT3P(m1
, ==, m2
);
214 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
216 metaslab_group_t
*mg
;
218 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
219 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
220 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
221 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
224 mg
->mg_activation_count
= 0;
230 metaslab_group_destroy(metaslab_group_t
*mg
)
232 ASSERT(mg
->mg_prev
== NULL
);
233 ASSERT(mg
->mg_next
== NULL
);
235 * We may have gone below zero with the activation count
236 * either because we never activated in the first place or
237 * because we're done, and possibly removing the vdev.
239 ASSERT(mg
->mg_activation_count
<= 0);
241 avl_destroy(&mg
->mg_metaslab_tree
);
242 mutex_destroy(&mg
->mg_lock
);
243 kmem_free(mg
, sizeof (metaslab_group_t
));
247 metaslab_group_activate(metaslab_group_t
*mg
)
249 metaslab_class_t
*mc
= mg
->mg_class
;
250 metaslab_group_t
*mgprev
, *mgnext
;
252 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
254 ASSERT(mc
->mc_rotor
!= mg
);
255 ASSERT(mg
->mg_prev
== NULL
);
256 ASSERT(mg
->mg_next
== NULL
);
257 ASSERT(mg
->mg_activation_count
<= 0);
259 if (++mg
->mg_activation_count
<= 0)
262 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
264 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
268 mgnext
= mgprev
->mg_next
;
269 mg
->mg_prev
= mgprev
;
270 mg
->mg_next
= mgnext
;
271 mgprev
->mg_next
= mg
;
272 mgnext
->mg_prev
= mg
;
278 metaslab_group_passivate(metaslab_group_t
*mg
)
280 metaslab_class_t
*mc
= mg
->mg_class
;
281 metaslab_group_t
*mgprev
, *mgnext
;
283 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
285 if (--mg
->mg_activation_count
!= 0) {
286 ASSERT(mc
->mc_rotor
!= mg
);
287 ASSERT(mg
->mg_prev
== NULL
);
288 ASSERT(mg
->mg_next
== NULL
);
289 ASSERT(mg
->mg_activation_count
< 0);
293 mgprev
= mg
->mg_prev
;
294 mgnext
= mg
->mg_next
;
299 mc
->mc_rotor
= mgnext
;
300 mgprev
->mg_next
= mgnext
;
301 mgnext
->mg_prev
= mgprev
;
309 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
311 mutex_enter(&mg
->mg_lock
);
312 ASSERT(msp
->ms_group
== NULL
);
315 avl_add(&mg
->mg_metaslab_tree
, msp
);
316 mutex_exit(&mg
->mg_lock
);
320 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
322 mutex_enter(&mg
->mg_lock
);
323 ASSERT(msp
->ms_group
== mg
);
324 avl_remove(&mg
->mg_metaslab_tree
, msp
);
325 msp
->ms_group
= NULL
;
326 mutex_exit(&mg
->mg_lock
);
330 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
333 * Although in principle the weight can be any value, in
334 * practice we do not use values in the range [1, 510].
336 ASSERT(weight
>= SPA_MINBLOCKSIZE
-1 || weight
== 0);
337 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
339 mutex_enter(&mg
->mg_lock
);
340 ASSERT(msp
->ms_group
== mg
);
341 avl_remove(&mg
->mg_metaslab_tree
, msp
);
342 msp
->ms_weight
= weight
;
343 avl_add(&mg
->mg_metaslab_tree
, msp
);
344 mutex_exit(&mg
->mg_lock
);
348 * ==========================================================================
349 * Common allocator routines
350 * ==========================================================================
353 metaslab_segsize_compare(const void *x1
, const void *x2
)
355 const space_seg_t
*s1
= x1
;
356 const space_seg_t
*s2
= x2
;
357 uint64_t ss_size1
= s1
->ss_end
- s1
->ss_start
;
358 uint64_t ss_size2
= s2
->ss_end
- s2
->ss_start
;
360 if (ss_size1
< ss_size2
)
362 if (ss_size1
> ss_size2
)
365 if (s1
->ss_start
< s2
->ss_start
)
367 if (s1
->ss_start
> s2
->ss_start
)
374 * This is a helper function that can be used by the allocator to find
375 * a suitable block to allocate. This will search the specified AVL
376 * tree looking for a block that matches the specified criteria.
379 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
382 space_seg_t
*ss
, ssearch
;
385 ssearch
.ss_start
= *cursor
;
386 ssearch
.ss_end
= *cursor
+ size
;
388 ss
= avl_find(t
, &ssearch
, &where
);
390 ss
= avl_nearest(t
, where
, AVL_AFTER
);
393 uint64_t offset
= P2ROUNDUP(ss
->ss_start
, align
);
395 if (offset
+ size
<= ss
->ss_end
) {
396 *cursor
= offset
+ size
;
399 ss
= AVL_NEXT(t
, ss
);
403 * If we know we've searched the whole map (*cursor == 0), give up.
404 * Otherwise, reset the cursor to the beginning and try again.
410 return (metaslab_block_picker(t
, cursor
, size
, align
));
414 metaslab_pp_load(space_map_t
*sm
)
418 ASSERT(sm
->sm_ppd
== NULL
);
419 sm
->sm_ppd
= kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP
);
421 sm
->sm_pp_root
= kmem_alloc(sizeof (avl_tree_t
), KM_SLEEP
);
422 avl_create(sm
->sm_pp_root
, metaslab_segsize_compare
,
423 sizeof (space_seg_t
), offsetof(struct space_seg
, ss_pp_node
));
425 for (ss
= avl_first(&sm
->sm_root
); ss
; ss
= AVL_NEXT(&sm
->sm_root
, ss
))
426 avl_add(sm
->sm_pp_root
, ss
);
430 metaslab_pp_unload(space_map_t
*sm
)
434 kmem_free(sm
->sm_ppd
, 64 * sizeof (uint64_t));
437 while (avl_destroy_nodes(sm
->sm_pp_root
, &cookie
) != NULL
) {
438 /* tear down the tree */
441 avl_destroy(sm
->sm_pp_root
);
442 kmem_free(sm
->sm_pp_root
, sizeof (avl_tree_t
));
443 sm
->sm_pp_root
= NULL
;
448 metaslab_pp_claim(space_map_t
*sm
, uint64_t start
, uint64_t size
)
450 /* No need to update cursor */
455 metaslab_pp_free(space_map_t
*sm
, uint64_t start
, uint64_t size
)
457 /* No need to update cursor */
461 * Return the maximum contiguous segment within the metaslab.
464 metaslab_pp_maxsize(space_map_t
*sm
)
466 avl_tree_t
*t
= sm
->sm_pp_root
;
469 if (t
== NULL
|| (ss
= avl_last(t
)) == NULL
)
472 return (ss
->ss_end
- ss
->ss_start
);
476 * ==========================================================================
477 * The first-fit block allocator
478 * ==========================================================================
481 metaslab_ff_alloc(space_map_t
*sm
, uint64_t size
)
483 avl_tree_t
*t
= &sm
->sm_root
;
484 uint64_t align
= size
& -size
;
485 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
+ highbit(align
) - 1;
487 return (metaslab_block_picker(t
, cursor
, size
, align
));
492 metaslab_ff_fragmented(space_map_t
*sm
)
497 static space_map_ops_t metaslab_ff_ops
= {
504 metaslab_ff_fragmented
508 * ==========================================================================
509 * Dynamic block allocator -
510 * Uses the first fit allocation scheme until space get low and then
511 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
512 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
513 * ==========================================================================
516 metaslab_df_alloc(space_map_t
*sm
, uint64_t size
)
518 avl_tree_t
*t
= &sm
->sm_root
;
519 uint64_t align
= size
& -size
;
520 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
+ highbit(align
) - 1;
521 uint64_t max_size
= metaslab_pp_maxsize(sm
);
522 int free_pct
= sm
->sm_space
* 100 / sm
->sm_size
;
524 ASSERT(MUTEX_HELD(sm
->sm_lock
));
525 ASSERT3U(avl_numnodes(&sm
->sm_root
), ==, avl_numnodes(sm
->sm_pp_root
));
531 * If we're running low on space switch to using the size
532 * sorted AVL tree (best-fit).
534 if (max_size
< metaslab_df_alloc_threshold
||
535 free_pct
< metaslab_df_free_pct
) {
540 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
544 metaslab_df_fragmented(space_map_t
*sm
)
546 uint64_t max_size
= metaslab_pp_maxsize(sm
);
547 int free_pct
= sm
->sm_space
* 100 / sm
->sm_size
;
549 if (max_size
>= metaslab_df_alloc_threshold
&&
550 free_pct
>= metaslab_df_free_pct
)
556 static space_map_ops_t metaslab_df_ops
= {
563 metaslab_df_fragmented
567 * ==========================================================================
568 * Other experimental allocators
569 * ==========================================================================
572 metaslab_cdf_alloc(space_map_t
*sm
, uint64_t size
)
574 avl_tree_t
*t
= &sm
->sm_root
;
575 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
;
576 uint64_t *extent_end
= (uint64_t *)sm
->sm_ppd
+ 1;
577 uint64_t max_size
= metaslab_pp_maxsize(sm
);
578 uint64_t rsize
= size
;
581 ASSERT(MUTEX_HELD(sm
->sm_lock
));
582 ASSERT3U(avl_numnodes(&sm
->sm_root
), ==, avl_numnodes(sm
->sm_pp_root
));
587 ASSERT3U(*extent_end
, >=, *cursor
);
590 * If we're running low on space switch to using the size
591 * sorted AVL tree (best-fit).
593 if ((*cursor
+ size
) > *extent_end
) {
596 *cursor
= *extent_end
= 0;
598 if (max_size
> 2 * SPA_MAXBLOCKSIZE
)
599 rsize
= MIN(metaslab_min_alloc_size
, max_size
);
600 offset
= metaslab_block_picker(t
, extent_end
, rsize
, 1ULL);
602 *cursor
= offset
+ size
;
604 offset
= metaslab_block_picker(t
, cursor
, rsize
, 1ULL);
606 ASSERT3U(*cursor
, <=, *extent_end
);
611 metaslab_cdf_fragmented(space_map_t
*sm
)
613 uint64_t max_size
= metaslab_pp_maxsize(sm
);
615 if (max_size
> (metaslab_min_alloc_size
* 10))
620 static space_map_ops_t metaslab_cdf_ops
= {
627 metaslab_cdf_fragmented
630 uint64_t metaslab_ndf_clump_shift
= 4;
633 metaslab_ndf_alloc(space_map_t
*sm
, uint64_t size
)
635 avl_tree_t
*t
= &sm
->sm_root
;
637 space_seg_t
*ss
, ssearch
;
638 uint64_t hbit
= highbit(size
);
639 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
+ hbit
- 1;
640 uint64_t max_size
= metaslab_pp_maxsize(sm
);
642 ASSERT(MUTEX_HELD(sm
->sm_lock
));
643 ASSERT3U(avl_numnodes(&sm
->sm_root
), ==, avl_numnodes(sm
->sm_pp_root
));
648 ssearch
.ss_start
= *cursor
;
649 ssearch
.ss_end
= *cursor
+ size
;
651 ss
= avl_find(t
, &ssearch
, &where
);
652 if (ss
== NULL
|| (ss
->ss_start
+ size
> ss
->ss_end
)) {
655 ssearch
.ss_start
= 0;
656 ssearch
.ss_end
= MIN(max_size
,
657 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
658 ss
= avl_find(t
, &ssearch
, &where
);
660 ss
= avl_nearest(t
, where
, AVL_AFTER
);
665 if (ss
->ss_start
+ size
<= ss
->ss_end
) {
666 *cursor
= ss
->ss_start
+ size
;
667 return (ss
->ss_start
);
674 metaslab_ndf_fragmented(space_map_t
*sm
)
676 uint64_t max_size
= metaslab_pp_maxsize(sm
);
678 if (max_size
> (metaslab_min_alloc_size
<< metaslab_ndf_clump_shift
))
684 static space_map_ops_t metaslab_ndf_ops
= {
691 metaslab_ndf_fragmented
694 space_map_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
697 * ==========================================================================
699 * ==========================================================================
702 metaslab_init(metaslab_group_t
*mg
, space_map_obj_t
*smo
,
703 uint64_t start
, uint64_t size
, uint64_t txg
)
705 vdev_t
*vd
= mg
->mg_vd
;
708 msp
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
709 mutex_init(&msp
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
711 msp
->ms_smo_syncing
= *smo
;
714 * We create the main space map here, but we don't create the
715 * allocmaps and freemaps until metaslab_sync_done(). This serves
716 * two purposes: it allows metaslab_sync_done() to detect the
717 * addition of new space; and for debugging, it ensures that we'd
718 * data fault on any attempt to use this metaslab before it's ready.
720 space_map_create(&msp
->ms_map
, start
, size
,
721 vd
->vdev_ashift
, &msp
->ms_lock
);
723 metaslab_group_add(mg
, msp
);
725 if (metaslab_debug
&& smo
->smo_object
!= 0) {
726 mutex_enter(&msp
->ms_lock
);
727 VERIFY(space_map_load(&msp
->ms_map
, mg
->mg_class
->mc_ops
,
728 SM_FREE
, smo
, spa_meta_objset(vd
->vdev_spa
)) == 0);
729 mutex_exit(&msp
->ms_lock
);
733 * If we're opening an existing pool (txg == 0) or creating
734 * a new one (txg == TXG_INITIAL), all space is available now.
735 * If we're adding space to an existing pool, the new space
736 * does not become available until after this txg has synced.
738 if (txg
<= TXG_INITIAL
)
739 metaslab_sync_done(msp
, 0);
742 vdev_dirty(vd
, 0, NULL
, txg
);
743 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
750 metaslab_fini(metaslab_t
*msp
)
752 metaslab_group_t
*mg
= msp
->ms_group
;
754 vdev_space_update(mg
->mg_vd
,
755 -msp
->ms_smo
.smo_alloc
, 0, -msp
->ms_map
.sm_size
);
757 metaslab_group_remove(mg
, msp
);
759 mutex_enter(&msp
->ms_lock
);
761 space_map_unload(&msp
->ms_map
);
762 space_map_destroy(&msp
->ms_map
);
764 for (int t
= 0; t
< TXG_SIZE
; t
++) {
765 space_map_destroy(&msp
->ms_allocmap
[t
]);
766 space_map_destroy(&msp
->ms_freemap
[t
]);
769 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++)
770 space_map_destroy(&msp
->ms_defermap
[t
]);
772 ASSERT0(msp
->ms_deferspace
);
774 mutex_exit(&msp
->ms_lock
);
775 mutex_destroy(&msp
->ms_lock
);
777 kmem_free(msp
, sizeof (metaslab_t
));
780 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
781 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
782 #define METASLAB_ACTIVE_MASK \
783 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
786 metaslab_weight(metaslab_t
*msp
)
788 metaslab_group_t
*mg
= msp
->ms_group
;
789 space_map_t
*sm
= &msp
->ms_map
;
790 space_map_obj_t
*smo
= &msp
->ms_smo
;
791 vdev_t
*vd
= mg
->mg_vd
;
792 uint64_t weight
, space
;
794 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
797 * The baseline weight is the metaslab's free space.
799 space
= sm
->sm_size
- smo
->smo_alloc
;
803 * Modern disks have uniform bit density and constant angular velocity.
804 * Therefore, the outer recording zones are faster (higher bandwidth)
805 * than the inner zones by the ratio of outer to inner track diameter,
806 * which is typically around 2:1. We account for this by assigning
807 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
808 * In effect, this means that we'll select the metaslab with the most
809 * free bandwidth rather than simply the one with the most free space.
811 weight
= 2 * weight
-
812 ((sm
->sm_start
>> vd
->vdev_ms_shift
) * weight
) / vd
->vdev_ms_count
;
813 ASSERT(weight
>= space
&& weight
<= 2 * space
);
816 * For locality, assign higher weight to metaslabs which have
817 * a lower offset than what we've already activated.
819 if (sm
->sm_start
<= mg
->mg_bonus_area
)
820 weight
*= (metaslab_smo_bonus_pct
/ 100);
821 ASSERT(weight
>= space
&&
822 weight
<= 2 * (metaslab_smo_bonus_pct
/ 100) * space
);
824 if (sm
->sm_loaded
&& !sm
->sm_ops
->smop_fragmented(sm
)) {
826 * If this metaslab is one we're actively using, adjust its
827 * weight to make it preferable to any inactive metaslab so
828 * we'll polish it off.
830 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
836 metaslab_prefetch(metaslab_group_t
*mg
)
838 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
840 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
843 mutex_enter(&mg
->mg_lock
);
846 * Prefetch the next potential metaslabs
848 for (msp
= avl_first(t
), m
= 0; msp
; msp
= AVL_NEXT(t
, msp
), m
++) {
849 space_map_t
*sm
= &msp
->ms_map
;
850 space_map_obj_t
*smo
= &msp
->ms_smo
;
852 /* If we have reached our prefetch limit then we're done */
853 if (m
>= metaslab_prefetch_limit
)
856 if (!sm
->sm_loaded
&& smo
->smo_object
!= 0) {
857 mutex_exit(&mg
->mg_lock
);
858 dmu_prefetch(spa_meta_objset(spa
), smo
->smo_object
,
859 0ULL, smo
->smo_objsize
);
860 mutex_enter(&mg
->mg_lock
);
863 mutex_exit(&mg
->mg_lock
);
867 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
869 metaslab_group_t
*mg
= msp
->ms_group
;
870 space_map_t
*sm
= &msp
->ms_map
;
871 space_map_ops_t
*sm_ops
= msp
->ms_group
->mg_class
->mc_ops
;
873 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
875 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
876 space_map_load_wait(sm
);
877 if (!sm
->sm_loaded
) {
878 space_map_obj_t
*smo
= &msp
->ms_smo
;
880 int error
= space_map_load(sm
, sm_ops
, SM_FREE
, smo
,
881 spa_meta_objset(msp
->ms_group
->mg_vd
->vdev_spa
));
883 metaslab_group_sort(msp
->ms_group
, msp
, 0);
886 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++)
887 space_map_walk(&msp
->ms_defermap
[t
],
888 space_map_claim
, sm
);
893 * Track the bonus area as we activate new metaslabs.
895 if (sm
->sm_start
> mg
->mg_bonus_area
) {
896 mutex_enter(&mg
->mg_lock
);
897 mg
->mg_bonus_area
= sm
->sm_start
;
898 mutex_exit(&mg
->mg_lock
);
901 metaslab_group_sort(msp
->ms_group
, msp
,
902 msp
->ms_weight
| activation_weight
);
904 ASSERT(sm
->sm_loaded
);
905 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
911 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
914 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
915 * this metaslab again. In that case, it had better be empty,
916 * or we would be leaving space on the table.
918 ASSERT(size
>= SPA_MINBLOCKSIZE
|| msp
->ms_map
.sm_space
== 0);
919 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
920 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
924 * Write a metaslab to disk in the context of the specified transaction group.
927 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
929 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
930 spa_t
*spa
= vd
->vdev_spa
;
931 objset_t
*mos
= spa_meta_objset(spa
);
932 space_map_t
*allocmap
= &msp
->ms_allocmap
[txg
& TXG_MASK
];
933 space_map_t
*freemap
= &msp
->ms_freemap
[txg
& TXG_MASK
];
934 space_map_t
*freed_map
= &msp
->ms_freemap
[TXG_CLEAN(txg
) & TXG_MASK
];
935 space_map_t
*sm
= &msp
->ms_map
;
936 space_map_obj_t
*smo
= &msp
->ms_smo_syncing
;
940 ASSERT(!vd
->vdev_ishole
);
942 if (allocmap
->sm_space
== 0 && freemap
->sm_space
== 0)
946 * The only state that can actually be changing concurrently with
947 * metaslab_sync() is the metaslab's ms_map. No other thread can
948 * be modifying this txg's allocmap, freemap, freed_map, or smo.
949 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
950 * We drop it whenever we call into the DMU, because the DMU
951 * can call down to us (e.g. via zio_free()) at any time.
954 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
956 if (smo
->smo_object
== 0) {
957 ASSERT(smo
->smo_objsize
== 0);
958 ASSERT(smo
->smo_alloc
== 0);
959 smo
->smo_object
= dmu_object_alloc(mos
,
960 DMU_OT_SPACE_MAP
, 1 << SPACE_MAP_BLOCKSHIFT
,
961 DMU_OT_SPACE_MAP_HEADER
, sizeof (*smo
), tx
);
962 ASSERT(smo
->smo_object
!= 0);
963 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
964 (sm
->sm_start
>> vd
->vdev_ms_shift
),
965 sizeof (uint64_t), &smo
->smo_object
, tx
);
968 mutex_enter(&msp
->ms_lock
);
970 space_map_walk(freemap
, space_map_add
, freed_map
);
972 if (sm
->sm_loaded
&& spa_sync_pass(spa
) == 1 && smo
->smo_objsize
>=
973 2 * sizeof (uint64_t) * avl_numnodes(&sm
->sm_root
)) {
975 * The in-core space map representation is twice as compact
976 * as the on-disk one, so it's time to condense the latter
977 * by generating a pure allocmap from first principles.
979 * This metaslab is 100% allocated,
980 * minus the content of the in-core map (sm),
981 * minus what's been freed this txg (freed_map),
982 * minus deferred frees (ms_defermap[]),
983 * minus allocations from txgs in the future
984 * (because they haven't been committed yet).
986 space_map_vacate(allocmap
, NULL
, NULL
);
987 space_map_vacate(freemap
, NULL
, NULL
);
989 space_map_add(allocmap
, allocmap
->sm_start
, allocmap
->sm_size
);
991 space_map_walk(sm
, space_map_remove
, allocmap
);
992 space_map_walk(freed_map
, space_map_remove
, allocmap
);
994 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++)
995 space_map_walk(&msp
->ms_defermap
[t
],
996 space_map_remove
, allocmap
);
998 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++)
999 space_map_walk(&msp
->ms_allocmap
[(txg
+ t
) & TXG_MASK
],
1000 space_map_remove
, allocmap
);
1002 mutex_exit(&msp
->ms_lock
);
1003 space_map_truncate(smo
, mos
, tx
);
1004 mutex_enter(&msp
->ms_lock
);
1007 space_map_sync(allocmap
, SM_ALLOC
, smo
, mos
, tx
);
1008 space_map_sync(freemap
, SM_FREE
, smo
, mos
, tx
);
1010 mutex_exit(&msp
->ms_lock
);
1012 VERIFY(0 == dmu_bonus_hold(mos
, smo
->smo_object
, FTAG
, &db
));
1013 dmu_buf_will_dirty(db
, tx
);
1014 ASSERT3U(db
->db_size
, >=, sizeof (*smo
));
1015 bcopy(smo
, db
->db_data
, sizeof (*smo
));
1016 dmu_buf_rele(db
, FTAG
);
1022 * Called after a transaction group has completely synced to mark
1023 * all of the metaslab's free space as usable.
1026 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1028 space_map_obj_t
*smo
= &msp
->ms_smo
;
1029 space_map_obj_t
*smosync
= &msp
->ms_smo_syncing
;
1030 space_map_t
*sm
= &msp
->ms_map
;
1031 space_map_t
*freed_map
= &msp
->ms_freemap
[TXG_CLEAN(txg
) & TXG_MASK
];
1032 space_map_t
*defer_map
= &msp
->ms_defermap
[txg
% TXG_DEFER_SIZE
];
1033 metaslab_group_t
*mg
= msp
->ms_group
;
1034 vdev_t
*vd
= mg
->mg_vd
;
1035 int64_t alloc_delta
, defer_delta
;
1037 ASSERT(!vd
->vdev_ishole
);
1039 mutex_enter(&msp
->ms_lock
);
1042 * If this metaslab is just becoming available, initialize its
1043 * allocmaps and freemaps and add its capacity to the vdev.
1045 if (freed_map
->sm_size
== 0) {
1046 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1047 space_map_create(&msp
->ms_allocmap
[t
], sm
->sm_start
,
1048 sm
->sm_size
, sm
->sm_shift
, sm
->sm_lock
);
1049 space_map_create(&msp
->ms_freemap
[t
], sm
->sm_start
,
1050 sm
->sm_size
, sm
->sm_shift
, sm
->sm_lock
);
1053 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++)
1054 space_map_create(&msp
->ms_defermap
[t
], sm
->sm_start
,
1055 sm
->sm_size
, sm
->sm_shift
, sm
->sm_lock
);
1057 vdev_space_update(vd
, 0, 0, sm
->sm_size
);
1060 alloc_delta
= smosync
->smo_alloc
- smo
->smo_alloc
;
1061 defer_delta
= freed_map
->sm_space
- defer_map
->sm_space
;
1063 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1065 ASSERT(msp
->ms_allocmap
[txg
& TXG_MASK
].sm_space
== 0);
1066 ASSERT(msp
->ms_freemap
[txg
& TXG_MASK
].sm_space
== 0);
1069 * If there's a space_map_load() in progress, wait for it to complete
1070 * so that we have a consistent view of the in-core space map.
1071 * Then, add defer_map (oldest deferred frees) to this map and
1072 * transfer freed_map (this txg's frees) to defer_map.
1074 space_map_load_wait(sm
);
1075 space_map_vacate(defer_map
, sm
->sm_loaded
? space_map_free
: NULL
, sm
);
1076 space_map_vacate(freed_map
, space_map_add
, defer_map
);
1080 msp
->ms_deferspace
+= defer_delta
;
1081 ASSERT3S(msp
->ms_deferspace
, >=, 0);
1082 ASSERT3S(msp
->ms_deferspace
, <=, sm
->sm_size
);
1083 if (msp
->ms_deferspace
!= 0) {
1085 * Keep syncing this metaslab until all deferred frees
1086 * are back in circulation.
1088 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1092 * If the map is loaded but no longer active, evict it as soon as all
1093 * future allocations have synced. (If we unloaded it now and then
1094 * loaded a moment later, the map wouldn't reflect those allocations.)
1096 if (sm
->sm_loaded
&& (msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1099 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++)
1100 if (msp
->ms_allocmap
[(txg
+ t
) & TXG_MASK
].sm_space
)
1103 if (evictable
&& !metaslab_debug
)
1104 space_map_unload(sm
);
1107 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
1109 mutex_exit(&msp
->ms_lock
);
1113 metaslab_sync_reassess(metaslab_group_t
*mg
)
1115 vdev_t
*vd
= mg
->mg_vd
;
1116 int64_t failures
= mg
->mg_alloc_failures
;
1119 * Re-evaluate all metaslabs which have lower offsets than the
1122 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
1123 metaslab_t
*msp
= vd
->vdev_ms
[m
];
1125 if (msp
->ms_map
.sm_start
> mg
->mg_bonus_area
)
1128 mutex_enter(&msp
->ms_lock
);
1129 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
1130 mutex_exit(&msp
->ms_lock
);
1133 atomic_add_64(&mg
->mg_alloc_failures
, -failures
);
1136 * Prefetch the next potential metaslabs
1138 metaslab_prefetch(mg
);
1142 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
1144 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
1145 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
1146 uint64_t start
= msp
->ms_map
.sm_start
>> ms_shift
;
1148 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
1149 return (1ULL << 63);
1152 return ((start
- offset
) << ms_shift
);
1154 return ((offset
- start
) << ms_shift
);
1159 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
1160 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
, int flags
)
1162 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1163 metaslab_t
*msp
= NULL
;
1164 uint64_t offset
= -1ULL;
1165 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1166 uint64_t activation_weight
;
1167 uint64_t target_distance
;
1170 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
1171 for (i
= 0; i
< d
; i
++) {
1172 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
1173 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
1179 boolean_t was_active
;
1181 mutex_enter(&mg
->mg_lock
);
1182 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
1183 if (msp
->ms_weight
< asize
) {
1184 spa_dbgmsg(spa
, "%s: failed to meet weight "
1185 "requirement: vdev %llu, txg %llu, mg %p, "
1186 "msp %p, psize %llu, asize %llu, "
1187 "failures %llu, weight %llu",
1188 spa_name(spa
), mg
->mg_vd
->vdev_id
, txg
,
1189 mg
, msp
, psize
, asize
,
1190 mg
->mg_alloc_failures
, msp
->ms_weight
);
1191 mutex_exit(&mg
->mg_lock
);
1194 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
1195 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
1198 target_distance
= min_distance
+
1199 (msp
->ms_smo
.smo_alloc
? 0 : min_distance
>> 1);
1201 for (i
= 0; i
< d
; i
++)
1202 if (metaslab_distance(msp
, &dva
[i
]) <
1208 mutex_exit(&mg
->mg_lock
);
1213 * If we've already reached the allowable number of failed
1214 * allocation attempts on this metaslab group then we
1215 * consider skipping it. We skip it only if we're allowed
1216 * to "fast" gang, the physical size is larger than
1217 * a gang block, and we're attempting to allocate from
1218 * the primary metaslab.
1220 if (mg
->mg_alloc_failures
> zfs_mg_alloc_failures
&&
1221 CAN_FASTGANG(flags
) && psize
> SPA_GANGBLOCKSIZE
&&
1222 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
1223 spa_dbgmsg(spa
, "%s: skipping metaslab group: "
1224 "vdev %llu, txg %llu, mg %p, psize %llu, "
1225 "asize %llu, failures %llu", spa_name(spa
),
1226 mg
->mg_vd
->vdev_id
, txg
, mg
, psize
, asize
,
1227 mg
->mg_alloc_failures
);
1231 mutex_enter(&msp
->ms_lock
);
1234 * Ensure that the metaslab we have selected is still
1235 * capable of handling our request. It's possible that
1236 * another thread may have changed the weight while we
1237 * were blocked on the metaslab lock.
1239 if (msp
->ms_weight
< asize
|| (was_active
&&
1240 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
1241 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
1242 mutex_exit(&msp
->ms_lock
);
1246 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
1247 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
1248 metaslab_passivate(msp
,
1249 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
1250 mutex_exit(&msp
->ms_lock
);
1254 if (metaslab_activate(msp
, activation_weight
) != 0) {
1255 mutex_exit(&msp
->ms_lock
);
1259 if ((offset
= space_map_alloc(&msp
->ms_map
, asize
)) != -1ULL)
1262 atomic_inc_64(&mg
->mg_alloc_failures
);
1264 metaslab_passivate(msp
, space_map_maxsize(&msp
->ms_map
));
1266 mutex_exit(&msp
->ms_lock
);
1269 if (msp
->ms_allocmap
[txg
& TXG_MASK
].sm_space
== 0)
1270 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
1272 space_map_add(&msp
->ms_allocmap
[txg
& TXG_MASK
], offset
, asize
);
1274 mutex_exit(&msp
->ms_lock
);
1280 * Allocate a block for the specified i/o.
1283 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
1284 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
1286 metaslab_group_t
*mg
, *rotor
;
1290 int zio_lock
= B_FALSE
;
1291 boolean_t allocatable
;
1292 uint64_t offset
= -1ULL;
1296 ASSERT(!DVA_IS_VALID(&dva
[d
]));
1299 * For testing, make some blocks above a certain size be gang blocks.
1301 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
1305 * Start at the rotor and loop through all mgs until we find something.
1306 * Note that there's no locking on mc_rotor or mc_aliquot because
1307 * nothing actually breaks if we miss a few updates -- we just won't
1308 * allocate quite as evenly. It all balances out over time.
1310 * If we are doing ditto or log blocks, try to spread them across
1311 * consecutive vdevs. If we're forced to reuse a vdev before we've
1312 * allocated all of our ditto blocks, then try and spread them out on
1313 * that vdev as much as possible. If it turns out to not be possible,
1314 * gradually lower our standards until anything becomes acceptable.
1315 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1316 * gives us hope of containing our fault domains to something we're
1317 * able to reason about. Otherwise, any two top-level vdev failures
1318 * will guarantee the loss of data. With consecutive allocation,
1319 * only two adjacent top-level vdev failures will result in data loss.
1321 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1322 * ourselves on the same vdev as our gang block header. That
1323 * way, we can hope for locality in vdev_cache, plus it makes our
1324 * fault domains something tractable.
1327 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
1330 * It's possible the vdev we're using as the hint no
1331 * longer exists (i.e. removed). Consult the rotor when
1337 if (flags
& METASLAB_HINTBP_AVOID
&&
1338 mg
->mg_next
!= NULL
)
1343 } else if (d
!= 0) {
1344 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
1345 mg
= vd
->vdev_mg
->mg_next
;
1351 * If the hint put us into the wrong metaslab class, or into a
1352 * metaslab group that has been passivated, just follow the rotor.
1354 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
1361 ASSERT(mg
->mg_activation_count
== 1);
1366 * Don't allocate from faulted devices.
1369 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
1370 allocatable
= vdev_allocatable(vd
);
1371 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
1373 allocatable
= vdev_allocatable(vd
);
1379 * Avoid writing single-copy data to a failing vdev
1381 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
1382 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
1383 d
== 0 && dshift
== 3) {
1388 ASSERT(mg
->mg_class
== mc
);
1390 distance
= vd
->vdev_asize
>> dshift
;
1391 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
1396 asize
= vdev_psize_to_asize(vd
, psize
);
1397 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
1399 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
1401 if (offset
!= -1ULL) {
1403 * If we've just selected this metaslab group,
1404 * figure out whether the corresponding vdev is
1405 * over- or under-used relative to the pool,
1406 * and set an allocation bias to even it out.
1408 if (mc
->mc_aliquot
== 0) {
1409 vdev_stat_t
*vs
= &vd
->vdev_stat
;
1412 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
1413 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
1416 * Calculate how much more or less we should
1417 * try to allocate from this device during
1418 * this iteration around the rotor.
1419 * For example, if a device is 80% full
1420 * and the pool is 20% full then we should
1421 * reduce allocations by 60% on this device.
1423 * mg_bias = (20 - 80) * 512K / 100 = -307K
1425 * This reduces allocations by 307K for this
1428 mg
->mg_bias
= ((cu
- vu
) *
1429 (int64_t)mg
->mg_aliquot
) / 100;
1432 if (atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
1433 mg
->mg_aliquot
+ mg
->mg_bias
) {
1434 mc
->mc_rotor
= mg
->mg_next
;
1438 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
1439 DVA_SET_OFFSET(&dva
[d
], offset
);
1440 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
1441 DVA_SET_ASIZE(&dva
[d
], asize
);
1446 mc
->mc_rotor
= mg
->mg_next
;
1448 } while ((mg
= mg
->mg_next
) != rotor
);
1452 ASSERT(dshift
< 64);
1456 if (!allocatable
&& !zio_lock
) {
1462 bzero(&dva
[d
], sizeof (dva_t
));
1468 * Free the block represented by DVA in the context of the specified
1469 * transaction group.
1472 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
1474 uint64_t vdev
= DVA_GET_VDEV(dva
);
1475 uint64_t offset
= DVA_GET_OFFSET(dva
);
1476 uint64_t size
= DVA_GET_ASIZE(dva
);
1480 ASSERT(DVA_IS_VALID(dva
));
1482 if (txg
> spa_freeze_txg(spa
))
1485 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
1486 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
1487 cmn_err(CE_WARN
, "metaslab_free_dva(): bad DVA %llu:%llu",
1488 (u_longlong_t
)vdev
, (u_longlong_t
)offset
);
1493 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
1495 if (DVA_GET_GANG(dva
))
1496 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1498 mutex_enter(&msp
->ms_lock
);
1501 space_map_remove(&msp
->ms_allocmap
[txg
& TXG_MASK
],
1503 space_map_free(&msp
->ms_map
, offset
, size
);
1505 if (msp
->ms_freemap
[txg
& TXG_MASK
].sm_space
== 0)
1506 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
1507 space_map_add(&msp
->ms_freemap
[txg
& TXG_MASK
], offset
, size
);
1510 mutex_exit(&msp
->ms_lock
);
1514 * Intent log support: upon opening the pool after a crash, notify the SPA
1515 * of blocks that the intent log has allocated for immediate write, but
1516 * which are still considered free by the SPA because the last transaction
1517 * group didn't commit yet.
1520 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
1522 uint64_t vdev
= DVA_GET_VDEV(dva
);
1523 uint64_t offset
= DVA_GET_OFFSET(dva
);
1524 uint64_t size
= DVA_GET_ASIZE(dva
);
1529 ASSERT(DVA_IS_VALID(dva
));
1531 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
1532 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
1535 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
1537 if (DVA_GET_GANG(dva
))
1538 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1540 mutex_enter(&msp
->ms_lock
);
1542 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_map
.sm_loaded
)
1543 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
1545 if (error
== 0 && !space_map_contains(&msp
->ms_map
, offset
, size
))
1548 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
1549 mutex_exit(&msp
->ms_lock
);
1553 space_map_claim(&msp
->ms_map
, offset
, size
);
1555 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
1556 if (msp
->ms_allocmap
[txg
& TXG_MASK
].sm_space
== 0)
1557 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
1558 space_map_add(&msp
->ms_allocmap
[txg
& TXG_MASK
], offset
, size
);
1561 mutex_exit(&msp
->ms_lock
);
1567 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
1568 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
1570 dva_t
*dva
= bp
->blk_dva
;
1571 dva_t
*hintdva
= hintbp
->blk_dva
;
1574 ASSERT(bp
->blk_birth
== 0);
1575 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
1577 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
1579 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
1580 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1584 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
1585 ASSERT(BP_GET_NDVAS(bp
) == 0);
1586 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
1588 for (int d
= 0; d
< ndvas
; d
++) {
1589 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
1592 for (d
--; d
>= 0; d
--) {
1593 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
1594 bzero(&dva
[d
], sizeof (dva_t
));
1596 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1601 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
1603 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1605 BP_SET_BIRTH(bp
, txg
, txg
);
1611 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
1613 const dva_t
*dva
= bp
->blk_dva
;
1614 int ndvas
= BP_GET_NDVAS(bp
);
1616 ASSERT(!BP_IS_HOLE(bp
));
1617 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
1619 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
1621 for (int d
= 0; d
< ndvas
; d
++)
1622 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
1624 spa_config_exit(spa
, SCL_FREE
, FTAG
);
1628 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
1630 const dva_t
*dva
= bp
->blk_dva
;
1631 int ndvas
= BP_GET_NDVAS(bp
);
1634 ASSERT(!BP_IS_HOLE(bp
));
1638 * First do a dry run to make sure all DVAs are claimable,
1639 * so we don't have to unwind from partial failures below.
1641 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
1645 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
1647 for (int d
= 0; d
< ndvas
; d
++)
1648 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
1651 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1653 ASSERT(error
== 0 || txg
== 0);