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) 2011, 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
27 #include <sys/zfs_context.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/space_map.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/vdev_impl.h>
34 #include <sys/spa_impl.h>
35 #include <sys/zfeature.h>
37 #define WITH_DF_BLOCK_ALLOCATOR
39 #define GANG_ALLOCATION(flags) \
40 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
43 * Metaslab granularity, in bytes. This is roughly similar to what would be
44 * referred to as the "stripe size" in traditional RAID arrays. In normal
45 * operation, we will try to write this amount of data to a top-level vdev
46 * before moving on to the next one.
48 unsigned long metaslab_aliquot
= 512 << 10;
50 uint64_t metaslab_gang_bang
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
53 * The in-core space map representation is more compact than its on-disk form.
54 * The zfs_condense_pct determines how much more compact the in-core
55 * space map representation must be before we compact it on-disk.
56 * Values should be greater than or equal to 100.
58 int zfs_condense_pct
= 200;
61 * Condensing a metaslab is not guaranteed to actually reduce the amount of
62 * space used on disk. In particular, a space map uses data in increments of
63 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
64 * same number of blocks after condensing. Since the goal of condensing is to
65 * reduce the number of IOPs required to read the space map, we only want to
66 * condense when we can be sure we will reduce the number of blocks used by the
67 * space map. Unfortunately, we cannot precisely compute whether or not this is
68 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
69 * we apply the following heuristic: do not condense a spacemap unless the
70 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
73 int zfs_metaslab_condense_block_threshold
= 4;
76 * The zfs_mg_noalloc_threshold defines which metaslab groups should
77 * be eligible for allocation. The value is defined as a percentage of
78 * free space. Metaslab groups that have more free space than
79 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
80 * a metaslab group's free space is less than or equal to the
81 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
82 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
83 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
84 * groups are allowed to accept allocations. Gang blocks are always
85 * eligible to allocate on any metaslab group. The default value of 0 means
86 * no metaslab group will be excluded based on this criterion.
88 int zfs_mg_noalloc_threshold
= 0;
91 * Metaslab groups are considered eligible for allocations if their
92 * fragmenation metric (measured as a percentage) is less than or equal to
93 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
94 * then it will be skipped unless all metaslab groups within the metaslab
95 * class have also crossed this threshold.
97 int zfs_mg_fragmentation_threshold
= 85;
100 * Allow metaslabs to keep their active state as long as their fragmentation
101 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
102 * active metaslab that exceeds this threshold will no longer keep its active
103 * status allowing better metaslabs to be selected.
105 int zfs_metaslab_fragmentation_threshold
= 70;
108 * When set will load all metaslabs when pool is first opened.
110 int metaslab_debug_load
= 0;
113 * When set will prevent metaslabs from being unloaded.
115 int metaslab_debug_unload
= 0;
118 * Minimum size which forces the dynamic allocator to change
119 * it's allocation strategy. Once the space map cannot satisfy
120 * an allocation of this size then it switches to using more
121 * aggressive strategy (i.e search by size rather than offset).
123 uint64_t metaslab_df_alloc_threshold
= SPA_OLD_MAXBLOCKSIZE
;
126 * The minimum free space, in percent, which must be available
127 * in a space map to continue allocations in a first-fit fashion.
128 * Once the space map's free space drops below this level we dynamically
129 * switch to using best-fit allocations.
131 int metaslab_df_free_pct
= 4;
134 * Percentage of all cpus that can be used by the metaslab taskq.
136 int metaslab_load_pct
= 50;
139 * Determines how many txgs a metaslab may remain loaded without having any
140 * allocations from it. As long as a metaslab continues to be used we will
143 int metaslab_unload_delay
= TXG_SIZE
* 2;
146 * Max number of metaslabs per group to preload.
148 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
151 * Enable/disable preloading of metaslab.
153 int metaslab_preload_enabled
= B_TRUE
;
156 * Enable/disable fragmentation weighting on metaslabs.
158 int metaslab_fragmentation_factor_enabled
= B_TRUE
;
161 * Enable/disable lba weighting (i.e. outer tracks are given preference).
163 int metaslab_lba_weighting_enabled
= B_TRUE
;
166 * Enable/disable metaslab group biasing.
168 int metaslab_bias_enabled
= B_TRUE
;
172 * Enable/disable segment-based metaslab selection.
174 int zfs_metaslab_segment_weight_enabled
= B_TRUE
;
177 * When using segment-based metaslab selection, we will continue
178 * allocating from the active metaslab until we have exhausted
179 * zfs_metaslab_switch_threshold of its buckets.
181 int zfs_metaslab_switch_threshold
= 2;
184 * Internal switch to enable/disable the metaslab allocation tracing
187 #ifdef _METASLAB_TRACING
188 boolean_t metaslab_trace_enabled
= B_TRUE
;
192 * Maximum entries that the metaslab allocation tracing facility will keep
193 * in a given list when running in non-debug mode. We limit the number
194 * of entries in non-debug mode to prevent us from using up too much memory.
195 * The limit should be sufficiently large that we don't expect any allocation
196 * to every exceed this value. In debug mode, the system will panic if this
197 * limit is ever reached allowing for further investigation.
199 #ifdef _METASLAB_TRACING
200 uint64_t metaslab_trace_max_entries
= 5000;
203 static uint64_t metaslab_weight(metaslab_t
*);
204 static void metaslab_set_fragmentation(metaslab_t
*);
206 #ifdef _METASLAB_TRACING
207 kmem_cache_t
*metaslab_alloc_trace_cache
;
211 * ==========================================================================
213 * ==========================================================================
216 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
218 metaslab_class_t
*mc
;
220 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
225 mutex_init(&mc
->mc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
226 refcount_create_tracked(&mc
->mc_alloc_slots
);
232 metaslab_class_destroy(metaslab_class_t
*mc
)
234 ASSERT(mc
->mc_rotor
== NULL
);
235 ASSERT(mc
->mc_alloc
== 0);
236 ASSERT(mc
->mc_deferred
== 0);
237 ASSERT(mc
->mc_space
== 0);
238 ASSERT(mc
->mc_dspace
== 0);
240 refcount_destroy(&mc
->mc_alloc_slots
);
241 mutex_destroy(&mc
->mc_lock
);
242 kmem_free(mc
, sizeof (metaslab_class_t
));
246 metaslab_class_validate(metaslab_class_t
*mc
)
248 metaslab_group_t
*mg
;
252 * Must hold one of the spa_config locks.
254 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
255 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
257 if ((mg
= mc
->mc_rotor
) == NULL
)
262 ASSERT(vd
->vdev_mg
!= NULL
);
263 ASSERT3P(vd
->vdev_top
, ==, vd
);
264 ASSERT3P(mg
->mg_class
, ==, mc
);
265 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
266 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
272 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
273 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
275 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
276 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
277 atomic_add_64(&mc
->mc_space
, space_delta
);
278 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
282 metaslab_class_get_alloc(metaslab_class_t
*mc
)
284 return (mc
->mc_alloc
);
288 metaslab_class_get_deferred(metaslab_class_t
*mc
)
290 return (mc
->mc_deferred
);
294 metaslab_class_get_space(metaslab_class_t
*mc
)
296 return (mc
->mc_space
);
300 metaslab_class_get_dspace(metaslab_class_t
*mc
)
302 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
306 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
308 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
312 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
315 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
318 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
319 vdev_t
*tvd
= rvd
->vdev_child
[c
];
320 metaslab_group_t
*mg
= tvd
->vdev_mg
;
323 * Skip any holes, uninitialized top-levels, or
324 * vdevs that are not in this metalab class.
326 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
327 mg
->mg_class
!= mc
) {
331 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
332 mc_hist
[i
] += mg
->mg_histogram
[i
];
335 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
336 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
338 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
342 * Calculate the metaslab class's fragmentation metric. The metric
343 * is weighted based on the space contribution of each metaslab group.
344 * The return value will be a number between 0 and 100 (inclusive), or
345 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
346 * zfs_frag_table for more information about the metric.
349 metaslab_class_fragmentation(metaslab_class_t
*mc
)
351 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
352 uint64_t fragmentation
= 0;
355 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
357 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
358 vdev_t
*tvd
= rvd
->vdev_child
[c
];
359 metaslab_group_t
*mg
= tvd
->vdev_mg
;
362 * Skip any holes, uninitialized top-levels, or
363 * vdevs that are not in this metalab class.
365 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
366 mg
->mg_class
!= mc
) {
371 * If a metaslab group does not contain a fragmentation
372 * metric then just bail out.
374 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
375 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
376 return (ZFS_FRAG_INVALID
);
380 * Determine how much this metaslab_group is contributing
381 * to the overall pool fragmentation metric.
383 fragmentation
+= mg
->mg_fragmentation
*
384 metaslab_group_get_space(mg
);
386 fragmentation
/= metaslab_class_get_space(mc
);
388 ASSERT3U(fragmentation
, <=, 100);
389 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
390 return (fragmentation
);
394 * Calculate the amount of expandable space that is available in
395 * this metaslab class. If a device is expanded then its expandable
396 * space will be the amount of allocatable space that is currently not
397 * part of this metaslab class.
400 metaslab_class_expandable_space(metaslab_class_t
*mc
)
402 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
406 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
407 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
408 vdev_t
*tvd
= rvd
->vdev_child
[c
];
409 metaslab_group_t
*mg
= tvd
->vdev_mg
;
411 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
412 mg
->mg_class
!= mc
) {
417 * Calculate if we have enough space to add additional
418 * metaslabs. We report the expandable space in terms
419 * of the metaslab size since that's the unit of expansion.
421 space
+= P2ALIGN(tvd
->vdev_max_asize
- tvd
->vdev_asize
,
422 1ULL << tvd
->vdev_ms_shift
);
424 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
429 metaslab_compare(const void *x1
, const void *x2
)
431 const metaslab_t
*m1
= (const metaslab_t
*)x1
;
432 const metaslab_t
*m2
= (const metaslab_t
*)x2
;
434 int cmp
= AVL_CMP(m2
->ms_weight
, m1
->ms_weight
);
438 IMPLY(AVL_CMP(m1
->ms_start
, m2
->ms_start
) == 0, m1
== m2
);
440 return (AVL_CMP(m1
->ms_start
, m2
->ms_start
));
444 * Verify that the space accounting on disk matches the in-core range_trees.
447 metaslab_verify_space(metaslab_t
*msp
, uint64_t txg
)
449 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
450 uint64_t allocated
= 0;
451 uint64_t sm_free_space
, msp_free_space
;
454 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
456 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
460 * We can only verify the metaslab space when we're called
461 * from syncing context with a loaded metaslab that has an allocated
462 * space map. Calling this in non-syncing context does not
463 * provide a consistent view of the metaslab since we're performing
464 * allocations in the future.
466 if (txg
!= spa_syncing_txg(spa
) || msp
->ms_sm
== NULL
||
470 sm_free_space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
) -
471 space_map_alloc_delta(msp
->ms_sm
);
474 * Account for future allocations since we would have already
475 * deducted that space from the ms_freetree.
477 for (t
= 0; t
< TXG_CONCURRENT_STATES
; t
++) {
479 range_tree_space(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]);
482 msp_free_space
= range_tree_space(msp
->ms_tree
) + allocated
+
483 msp
->ms_deferspace
+ range_tree_space(msp
->ms_freedtree
);
485 VERIFY3U(sm_free_space
, ==, msp_free_space
);
489 * ==========================================================================
491 * ==========================================================================
494 * Update the allocatable flag and the metaslab group's capacity.
495 * The allocatable flag is set to true if the capacity is below
496 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
497 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
498 * transitions from allocatable to non-allocatable or vice versa then the
499 * metaslab group's class is updated to reflect the transition.
502 metaslab_group_alloc_update(metaslab_group_t
*mg
)
504 vdev_t
*vd
= mg
->mg_vd
;
505 metaslab_class_t
*mc
= mg
->mg_class
;
506 vdev_stat_t
*vs
= &vd
->vdev_stat
;
507 boolean_t was_allocatable
;
508 boolean_t was_initialized
;
510 ASSERT(vd
== vd
->vdev_top
);
512 mutex_enter(&mg
->mg_lock
);
513 was_allocatable
= mg
->mg_allocatable
;
514 was_initialized
= mg
->mg_initialized
;
516 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
519 mutex_enter(&mc
->mc_lock
);
522 * If the metaslab group was just added then it won't
523 * have any space until we finish syncing out this txg.
524 * At that point we will consider it initialized and available
525 * for allocations. We also don't consider non-activated
526 * metaslab groups (e.g. vdevs that are in the middle of being removed)
527 * to be initialized, because they can't be used for allocation.
529 mg
->mg_initialized
= metaslab_group_initialized(mg
);
530 if (!was_initialized
&& mg
->mg_initialized
) {
532 } else if (was_initialized
&& !mg
->mg_initialized
) {
533 ASSERT3U(mc
->mc_groups
, >, 0);
536 if (mg
->mg_initialized
)
537 mg
->mg_no_free_space
= B_FALSE
;
540 * A metaslab group is considered allocatable if it has plenty
541 * of free space or is not heavily fragmented. We only take
542 * fragmentation into account if the metaslab group has a valid
543 * fragmentation metric (i.e. a value between 0 and 100).
545 mg
->mg_allocatable
= (mg
->mg_activation_count
> 0 &&
546 mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
547 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
548 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
551 * The mc_alloc_groups maintains a count of the number of
552 * groups in this metaslab class that are still above the
553 * zfs_mg_noalloc_threshold. This is used by the allocating
554 * threads to determine if they should avoid allocations to
555 * a given group. The allocator will avoid allocations to a group
556 * if that group has reached or is below the zfs_mg_noalloc_threshold
557 * and there are still other groups that are above the threshold.
558 * When a group transitions from allocatable to non-allocatable or
559 * vice versa we update the metaslab class to reflect that change.
560 * When the mc_alloc_groups value drops to 0 that means that all
561 * groups have reached the zfs_mg_noalloc_threshold making all groups
562 * eligible for allocations. This effectively means that all devices
563 * are balanced again.
565 if (was_allocatable
&& !mg
->mg_allocatable
)
566 mc
->mc_alloc_groups
--;
567 else if (!was_allocatable
&& mg
->mg_allocatable
)
568 mc
->mc_alloc_groups
++;
569 mutex_exit(&mc
->mc_lock
);
571 mutex_exit(&mg
->mg_lock
);
575 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
577 metaslab_group_t
*mg
;
579 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
580 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
581 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
582 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
585 mg
->mg_activation_count
= 0;
586 mg
->mg_initialized
= B_FALSE
;
587 mg
->mg_no_free_space
= B_TRUE
;
588 refcount_create_tracked(&mg
->mg_alloc_queue_depth
);
590 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
591 maxclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
| TASKQ_DYNAMIC
);
597 metaslab_group_destroy(metaslab_group_t
*mg
)
599 ASSERT(mg
->mg_prev
== NULL
);
600 ASSERT(mg
->mg_next
== NULL
);
602 * We may have gone below zero with the activation count
603 * either because we never activated in the first place or
604 * because we're done, and possibly removing the vdev.
606 ASSERT(mg
->mg_activation_count
<= 0);
608 taskq_destroy(mg
->mg_taskq
);
609 avl_destroy(&mg
->mg_metaslab_tree
);
610 mutex_destroy(&mg
->mg_lock
);
611 refcount_destroy(&mg
->mg_alloc_queue_depth
);
612 kmem_free(mg
, sizeof (metaslab_group_t
));
616 metaslab_group_activate(metaslab_group_t
*mg
)
618 metaslab_class_t
*mc
= mg
->mg_class
;
619 metaslab_group_t
*mgprev
, *mgnext
;
621 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
623 ASSERT(mc
->mc_rotor
!= mg
);
624 ASSERT(mg
->mg_prev
== NULL
);
625 ASSERT(mg
->mg_next
== NULL
);
626 ASSERT(mg
->mg_activation_count
<= 0);
628 if (++mg
->mg_activation_count
<= 0)
631 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
632 metaslab_group_alloc_update(mg
);
634 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
638 mgnext
= mgprev
->mg_next
;
639 mg
->mg_prev
= mgprev
;
640 mg
->mg_next
= mgnext
;
641 mgprev
->mg_next
= mg
;
642 mgnext
->mg_prev
= mg
;
648 metaslab_group_passivate(metaslab_group_t
*mg
)
650 metaslab_class_t
*mc
= mg
->mg_class
;
651 metaslab_group_t
*mgprev
, *mgnext
;
653 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
655 if (--mg
->mg_activation_count
!= 0) {
656 ASSERT(mc
->mc_rotor
!= mg
);
657 ASSERT(mg
->mg_prev
== NULL
);
658 ASSERT(mg
->mg_next
== NULL
);
659 ASSERT(mg
->mg_activation_count
< 0);
663 taskq_wait_outstanding(mg
->mg_taskq
, 0);
664 metaslab_group_alloc_update(mg
);
666 mgprev
= mg
->mg_prev
;
667 mgnext
= mg
->mg_next
;
672 mc
->mc_rotor
= mgnext
;
673 mgprev
->mg_next
= mgnext
;
674 mgnext
->mg_prev
= mgprev
;
682 metaslab_group_initialized(metaslab_group_t
*mg
)
684 vdev_t
*vd
= mg
->mg_vd
;
685 vdev_stat_t
*vs
= &vd
->vdev_stat
;
687 return (vs
->vs_space
!= 0 && mg
->mg_activation_count
> 0);
691 metaslab_group_get_space(metaslab_group_t
*mg
)
693 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
697 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
700 vdev_t
*vd
= mg
->mg_vd
;
701 uint64_t ashift
= vd
->vdev_ashift
;
704 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
707 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
710 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
711 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
713 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
714 metaslab_t
*msp
= vd
->vdev_ms
[m
];
716 if (msp
->ms_sm
== NULL
)
719 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
720 mg_hist
[i
+ ashift
] +=
721 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
724 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
725 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
727 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
731 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
733 metaslab_class_t
*mc
= mg
->mg_class
;
734 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
737 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
738 if (msp
->ms_sm
== NULL
)
741 mutex_enter(&mg
->mg_lock
);
742 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
743 mg
->mg_histogram
[i
+ ashift
] +=
744 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
745 mc
->mc_histogram
[i
+ ashift
] +=
746 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
748 mutex_exit(&mg
->mg_lock
);
752 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
754 metaslab_class_t
*mc
= mg
->mg_class
;
755 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
758 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
759 if (msp
->ms_sm
== NULL
)
762 mutex_enter(&mg
->mg_lock
);
763 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
764 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
765 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
766 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
767 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
769 mg
->mg_histogram
[i
+ ashift
] -=
770 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
771 mc
->mc_histogram
[i
+ ashift
] -=
772 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
774 mutex_exit(&mg
->mg_lock
);
778 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
780 ASSERT(msp
->ms_group
== NULL
);
781 mutex_enter(&mg
->mg_lock
);
784 avl_add(&mg
->mg_metaslab_tree
, msp
);
785 mutex_exit(&mg
->mg_lock
);
787 mutex_enter(&msp
->ms_lock
);
788 metaslab_group_histogram_add(mg
, msp
);
789 mutex_exit(&msp
->ms_lock
);
793 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
795 mutex_enter(&msp
->ms_lock
);
796 metaslab_group_histogram_remove(mg
, msp
);
797 mutex_exit(&msp
->ms_lock
);
799 mutex_enter(&mg
->mg_lock
);
800 ASSERT(msp
->ms_group
== mg
);
801 avl_remove(&mg
->mg_metaslab_tree
, msp
);
802 msp
->ms_group
= NULL
;
803 mutex_exit(&mg
->mg_lock
);
807 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
810 * Although in principle the weight can be any value, in
811 * practice we do not use values in the range [1, 511].
813 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
814 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
816 mutex_enter(&mg
->mg_lock
);
817 ASSERT(msp
->ms_group
== mg
);
818 avl_remove(&mg
->mg_metaslab_tree
, msp
);
819 msp
->ms_weight
= weight
;
820 avl_add(&mg
->mg_metaslab_tree
, msp
);
821 mutex_exit(&mg
->mg_lock
);
825 * Calculate the fragmentation for a given metaslab group. We can use
826 * a simple average here since all metaslabs within the group must have
827 * the same size. The return value will be a value between 0 and 100
828 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
829 * group have a fragmentation metric.
832 metaslab_group_fragmentation(metaslab_group_t
*mg
)
834 vdev_t
*vd
= mg
->mg_vd
;
835 uint64_t fragmentation
= 0;
836 uint64_t valid_ms
= 0;
839 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
840 metaslab_t
*msp
= vd
->vdev_ms
[m
];
842 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
846 fragmentation
+= msp
->ms_fragmentation
;
849 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
850 return (ZFS_FRAG_INVALID
);
852 fragmentation
/= valid_ms
;
853 ASSERT3U(fragmentation
, <=, 100);
854 return (fragmentation
);
858 * Determine if a given metaslab group should skip allocations. A metaslab
859 * group should avoid allocations if its free capacity is less than the
860 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
861 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
862 * that can still handle allocations. If the allocation throttle is enabled
863 * then we skip allocations to devices that have reached their maximum
864 * allocation queue depth unless the selected metaslab group is the only
865 * eligible group remaining.
868 metaslab_group_allocatable(metaslab_group_t
*mg
, metaslab_group_t
*rotor
,
871 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
872 metaslab_class_t
*mc
= mg
->mg_class
;
875 * We can only consider skipping this metaslab group if it's
876 * in the normal metaslab class and there are other metaslab
877 * groups to select from. Otherwise, we always consider it eligible
880 if (mc
!= spa_normal_class(spa
) || mc
->mc_groups
<= 1)
884 * If the metaslab group's mg_allocatable flag is set (see comments
885 * in metaslab_group_alloc_update() for more information) and
886 * the allocation throttle is disabled then allow allocations to this
887 * device. However, if the allocation throttle is enabled then
888 * check if we have reached our allocation limit (mg_alloc_queue_depth)
889 * to determine if we should allow allocations to this metaslab group.
890 * If all metaslab groups are no longer considered allocatable
891 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
892 * gang block size then we allow allocations on this metaslab group
893 * regardless of the mg_allocatable or throttle settings.
895 if (mg
->mg_allocatable
) {
896 metaslab_group_t
*mgp
;
898 uint64_t qmax
= mg
->mg_max_alloc_queue_depth
;
900 if (!mc
->mc_alloc_throttle_enabled
)
904 * If this metaslab group does not have any free space, then
905 * there is no point in looking further.
907 if (mg
->mg_no_free_space
)
910 qdepth
= refcount_count(&mg
->mg_alloc_queue_depth
);
913 * If this metaslab group is below its qmax or it's
914 * the only allocatable metasable group, then attempt
915 * to allocate from it.
917 if (qdepth
< qmax
|| mc
->mc_alloc_groups
== 1)
919 ASSERT3U(mc
->mc_alloc_groups
, >, 1);
922 * Since this metaslab group is at or over its qmax, we
923 * need to determine if there are metaslab groups after this
924 * one that might be able to handle this allocation. This is
925 * racy since we can't hold the locks for all metaslab
926 * groups at the same time when we make this check.
928 for (mgp
= mg
->mg_next
; mgp
!= rotor
; mgp
= mgp
->mg_next
) {
929 qmax
= mgp
->mg_max_alloc_queue_depth
;
931 qdepth
= refcount_count(&mgp
->mg_alloc_queue_depth
);
934 * If there is another metaslab group that
935 * might be able to handle the allocation, then
936 * we return false so that we skip this group.
938 if (qdepth
< qmax
&& !mgp
->mg_no_free_space
)
943 * We didn't find another group to handle the allocation
944 * so we can't skip this metaslab group even though
945 * we are at or over our qmax.
949 } else if (mc
->mc_alloc_groups
== 0 || psize
== SPA_MINBLOCKSIZE
) {
956 * ==========================================================================
957 * Range tree callbacks
958 * ==========================================================================
962 * Comparison function for the private size-ordered tree. Tree is sorted
963 * by size, larger sizes at the end of the tree.
966 metaslab_rangesize_compare(const void *x1
, const void *x2
)
968 const range_seg_t
*r1
= x1
;
969 const range_seg_t
*r2
= x2
;
970 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
971 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
973 int cmp
= AVL_CMP(rs_size1
, rs_size2
);
977 return (AVL_CMP(r1
->rs_start
, r2
->rs_start
));
981 * Create any block allocator specific components. The current allocators
982 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
985 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
987 metaslab_t
*msp
= arg
;
989 ASSERT3P(rt
->rt_arg
, ==, msp
);
990 ASSERT(msp
->ms_tree
== NULL
);
992 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
993 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
997 * Destroy the block allocator specific components.
1000 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
1002 metaslab_t
*msp
= arg
;
1004 ASSERT3P(rt
->rt_arg
, ==, msp
);
1005 ASSERT3P(msp
->ms_tree
, ==, rt
);
1006 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
1008 avl_destroy(&msp
->ms_size_tree
);
1012 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
1014 metaslab_t
*msp
= arg
;
1016 ASSERT3P(rt
->rt_arg
, ==, msp
);
1017 ASSERT3P(msp
->ms_tree
, ==, rt
);
1018 VERIFY(!msp
->ms_condensing
);
1019 avl_add(&msp
->ms_size_tree
, rs
);
1023 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
1025 metaslab_t
*msp
= arg
;
1027 ASSERT3P(rt
->rt_arg
, ==, msp
);
1028 ASSERT3P(msp
->ms_tree
, ==, rt
);
1029 VERIFY(!msp
->ms_condensing
);
1030 avl_remove(&msp
->ms_size_tree
, rs
);
1034 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
1036 metaslab_t
*msp
= arg
;
1038 ASSERT3P(rt
->rt_arg
, ==, msp
);
1039 ASSERT3P(msp
->ms_tree
, ==, rt
);
1042 * Normally one would walk the tree freeing nodes along the way.
1043 * Since the nodes are shared with the range trees we can avoid
1044 * walking all nodes and just reinitialize the avl tree. The nodes
1045 * will be freed by the range tree, so we don't want to free them here.
1047 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
1048 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
1051 static range_tree_ops_t metaslab_rt_ops
= {
1053 metaslab_rt_destroy
,
1060 * ==========================================================================
1061 * Common allocator routines
1062 * ==========================================================================
1066 * Return the maximum contiguous segment within the metaslab.
1069 metaslab_block_maxsize(metaslab_t
*msp
)
1071 avl_tree_t
*t
= &msp
->ms_size_tree
;
1074 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
1077 return (rs
->rs_end
- rs
->rs_start
);
1080 static range_seg_t
*
1081 metaslab_block_find(avl_tree_t
*t
, uint64_t start
, uint64_t size
)
1083 range_seg_t
*rs
, rsearch
;
1086 rsearch
.rs_start
= start
;
1087 rsearch
.rs_end
= start
+ size
;
1089 rs
= avl_find(t
, &rsearch
, &where
);
1091 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1097 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
1098 defined(WITH_DF_BLOCK_ALLOCATOR) || \
1099 defined(WITH_CF_BLOCK_ALLOCATOR)
1101 * This is a helper function that can be used by the allocator to find
1102 * a suitable block to allocate. This will search the specified AVL
1103 * tree looking for a block that matches the specified criteria.
1106 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
1109 range_seg_t
*rs
= metaslab_block_find(t
, *cursor
, size
);
1111 while (rs
!= NULL
) {
1112 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
1114 if (offset
+ size
<= rs
->rs_end
) {
1115 *cursor
= offset
+ size
;
1118 rs
= AVL_NEXT(t
, rs
);
1122 * If we know we've searched the whole map (*cursor == 0), give up.
1123 * Otherwise, reset the cursor to the beginning and try again.
1129 return (metaslab_block_picker(t
, cursor
, size
, align
));
1131 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
1133 #if defined(WITH_FF_BLOCK_ALLOCATOR)
1135 * ==========================================================================
1136 * The first-fit block allocator
1137 * ==========================================================================
1140 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
1143 * Find the largest power of 2 block size that evenly divides the
1144 * requested size. This is used to try to allocate blocks with similar
1145 * alignment from the same area of the metaslab (i.e. same cursor
1146 * bucket) but it does not guarantee that other allocations sizes
1147 * may exist in the same region.
1149 uint64_t align
= size
& -size
;
1150 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1151 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1153 return (metaslab_block_picker(t
, cursor
, size
, align
));
1156 static metaslab_ops_t metaslab_ff_ops
= {
1160 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
1161 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1163 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1165 * ==========================================================================
1166 * Dynamic block allocator -
1167 * Uses the first fit allocation scheme until space get low and then
1168 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1169 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1170 * ==========================================================================
1173 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1176 * Find the largest power of 2 block size that evenly divides the
1177 * requested size. This is used to try to allocate blocks with similar
1178 * alignment from the same area of the metaslab (i.e. same cursor
1179 * bucket) but it does not guarantee that other allocations sizes
1180 * may exist in the same region.
1182 uint64_t align
= size
& -size
;
1183 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1184 range_tree_t
*rt
= msp
->ms_tree
;
1185 avl_tree_t
*t
= &rt
->rt_root
;
1186 uint64_t max_size
= metaslab_block_maxsize(msp
);
1187 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1189 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1190 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1192 if (max_size
< size
)
1196 * If we're running low on space switch to using the size
1197 * sorted AVL tree (best-fit).
1199 if (max_size
< metaslab_df_alloc_threshold
||
1200 free_pct
< metaslab_df_free_pct
) {
1201 t
= &msp
->ms_size_tree
;
1205 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1208 static metaslab_ops_t metaslab_df_ops
= {
1212 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1213 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1215 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1217 * ==========================================================================
1218 * Cursor fit block allocator -
1219 * Select the largest region in the metaslab, set the cursor to the beginning
1220 * of the range and the cursor_end to the end of the range. As allocations
1221 * are made advance the cursor. Continue allocating from the cursor until
1222 * the range is exhausted and then find a new range.
1223 * ==========================================================================
1226 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1228 range_tree_t
*rt
= msp
->ms_tree
;
1229 avl_tree_t
*t
= &msp
->ms_size_tree
;
1230 uint64_t *cursor
= &msp
->ms_lbas
[0];
1231 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1232 uint64_t offset
= 0;
1234 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1235 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1237 ASSERT3U(*cursor_end
, >=, *cursor
);
1239 if ((*cursor
+ size
) > *cursor_end
) {
1242 rs
= avl_last(&msp
->ms_size_tree
);
1243 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1246 *cursor
= rs
->rs_start
;
1247 *cursor_end
= rs
->rs_end
;
1256 static metaslab_ops_t metaslab_cf_ops
= {
1260 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1261 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1263 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1265 * ==========================================================================
1266 * New dynamic fit allocator -
1267 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1268 * contiguous blocks. If no region is found then just use the largest segment
1270 * ==========================================================================
1274 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1275 * to request from the allocator.
1277 uint64_t metaslab_ndf_clump_shift
= 4;
1280 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1282 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1284 range_seg_t
*rs
, rsearch
;
1285 uint64_t hbit
= highbit64(size
);
1286 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1287 uint64_t max_size
= metaslab_block_maxsize(msp
);
1289 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1290 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1292 if (max_size
< size
)
1295 rsearch
.rs_start
= *cursor
;
1296 rsearch
.rs_end
= *cursor
+ size
;
1298 rs
= avl_find(t
, &rsearch
, &where
);
1299 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1300 t
= &msp
->ms_size_tree
;
1302 rsearch
.rs_start
= 0;
1303 rsearch
.rs_end
= MIN(max_size
,
1304 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1305 rs
= avl_find(t
, &rsearch
, &where
);
1307 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1311 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1312 *cursor
= rs
->rs_start
+ size
;
1313 return (rs
->rs_start
);
1318 static metaslab_ops_t metaslab_ndf_ops
= {
1322 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1323 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1327 * ==========================================================================
1329 * ==========================================================================
1333 * Wait for any in-progress metaslab loads to complete.
1336 metaslab_load_wait(metaslab_t
*msp
)
1338 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1340 while (msp
->ms_loading
) {
1341 ASSERT(!msp
->ms_loaded
);
1342 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1347 metaslab_load(metaslab_t
*msp
)
1351 boolean_t success
= B_FALSE
;
1353 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1354 ASSERT(!msp
->ms_loaded
);
1355 ASSERT(!msp
->ms_loading
);
1357 msp
->ms_loading
= B_TRUE
;
1360 * If the space map has not been allocated yet, then treat
1361 * all the space in the metaslab as free and add it to the
1364 if (msp
->ms_sm
!= NULL
)
1365 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1367 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1369 success
= (error
== 0);
1370 msp
->ms_loading
= B_FALSE
;
1373 ASSERT3P(msp
->ms_group
, !=, NULL
);
1374 msp
->ms_loaded
= B_TRUE
;
1376 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1377 range_tree_walk(msp
->ms_defertree
[t
],
1378 range_tree_remove
, msp
->ms_tree
);
1380 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1382 cv_broadcast(&msp
->ms_load_cv
);
1387 metaslab_unload(metaslab_t
*msp
)
1389 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1390 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1391 msp
->ms_loaded
= B_FALSE
;
1392 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1393 msp
->ms_max_size
= 0;
1397 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1400 vdev_t
*vd
= mg
->mg_vd
;
1401 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1405 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1406 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1407 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1409 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1410 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1413 * We only open space map objects that already exist. All others
1414 * will be opened when we finally allocate an object for it.
1417 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1418 ms
->ms_size
, vd
->vdev_ashift
, &ms
->ms_lock
);
1421 kmem_free(ms
, sizeof (metaslab_t
));
1425 ASSERT(ms
->ms_sm
!= NULL
);
1429 * We create the main range tree here, but we don't create the
1430 * other range trees until metaslab_sync_done(). This serves
1431 * two purposes: it allows metaslab_sync_done() to detect the
1432 * addition of new space; and for debugging, it ensures that we'd
1433 * data fault on any attempt to use this metaslab before it's ready.
1435 ms
->ms_tree
= range_tree_create(&metaslab_rt_ops
, ms
, &ms
->ms_lock
);
1436 metaslab_group_add(mg
, ms
);
1438 metaslab_set_fragmentation(ms
);
1441 * If we're opening an existing pool (txg == 0) or creating
1442 * a new one (txg == TXG_INITIAL), all space is available now.
1443 * If we're adding space to an existing pool, the new space
1444 * does not become available until after this txg has synced.
1445 * The metaslab's weight will also be initialized when we sync
1446 * out this txg. This ensures that we don't attempt to allocate
1447 * from it before we have initialized it completely.
1449 if (txg
<= TXG_INITIAL
)
1450 metaslab_sync_done(ms
, 0);
1453 * If metaslab_debug_load is set and we're initializing a metaslab
1454 * that has an allocated space map object then load the its space
1455 * map so that can verify frees.
1457 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1458 mutex_enter(&ms
->ms_lock
);
1459 VERIFY0(metaslab_load(ms
));
1460 mutex_exit(&ms
->ms_lock
);
1464 vdev_dirty(vd
, 0, NULL
, txg
);
1465 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1474 metaslab_fini(metaslab_t
*msp
)
1478 metaslab_group_t
*mg
= msp
->ms_group
;
1480 metaslab_group_remove(mg
, msp
);
1482 mutex_enter(&msp
->ms_lock
);
1483 VERIFY(msp
->ms_group
== NULL
);
1484 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1486 space_map_close(msp
->ms_sm
);
1488 metaslab_unload(msp
);
1489 range_tree_destroy(msp
->ms_tree
);
1490 range_tree_destroy(msp
->ms_freeingtree
);
1491 range_tree_destroy(msp
->ms_freedtree
);
1493 for (t
= 0; t
< TXG_SIZE
; t
++) {
1494 range_tree_destroy(msp
->ms_alloctree
[t
]);
1497 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1498 range_tree_destroy(msp
->ms_defertree
[t
]);
1501 ASSERT0(msp
->ms_deferspace
);
1503 mutex_exit(&msp
->ms_lock
);
1504 cv_destroy(&msp
->ms_load_cv
);
1505 mutex_destroy(&msp
->ms_lock
);
1507 kmem_free(msp
, sizeof (metaslab_t
));
1510 #define FRAGMENTATION_TABLE_SIZE 17
1513 * This table defines a segment size based fragmentation metric that will
1514 * allow each metaslab to derive its own fragmentation value. This is done
1515 * by calculating the space in each bucket of the spacemap histogram and
1516 * multiplying that by the fragmetation metric in this table. Doing
1517 * this for all buckets and dividing it by the total amount of free
1518 * space in this metaslab (i.e. the total free space in all buckets) gives
1519 * us the fragmentation metric. This means that a high fragmentation metric
1520 * equates to most of the free space being comprised of small segments.
1521 * Conversely, if the metric is low, then most of the free space is in
1522 * large segments. A 10% change in fragmentation equates to approximately
1523 * double the number of segments.
1525 * This table defines 0% fragmented space using 16MB segments. Testing has
1526 * shown that segments that are greater than or equal to 16MB do not suffer
1527 * from drastic performance problems. Using this value, we derive the rest
1528 * of the table. Since the fragmentation value is never stored on disk, it
1529 * is possible to change these calculations in the future.
1531 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1551 * Calclate the metaslab's fragmentation metric. A return value
1552 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1553 * not support this metric. Otherwise, the return value should be in the
1557 metaslab_set_fragmentation(metaslab_t
*msp
)
1559 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1560 uint64_t fragmentation
= 0;
1562 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1563 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1566 if (!feature_enabled
) {
1567 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1572 * A null space map means that the entire metaslab is free
1573 * and thus is not fragmented.
1575 if (msp
->ms_sm
== NULL
) {
1576 msp
->ms_fragmentation
= 0;
1581 * If this metaslab's space map has not been upgraded, flag it
1582 * so that we upgrade next time we encounter it.
1584 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1585 uint64_t txg
= spa_syncing_txg(spa
);
1586 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1589 * If we've reached the final dirty txg, then we must
1590 * be shutting down the pool. We don't want to dirty
1591 * any data past this point so skip setting the condense
1592 * flag. We can retry this action the next time the pool
1595 if (spa_writeable(spa
) && txg
< spa_final_dirty_txg(spa
)) {
1596 msp
->ms_condense_wanted
= B_TRUE
;
1597 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1598 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1599 "ms_id %llu, vdev_id %llu", txg
, msp
->ms_id
,
1602 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1606 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1608 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1610 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1611 FRAGMENTATION_TABLE_SIZE
- 1);
1613 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1616 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1619 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1620 fragmentation
+= space
* zfs_frag_table
[idx
];
1624 fragmentation
/= total
;
1625 ASSERT3U(fragmentation
, <=, 100);
1627 msp
->ms_fragmentation
= fragmentation
;
1631 * Compute a weight -- a selection preference value -- for the given metaslab.
1632 * This is based on the amount of free space, the level of fragmentation,
1633 * the LBA range, and whether the metaslab is loaded.
1636 metaslab_space_weight(metaslab_t
*msp
)
1638 metaslab_group_t
*mg
= msp
->ms_group
;
1639 vdev_t
*vd
= mg
->mg_vd
;
1640 uint64_t weight
, space
;
1642 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1643 ASSERT(!vd
->vdev_removing
);
1646 * The baseline weight is the metaslab's free space.
1648 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1650 if (metaslab_fragmentation_factor_enabled
&&
1651 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1653 * Use the fragmentation information to inversely scale
1654 * down the baseline weight. We need to ensure that we
1655 * don't exclude this metaslab completely when it's 100%
1656 * fragmented. To avoid this we reduce the fragmented value
1659 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1662 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1663 * this metaslab again. The fragmentation metric may have
1664 * decreased the space to something smaller than
1665 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1666 * so that we can consume any remaining space.
1668 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1669 space
= SPA_MINBLOCKSIZE
;
1674 * Modern disks have uniform bit density and constant angular velocity.
1675 * Therefore, the outer recording zones are faster (higher bandwidth)
1676 * than the inner zones by the ratio of outer to inner track diameter,
1677 * which is typically around 2:1. We account for this by assigning
1678 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1679 * In effect, this means that we'll select the metaslab with the most
1680 * free bandwidth rather than simply the one with the most free space.
1682 if (!vd
->vdev_nonrot
&& metaslab_lba_weighting_enabled
) {
1683 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1684 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1688 * If this metaslab is one we're actively using, adjust its
1689 * weight to make it preferable to any inactive metaslab so
1690 * we'll polish it off. If the fragmentation on this metaslab
1691 * has exceed our threshold, then don't mark it active.
1693 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1694 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1695 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1698 WEIGHT_SET_SPACEBASED(weight
);
1703 * Return the weight of the specified metaslab, according to the segment-based
1704 * weighting algorithm. The metaslab must be loaded. This function can
1705 * be called within a sync pass since it relies only on the metaslab's
1706 * range tree which is always accurate when the metaslab is loaded.
1709 metaslab_weight_from_range_tree(metaslab_t
*msp
)
1711 uint64_t weight
= 0;
1712 uint32_t segments
= 0;
1715 ASSERT(msp
->ms_loaded
);
1717 for (i
= RANGE_TREE_HISTOGRAM_SIZE
- 1; i
>= SPA_MINBLOCKSHIFT
; i
--) {
1718 uint8_t shift
= msp
->ms_group
->mg_vd
->vdev_ashift
;
1719 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1722 segments
+= msp
->ms_tree
->rt_histogram
[i
];
1725 * The range tree provides more precision than the space map
1726 * and must be downgraded so that all values fit within the
1727 * space map's histogram. This allows us to compare loaded
1728 * vs. unloaded metaslabs to determine which metaslab is
1729 * considered "best".
1734 if (segments
!= 0) {
1735 WEIGHT_SET_COUNT(weight
, segments
);
1736 WEIGHT_SET_INDEX(weight
, i
);
1737 WEIGHT_SET_ACTIVE(weight
, 0);
1745 * Calculate the weight based on the on-disk histogram. This should only
1746 * be called after a sync pass has completely finished since the on-disk
1747 * information is updated in metaslab_sync().
1750 metaslab_weight_from_spacemap(metaslab_t
*msp
)
1752 uint64_t weight
= 0;
1755 for (i
= SPACE_MAP_HISTOGRAM_SIZE
- 1; i
>= 0; i
--) {
1756 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] != 0) {
1757 WEIGHT_SET_COUNT(weight
,
1758 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
1759 WEIGHT_SET_INDEX(weight
, i
+
1760 msp
->ms_sm
->sm_shift
);
1761 WEIGHT_SET_ACTIVE(weight
, 0);
1769 * Compute a segment-based weight for the specified metaslab. The weight
1770 * is determined by highest bucket in the histogram. The information
1771 * for the highest bucket is encoded into the weight value.
1774 metaslab_segment_weight(metaslab_t
*msp
)
1776 metaslab_group_t
*mg
= msp
->ms_group
;
1777 uint64_t weight
= 0;
1778 uint8_t shift
= mg
->mg_vd
->vdev_ashift
;
1780 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1783 * The metaslab is completely free.
1785 if (space_map_allocated(msp
->ms_sm
) == 0) {
1786 int idx
= highbit64(msp
->ms_size
) - 1;
1787 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1789 if (idx
< max_idx
) {
1790 WEIGHT_SET_COUNT(weight
, 1ULL);
1791 WEIGHT_SET_INDEX(weight
, idx
);
1793 WEIGHT_SET_COUNT(weight
, 1ULL << (idx
- max_idx
));
1794 WEIGHT_SET_INDEX(weight
, max_idx
);
1796 WEIGHT_SET_ACTIVE(weight
, 0);
1797 ASSERT(!WEIGHT_IS_SPACEBASED(weight
));
1802 ASSERT3U(msp
->ms_sm
->sm_dbuf
->db_size
, ==, sizeof (space_map_phys_t
));
1805 * If the metaslab is fully allocated then just make the weight 0.
1807 if (space_map_allocated(msp
->ms_sm
) == msp
->ms_size
)
1810 * If the metaslab is already loaded, then use the range tree to
1811 * determine the weight. Otherwise, we rely on the space map information
1812 * to generate the weight.
1814 if (msp
->ms_loaded
) {
1815 weight
= metaslab_weight_from_range_tree(msp
);
1817 weight
= metaslab_weight_from_spacemap(msp
);
1821 * If the metaslab was active the last time we calculated its weight
1822 * then keep it active. We want to consume the entire region that
1823 * is associated with this weight.
1825 if (msp
->ms_activation_weight
!= 0 && weight
!= 0)
1826 WEIGHT_SET_ACTIVE(weight
, WEIGHT_GET_ACTIVE(msp
->ms_weight
));
1831 * Determine if we should attempt to allocate from this metaslab. If the
1832 * metaslab has a maximum size then we can quickly determine if the desired
1833 * allocation size can be satisfied. Otherwise, if we're using segment-based
1834 * weighting then we can determine the maximum allocation that this metaslab
1835 * can accommodate based on the index encoded in the weight. If we're using
1836 * space-based weights then rely on the entire weight (excluding the weight
1840 metaslab_should_allocate(metaslab_t
*msp
, uint64_t asize
)
1842 boolean_t should_allocate
;
1844 if (msp
->ms_max_size
!= 0)
1845 return (msp
->ms_max_size
>= asize
);
1847 if (!WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
1849 * The metaslab segment weight indicates segments in the
1850 * range [2^i, 2^(i+1)), where i is the index in the weight.
1851 * Since the asize might be in the middle of the range, we
1852 * should attempt the allocation if asize < 2^(i+1).
1854 should_allocate
= (asize
<
1855 1ULL << (WEIGHT_GET_INDEX(msp
->ms_weight
) + 1));
1857 should_allocate
= (asize
<=
1858 (msp
->ms_weight
& ~METASLAB_WEIGHT_TYPE
));
1860 return (should_allocate
);
1863 metaslab_weight(metaslab_t
*msp
)
1865 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1866 spa_t
*spa
= vd
->vdev_spa
;
1869 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1872 * This vdev is in the process of being removed so there is nothing
1873 * for us to do here.
1875 if (vd
->vdev_removing
) {
1876 ASSERT0(space_map_allocated(msp
->ms_sm
));
1877 ASSERT0(vd
->vdev_ms_shift
);
1881 metaslab_set_fragmentation(msp
);
1884 * Update the maximum size if the metaslab is loaded. This will
1885 * ensure that we get an accurate maximum size if newly freed space
1886 * has been added back into the free tree.
1889 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1892 * Segment-based weighting requires space map histogram support.
1894 if (zfs_metaslab_segment_weight_enabled
&&
1895 spa_feature_is_enabled(spa
, SPA_FEATURE_SPACEMAP_HISTOGRAM
) &&
1896 (msp
->ms_sm
== NULL
|| msp
->ms_sm
->sm_dbuf
->db_size
==
1897 sizeof (space_map_phys_t
))) {
1898 weight
= metaslab_segment_weight(msp
);
1900 weight
= metaslab_space_weight(msp
);
1906 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1908 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1910 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1911 metaslab_load_wait(msp
);
1912 if (!msp
->ms_loaded
) {
1913 int error
= metaslab_load(msp
);
1915 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1920 msp
->ms_activation_weight
= msp
->ms_weight
;
1921 metaslab_group_sort(msp
->ms_group
, msp
,
1922 msp
->ms_weight
| activation_weight
);
1924 ASSERT(msp
->ms_loaded
);
1925 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1931 metaslab_passivate(metaslab_t
*msp
, uint64_t weight
)
1933 ASSERTV(uint64_t size
= weight
& ~METASLAB_WEIGHT_TYPE
);
1936 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1937 * this metaslab again. In that case, it had better be empty,
1938 * or we would be leaving space on the table.
1940 ASSERT(!WEIGHT_IS_SPACEBASED(msp
->ms_weight
) ||
1941 size
>= SPA_MINBLOCKSIZE
||
1942 range_tree_space(msp
->ms_tree
) == 0);
1943 ASSERT0(weight
& METASLAB_ACTIVE_MASK
);
1945 msp
->ms_activation_weight
= 0;
1946 metaslab_group_sort(msp
->ms_group
, msp
, weight
);
1947 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1951 * Segment-based metaslabs are activated once and remain active until
1952 * we either fail an allocation attempt (similar to space-based metaslabs)
1953 * or have exhausted the free space in zfs_metaslab_switch_threshold
1954 * buckets since the metaslab was activated. This function checks to see
1955 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
1956 * metaslab and passivates it proactively. This will allow us to select a
1957 * metaslab with a larger contiguous region, if any, remaining within this
1958 * metaslab group. If we're in sync pass > 1, then we continue using this
1959 * metaslab so that we don't dirty more block and cause more sync passes.
1962 metaslab_segment_may_passivate(metaslab_t
*msp
)
1964 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1966 int activation_idx
, current_idx
;
1968 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
) || spa_sync_pass(spa
) > 1)
1972 * Since we are in the middle of a sync pass, the most accurate
1973 * information that is accessible to us is the in-core range tree
1974 * histogram; calculate the new weight based on that information.
1976 weight
= metaslab_weight_from_range_tree(msp
);
1977 activation_idx
= WEIGHT_GET_INDEX(msp
->ms_activation_weight
);
1978 current_idx
= WEIGHT_GET_INDEX(weight
);
1980 if (current_idx
<= activation_idx
- zfs_metaslab_switch_threshold
)
1981 metaslab_passivate(msp
, weight
);
1985 metaslab_preload(void *arg
)
1987 metaslab_t
*msp
= arg
;
1988 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1989 fstrans_cookie_t cookie
= spl_fstrans_mark();
1991 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1993 mutex_enter(&msp
->ms_lock
);
1994 metaslab_load_wait(msp
);
1995 if (!msp
->ms_loaded
)
1996 (void) metaslab_load(msp
);
1997 msp
->ms_selected_txg
= spa_syncing_txg(spa
);
1998 mutex_exit(&msp
->ms_lock
);
1999 spl_fstrans_unmark(cookie
);
2003 metaslab_group_preload(metaslab_group_t
*mg
)
2005 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2007 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2010 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
2011 taskq_wait_outstanding(mg
->mg_taskq
, 0);
2015 mutex_enter(&mg
->mg_lock
);
2017 * Load the next potential metaslabs
2019 for (msp
= avl_first(t
); msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
2021 * We preload only the maximum number of metaslabs specified
2022 * by metaslab_preload_limit. If a metaslab is being forced
2023 * to condense then we preload it too. This will ensure
2024 * that force condensing happens in the next txg.
2026 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
2030 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
2031 msp
, TQ_SLEEP
) != TASKQID_INVALID
);
2033 mutex_exit(&mg
->mg_lock
);
2037 * Determine if the space map's on-disk footprint is past our tolerance
2038 * for inefficiency. We would like to use the following criteria to make
2041 * 1. The size of the space map object should not dramatically increase as a
2042 * result of writing out the free space range tree.
2044 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2045 * times the size than the free space range tree representation
2046 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
2048 * 3. The on-disk size of the space map should actually decrease.
2050 * Checking the first condition is tricky since we don't want to walk
2051 * the entire AVL tree calculating the estimated on-disk size. Instead we
2052 * use the size-ordered range tree in the metaslab and calculate the
2053 * size required to write out the largest segment in our free tree. If the
2054 * size required to represent that segment on disk is larger than the space
2055 * map object then we avoid condensing this map.
2057 * To determine the second criterion we use a best-case estimate and assume
2058 * each segment can be represented on-disk as a single 64-bit entry. We refer
2059 * to this best-case estimate as the space map's minimal form.
2061 * Unfortunately, we cannot compute the on-disk size of the space map in this
2062 * context because we cannot accurately compute the effects of compression, etc.
2063 * Instead, we apply the heuristic described in the block comment for
2064 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2065 * is greater than a threshold number of blocks.
2068 metaslab_should_condense(metaslab_t
*msp
)
2070 space_map_t
*sm
= msp
->ms_sm
;
2072 uint64_t size
, entries
, segsz
, object_size
, optimal_size
, record_size
;
2073 dmu_object_info_t doi
;
2074 uint64_t vdev_blocksize
= 1ULL << msp
->ms_group
->mg_vd
->vdev_ashift
;
2076 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2077 ASSERT(msp
->ms_loaded
);
2080 * Use the ms_size_tree range tree, which is ordered by size, to
2081 * obtain the largest segment in the free tree. We always condense
2082 * metaslabs that are empty and metaslabs for which a condense
2083 * request has been made.
2085 rs
= avl_last(&msp
->ms_size_tree
);
2086 if (rs
== NULL
|| msp
->ms_condense_wanted
)
2090 * Calculate the number of 64-bit entries this segment would
2091 * require when written to disk. If this single segment would be
2092 * larger on-disk than the entire current on-disk structure, then
2093 * clearly condensing will increase the on-disk structure size.
2095 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
2096 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
2097 segsz
= entries
* sizeof (uint64_t);
2099 optimal_size
= sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
);
2100 object_size
= space_map_length(msp
->ms_sm
);
2102 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
2103 record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
2105 return (segsz
<= object_size
&&
2106 object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
2107 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
2111 * Condense the on-disk space map representation to its minimized form.
2112 * The minimized form consists of a small number of allocations followed by
2113 * the entries of the free range tree.
2116 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
2118 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2119 range_tree_t
*condense_tree
;
2120 space_map_t
*sm
= msp
->ms_sm
;
2123 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2124 ASSERT3U(spa_sync_pass(spa
), ==, 1);
2125 ASSERT(msp
->ms_loaded
);
2128 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2129 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
2130 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
2131 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
2132 space_map_length(msp
->ms_sm
), avl_numnodes(&msp
->ms_tree
->rt_root
),
2133 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
2135 msp
->ms_condense_wanted
= B_FALSE
;
2138 * Create an range tree that is 100% allocated. We remove segments
2139 * that have been freed in this txg, any deferred frees that exist,
2140 * and any allocation in the future. Removing segments should be
2141 * a relatively inexpensive operation since we expect these trees to
2142 * have a small number of nodes.
2144 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
2145 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
2148 * Remove what's been freed in this txg from the condense_tree.
2149 * Since we're in sync_pass 1, we know that all the frees from
2150 * this txg are in the freeingtree.
2152 range_tree_walk(msp
->ms_freeingtree
, range_tree_remove
, condense_tree
);
2154 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2155 range_tree_walk(msp
->ms_defertree
[t
],
2156 range_tree_remove
, condense_tree
);
2159 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2160 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
2161 range_tree_remove
, condense_tree
);
2165 * We're about to drop the metaslab's lock thus allowing
2166 * other consumers to change it's content. Set the
2167 * metaslab's ms_condensing flag to ensure that
2168 * allocations on this metaslab do not occur while we're
2169 * in the middle of committing it to disk. This is only critical
2170 * for the ms_tree as all other range trees use per txg
2171 * views of their content.
2173 msp
->ms_condensing
= B_TRUE
;
2175 mutex_exit(&msp
->ms_lock
);
2176 space_map_truncate(sm
, tx
);
2177 mutex_enter(&msp
->ms_lock
);
2180 * While we would ideally like to create a space map representation
2181 * that consists only of allocation records, doing so can be
2182 * prohibitively expensive because the in-core free tree can be
2183 * large, and therefore computationally expensive to subtract
2184 * from the condense_tree. Instead we sync out two trees, a cheap
2185 * allocation only tree followed by the in-core free tree. While not
2186 * optimal, this is typically close to optimal, and much cheaper to
2189 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
2190 range_tree_vacate(condense_tree
, NULL
, NULL
);
2191 range_tree_destroy(condense_tree
);
2193 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
2194 msp
->ms_condensing
= B_FALSE
;
2198 * Write a metaslab to disk in the context of the specified transaction group.
2201 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
2203 metaslab_group_t
*mg
= msp
->ms_group
;
2204 vdev_t
*vd
= mg
->mg_vd
;
2205 spa_t
*spa
= vd
->vdev_spa
;
2206 objset_t
*mos
= spa_meta_objset(spa
);
2207 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
2209 uint64_t object
= space_map_object(msp
->ms_sm
);
2211 ASSERT(!vd
->vdev_ishole
);
2214 * This metaslab has just been added so there's no work to do now.
2216 if (msp
->ms_freeingtree
== NULL
) {
2217 ASSERT3P(alloctree
, ==, NULL
);
2221 ASSERT3P(alloctree
, !=, NULL
);
2222 ASSERT3P(msp
->ms_freeingtree
, !=, NULL
);
2223 ASSERT3P(msp
->ms_freedtree
, !=, NULL
);
2226 * Normally, we don't want to process a metaslab if there
2227 * are no allocations or frees to perform. However, if the metaslab
2228 * is being forced to condense and it's loaded, we need to let it
2231 if (range_tree_space(alloctree
) == 0 &&
2232 range_tree_space(msp
->ms_freeingtree
) == 0 &&
2233 !(msp
->ms_loaded
&& msp
->ms_condense_wanted
))
2237 VERIFY(txg
<= spa_final_dirty_txg(spa
));
2240 * The only state that can actually be changing concurrently with
2241 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2242 * be modifying this txg's alloctree, freeingtree, freedtree, or
2243 * space_map_phys_t. Therefore, we only hold ms_lock to satify
2244 * space map ASSERTs. We drop it whenever we call into the DMU,
2245 * because the DMU can call down to us (e.g. via zio_free()) at
2249 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
2251 if (msp
->ms_sm
== NULL
) {
2252 uint64_t new_object
;
2254 new_object
= space_map_alloc(mos
, tx
);
2255 VERIFY3U(new_object
, !=, 0);
2257 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
2258 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
2260 ASSERT(msp
->ms_sm
!= NULL
);
2263 mutex_enter(&msp
->ms_lock
);
2266 * Note: metaslab_condense() clears the space map's histogram.
2267 * Therefore we must verify and remove this histogram before
2270 metaslab_group_histogram_verify(mg
);
2271 metaslab_class_histogram_verify(mg
->mg_class
);
2272 metaslab_group_histogram_remove(mg
, msp
);
2274 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
2275 metaslab_should_condense(msp
)) {
2276 metaslab_condense(msp
, txg
, tx
);
2278 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
2279 space_map_write(msp
->ms_sm
, msp
->ms_freeingtree
, SM_FREE
, tx
);
2282 if (msp
->ms_loaded
) {
2286 * When the space map is loaded, we have an accruate
2287 * histogram in the range tree. This gives us an opportunity
2288 * to bring the space map's histogram up-to-date so we clear
2289 * it first before updating it.
2291 space_map_histogram_clear(msp
->ms_sm
);
2292 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
2295 * Since we've cleared the histogram we need to add back
2296 * any free space that has already been processed, plus
2297 * any deferred space. This allows the on-disk histogram
2298 * to accurately reflect all free space even if some space
2299 * is not yet available for allocation (i.e. deferred).
2301 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freedtree
, tx
);
2304 * Add back any deferred free space that has not been
2305 * added back into the in-core free tree yet. This will
2306 * ensure that we don't end up with a space map histogram
2307 * that is completely empty unless the metaslab is fully
2310 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2311 space_map_histogram_add(msp
->ms_sm
,
2312 msp
->ms_defertree
[t
], tx
);
2317 * Always add the free space from this sync pass to the space
2318 * map histogram. We want to make sure that the on-disk histogram
2319 * accounts for all free space. If the space map is not loaded,
2320 * then we will lose some accuracy but will correct it the next
2321 * time we load the space map.
2323 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freeingtree
, tx
);
2325 metaslab_group_histogram_add(mg
, msp
);
2326 metaslab_group_histogram_verify(mg
);
2327 metaslab_class_histogram_verify(mg
->mg_class
);
2330 * For sync pass 1, we avoid traversing this txg's free range tree
2331 * and instead will just swap the pointers for freeingtree and
2332 * freedtree. We can safely do this since the freed_tree is
2333 * guaranteed to be empty on the initial pass.
2335 if (spa_sync_pass(spa
) == 1) {
2336 range_tree_swap(&msp
->ms_freeingtree
, &msp
->ms_freedtree
);
2338 range_tree_vacate(msp
->ms_freeingtree
,
2339 range_tree_add
, msp
->ms_freedtree
);
2341 range_tree_vacate(alloctree
, NULL
, NULL
);
2343 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
2344 ASSERT0(range_tree_space(msp
->ms_alloctree
[TXG_CLEAN(txg
) & TXG_MASK
]));
2345 ASSERT0(range_tree_space(msp
->ms_freeingtree
));
2347 mutex_exit(&msp
->ms_lock
);
2349 if (object
!= space_map_object(msp
->ms_sm
)) {
2350 object
= space_map_object(msp
->ms_sm
);
2351 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
2352 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
2358 * Called after a transaction group has completely synced to mark
2359 * all of the metaslab's free space as usable.
2362 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
2364 metaslab_group_t
*mg
= msp
->ms_group
;
2365 vdev_t
*vd
= mg
->mg_vd
;
2366 spa_t
*spa
= vd
->vdev_spa
;
2367 range_tree_t
**defer_tree
;
2368 int64_t alloc_delta
, defer_delta
;
2369 uint64_t free_space
;
2370 boolean_t defer_allowed
= B_TRUE
;
2373 ASSERT(!vd
->vdev_ishole
);
2375 mutex_enter(&msp
->ms_lock
);
2378 * If this metaslab is just becoming available, initialize its
2379 * range trees and add its capacity to the vdev.
2381 if (msp
->ms_freedtree
== NULL
) {
2382 for (t
= 0; t
< TXG_SIZE
; t
++) {
2383 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
2385 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
2389 ASSERT3P(msp
->ms_freeingtree
, ==, NULL
);
2390 msp
->ms_freeingtree
= range_tree_create(NULL
, msp
,
2393 ASSERT3P(msp
->ms_freedtree
, ==, NULL
);
2394 msp
->ms_freedtree
= range_tree_create(NULL
, msp
,
2397 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2398 ASSERT(msp
->ms_defertree
[t
] == NULL
);
2400 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
2404 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
2407 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
2409 free_space
= metaslab_class_get_space(spa_normal_class(spa
)) -
2410 metaslab_class_get_alloc(spa_normal_class(spa
));
2411 if (free_space
<= spa_get_slop_space(spa
)) {
2412 defer_allowed
= B_FALSE
;
2416 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
2417 if (defer_allowed
) {
2418 defer_delta
= range_tree_space(msp
->ms_freedtree
) -
2419 range_tree_space(*defer_tree
);
2421 defer_delta
-= range_tree_space(*defer_tree
);
2424 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
2427 * If there's a metaslab_load() in progress, wait for it to complete
2428 * so that we have a consistent view of the in-core space map.
2430 metaslab_load_wait(msp
);
2433 * Move the frees from the defer_tree back to the free
2434 * range tree (if it's loaded). Swap the freed_tree and the
2435 * defer_tree -- this is safe to do because we've just emptied out
2438 range_tree_vacate(*defer_tree
,
2439 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
2440 if (defer_allowed
) {
2441 range_tree_swap(&msp
->ms_freedtree
, defer_tree
);
2443 range_tree_vacate(msp
->ms_freedtree
,
2444 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
2447 space_map_update(msp
->ms_sm
);
2449 msp
->ms_deferspace
+= defer_delta
;
2450 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2451 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2452 if (msp
->ms_deferspace
!= 0) {
2454 * Keep syncing this metaslab until all deferred frees
2455 * are back in circulation.
2457 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2461 * Calculate the new weights before unloading any metaslabs.
2462 * This will give us the most accurate weighting.
2464 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
2467 * If the metaslab is loaded and we've not tried to load or allocate
2468 * from it in 'metaslab_unload_delay' txgs, then unload it.
2470 if (msp
->ms_loaded
&&
2471 msp
->ms_selected_txg
+ metaslab_unload_delay
< txg
) {
2473 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2474 VERIFY0(range_tree_space(
2475 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
2478 if (!metaslab_debug_unload
)
2479 metaslab_unload(msp
);
2482 mutex_exit(&msp
->ms_lock
);
2486 metaslab_sync_reassess(metaslab_group_t
*mg
)
2488 metaslab_group_alloc_update(mg
);
2489 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2492 * Preload the next potential metaslabs
2494 metaslab_group_preload(mg
);
2498 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2500 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2501 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2502 uint64_t start
= msp
->ms_id
;
2504 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2505 return (1ULL << 63);
2508 return ((start
- offset
) << ms_shift
);
2510 return ((offset
- start
) << ms_shift
);
2515 * ==========================================================================
2516 * Metaslab allocation tracing facility
2517 * ==========================================================================
2519 #ifdef _METASLAB_TRACING
2520 kstat_t
*metaslab_trace_ksp
;
2521 kstat_named_t metaslab_trace_over_limit
;
2524 metaslab_alloc_trace_init(void)
2526 ASSERT(metaslab_alloc_trace_cache
== NULL
);
2527 metaslab_alloc_trace_cache
= kmem_cache_create(
2528 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t
),
2529 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
2530 metaslab_trace_ksp
= kstat_create("zfs", 0, "metaslab_trace_stats",
2531 "misc", KSTAT_TYPE_NAMED
, 1, KSTAT_FLAG_VIRTUAL
);
2532 if (metaslab_trace_ksp
!= NULL
) {
2533 metaslab_trace_ksp
->ks_data
= &metaslab_trace_over_limit
;
2534 kstat_named_init(&metaslab_trace_over_limit
,
2535 "metaslab_trace_over_limit", KSTAT_DATA_UINT64
);
2536 kstat_install(metaslab_trace_ksp
);
2541 metaslab_alloc_trace_fini(void)
2543 if (metaslab_trace_ksp
!= NULL
) {
2544 kstat_delete(metaslab_trace_ksp
);
2545 metaslab_trace_ksp
= NULL
;
2547 kmem_cache_destroy(metaslab_alloc_trace_cache
);
2548 metaslab_alloc_trace_cache
= NULL
;
2552 * Add an allocation trace element to the allocation tracing list.
2555 metaslab_trace_add(zio_alloc_list_t
*zal
, metaslab_group_t
*mg
,
2556 metaslab_t
*msp
, uint64_t psize
, uint32_t dva_id
, uint64_t offset
)
2558 metaslab_alloc_trace_t
*mat
;
2560 if (!metaslab_trace_enabled
)
2564 * When the tracing list reaches its maximum we remove
2565 * the second element in the list before adding a new one.
2566 * By removing the second element we preserve the original
2567 * entry as a clue to what allocations steps have already been
2570 if (zal
->zal_size
== metaslab_trace_max_entries
) {
2571 metaslab_alloc_trace_t
*mat_next
;
2573 panic("too many entries in allocation list");
2575 atomic_inc_64(&metaslab_trace_over_limit
.value
.ui64
);
2577 mat_next
= list_next(&zal
->zal_list
, list_head(&zal
->zal_list
));
2578 list_remove(&zal
->zal_list
, mat_next
);
2579 kmem_cache_free(metaslab_alloc_trace_cache
, mat_next
);
2582 mat
= kmem_cache_alloc(metaslab_alloc_trace_cache
, KM_SLEEP
);
2583 list_link_init(&mat
->mat_list_node
);
2586 mat
->mat_size
= psize
;
2587 mat
->mat_dva_id
= dva_id
;
2588 mat
->mat_offset
= offset
;
2589 mat
->mat_weight
= 0;
2592 mat
->mat_weight
= msp
->ms_weight
;
2595 * The list is part of the zio so locking is not required. Only
2596 * a single thread will perform allocations for a given zio.
2598 list_insert_tail(&zal
->zal_list
, mat
);
2601 ASSERT3U(zal
->zal_size
, <=, metaslab_trace_max_entries
);
2605 metaslab_trace_init(zio_alloc_list_t
*zal
)
2607 list_create(&zal
->zal_list
, sizeof (metaslab_alloc_trace_t
),
2608 offsetof(metaslab_alloc_trace_t
, mat_list_node
));
2613 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2615 metaslab_alloc_trace_t
*mat
;
2617 while ((mat
= list_remove_head(&zal
->zal_list
)) != NULL
)
2618 kmem_cache_free(metaslab_alloc_trace_cache
, mat
);
2619 list_destroy(&zal
->zal_list
);
2624 #define metaslab_trace_add(zal, mg, msp, psize, id, off)
2627 metaslab_alloc_trace_init(void)
2632 metaslab_alloc_trace_fini(void)
2637 metaslab_trace_init(zio_alloc_list_t
*zal
)
2642 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2646 #endif /* _METASLAB_TRACING */
2649 * ==========================================================================
2650 * Metaslab block operations
2651 * ==========================================================================
2655 metaslab_group_alloc_increment(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
)
2657 metaslab_group_t
*mg
;
2659 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2660 flags
& METASLAB_DONT_THROTTLE
)
2663 mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2664 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2667 (void) refcount_add(&mg
->mg_alloc_queue_depth
, tag
);
2671 metaslab_group_alloc_decrement(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
)
2673 metaslab_group_t
*mg
;
2675 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2676 flags
& METASLAB_DONT_THROTTLE
)
2679 mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2680 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2683 (void) refcount_remove(&mg
->mg_alloc_queue_depth
, tag
);
2687 metaslab_group_alloc_verify(spa_t
*spa
, const blkptr_t
*bp
, void *tag
)
2690 const dva_t
*dva
= bp
->blk_dva
;
2691 int ndvas
= BP_GET_NDVAS(bp
);
2694 for (d
= 0; d
< ndvas
; d
++) {
2695 uint64_t vdev
= DVA_GET_VDEV(&dva
[d
]);
2696 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2697 VERIFY(refcount_not_held(&mg
->mg_alloc_queue_depth
, tag
));
2703 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
, uint64_t txg
)
2706 range_tree_t
*rt
= msp
->ms_tree
;
2707 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
2709 VERIFY(!msp
->ms_condensing
);
2711 start
= mc
->mc_ops
->msop_alloc(msp
, size
);
2712 if (start
!= -1ULL) {
2713 metaslab_group_t
*mg
= msp
->ms_group
;
2714 vdev_t
*vd
= mg
->mg_vd
;
2716 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
2717 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2718 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
2719 range_tree_remove(rt
, start
, size
);
2721 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2722 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2724 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], start
, size
);
2726 /* Track the last successful allocation */
2727 msp
->ms_alloc_txg
= txg
;
2728 metaslab_verify_space(msp
, txg
);
2732 * Now that we've attempted the allocation we need to update the
2733 * metaslab's maximum block size since it may have changed.
2735 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
2740 metaslab_group_alloc_normal(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
2741 uint64_t asize
, uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2743 metaslab_t
*msp
= NULL
;
2745 uint64_t offset
= -1ULL;
2746 uint64_t activation_weight
;
2747 uint64_t target_distance
;
2750 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2751 for (i
= 0; i
< d
; i
++) {
2752 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2753 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2758 search
= kmem_alloc(sizeof (*search
), KM_SLEEP
);
2759 search
->ms_weight
= UINT64_MAX
;
2760 search
->ms_start
= 0;
2762 boolean_t was_active
;
2763 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2766 mutex_enter(&mg
->mg_lock
);
2769 * Find the metaslab with the highest weight that is less
2770 * than what we've already tried. In the common case, this
2771 * means that we will examine each metaslab at most once.
2772 * Note that concurrent callers could reorder metaslabs
2773 * by activation/passivation once we have dropped the mg_lock.
2774 * If a metaslab is activated by another thread, and we fail
2775 * to allocate from the metaslab we have selected, we may
2776 * not try the newly-activated metaslab, and instead activate
2777 * another metaslab. This is not optimal, but generally
2778 * does not cause any problems (a possible exception being
2779 * if every metaslab is completely full except for the
2780 * the newly-activated metaslab which we fail to examine).
2782 msp
= avl_find(t
, search
, &idx
);
2784 msp
= avl_nearest(t
, idx
, AVL_AFTER
);
2785 for (; msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
2787 if (!metaslab_should_allocate(msp
, asize
)) {
2788 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
2794 * If the selected metaslab is condensing, skip it.
2796 if (msp
->ms_condensing
)
2799 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2800 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2803 target_distance
= min_distance
+
2804 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2807 for (i
= 0; i
< d
; i
++) {
2808 if (metaslab_distance(msp
, &dva
[i
]) <
2815 mutex_exit(&mg
->mg_lock
);
2817 kmem_free(search
, sizeof (*search
));
2820 search
->ms_weight
= msp
->ms_weight
;
2821 search
->ms_start
= msp
->ms_start
+ 1;
2823 mutex_enter(&msp
->ms_lock
);
2826 * Ensure that the metaslab we have selected is still
2827 * capable of handling our request. It's possible that
2828 * another thread may have changed the weight while we
2829 * were blocked on the metaslab lock. We check the
2830 * active status first to see if we need to reselect
2833 if (was_active
&& !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
)) {
2834 mutex_exit(&msp
->ms_lock
);
2838 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2839 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2840 metaslab_passivate(msp
,
2841 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2842 mutex_exit(&msp
->ms_lock
);
2846 if (metaslab_activate(msp
, activation_weight
) != 0) {
2847 mutex_exit(&msp
->ms_lock
);
2850 msp
->ms_selected_txg
= txg
;
2853 * Now that we have the lock, recheck to see if we should
2854 * continue to use this metaslab for this allocation. The
2855 * the metaslab is now loaded so metaslab_should_allocate() can
2856 * accurately determine if the allocation attempt should
2859 if (!metaslab_should_allocate(msp
, asize
)) {
2860 /* Passivate this metaslab and select a new one. */
2861 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
2868 * If this metaslab is currently condensing then pick again as
2869 * we can't manipulate this metaslab until it's committed
2872 if (msp
->ms_condensing
) {
2873 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
2875 mutex_exit(&msp
->ms_lock
);
2879 offset
= metaslab_block_alloc(msp
, asize
, txg
);
2880 metaslab_trace_add(zal
, mg
, msp
, asize
, d
, offset
);
2882 if (offset
!= -1ULL) {
2883 /* Proactively passivate the metaslab, if needed */
2884 metaslab_segment_may_passivate(msp
);
2888 ASSERT(msp
->ms_loaded
);
2891 * We were unable to allocate from this metaslab so determine
2892 * a new weight for this metaslab. Now that we have loaded
2893 * the metaslab we can provide a better hint to the metaslab
2896 * For space-based metaslabs, we use the maximum block size.
2897 * This information is only available when the metaslab
2898 * is loaded and is more accurate than the generic free
2899 * space weight that was calculated by metaslab_weight().
2900 * This information allows us to quickly compare the maximum
2901 * available allocation in the metaslab to the allocation
2902 * size being requested.
2904 * For segment-based metaslabs, determine the new weight
2905 * based on the highest bucket in the range tree. We
2906 * explicitly use the loaded segment weight (i.e. the range
2907 * tree histogram) since it contains the space that is
2908 * currently available for allocation and is accurate
2909 * even within a sync pass.
2911 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
2912 uint64_t weight
= metaslab_block_maxsize(msp
);
2913 WEIGHT_SET_SPACEBASED(weight
);
2914 metaslab_passivate(msp
, weight
);
2916 metaslab_passivate(msp
,
2917 metaslab_weight_from_range_tree(msp
));
2921 * We have just failed an allocation attempt, check
2922 * that metaslab_should_allocate() agrees. Otherwise,
2923 * we may end up in an infinite loop retrying the same
2926 ASSERT(!metaslab_should_allocate(msp
, asize
));
2927 mutex_exit(&msp
->ms_lock
);
2929 mutex_exit(&msp
->ms_lock
);
2930 kmem_free(search
, sizeof (*search
));
2935 metaslab_group_alloc(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
2936 uint64_t asize
, uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2939 ASSERT(mg
->mg_initialized
);
2941 offset
= metaslab_group_alloc_normal(mg
, zal
, asize
, txg
,
2942 min_distance
, dva
, d
);
2944 mutex_enter(&mg
->mg_lock
);
2945 if (offset
== -1ULL) {
2946 mg
->mg_failed_allocations
++;
2947 metaslab_trace_add(zal
, mg
, NULL
, asize
, d
,
2948 TRACE_GROUP_FAILURE
);
2949 if (asize
== SPA_GANGBLOCKSIZE
) {
2951 * This metaslab group was unable to allocate
2952 * the minimum gang block size so it must be out of
2953 * space. We must notify the allocation throttle
2954 * to start skipping allocation attempts to this
2955 * metaslab group until more space becomes available.
2956 * Note: this failure cannot be caused by the
2957 * allocation throttle since the allocation throttle
2958 * is only responsible for skipping devices and
2959 * not failing block allocations.
2961 mg
->mg_no_free_space
= B_TRUE
;
2964 mg
->mg_allocations
++;
2965 mutex_exit(&mg
->mg_lock
);
2970 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
2971 * on the same vdev as an existing DVA of this BP, then try to allocate it
2972 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
2975 int ditto_same_vdev_distance_shift
= 3;
2978 * Allocate a block for the specified i/o.
2981 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2982 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
,
2983 zio_alloc_list_t
*zal
)
2985 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
2987 boolean_t try_hard
= B_FALSE
;
2989 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2992 * For testing, make some blocks above a certain size be gang blocks.
2994 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0) {
2995 metaslab_trace_add(zal
, NULL
, NULL
, psize
, d
, TRACE_FORCE_GANG
);
2996 return (SET_ERROR(ENOSPC
));
3000 * Start at the rotor and loop through all mgs until we find something.
3001 * Note that there's no locking on mc_rotor or mc_aliquot because
3002 * nothing actually breaks if we miss a few updates -- we just won't
3003 * allocate quite as evenly. It all balances out over time.
3005 * If we are doing ditto or log blocks, try to spread them across
3006 * consecutive vdevs. If we're forced to reuse a vdev before we've
3007 * allocated all of our ditto blocks, then try and spread them out on
3008 * that vdev as much as possible. If it turns out to not be possible,
3009 * gradually lower our standards until anything becomes acceptable.
3010 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3011 * gives us hope of containing our fault domains to something we're
3012 * able to reason about. Otherwise, any two top-level vdev failures
3013 * will guarantee the loss of data. With consecutive allocation,
3014 * only two adjacent top-level vdev failures will result in data loss.
3016 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3017 * ourselves on the same vdev as our gang block header. That
3018 * way, we can hope for locality in vdev_cache, plus it makes our
3019 * fault domains something tractable.
3022 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
3025 * It's possible the vdev we're using as the hint no
3026 * longer exists (i.e. removed). Consult the rotor when
3032 if (flags
& METASLAB_HINTBP_AVOID
&&
3033 mg
->mg_next
!= NULL
)
3038 } else if (d
!= 0) {
3039 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
3040 mg
= vd
->vdev_mg
->mg_next
;
3041 } else if (flags
& METASLAB_FASTWRITE
) {
3042 mg
= fast_mg
= mc
->mc_rotor
;
3045 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
3046 mg
->mg_vd
->vdev_pending_fastwrite
)
3048 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
3055 * If the hint put us into the wrong metaslab class, or into a
3056 * metaslab group that has been passivated, just follow the rotor.
3058 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
3064 boolean_t allocatable
;
3066 uint64_t distance
, asize
;
3068 ASSERT(mg
->mg_activation_count
== 1);
3072 * Don't allocate from faulted devices.
3075 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
3076 allocatable
= vdev_allocatable(vd
);
3077 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
3079 allocatable
= vdev_allocatable(vd
);
3083 * Determine if the selected metaslab group is eligible
3084 * for allocations. If we're ganging then don't allow
3085 * this metaslab group to skip allocations since that would
3086 * inadvertently return ENOSPC and suspend the pool
3087 * even though space is still available.
3089 if (allocatable
&& !GANG_ALLOCATION(flags
) && !try_hard
) {
3090 allocatable
= metaslab_group_allocatable(mg
, rotor
,
3095 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3096 TRACE_NOT_ALLOCATABLE
);
3100 ASSERT(mg
->mg_initialized
);
3103 * Avoid writing single-copy data to a failing,
3104 * non-redundant vdev, unless we've already tried all
3107 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
3108 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
3109 d
== 0 && !try_hard
&& vd
->vdev_children
== 0) {
3110 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3115 ASSERT(mg
->mg_class
== mc
);
3118 * If we don't need to try hard, then require that the
3119 * block be 1/8th of the device away from any other DVAs
3120 * in this BP. If we are trying hard, allow any offset
3121 * to be used (distance=0).
3125 distance
= vd
->vdev_asize
>>
3126 ditto_same_vdev_distance_shift
;
3127 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
3131 asize
= vdev_psize_to_asize(vd
, psize
);
3132 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
3134 offset
= metaslab_group_alloc(mg
, zal
, asize
, txg
, distance
,
3137 if (offset
!= -1ULL) {
3139 * If we've just selected this metaslab group,
3140 * figure out whether the corresponding vdev is
3141 * over- or under-used relative to the pool,
3142 * and set an allocation bias to even it out.
3144 * Bias is also used to compensate for unequally
3145 * sized vdevs so that space is allocated fairly.
3147 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
3148 vdev_stat_t
*vs
= &vd
->vdev_stat
;
3149 int64_t vs_free
= vs
->vs_space
- vs
->vs_alloc
;
3150 int64_t mc_free
= mc
->mc_space
- mc
->mc_alloc
;
3154 * Calculate how much more or less we should
3155 * try to allocate from this device during
3156 * this iteration around the rotor.
3158 * This basically introduces a zero-centered
3159 * bias towards the devices with the most
3160 * free space, while compensating for vdev
3164 * vdev V1 = 16M/128M
3165 * vdev V2 = 16M/128M
3166 * ratio(V1) = 100% ratio(V2) = 100%
3168 * vdev V1 = 16M/128M
3169 * vdev V2 = 64M/128M
3170 * ratio(V1) = 127% ratio(V2) = 72%
3172 * vdev V1 = 16M/128M
3173 * vdev V2 = 64M/512M
3174 * ratio(V1) = 40% ratio(V2) = 160%
3176 ratio
= (vs_free
* mc
->mc_alloc_groups
* 100) /
3178 mg
->mg_bias
= ((ratio
- 100) *
3179 (int64_t)mg
->mg_aliquot
) / 100;
3180 } else if (!metaslab_bias_enabled
) {
3184 if ((flags
& METASLAB_FASTWRITE
) ||
3185 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
3186 mg
->mg_aliquot
+ mg
->mg_bias
) {
3187 mc
->mc_rotor
= mg
->mg_next
;
3191 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
3192 DVA_SET_OFFSET(&dva
[d
], offset
);
3193 DVA_SET_GANG(&dva
[d
],
3194 ((flags
& METASLAB_GANG_HEADER
) ? 1 : 0));
3195 DVA_SET_ASIZE(&dva
[d
], asize
);
3197 if (flags
& METASLAB_FASTWRITE
) {
3198 atomic_add_64(&vd
->vdev_pending_fastwrite
,
3205 mc
->mc_rotor
= mg
->mg_next
;
3207 } while ((mg
= mg
->mg_next
) != rotor
);
3210 * If we haven't tried hard, do so now.
3217 bzero(&dva
[d
], sizeof (dva_t
));
3219 metaslab_trace_add(zal
, rotor
, NULL
, psize
, d
, TRACE_ENOSPC
);
3220 return (SET_ERROR(ENOSPC
));
3224 * Free the block represented by DVA in the context of the specified
3225 * transaction group.
3228 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
3230 uint64_t vdev
= DVA_GET_VDEV(dva
);
3231 uint64_t offset
= DVA_GET_OFFSET(dva
);
3232 uint64_t size
= DVA_GET_ASIZE(dva
);
3236 if (txg
> spa_freeze_txg(spa
))
3239 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
3240 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
3241 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
3242 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
3243 (u_longlong_t
)size
);
3247 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3249 if (DVA_GET_GANG(dva
))
3250 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3252 mutex_enter(&msp
->ms_lock
);
3255 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
3258 VERIFY(!msp
->ms_condensing
);
3259 VERIFY3U(offset
, >=, msp
->ms_start
);
3260 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
3261 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
3263 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3264 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3265 range_tree_add(msp
->ms_tree
, offset
, size
);
3266 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
3268 VERIFY3U(txg
, ==, spa
->spa_syncing_txg
);
3269 if (range_tree_space(msp
->ms_freeingtree
) == 0)
3270 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
3271 range_tree_add(msp
->ms_freeingtree
, offset
, size
);
3274 mutex_exit(&msp
->ms_lock
);
3278 * Intent log support: upon opening the pool after a crash, notify the SPA
3279 * of blocks that the intent log has allocated for immediate write, but
3280 * which are still considered free by the SPA because the last transaction
3281 * group didn't commit yet.
3284 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
3286 uint64_t vdev
= DVA_GET_VDEV(dva
);
3287 uint64_t offset
= DVA_GET_OFFSET(dva
);
3288 uint64_t size
= DVA_GET_ASIZE(dva
);
3293 ASSERT(DVA_IS_VALID(dva
));
3295 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
3296 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
3297 return (SET_ERROR(ENXIO
));
3299 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3301 if (DVA_GET_GANG(dva
))
3302 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3304 mutex_enter(&msp
->ms_lock
);
3306 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
3307 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
3309 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
3310 error
= SET_ERROR(ENOENT
);
3312 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
3313 mutex_exit(&msp
->ms_lock
);
3317 VERIFY(!msp
->ms_condensing
);
3318 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3319 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3320 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
3321 range_tree_remove(msp
->ms_tree
, offset
, size
);
3323 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
3324 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
3325 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
3326 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
3329 mutex_exit(&msp
->ms_lock
);
3335 * Reserve some allocation slots. The reservation system must be called
3336 * before we call into the allocator. If there aren't any available slots
3337 * then the I/O will be throttled until an I/O completes and its slots are
3338 * freed up. The function returns true if it was successful in placing
3342 metaslab_class_throttle_reserve(metaslab_class_t
*mc
, int slots
, zio_t
*zio
,
3345 uint64_t available_slots
= 0;
3346 uint64_t reserved_slots
;
3347 boolean_t slot_reserved
= B_FALSE
;
3349 ASSERT(mc
->mc_alloc_throttle_enabled
);
3350 mutex_enter(&mc
->mc_lock
);
3352 reserved_slots
= refcount_count(&mc
->mc_alloc_slots
);
3353 if (reserved_slots
< mc
->mc_alloc_max_slots
)
3354 available_slots
= mc
->mc_alloc_max_slots
- reserved_slots
;
3356 if (slots
<= available_slots
|| GANG_ALLOCATION(flags
)) {
3360 * We reserve the slots individually so that we can unreserve
3361 * them individually when an I/O completes.
3363 for (d
= 0; d
< slots
; d
++) {
3364 reserved_slots
= refcount_add(&mc
->mc_alloc_slots
, zio
);
3366 zio
->io_flags
|= ZIO_FLAG_IO_ALLOCATING
;
3367 slot_reserved
= B_TRUE
;
3370 mutex_exit(&mc
->mc_lock
);
3371 return (slot_reserved
);
3375 metaslab_class_throttle_unreserve(metaslab_class_t
*mc
, int slots
, zio_t
*zio
)
3379 ASSERT(mc
->mc_alloc_throttle_enabled
);
3380 mutex_enter(&mc
->mc_lock
);
3381 for (d
= 0; d
< slots
; d
++) {
3382 (void) refcount_remove(&mc
->mc_alloc_slots
, zio
);
3384 mutex_exit(&mc
->mc_lock
);
3388 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
3389 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
,
3390 zio_alloc_list_t
*zal
, zio_t
*zio
)
3392 dva_t
*dva
= bp
->blk_dva
;
3393 dva_t
*hintdva
= hintbp
->blk_dva
;
3396 ASSERT(bp
->blk_birth
== 0);
3397 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
3399 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
3401 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
3402 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3403 return (SET_ERROR(ENOSPC
));
3406 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
3407 ASSERT(BP_GET_NDVAS(bp
) == 0);
3408 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
3409 ASSERT3P(zal
, !=, NULL
);
3411 for (d
= 0; d
< ndvas
; d
++) {
3412 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
3415 for (d
--; d
>= 0; d
--) {
3416 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
3417 metaslab_group_alloc_decrement(spa
,
3418 DVA_GET_VDEV(&dva
[d
]), zio
, flags
);
3419 bzero(&dva
[d
], sizeof (dva_t
));
3421 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3425 * Update the metaslab group's queue depth
3426 * based on the newly allocated dva.
3428 metaslab_group_alloc_increment(spa
,
3429 DVA_GET_VDEV(&dva
[d
]), zio
, flags
);
3434 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
3436 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3438 BP_SET_BIRTH(bp
, txg
, 0);
3444 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
3446 const dva_t
*dva
= bp
->blk_dva
;
3447 int d
, ndvas
= BP_GET_NDVAS(bp
);
3449 ASSERT(!BP_IS_HOLE(bp
));
3450 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
3452 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
3454 for (d
= 0; d
< ndvas
; d
++)
3455 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
3457 spa_config_exit(spa
, SCL_FREE
, FTAG
);
3461 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
3463 const dva_t
*dva
= bp
->blk_dva
;
3464 int ndvas
= BP_GET_NDVAS(bp
);
3467 ASSERT(!BP_IS_HOLE(bp
));
3471 * First do a dry run to make sure all DVAs are claimable,
3472 * so we don't have to unwind from partial failures below.
3474 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
3478 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
3480 for (d
= 0; d
< ndvas
; d
++)
3481 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
3484 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
3486 ASSERT(error
== 0 || txg
== 0);
3492 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
3494 const dva_t
*dva
= bp
->blk_dva
;
3495 int ndvas
= BP_GET_NDVAS(bp
);
3496 uint64_t psize
= BP_GET_PSIZE(bp
);
3500 ASSERT(!BP_IS_HOLE(bp
));
3501 ASSERT(!BP_IS_EMBEDDED(bp
));
3504 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
3506 for (d
= 0; d
< ndvas
; d
++) {
3507 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
3509 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
3512 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
3516 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
3518 const dva_t
*dva
= bp
->blk_dva
;
3519 int ndvas
= BP_GET_NDVAS(bp
);
3520 uint64_t psize
= BP_GET_PSIZE(bp
);
3524 ASSERT(!BP_IS_HOLE(bp
));
3525 ASSERT(!BP_IS_EMBEDDED(bp
));
3528 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
3530 for (d
= 0; d
< ndvas
; d
++) {
3531 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
3533 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
3534 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
3537 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
3541 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
3545 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
3548 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
3549 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
3550 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
3551 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
3552 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
3553 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
3554 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3557 range_tree_verify(msp
->ms_tree
, offset
, size
);
3559 range_tree_verify(msp
->ms_freeingtree
, offset
, size
);
3560 range_tree_verify(msp
->ms_freedtree
, offset
, size
);
3561 for (j
= 0; j
< TXG_DEFER_SIZE
; j
++)
3562 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
3564 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
3567 #if defined(_KERNEL) && defined(HAVE_SPL)
3569 module_param(metaslab_aliquot
, ulong
, 0644);
3570 MODULE_PARM_DESC(metaslab_aliquot
,
3571 "allocation granularity (a.k.a. stripe size)");
3573 module_param(metaslab_debug_load
, int, 0644);
3574 MODULE_PARM_DESC(metaslab_debug_load
,
3575 "load all metaslabs when pool is first opened");
3577 module_param(metaslab_debug_unload
, int, 0644);
3578 MODULE_PARM_DESC(metaslab_debug_unload
,
3579 "prevent metaslabs from being unloaded");
3581 module_param(metaslab_preload_enabled
, int, 0644);
3582 MODULE_PARM_DESC(metaslab_preload_enabled
,
3583 "preload potential metaslabs during reassessment");
3585 module_param(zfs_mg_noalloc_threshold
, int, 0644);
3586 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
3587 "percentage of free space for metaslab group to allow allocation");
3589 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
3590 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
3591 "fragmentation for metaslab group to allow allocation");
3593 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
3594 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
3595 "fragmentation for metaslab to allow allocation");
3597 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
3598 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
3599 "use the fragmentation metric to prefer less fragmented metaslabs");
3601 module_param(metaslab_lba_weighting_enabled
, int, 0644);
3602 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
3603 "prefer metaslabs with lower LBAs");
3605 module_param(metaslab_bias_enabled
, int, 0644);
3606 MODULE_PARM_DESC(metaslab_bias_enabled
,
3607 "enable metaslab group biasing");
3609 module_param(zfs_metaslab_segment_weight_enabled
, int, 0644);
3610 MODULE_PARM_DESC(zfs_metaslab_segment_weight_enabled
,
3611 "enable segment-based metaslab selection");
3613 module_param(zfs_metaslab_switch_threshold
, int, 0644);
3614 MODULE_PARM_DESC(zfs_metaslab_switch_threshold
,
3615 "segment-based metaslab selection maximum buckets before switching");
3616 #endif /* _KERNEL && HAVE_SPL */