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 * This value defines the number of allowed allocation failures per vdev.
53 * If a device reaches this threshold in a given txg then we consider skipping
54 * allocations on that device.
56 int zfs_mg_alloc_failures
;
59 * Metaslab debugging: when set, keeps all space maps in core to verify frees.
61 static int metaslab_debug
= 0;
64 * Minimum size which forces the dynamic allocator to change
65 * it's allocation strategy. Once the space map cannot satisfy
66 * an allocation of this size then it switches to using more
67 * aggressive strategy (i.e search by size rather than offset).
69 uint64_t metaslab_df_alloc_threshold
= SPA_MAXBLOCKSIZE
;
72 * The minimum free space, in percent, which must be available
73 * in a space map to continue allocations in a first-fit fashion.
74 * Once the space_map's free space drops below this level we dynamically
75 * switch to using best-fit allocations.
77 int metaslab_df_free_pct
= 4;
80 * A metaslab is considered "free" if it contains a contiguous
81 * segment which is greater than metaslab_min_alloc_size.
83 uint64_t metaslab_min_alloc_size
= DMU_MAX_ACCESS
;
86 * Max number of space_maps to prefetch.
88 int metaslab_prefetch_limit
= SPA_DVAS_PER_BP
;
91 * Percentage bonus multiplier for metaslabs that are in the bonus area.
93 int metaslab_smo_bonus_pct
= 150;
96 * ==========================================================================
98 * ==========================================================================
101 metaslab_class_create(spa_t
*spa
, space_map_ops_t
*ops
)
103 metaslab_class_t
*mc
;
105 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_PUSHPAGE
);
115 metaslab_class_destroy(metaslab_class_t
*mc
)
117 ASSERT(mc
->mc_rotor
== NULL
);
118 ASSERT(mc
->mc_alloc
== 0);
119 ASSERT(mc
->mc_deferred
== 0);
120 ASSERT(mc
->mc_space
== 0);
121 ASSERT(mc
->mc_dspace
== 0);
123 kmem_free(mc
, sizeof (metaslab_class_t
));
127 metaslab_class_validate(metaslab_class_t
*mc
)
129 metaslab_group_t
*mg
;
133 * Must hold one of the spa_config locks.
135 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
136 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
138 if ((mg
= mc
->mc_rotor
) == NULL
)
143 ASSERT(vd
->vdev_mg
!= NULL
);
144 ASSERT3P(vd
->vdev_top
, ==, vd
);
145 ASSERT3P(mg
->mg_class
, ==, mc
);
146 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
147 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
153 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
154 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
156 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
157 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
158 atomic_add_64(&mc
->mc_space
, space_delta
);
159 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
163 metaslab_class_get_alloc(metaslab_class_t
*mc
)
165 return (mc
->mc_alloc
);
169 metaslab_class_get_deferred(metaslab_class_t
*mc
)
171 return (mc
->mc_deferred
);
175 metaslab_class_get_space(metaslab_class_t
*mc
)
177 return (mc
->mc_space
);
181 metaslab_class_get_dspace(metaslab_class_t
*mc
)
183 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
187 * ==========================================================================
189 * ==========================================================================
192 metaslab_compare(const void *x1
, const void *x2
)
194 const metaslab_t
*m1
= x1
;
195 const metaslab_t
*m2
= x2
;
197 if (m1
->ms_weight
< m2
->ms_weight
)
199 if (m1
->ms_weight
> m2
->ms_weight
)
203 * If the weights are identical, use the offset to force uniqueness.
205 if (m1
->ms_map
.sm_start
< m2
->ms_map
.sm_start
)
207 if (m1
->ms_map
.sm_start
> m2
->ms_map
.sm_start
)
210 ASSERT3P(m1
, ==, m2
);
216 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
218 metaslab_group_t
*mg
;
220 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_PUSHPAGE
);
221 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
222 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
223 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
226 mg
->mg_activation_count
= 0;
232 metaslab_group_destroy(metaslab_group_t
*mg
)
234 ASSERT(mg
->mg_prev
== NULL
);
235 ASSERT(mg
->mg_next
== NULL
);
237 * We may have gone below zero with the activation count
238 * either because we never activated in the first place or
239 * because we're done, and possibly removing the vdev.
241 ASSERT(mg
->mg_activation_count
<= 0);
243 avl_destroy(&mg
->mg_metaslab_tree
);
244 mutex_destroy(&mg
->mg_lock
);
245 kmem_free(mg
, sizeof (metaslab_group_t
));
249 metaslab_group_activate(metaslab_group_t
*mg
)
251 metaslab_class_t
*mc
= mg
->mg_class
;
252 metaslab_group_t
*mgprev
, *mgnext
;
254 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
256 ASSERT(mc
->mc_rotor
!= mg
);
257 ASSERT(mg
->mg_prev
== NULL
);
258 ASSERT(mg
->mg_next
== NULL
);
259 ASSERT(mg
->mg_activation_count
<= 0);
261 if (++mg
->mg_activation_count
<= 0)
264 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
266 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
270 mgnext
= mgprev
->mg_next
;
271 mg
->mg_prev
= mgprev
;
272 mg
->mg_next
= mgnext
;
273 mgprev
->mg_next
= mg
;
274 mgnext
->mg_prev
= mg
;
280 metaslab_group_passivate(metaslab_group_t
*mg
)
282 metaslab_class_t
*mc
= mg
->mg_class
;
283 metaslab_group_t
*mgprev
, *mgnext
;
285 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
287 if (--mg
->mg_activation_count
!= 0) {
288 ASSERT(mc
->mc_rotor
!= mg
);
289 ASSERT(mg
->mg_prev
== NULL
);
290 ASSERT(mg
->mg_next
== NULL
);
291 ASSERT(mg
->mg_activation_count
< 0);
295 mgprev
= mg
->mg_prev
;
296 mgnext
= mg
->mg_next
;
301 mc
->mc_rotor
= mgnext
;
302 mgprev
->mg_next
= mgnext
;
303 mgnext
->mg_prev
= mgprev
;
311 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
313 mutex_enter(&mg
->mg_lock
);
314 ASSERT(msp
->ms_group
== NULL
);
317 avl_add(&mg
->mg_metaslab_tree
, msp
);
318 mutex_exit(&mg
->mg_lock
);
322 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
324 mutex_enter(&mg
->mg_lock
);
325 ASSERT(msp
->ms_group
== mg
);
326 avl_remove(&mg
->mg_metaslab_tree
, msp
);
327 msp
->ms_group
= NULL
;
328 mutex_exit(&mg
->mg_lock
);
332 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
335 * Although in principle the weight can be any value, in
336 * practice we do not use values in the range [1, 510].
338 ASSERT(weight
>= SPA_MINBLOCKSIZE
-1 || weight
== 0);
339 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
341 mutex_enter(&mg
->mg_lock
);
342 ASSERT(msp
->ms_group
== mg
);
343 avl_remove(&mg
->mg_metaslab_tree
, msp
);
344 msp
->ms_weight
= weight
;
345 avl_add(&mg
->mg_metaslab_tree
, msp
);
346 mutex_exit(&mg
->mg_lock
);
350 * ==========================================================================
351 * Common allocator routines
352 * ==========================================================================
355 metaslab_segsize_compare(const void *x1
, const void *x2
)
357 const space_seg_t
*s1
= x1
;
358 const space_seg_t
*s2
= x2
;
359 uint64_t ss_size1
= s1
->ss_end
- s1
->ss_start
;
360 uint64_t ss_size2
= s2
->ss_end
- s2
->ss_start
;
362 if (ss_size1
< ss_size2
)
364 if (ss_size1
> ss_size2
)
367 if (s1
->ss_start
< s2
->ss_start
)
369 if (s1
->ss_start
> s2
->ss_start
)
375 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
376 defined(WITH_DF_BLOCK_ALLOCATOR) || \
377 defined(WITH_CDF_BLOCK_ALLOCATOR)
379 * This is a helper function that can be used by the allocator to find
380 * a suitable block to allocate. This will search the specified AVL
381 * tree looking for a block that matches the specified criteria.
384 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
387 space_seg_t
*ss
, ssearch
;
390 ssearch
.ss_start
= *cursor
;
391 ssearch
.ss_end
= *cursor
+ size
;
393 ss
= avl_find(t
, &ssearch
, &where
);
395 ss
= avl_nearest(t
, where
, AVL_AFTER
);
398 uint64_t offset
= P2ROUNDUP(ss
->ss_start
, align
);
400 if (offset
+ size
<= ss
->ss_end
) {
401 *cursor
= offset
+ size
;
404 ss
= AVL_NEXT(t
, ss
);
408 * If we know we've searched the whole map (*cursor == 0), give up.
409 * Otherwise, reset the cursor to the beginning and try again.
415 return (metaslab_block_picker(t
, cursor
, size
, align
));
417 #endif /* WITH_FF/DF/CDF_BLOCK_ALLOCATOR */
420 metaslab_pp_load(space_map_t
*sm
)
424 ASSERT(sm
->sm_ppd
== NULL
);
425 sm
->sm_ppd
= kmem_zalloc(64 * sizeof (uint64_t), KM_PUSHPAGE
);
427 sm
->sm_pp_root
= kmem_alloc(sizeof (avl_tree_t
), KM_PUSHPAGE
);
428 avl_create(sm
->sm_pp_root
, metaslab_segsize_compare
,
429 sizeof (space_seg_t
), offsetof(struct space_seg
, ss_pp_node
));
431 for (ss
= avl_first(&sm
->sm_root
); ss
; ss
= AVL_NEXT(&sm
->sm_root
, ss
))
432 avl_add(sm
->sm_pp_root
, ss
);
436 metaslab_pp_unload(space_map_t
*sm
)
440 kmem_free(sm
->sm_ppd
, 64 * sizeof (uint64_t));
443 while (avl_destroy_nodes(sm
->sm_pp_root
, &cookie
) != NULL
) {
444 /* tear down the tree */
447 avl_destroy(sm
->sm_pp_root
);
448 kmem_free(sm
->sm_pp_root
, sizeof (avl_tree_t
));
449 sm
->sm_pp_root
= NULL
;
454 metaslab_pp_claim(space_map_t
*sm
, uint64_t start
, uint64_t size
)
456 /* No need to update cursor */
461 metaslab_pp_free(space_map_t
*sm
, uint64_t start
, uint64_t size
)
463 /* No need to update cursor */
467 * Return the maximum contiguous segment within the metaslab.
470 metaslab_pp_maxsize(space_map_t
*sm
)
472 avl_tree_t
*t
= sm
->sm_pp_root
;
475 if (t
== NULL
|| (ss
= avl_last(t
)) == NULL
)
478 return (ss
->ss_end
- ss
->ss_start
);
481 #if defined(WITH_FF_BLOCK_ALLOCATOR)
483 * ==========================================================================
484 * The first-fit block allocator
485 * ==========================================================================
488 metaslab_ff_alloc(space_map_t
*sm
, uint64_t size
)
490 avl_tree_t
*t
= &sm
->sm_root
;
491 uint64_t align
= size
& -size
;
492 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
+ highbit(align
) - 1;
494 return (metaslab_block_picker(t
, cursor
, size
, align
));
499 metaslab_ff_fragmented(space_map_t
*sm
)
504 static space_map_ops_t metaslab_ff_ops
= {
511 metaslab_ff_fragmented
514 space_map_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
515 #endif /* WITH_FF_BLOCK_ALLOCATOR */
517 #if defined(WITH_DF_BLOCK_ALLOCATOR)
519 * ==========================================================================
520 * Dynamic block allocator -
521 * Uses the first fit allocation scheme until space get low and then
522 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
523 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
524 * ==========================================================================
527 metaslab_df_alloc(space_map_t
*sm
, uint64_t size
)
529 avl_tree_t
*t
= &sm
->sm_root
;
530 uint64_t align
= size
& -size
;
531 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
+ highbit(align
) - 1;
532 uint64_t max_size
= metaslab_pp_maxsize(sm
);
533 int free_pct
= sm
->sm_space
* 100 / sm
->sm_size
;
535 ASSERT(MUTEX_HELD(sm
->sm_lock
));
536 ASSERT3U(avl_numnodes(&sm
->sm_root
), ==, avl_numnodes(sm
->sm_pp_root
));
542 * If we're running low on space switch to using the size
543 * sorted AVL tree (best-fit).
545 if (max_size
< metaslab_df_alloc_threshold
||
546 free_pct
< metaslab_df_free_pct
) {
551 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
555 metaslab_df_fragmented(space_map_t
*sm
)
557 uint64_t max_size
= metaslab_pp_maxsize(sm
);
558 int free_pct
= sm
->sm_space
* 100 / sm
->sm_size
;
560 if (max_size
>= metaslab_df_alloc_threshold
&&
561 free_pct
>= metaslab_df_free_pct
)
567 static space_map_ops_t metaslab_df_ops
= {
574 metaslab_df_fragmented
577 space_map_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
578 #endif /* WITH_DF_BLOCK_ALLOCATOR */
581 * ==========================================================================
582 * Other experimental allocators
583 * ==========================================================================
585 #if defined(WITH_CDF_BLOCK_ALLOCATOR)
587 metaslab_cdf_alloc(space_map_t
*sm
, uint64_t size
)
589 avl_tree_t
*t
= &sm
->sm_root
;
590 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
;
591 uint64_t *extent_end
= (uint64_t *)sm
->sm_ppd
+ 1;
592 uint64_t max_size
= metaslab_pp_maxsize(sm
);
593 uint64_t rsize
= size
;
596 ASSERT(MUTEX_HELD(sm
->sm_lock
));
597 ASSERT3U(avl_numnodes(&sm
->sm_root
), ==, avl_numnodes(sm
->sm_pp_root
));
602 ASSERT3U(*extent_end
, >=, *cursor
);
605 * If we're running low on space switch to using the size
606 * sorted AVL tree (best-fit).
608 if ((*cursor
+ size
) > *extent_end
) {
611 *cursor
= *extent_end
= 0;
613 if (max_size
> 2 * SPA_MAXBLOCKSIZE
)
614 rsize
= MIN(metaslab_min_alloc_size
, max_size
);
615 offset
= metaslab_block_picker(t
, extent_end
, rsize
, 1ULL);
617 *cursor
= offset
+ size
;
619 offset
= metaslab_block_picker(t
, cursor
, rsize
, 1ULL);
621 ASSERT3U(*cursor
, <=, *extent_end
);
626 metaslab_cdf_fragmented(space_map_t
*sm
)
628 uint64_t max_size
= metaslab_pp_maxsize(sm
);
630 if (max_size
> (metaslab_min_alloc_size
* 10))
635 static space_map_ops_t metaslab_cdf_ops
= {
642 metaslab_cdf_fragmented
645 space_map_ops_t
*zfs_metaslab_ops
= &metaslab_cdf_ops
;
646 #endif /* WITH_CDF_BLOCK_ALLOCATOR */
648 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
649 uint64_t metaslab_ndf_clump_shift
= 4;
652 metaslab_ndf_alloc(space_map_t
*sm
, uint64_t size
)
654 avl_tree_t
*t
= &sm
->sm_root
;
656 space_seg_t
*ss
, ssearch
;
657 uint64_t hbit
= highbit(size
);
658 uint64_t *cursor
= (uint64_t *)sm
->sm_ppd
+ hbit
- 1;
659 uint64_t max_size
= metaslab_pp_maxsize(sm
);
661 ASSERT(MUTEX_HELD(sm
->sm_lock
));
662 ASSERT3U(avl_numnodes(&sm
->sm_root
), ==, avl_numnodes(sm
->sm_pp_root
));
667 ssearch
.ss_start
= *cursor
;
668 ssearch
.ss_end
= *cursor
+ size
;
670 ss
= avl_find(t
, &ssearch
, &where
);
671 if (ss
== NULL
|| (ss
->ss_start
+ size
> ss
->ss_end
)) {
674 ssearch
.ss_start
= 0;
675 ssearch
.ss_end
= MIN(max_size
,
676 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
677 ss
= avl_find(t
, &ssearch
, &where
);
679 ss
= avl_nearest(t
, where
, AVL_AFTER
);
684 if (ss
->ss_start
+ size
<= ss
->ss_end
) {
685 *cursor
= ss
->ss_start
+ size
;
686 return (ss
->ss_start
);
693 metaslab_ndf_fragmented(space_map_t
*sm
)
695 uint64_t max_size
= metaslab_pp_maxsize(sm
);
697 if (max_size
> (metaslab_min_alloc_size
<< metaslab_ndf_clump_shift
))
703 static space_map_ops_t metaslab_ndf_ops
= {
710 metaslab_ndf_fragmented
713 space_map_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
714 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
717 * ==========================================================================
719 * ==========================================================================
722 metaslab_init(metaslab_group_t
*mg
, space_map_obj_t
*smo
,
723 uint64_t start
, uint64_t size
, uint64_t txg
)
725 vdev_t
*vd
= mg
->mg_vd
;
728 msp
= kmem_zalloc(sizeof (metaslab_t
), KM_PUSHPAGE
);
729 mutex_init(&msp
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
731 msp
->ms_smo_syncing
= *smo
;
734 * We create the main space map here, but we don't create the
735 * allocmaps and freemaps until metaslab_sync_done(). This serves
736 * two purposes: it allows metaslab_sync_done() to detect the
737 * addition of new space; and for debugging, it ensures that we'd
738 * data fault on any attempt to use this metaslab before it's ready.
740 space_map_create(&msp
->ms_map
, start
, size
,
741 vd
->vdev_ashift
, &msp
->ms_lock
);
743 metaslab_group_add(mg
, msp
);
745 if (metaslab_debug
&& smo
->smo_object
!= 0) {
746 mutex_enter(&msp
->ms_lock
);
747 VERIFY(space_map_load(&msp
->ms_map
, mg
->mg_class
->mc_ops
,
748 SM_FREE
, smo
, spa_meta_objset(vd
->vdev_spa
)) == 0);
749 mutex_exit(&msp
->ms_lock
);
753 * If we're opening an existing pool (txg == 0) or creating
754 * a new one (txg == TXG_INITIAL), all space is available now.
755 * If we're adding space to an existing pool, the new space
756 * does not become available until after this txg has synced.
758 if (txg
<= TXG_INITIAL
)
759 metaslab_sync_done(msp
, 0);
762 vdev_dirty(vd
, 0, NULL
, txg
);
763 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
770 metaslab_fini(metaslab_t
*msp
)
772 metaslab_group_t
*mg
= msp
->ms_group
;
775 vdev_space_update(mg
->mg_vd
,
776 -msp
->ms_smo
.smo_alloc
, 0, -msp
->ms_map
.sm_size
);
778 metaslab_group_remove(mg
, msp
);
780 mutex_enter(&msp
->ms_lock
);
782 space_map_unload(&msp
->ms_map
);
783 space_map_destroy(&msp
->ms_map
);
785 for (t
= 0; t
< TXG_SIZE
; t
++) {
786 space_map_destroy(&msp
->ms_allocmap
[t
]);
787 space_map_destroy(&msp
->ms_freemap
[t
]);
790 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++)
791 space_map_destroy(&msp
->ms_defermap
[t
]);
793 ASSERT3S(msp
->ms_deferspace
, ==, 0);
795 mutex_exit(&msp
->ms_lock
);
796 mutex_destroy(&msp
->ms_lock
);
798 kmem_free(msp
, sizeof (metaslab_t
));
801 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
802 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
803 #define METASLAB_ACTIVE_MASK \
804 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
807 metaslab_weight(metaslab_t
*msp
)
809 metaslab_group_t
*mg
= msp
->ms_group
;
810 space_map_t
*sm
= &msp
->ms_map
;
811 space_map_obj_t
*smo
= &msp
->ms_smo
;
812 vdev_t
*vd
= mg
->mg_vd
;
813 uint64_t weight
, space
;
815 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
818 * The baseline weight is the metaslab's free space.
820 space
= sm
->sm_size
- smo
->smo_alloc
;
824 * Modern disks have uniform bit density and constant angular velocity.
825 * Therefore, the outer recording zones are faster (higher bandwidth)
826 * than the inner zones by the ratio of outer to inner track diameter,
827 * which is typically around 2:1. We account for this by assigning
828 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
829 * In effect, this means that we'll select the metaslab with the most
830 * free bandwidth rather than simply the one with the most free space.
832 weight
= 2 * weight
-
833 ((sm
->sm_start
>> vd
->vdev_ms_shift
) * weight
) / vd
->vdev_ms_count
;
834 ASSERT(weight
>= space
&& weight
<= 2 * space
);
837 * For locality, assign higher weight to metaslabs which have
838 * a lower offset than what we've already activated.
840 if (sm
->sm_start
<= mg
->mg_bonus_area
)
841 weight
*= (metaslab_smo_bonus_pct
/ 100);
842 ASSERT(weight
>= space
&&
843 weight
<= 2 * (metaslab_smo_bonus_pct
/ 100) * space
);
845 if (sm
->sm_loaded
&& !sm
->sm_ops
->smop_fragmented(sm
)) {
847 * If this metaslab is one we're actively using, adjust its
848 * weight to make it preferable to any inactive metaslab so
849 * we'll polish it off.
851 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
857 metaslab_prefetch(metaslab_group_t
*mg
)
859 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
861 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
864 mutex_enter(&mg
->mg_lock
);
867 * Prefetch the next potential metaslabs
869 for (msp
= avl_first(t
), m
= 0; msp
; msp
= AVL_NEXT(t
, msp
), m
++) {
870 space_map_t
*sm
= &msp
->ms_map
;
871 space_map_obj_t
*smo
= &msp
->ms_smo
;
873 /* If we have reached our prefetch limit then we're done */
874 if (m
>= metaslab_prefetch_limit
)
877 if (!sm
->sm_loaded
&& smo
->smo_object
!= 0) {
878 mutex_exit(&mg
->mg_lock
);
879 dmu_prefetch(spa_meta_objset(spa
), smo
->smo_object
,
880 0ULL, smo
->smo_objsize
);
881 mutex_enter(&mg
->mg_lock
);
884 mutex_exit(&mg
->mg_lock
);
888 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
890 metaslab_group_t
*mg
= msp
->ms_group
;
891 space_map_t
*sm
= &msp
->ms_map
;
892 space_map_ops_t
*sm_ops
= msp
->ms_group
->mg_class
->mc_ops
;
895 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
897 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
898 space_map_load_wait(sm
);
899 if (!sm
->sm_loaded
) {
900 int error
= space_map_load(sm
, sm_ops
, SM_FREE
,
902 spa_meta_objset(msp
->ms_group
->mg_vd
->vdev_spa
));
904 metaslab_group_sort(msp
->ms_group
, msp
, 0);
907 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++)
908 space_map_walk(&msp
->ms_defermap
[t
],
909 space_map_claim
, sm
);
914 * Track the bonus area as we activate new metaslabs.
916 if (sm
->sm_start
> mg
->mg_bonus_area
) {
917 mutex_enter(&mg
->mg_lock
);
918 mg
->mg_bonus_area
= sm
->sm_start
;
919 mutex_exit(&mg
->mg_lock
);
922 metaslab_group_sort(msp
->ms_group
, msp
,
923 msp
->ms_weight
| activation_weight
);
925 ASSERT(sm
->sm_loaded
);
926 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
932 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
935 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
936 * this metaslab again. In that case, it had better be empty,
937 * or we would be leaving space on the table.
939 ASSERT(size
>= SPA_MINBLOCKSIZE
|| msp
->ms_map
.sm_space
== 0);
940 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
941 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
945 * Write a metaslab to disk in the context of the specified transaction group.
948 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
950 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
951 spa_t
*spa
= vd
->vdev_spa
;
952 objset_t
*mos
= spa_meta_objset(spa
);
953 space_map_t
*allocmap
= &msp
->ms_allocmap
[txg
& TXG_MASK
];
954 space_map_t
*freemap
= &msp
->ms_freemap
[txg
& TXG_MASK
];
955 space_map_t
*freed_map
= &msp
->ms_freemap
[TXG_CLEAN(txg
) & TXG_MASK
];
956 space_map_t
*sm
= &msp
->ms_map
;
957 space_map_obj_t
*smo
= &msp
->ms_smo_syncing
;
962 ASSERT(!vd
->vdev_ishole
);
964 if (allocmap
->sm_space
== 0 && freemap
->sm_space
== 0)
968 * The only state that can actually be changing concurrently with
969 * metaslab_sync() is the metaslab's ms_map. No other thread can
970 * be modifying this txg's allocmap, freemap, freed_map, or smo.
971 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
972 * We drop it whenever we call into the DMU, because the DMU
973 * can call down to us (e.g. via zio_free()) at any time.
976 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
978 if (smo
->smo_object
== 0) {
979 ASSERT(smo
->smo_objsize
== 0);
980 ASSERT(smo
->smo_alloc
== 0);
981 smo
->smo_object
= dmu_object_alloc(mos
,
982 DMU_OT_SPACE_MAP
, 1 << SPACE_MAP_BLOCKSHIFT
,
983 DMU_OT_SPACE_MAP_HEADER
, sizeof (*smo
), tx
);
984 ASSERT(smo
->smo_object
!= 0);
985 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
986 (sm
->sm_start
>> vd
->vdev_ms_shift
),
987 sizeof (uint64_t), &smo
->smo_object
, tx
);
990 mutex_enter(&msp
->ms_lock
);
992 space_map_walk(freemap
, space_map_add
, freed_map
);
994 if (sm
->sm_loaded
&& spa_sync_pass(spa
) == 1 && smo
->smo_objsize
>=
995 2 * sizeof (uint64_t) * avl_numnodes(&sm
->sm_root
)) {
997 * The in-core space map representation is twice as compact
998 * as the on-disk one, so it's time to condense the latter
999 * by generating a pure allocmap from first principles.
1001 * This metaslab is 100% allocated,
1002 * minus the content of the in-core map (sm),
1003 * minus what's been freed this txg (freed_map),
1004 * minus deferred frees (ms_defermap[]),
1005 * minus allocations from txgs in the future
1006 * (because they haven't been committed yet).
1008 space_map_vacate(allocmap
, NULL
, NULL
);
1009 space_map_vacate(freemap
, NULL
, NULL
);
1011 space_map_add(allocmap
, allocmap
->sm_start
, allocmap
->sm_size
);
1013 space_map_walk(sm
, space_map_remove
, allocmap
);
1014 space_map_walk(freed_map
, space_map_remove
, allocmap
);
1016 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++)
1017 space_map_walk(&msp
->ms_defermap
[t
],
1018 space_map_remove
, allocmap
);
1020 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++)
1021 space_map_walk(&msp
->ms_allocmap
[(txg
+ t
) & TXG_MASK
],
1022 space_map_remove
, allocmap
);
1024 mutex_exit(&msp
->ms_lock
);
1025 space_map_truncate(smo
, mos
, tx
);
1026 mutex_enter(&msp
->ms_lock
);
1029 space_map_sync(allocmap
, SM_ALLOC
, smo
, mos
, tx
);
1030 space_map_sync(freemap
, SM_FREE
, smo
, mos
, tx
);
1032 mutex_exit(&msp
->ms_lock
);
1034 VERIFY(0 == dmu_bonus_hold(mos
, smo
->smo_object
, FTAG
, &db
));
1035 dmu_buf_will_dirty(db
, tx
);
1036 ASSERT3U(db
->db_size
, >=, sizeof (*smo
));
1037 bcopy(smo
, db
->db_data
, sizeof (*smo
));
1038 dmu_buf_rele(db
, FTAG
);
1044 * Called after a transaction group has completely synced to mark
1045 * all of the metaslab's free space as usable.
1048 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1050 space_map_obj_t
*smo
= &msp
->ms_smo
;
1051 space_map_obj_t
*smosync
= &msp
->ms_smo_syncing
;
1052 space_map_t
*sm
= &msp
->ms_map
;
1053 space_map_t
*freed_map
= &msp
->ms_freemap
[TXG_CLEAN(txg
) & TXG_MASK
];
1054 space_map_t
*defer_map
= &msp
->ms_defermap
[txg
% TXG_DEFER_SIZE
];
1055 metaslab_group_t
*mg
= msp
->ms_group
;
1056 vdev_t
*vd
= mg
->mg_vd
;
1057 int64_t alloc_delta
, defer_delta
;
1060 ASSERT(!vd
->vdev_ishole
);
1062 mutex_enter(&msp
->ms_lock
);
1065 * If this metaslab is just becoming available, initialize its
1066 * allocmaps and freemaps and add its capacity to the vdev.
1068 if (freed_map
->sm_size
== 0) {
1069 for (t
= 0; t
< TXG_SIZE
; t
++) {
1070 space_map_create(&msp
->ms_allocmap
[t
], sm
->sm_start
,
1071 sm
->sm_size
, sm
->sm_shift
, sm
->sm_lock
);
1072 space_map_create(&msp
->ms_freemap
[t
], sm
->sm_start
,
1073 sm
->sm_size
, sm
->sm_shift
, sm
->sm_lock
);
1076 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++)
1077 space_map_create(&msp
->ms_defermap
[t
], sm
->sm_start
,
1078 sm
->sm_size
, sm
->sm_shift
, sm
->sm_lock
);
1080 vdev_space_update(vd
, 0, 0, sm
->sm_size
);
1083 alloc_delta
= smosync
->smo_alloc
- smo
->smo_alloc
;
1084 defer_delta
= freed_map
->sm_space
- defer_map
->sm_space
;
1086 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1088 ASSERT(msp
->ms_allocmap
[txg
& TXG_MASK
].sm_space
== 0);
1089 ASSERT(msp
->ms_freemap
[txg
& TXG_MASK
].sm_space
== 0);
1092 * If there's a space_map_load() in progress, wait for it to complete
1093 * so that we have a consistent view of the in-core space map.
1094 * Then, add defer_map (oldest deferred frees) to this map and
1095 * transfer freed_map (this txg's frees) to defer_map.
1097 space_map_load_wait(sm
);
1098 space_map_vacate(defer_map
, sm
->sm_loaded
? space_map_free
: NULL
, sm
);
1099 space_map_vacate(freed_map
, space_map_add
, defer_map
);
1103 msp
->ms_deferspace
+= defer_delta
;
1104 ASSERT3S(msp
->ms_deferspace
, >=, 0);
1105 ASSERT3S(msp
->ms_deferspace
, <=, sm
->sm_size
);
1106 if (msp
->ms_deferspace
!= 0) {
1108 * Keep syncing this metaslab until all deferred frees
1109 * are back in circulation.
1111 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1115 * If the map is loaded but no longer active, evict it as soon as all
1116 * future allocations have synced. (If we unloaded it now and then
1117 * loaded a moment later, the map wouldn't reflect those allocations.)
1119 if (sm
->sm_loaded
&& (msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1122 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++)
1123 if (msp
->ms_allocmap
[(txg
+ t
) & TXG_MASK
].sm_space
)
1126 if (evictable
&& !metaslab_debug
)
1127 space_map_unload(sm
);
1130 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
1132 mutex_exit(&msp
->ms_lock
);
1136 metaslab_sync_reassess(metaslab_group_t
*mg
)
1138 vdev_t
*vd
= mg
->mg_vd
;
1139 int64_t failures
= mg
->mg_alloc_failures
;
1143 * Re-evaluate all metaslabs which have lower offsets than the
1146 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
1147 metaslab_t
*msp
= vd
->vdev_ms
[m
];
1149 if (msp
->ms_map
.sm_start
> mg
->mg_bonus_area
)
1152 mutex_enter(&msp
->ms_lock
);
1153 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
1154 mutex_exit(&msp
->ms_lock
);
1157 atomic_add_64(&mg
->mg_alloc_failures
, -failures
);
1160 * Prefetch the next potential metaslabs
1162 metaslab_prefetch(mg
);
1166 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
1168 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
1169 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
1170 uint64_t start
= msp
->ms_map
.sm_start
>> ms_shift
;
1172 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
1173 return (1ULL << 63);
1176 return ((start
- offset
) << ms_shift
);
1178 return ((offset
- start
) << ms_shift
);
1183 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
1184 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
, int flags
)
1186 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1187 metaslab_t
*msp
= NULL
;
1188 uint64_t offset
= -1ULL;
1189 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1190 uint64_t activation_weight
;
1191 uint64_t target_distance
;
1194 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
1195 for (i
= 0; i
< d
; i
++) {
1196 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
1197 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
1203 boolean_t was_active
;
1205 mutex_enter(&mg
->mg_lock
);
1206 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
1207 if (msp
->ms_weight
< asize
) {
1208 spa_dbgmsg(spa
, "%s: failed to meet weight "
1209 "requirement: vdev %llu, txg %llu, mg %p, "
1210 "msp %p, psize %llu, asize %llu, "
1211 "failures %llu, weight %llu",
1212 spa_name(spa
), mg
->mg_vd
->vdev_id
, txg
,
1213 mg
, msp
, psize
, asize
,
1214 mg
->mg_alloc_failures
, msp
->ms_weight
);
1215 mutex_exit(&mg
->mg_lock
);
1218 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
1219 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
1222 target_distance
= min_distance
+
1223 (msp
->ms_smo
.smo_alloc
? 0 : min_distance
>> 1);
1225 for (i
= 0; i
< d
; i
++)
1226 if (metaslab_distance(msp
, &dva
[i
]) <
1232 mutex_exit(&mg
->mg_lock
);
1237 * If we've already reached the allowable number of failed
1238 * allocation attempts on this metaslab group then we
1239 * consider skipping it. We skip it only if we're allowed
1240 * to "fast" gang, the physical size is larger than
1241 * a gang block, and we're attempting to allocate from
1242 * the primary metaslab.
1244 if (mg
->mg_alloc_failures
> zfs_mg_alloc_failures
&&
1245 CAN_FASTGANG(flags
) && psize
> SPA_GANGBLOCKSIZE
&&
1246 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
1247 spa_dbgmsg(spa
, "%s: skipping metaslab group: "
1248 "vdev %llu, txg %llu, mg %p, psize %llu, "
1249 "asize %llu, failures %llu", spa_name(spa
),
1250 mg
->mg_vd
->vdev_id
, txg
, mg
, psize
, asize
,
1251 mg
->mg_alloc_failures
);
1255 mutex_enter(&msp
->ms_lock
);
1258 * Ensure that the metaslab we have selected is still
1259 * capable of handling our request. It's possible that
1260 * another thread may have changed the weight while we
1261 * were blocked on the metaslab lock.
1263 if (msp
->ms_weight
< asize
|| (was_active
&&
1264 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
1265 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
1266 mutex_exit(&msp
->ms_lock
);
1270 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
1271 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
1272 metaslab_passivate(msp
,
1273 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
1274 mutex_exit(&msp
->ms_lock
);
1278 if (metaslab_activate(msp
, activation_weight
) != 0) {
1279 mutex_exit(&msp
->ms_lock
);
1283 if ((offset
= space_map_alloc(&msp
->ms_map
, asize
)) != -1ULL)
1286 atomic_inc_64(&mg
->mg_alloc_failures
);
1288 metaslab_passivate(msp
, space_map_maxsize(&msp
->ms_map
));
1290 mutex_exit(&msp
->ms_lock
);
1293 if (msp
->ms_allocmap
[txg
& TXG_MASK
].sm_space
== 0)
1294 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
1296 space_map_add(&msp
->ms_allocmap
[txg
& TXG_MASK
], offset
, asize
);
1298 mutex_exit(&msp
->ms_lock
);
1304 * Allocate a block for the specified i/o.
1307 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
1308 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
1310 metaslab_group_t
*mg
, *rotor
;
1314 int zio_lock
= B_FALSE
;
1315 boolean_t allocatable
;
1316 uint64_t offset
= -1ULL;
1320 ASSERT(!DVA_IS_VALID(&dva
[d
]));
1323 * For testing, make some blocks above a certain size be gang blocks.
1325 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
1329 * Start at the rotor and loop through all mgs until we find something.
1330 * Note that there's no locking on mc_rotor or mc_aliquot because
1331 * nothing actually breaks if we miss a few updates -- we just won't
1332 * allocate quite as evenly. It all balances out over time.
1334 * If we are doing ditto or log blocks, try to spread them across
1335 * consecutive vdevs. If we're forced to reuse a vdev before we've
1336 * allocated all of our ditto blocks, then try and spread them out on
1337 * that vdev as much as possible. If it turns out to not be possible,
1338 * gradually lower our standards until anything becomes acceptable.
1339 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
1340 * gives us hope of containing our fault domains to something we're
1341 * able to reason about. Otherwise, any two top-level vdev failures
1342 * will guarantee the loss of data. With consecutive allocation,
1343 * only two adjacent top-level vdev failures will result in data loss.
1345 * If we are doing gang blocks (hintdva is non-NULL), try to keep
1346 * ourselves on the same vdev as our gang block header. That
1347 * way, we can hope for locality in vdev_cache, plus it makes our
1348 * fault domains something tractable.
1351 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
1354 * It's possible the vdev we're using as the hint no
1355 * longer exists (i.e. removed). Consult the rotor when
1361 if (flags
& METASLAB_HINTBP_AVOID
&&
1362 mg
->mg_next
!= NULL
)
1367 } else if (d
!= 0) {
1368 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
1369 mg
= vd
->vdev_mg
->mg_next
;
1375 * If the hint put us into the wrong metaslab class, or into a
1376 * metaslab group that has been passivated, just follow the rotor.
1378 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
1385 ASSERT(mg
->mg_activation_count
== 1);
1390 * Don't allocate from faulted devices.
1393 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
1394 allocatable
= vdev_allocatable(vd
);
1395 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
1397 allocatable
= vdev_allocatable(vd
);
1403 * Avoid writing single-copy data to a failing vdev
1405 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
1406 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
1407 d
== 0 && dshift
== 3) {
1412 ASSERT(mg
->mg_class
== mc
);
1414 distance
= vd
->vdev_asize
>> dshift
;
1415 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
1420 asize
= vdev_psize_to_asize(vd
, psize
);
1421 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
1423 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
1425 if (offset
!= -1ULL) {
1427 * If we've just selected this metaslab group,
1428 * figure out whether the corresponding vdev is
1429 * over- or under-used relative to the pool,
1430 * and set an allocation bias to even it out.
1432 if (mc
->mc_aliquot
== 0) {
1433 vdev_stat_t
*vs
= &vd
->vdev_stat
;
1436 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
1437 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
1440 * Calculate how much more or less we should
1441 * try to allocate from this device during
1442 * this iteration around the rotor.
1443 * For example, if a device is 80% full
1444 * and the pool is 20% full then we should
1445 * reduce allocations by 60% on this device.
1447 * mg_bias = (20 - 80) * 512K / 100 = -307K
1449 * This reduces allocations by 307K for this
1452 mg
->mg_bias
= ((cu
- vu
) *
1453 (int64_t)mg
->mg_aliquot
) / 100;
1456 if (atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
1457 mg
->mg_aliquot
+ mg
->mg_bias
) {
1458 mc
->mc_rotor
= mg
->mg_next
;
1462 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
1463 DVA_SET_OFFSET(&dva
[d
], offset
);
1464 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
1465 DVA_SET_ASIZE(&dva
[d
], asize
);
1470 mc
->mc_rotor
= mg
->mg_next
;
1472 } while ((mg
= mg
->mg_next
) != rotor
);
1476 ASSERT(dshift
< 64);
1480 if (!allocatable
&& !zio_lock
) {
1486 bzero(&dva
[d
], sizeof (dva_t
));
1492 * Free the block represented by DVA in the context of the specified
1493 * transaction group.
1496 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
1498 uint64_t vdev
= DVA_GET_VDEV(dva
);
1499 uint64_t offset
= DVA_GET_OFFSET(dva
);
1500 uint64_t size
= DVA_GET_ASIZE(dva
);
1504 ASSERT(DVA_IS_VALID(dva
));
1506 if (txg
> spa_freeze_txg(spa
))
1509 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
1510 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
1511 cmn_err(CE_WARN
, "metaslab_free_dva(): bad DVA %llu:%llu",
1512 (u_longlong_t
)vdev
, (u_longlong_t
)offset
);
1517 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
1519 if (DVA_GET_GANG(dva
))
1520 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1522 mutex_enter(&msp
->ms_lock
);
1525 space_map_remove(&msp
->ms_allocmap
[txg
& TXG_MASK
],
1527 space_map_free(&msp
->ms_map
, offset
, size
);
1529 if (msp
->ms_freemap
[txg
& TXG_MASK
].sm_space
== 0)
1530 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
1531 space_map_add(&msp
->ms_freemap
[txg
& TXG_MASK
], offset
, size
);
1534 mutex_exit(&msp
->ms_lock
);
1538 * Intent log support: upon opening the pool after a crash, notify the SPA
1539 * of blocks that the intent log has allocated for immediate write, but
1540 * which are still considered free by the SPA because the last transaction
1541 * group didn't commit yet.
1544 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
1546 uint64_t vdev
= DVA_GET_VDEV(dva
);
1547 uint64_t offset
= DVA_GET_OFFSET(dva
);
1548 uint64_t size
= DVA_GET_ASIZE(dva
);
1553 ASSERT(DVA_IS_VALID(dva
));
1555 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
1556 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
1559 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
1561 if (DVA_GET_GANG(dva
))
1562 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
1564 mutex_enter(&msp
->ms_lock
);
1566 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_map
.sm_loaded
)
1567 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
1569 if (error
== 0 && !space_map_contains(&msp
->ms_map
, offset
, size
))
1572 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
1573 mutex_exit(&msp
->ms_lock
);
1577 space_map_claim(&msp
->ms_map
, offset
, size
);
1579 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
1580 if (msp
->ms_allocmap
[txg
& TXG_MASK
].sm_space
== 0)
1581 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
1582 space_map_add(&msp
->ms_allocmap
[txg
& TXG_MASK
], offset
, size
);
1585 mutex_exit(&msp
->ms_lock
);
1591 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
1592 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
1594 dva_t
*dva
= bp
->blk_dva
;
1595 dva_t
*hintdva
= hintbp
->blk_dva
;
1598 ASSERT(bp
->blk_birth
== 0);
1599 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
1601 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
1603 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
1604 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1608 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
1609 ASSERT(BP_GET_NDVAS(bp
) == 0);
1610 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
1612 for (d
= 0; d
< ndvas
; d
++) {
1613 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
1616 for (d
--; d
>= 0; d
--) {
1617 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
1618 bzero(&dva
[d
], sizeof (dva_t
));
1620 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1625 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
1627 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1629 BP_SET_BIRTH(bp
, txg
, txg
);
1635 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
1637 const dva_t
*dva
= bp
->blk_dva
;
1638 int d
, ndvas
= BP_GET_NDVAS(bp
);
1640 ASSERT(!BP_IS_HOLE(bp
));
1641 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
1643 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
1645 for (d
= 0; d
< ndvas
; d
++)
1646 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
1648 spa_config_exit(spa
, SCL_FREE
, FTAG
);
1652 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
1654 const dva_t
*dva
= bp
->blk_dva
;
1655 int ndvas
= BP_GET_NDVAS(bp
);
1658 ASSERT(!BP_IS_HOLE(bp
));
1662 * First do a dry run to make sure all DVAs are claimable,
1663 * so we don't have to unwind from partial failures below.
1665 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
1669 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
1671 for (d
= 0; d
< ndvas
; d
++)
1672 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
1675 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
1677 ASSERT(error
== 0 || txg
== 0);