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, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
28 #include <sys/zfs_context.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
37 #include <sys/vdev_indirect_mapping.h>
40 #define WITH_DF_BLOCK_ALLOCATOR
42 #define GANG_ALLOCATION(flags) \
43 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
46 * Metaslab granularity, in bytes. This is roughly similar to what would be
47 * referred to as the "stripe size" in traditional RAID arrays. In normal
48 * operation, we will try to write this amount of data to a top-level vdev
49 * before moving on to the next one.
51 unsigned long metaslab_aliquot
= 512 << 10;
54 * For testing, make some blocks above a certain size be gang blocks.
56 unsigned long metaslab_force_ganging
= SPA_MAXBLOCKSIZE
+ 1;
59 * Since we can touch multiple metaslabs (and their respective space maps)
60 * with each transaction group, we benefit from having a smaller space map
61 * block size since it allows us to issue more I/O operations scattered
64 int zfs_metaslab_sm_blksz
= (1 << 12);
67 * The in-core space map representation is more compact than its on-disk form.
68 * The zfs_condense_pct determines how much more compact the in-core
69 * space map representation must be before we compact it on-disk.
70 * Values should be greater than or equal to 100.
72 int zfs_condense_pct
= 200;
75 * Condensing a metaslab is not guaranteed to actually reduce the amount of
76 * space used on disk. In particular, a space map uses data in increments of
77 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
78 * same number of blocks after condensing. Since the goal of condensing is to
79 * reduce the number of IOPs required to read the space map, we only want to
80 * condense when we can be sure we will reduce the number of blocks used by the
81 * space map. Unfortunately, we cannot precisely compute whether or not this is
82 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
83 * we apply the following heuristic: do not condense a spacemap unless the
84 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
87 int zfs_metaslab_condense_block_threshold
= 4;
90 * The zfs_mg_noalloc_threshold defines which metaslab groups should
91 * be eligible for allocation. The value is defined as a percentage of
92 * free space. Metaslab groups that have more free space than
93 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
94 * a metaslab group's free space is less than or equal to the
95 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
96 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
97 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
98 * groups are allowed to accept allocations. Gang blocks are always
99 * eligible to allocate on any metaslab group. The default value of 0 means
100 * no metaslab group will be excluded based on this criterion.
102 int zfs_mg_noalloc_threshold
= 0;
105 * Metaslab groups are considered eligible for allocations if their
106 * fragmenation metric (measured as a percentage) is less than or equal to
107 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
108 * then it will be skipped unless all metaslab groups within the metaslab
109 * class have also crossed this threshold.
111 int zfs_mg_fragmentation_threshold
= 85;
114 * Allow metaslabs to keep their active state as long as their fragmentation
115 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
116 * active metaslab that exceeds this threshold will no longer keep its active
117 * status allowing better metaslabs to be selected.
119 int zfs_metaslab_fragmentation_threshold
= 70;
122 * When set will load all metaslabs when pool is first opened.
124 int metaslab_debug_load
= 0;
127 * When set will prevent metaslabs from being unloaded.
129 int metaslab_debug_unload
= 0;
132 * Minimum size which forces the dynamic allocator to change
133 * it's allocation strategy. Once the space map cannot satisfy
134 * an allocation of this size then it switches to using more
135 * aggressive strategy (i.e search by size rather than offset).
137 uint64_t metaslab_df_alloc_threshold
= SPA_OLD_MAXBLOCKSIZE
;
140 * The minimum free space, in percent, which must be available
141 * in a space map to continue allocations in a first-fit fashion.
142 * Once the space map's free space drops below this level we dynamically
143 * switch to using best-fit allocations.
145 int metaslab_df_free_pct
= 4;
148 * Percentage of all cpus that can be used by the metaslab taskq.
150 int metaslab_load_pct
= 50;
153 * Determines how many txgs a metaslab may remain loaded without having any
154 * allocations from it. As long as a metaslab continues to be used we will
157 int metaslab_unload_delay
= TXG_SIZE
* 2;
160 * Max number of metaslabs per group to preload.
162 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
165 * Enable/disable preloading of metaslab.
167 int metaslab_preload_enabled
= B_TRUE
;
170 * Enable/disable fragmentation weighting on metaslabs.
172 int metaslab_fragmentation_factor_enabled
= B_TRUE
;
175 * Enable/disable lba weighting (i.e. outer tracks are given preference).
177 int metaslab_lba_weighting_enabled
= B_TRUE
;
180 * Enable/disable metaslab group biasing.
182 int metaslab_bias_enabled
= B_TRUE
;
186 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
188 boolean_t zfs_remap_blkptr_enable
= B_TRUE
;
191 * Enable/disable segment-based metaslab selection.
193 int zfs_metaslab_segment_weight_enabled
= B_TRUE
;
196 * When using segment-based metaslab selection, we will continue
197 * allocating from the active metaslab until we have exhausted
198 * zfs_metaslab_switch_threshold of its buckets.
200 int zfs_metaslab_switch_threshold
= 2;
203 * Internal switch to enable/disable the metaslab allocation tracing
206 #ifdef _METASLAB_TRACING
207 boolean_t metaslab_trace_enabled
= B_TRUE
;
211 * Maximum entries that the metaslab allocation tracing facility will keep
212 * in a given list when running in non-debug mode. We limit the number
213 * of entries in non-debug mode to prevent us from using up too much memory.
214 * The limit should be sufficiently large that we don't expect any allocation
215 * to every exceed this value. In debug mode, the system will panic if this
216 * limit is ever reached allowing for further investigation.
218 #ifdef _METASLAB_TRACING
219 uint64_t metaslab_trace_max_entries
= 5000;
222 static uint64_t metaslab_weight(metaslab_t
*);
223 static void metaslab_set_fragmentation(metaslab_t
*);
224 static void metaslab_free_impl(vdev_t
*, uint64_t, uint64_t, boolean_t
);
225 static void metaslab_check_free_impl(vdev_t
*, uint64_t, uint64_t);
227 static void metaslab_passivate(metaslab_t
*msp
, uint64_t weight
);
228 static uint64_t metaslab_weight_from_range_tree(metaslab_t
*msp
);
229 #ifdef _METASLAB_TRACING
230 kmem_cache_t
*metaslab_alloc_trace_cache
;
234 * ==========================================================================
236 * ==========================================================================
239 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
241 metaslab_class_t
*mc
;
243 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
248 mutex_init(&mc
->mc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
249 mc
->mc_alloc_slots
= kmem_zalloc(spa
->spa_alloc_count
*
250 sizeof (zfs_refcount_t
), KM_SLEEP
);
251 mc
->mc_alloc_max_slots
= kmem_zalloc(spa
->spa_alloc_count
*
252 sizeof (uint64_t), KM_SLEEP
);
253 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++)
254 zfs_refcount_create_tracked(&mc
->mc_alloc_slots
[i
]);
260 metaslab_class_destroy(metaslab_class_t
*mc
)
262 ASSERT(mc
->mc_rotor
== NULL
);
263 ASSERT(mc
->mc_alloc
== 0);
264 ASSERT(mc
->mc_deferred
== 0);
265 ASSERT(mc
->mc_space
== 0);
266 ASSERT(mc
->mc_dspace
== 0);
268 for (int i
= 0; i
< mc
->mc_spa
->spa_alloc_count
; i
++)
269 zfs_refcount_destroy(&mc
->mc_alloc_slots
[i
]);
270 kmem_free(mc
->mc_alloc_slots
, mc
->mc_spa
->spa_alloc_count
*
271 sizeof (zfs_refcount_t
));
272 kmem_free(mc
->mc_alloc_max_slots
, mc
->mc_spa
->spa_alloc_count
*
274 mutex_destroy(&mc
->mc_lock
);
275 kmem_free(mc
, sizeof (metaslab_class_t
));
279 metaslab_class_validate(metaslab_class_t
*mc
)
281 metaslab_group_t
*mg
;
285 * Must hold one of the spa_config locks.
287 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
288 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
290 if ((mg
= mc
->mc_rotor
) == NULL
)
295 ASSERT(vd
->vdev_mg
!= NULL
);
296 ASSERT3P(vd
->vdev_top
, ==, vd
);
297 ASSERT3P(mg
->mg_class
, ==, mc
);
298 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
299 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
305 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
306 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
308 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
309 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
310 atomic_add_64(&mc
->mc_space
, space_delta
);
311 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
315 metaslab_class_get_alloc(metaslab_class_t
*mc
)
317 return (mc
->mc_alloc
);
321 metaslab_class_get_deferred(metaslab_class_t
*mc
)
323 return (mc
->mc_deferred
);
327 metaslab_class_get_space(metaslab_class_t
*mc
)
329 return (mc
->mc_space
);
333 metaslab_class_get_dspace(metaslab_class_t
*mc
)
335 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
339 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
341 spa_t
*spa
= mc
->mc_spa
;
342 vdev_t
*rvd
= spa
->spa_root_vdev
;
346 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
349 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
352 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
353 vdev_t
*tvd
= rvd
->vdev_child
[c
];
354 metaslab_group_t
*mg
= tvd
->vdev_mg
;
357 * Skip any holes, uninitialized top-levels, or
358 * vdevs that are not in this metalab class.
360 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
361 mg
->mg_class
!= mc
) {
365 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
366 mc_hist
[i
] += mg
->mg_histogram
[i
];
369 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
370 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
372 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
376 * Calculate the metaslab class's fragmentation metric. The metric
377 * is weighted based on the space contribution of each metaslab group.
378 * The return value will be a number between 0 and 100 (inclusive), or
379 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
380 * zfs_frag_table for more information about the metric.
383 metaslab_class_fragmentation(metaslab_class_t
*mc
)
385 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
386 uint64_t fragmentation
= 0;
388 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
390 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
391 vdev_t
*tvd
= rvd
->vdev_child
[c
];
392 metaslab_group_t
*mg
= tvd
->vdev_mg
;
395 * Skip any holes, uninitialized top-levels,
396 * or vdevs that are not in this metalab class.
398 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
399 mg
->mg_class
!= mc
) {
404 * If a metaslab group does not contain a fragmentation
405 * metric then just bail out.
407 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
408 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
409 return (ZFS_FRAG_INVALID
);
413 * Determine how much this metaslab_group is contributing
414 * to the overall pool fragmentation metric.
416 fragmentation
+= mg
->mg_fragmentation
*
417 metaslab_group_get_space(mg
);
419 fragmentation
/= metaslab_class_get_space(mc
);
421 ASSERT3U(fragmentation
, <=, 100);
422 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
423 return (fragmentation
);
427 * Calculate the amount of expandable space that is available in
428 * this metaslab class. If a device is expanded then its expandable
429 * space will be the amount of allocatable space that is currently not
430 * part of this metaslab class.
433 metaslab_class_expandable_space(metaslab_class_t
*mc
)
435 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
438 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
439 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
440 vdev_t
*tvd
= rvd
->vdev_child
[c
];
441 metaslab_group_t
*mg
= tvd
->vdev_mg
;
443 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
444 mg
->mg_class
!= mc
) {
449 * Calculate if we have enough space to add additional
450 * metaslabs. We report the expandable space in terms
451 * of the metaslab size since that's the unit of expansion.
453 space
+= P2ALIGN(tvd
->vdev_max_asize
- tvd
->vdev_asize
,
454 1ULL << tvd
->vdev_ms_shift
);
456 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
461 metaslab_compare(const void *x1
, const void *x2
)
463 const metaslab_t
*m1
= (const metaslab_t
*)x1
;
464 const metaslab_t
*m2
= (const metaslab_t
*)x2
;
468 if (m1
->ms_allocator
!= -1 && m1
->ms_primary
)
470 else if (m1
->ms_allocator
!= -1 && !m1
->ms_primary
)
472 if (m2
->ms_allocator
!= -1 && m2
->ms_primary
)
474 else if (m2
->ms_allocator
!= -1 && !m2
->ms_primary
)
478 * Sort inactive metaslabs first, then primaries, then secondaries. When
479 * selecting a metaslab to allocate from, an allocator first tries its
480 * primary, then secondary active metaslab. If it doesn't have active
481 * metaslabs, or can't allocate from them, it searches for an inactive
482 * metaslab to activate. If it can't find a suitable one, it will steal
483 * a primary or secondary metaslab from another allocator.
490 int cmp
= AVL_CMP(m2
->ms_weight
, m1
->ms_weight
);
494 IMPLY(AVL_CMP(m1
->ms_start
, m2
->ms_start
) == 0, m1
== m2
);
496 return (AVL_CMP(m1
->ms_start
, m2
->ms_start
));
500 metaslab_allocated_space(metaslab_t
*msp
)
502 return (msp
->ms_allocated_space
);
506 * Verify that the space accounting on disk matches the in-core range_trees.
509 metaslab_verify_space(metaslab_t
*msp
, uint64_t txg
)
511 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
512 uint64_t allocating
= 0;
513 uint64_t sm_free_space
, msp_free_space
;
515 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
516 ASSERT(!msp
->ms_condensing
);
518 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
522 * We can only verify the metaslab space when we're called
523 * from syncing context with a loaded metaslab that has an
524 * allocated space map. Calling this in non-syncing context
525 * does not provide a consistent view of the metaslab since
526 * we're performing allocations in the future.
528 if (txg
!= spa_syncing_txg(spa
) || msp
->ms_sm
== NULL
||
533 * Even though the smp_alloc field can get negative (e.g.
534 * see vdev_checkpoint_sm), that should never be the case
535 * when it come's to a metaslab's space map.
537 ASSERT3S(space_map_allocated(msp
->ms_sm
), >=, 0);
539 sm_free_space
= msp
->ms_size
- metaslab_allocated_space(msp
);
542 * Account for future allocations since we would have
543 * already deducted that space from the ms_allocatable.
545 for (int t
= 0; t
< TXG_CONCURRENT_STATES
; t
++) {
547 range_tree_space(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]);
550 ASSERT3U(msp
->ms_deferspace
, ==,
551 range_tree_space(msp
->ms_defer
[0]) +
552 range_tree_space(msp
->ms_defer
[1]));
554 msp_free_space
= range_tree_space(msp
->ms_allocatable
) + allocating
+
555 msp
->ms_deferspace
+ range_tree_space(msp
->ms_freed
);
557 VERIFY3U(sm_free_space
, ==, msp_free_space
);
561 * ==========================================================================
563 * ==========================================================================
566 * Update the allocatable flag and the metaslab group's capacity.
567 * The allocatable flag is set to true if the capacity is below
568 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
569 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
570 * transitions from allocatable to non-allocatable or vice versa then the
571 * metaslab group's class is updated to reflect the transition.
574 metaslab_group_alloc_update(metaslab_group_t
*mg
)
576 vdev_t
*vd
= mg
->mg_vd
;
577 metaslab_class_t
*mc
= mg
->mg_class
;
578 vdev_stat_t
*vs
= &vd
->vdev_stat
;
579 boolean_t was_allocatable
;
580 boolean_t was_initialized
;
582 ASSERT(vd
== vd
->vdev_top
);
583 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_READER
), ==,
586 mutex_enter(&mg
->mg_lock
);
587 was_allocatable
= mg
->mg_allocatable
;
588 was_initialized
= mg
->mg_initialized
;
590 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
593 mutex_enter(&mc
->mc_lock
);
596 * If the metaslab group was just added then it won't
597 * have any space until we finish syncing out this txg.
598 * At that point we will consider it initialized and available
599 * for allocations. We also don't consider non-activated
600 * metaslab groups (e.g. vdevs that are in the middle of being removed)
601 * to be initialized, because they can't be used for allocation.
603 mg
->mg_initialized
= metaslab_group_initialized(mg
);
604 if (!was_initialized
&& mg
->mg_initialized
) {
606 } else if (was_initialized
&& !mg
->mg_initialized
) {
607 ASSERT3U(mc
->mc_groups
, >, 0);
610 if (mg
->mg_initialized
)
611 mg
->mg_no_free_space
= B_FALSE
;
614 * A metaslab group is considered allocatable if it has plenty
615 * of free space or is not heavily fragmented. We only take
616 * fragmentation into account if the metaslab group has a valid
617 * fragmentation metric (i.e. a value between 0 and 100).
619 mg
->mg_allocatable
= (mg
->mg_activation_count
> 0 &&
620 mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
621 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
622 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
625 * The mc_alloc_groups maintains a count of the number of
626 * groups in this metaslab class that are still above the
627 * zfs_mg_noalloc_threshold. This is used by the allocating
628 * threads to determine if they should avoid allocations to
629 * a given group. The allocator will avoid allocations to a group
630 * if that group has reached or is below the zfs_mg_noalloc_threshold
631 * and there are still other groups that are above the threshold.
632 * When a group transitions from allocatable to non-allocatable or
633 * vice versa we update the metaslab class to reflect that change.
634 * When the mc_alloc_groups value drops to 0 that means that all
635 * groups have reached the zfs_mg_noalloc_threshold making all groups
636 * eligible for allocations. This effectively means that all devices
637 * are balanced again.
639 if (was_allocatable
&& !mg
->mg_allocatable
)
640 mc
->mc_alloc_groups
--;
641 else if (!was_allocatable
&& mg
->mg_allocatable
)
642 mc
->mc_alloc_groups
++;
643 mutex_exit(&mc
->mc_lock
);
645 mutex_exit(&mg
->mg_lock
);
649 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
, int allocators
)
651 metaslab_group_t
*mg
;
653 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
654 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
655 mutex_init(&mg
->mg_ms_initialize_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
656 cv_init(&mg
->mg_ms_initialize_cv
, NULL
, CV_DEFAULT
, NULL
);
657 mg
->mg_primaries
= kmem_zalloc(allocators
* sizeof (metaslab_t
*),
659 mg
->mg_secondaries
= kmem_zalloc(allocators
* sizeof (metaslab_t
*),
661 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
662 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
665 mg
->mg_activation_count
= 0;
666 mg
->mg_initialized
= B_FALSE
;
667 mg
->mg_no_free_space
= B_TRUE
;
668 mg
->mg_allocators
= allocators
;
670 mg
->mg_alloc_queue_depth
= kmem_zalloc(allocators
*
671 sizeof (zfs_refcount_t
), KM_SLEEP
);
672 mg
->mg_cur_max_alloc_queue_depth
= kmem_zalloc(allocators
*
673 sizeof (uint64_t), KM_SLEEP
);
674 for (int i
= 0; i
< allocators
; i
++) {
675 zfs_refcount_create_tracked(&mg
->mg_alloc_queue_depth
[i
]);
676 mg
->mg_cur_max_alloc_queue_depth
[i
] = 0;
679 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
680 maxclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
| TASKQ_DYNAMIC
);
686 metaslab_group_destroy(metaslab_group_t
*mg
)
688 ASSERT(mg
->mg_prev
== NULL
);
689 ASSERT(mg
->mg_next
== NULL
);
691 * We may have gone below zero with the activation count
692 * either because we never activated in the first place or
693 * because we're done, and possibly removing the vdev.
695 ASSERT(mg
->mg_activation_count
<= 0);
697 taskq_destroy(mg
->mg_taskq
);
698 avl_destroy(&mg
->mg_metaslab_tree
);
699 kmem_free(mg
->mg_primaries
, mg
->mg_allocators
* sizeof (metaslab_t
*));
700 kmem_free(mg
->mg_secondaries
, mg
->mg_allocators
*
701 sizeof (metaslab_t
*));
702 mutex_destroy(&mg
->mg_lock
);
703 mutex_destroy(&mg
->mg_ms_initialize_lock
);
704 cv_destroy(&mg
->mg_ms_initialize_cv
);
706 for (int i
= 0; i
< mg
->mg_allocators
; i
++) {
707 zfs_refcount_destroy(&mg
->mg_alloc_queue_depth
[i
]);
708 mg
->mg_cur_max_alloc_queue_depth
[i
] = 0;
710 kmem_free(mg
->mg_alloc_queue_depth
, mg
->mg_allocators
*
711 sizeof (zfs_refcount_t
));
712 kmem_free(mg
->mg_cur_max_alloc_queue_depth
, mg
->mg_allocators
*
715 kmem_free(mg
, sizeof (metaslab_group_t
));
719 metaslab_group_activate(metaslab_group_t
*mg
)
721 metaslab_class_t
*mc
= mg
->mg_class
;
722 metaslab_group_t
*mgprev
, *mgnext
;
724 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
), !=, 0);
726 ASSERT(mc
->mc_rotor
!= mg
);
727 ASSERT(mg
->mg_prev
== NULL
);
728 ASSERT(mg
->mg_next
== NULL
);
729 ASSERT(mg
->mg_activation_count
<= 0);
731 if (++mg
->mg_activation_count
<= 0)
734 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
735 metaslab_group_alloc_update(mg
);
737 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
741 mgnext
= mgprev
->mg_next
;
742 mg
->mg_prev
= mgprev
;
743 mg
->mg_next
= mgnext
;
744 mgprev
->mg_next
= mg
;
745 mgnext
->mg_prev
= mg
;
751 * Passivate a metaslab group and remove it from the allocation rotor.
752 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
753 * a metaslab group. This function will momentarily drop spa_config_locks
754 * that are lower than the SCL_ALLOC lock (see comment below).
757 metaslab_group_passivate(metaslab_group_t
*mg
)
759 metaslab_class_t
*mc
= mg
->mg_class
;
760 spa_t
*spa
= mc
->mc_spa
;
761 metaslab_group_t
*mgprev
, *mgnext
;
762 int locks
= spa_config_held(spa
, SCL_ALL
, RW_WRITER
);
764 ASSERT3U(spa_config_held(spa
, SCL_ALLOC
| SCL_ZIO
, RW_WRITER
), ==,
765 (SCL_ALLOC
| SCL_ZIO
));
767 if (--mg
->mg_activation_count
!= 0) {
768 ASSERT(mc
->mc_rotor
!= mg
);
769 ASSERT(mg
->mg_prev
== NULL
);
770 ASSERT(mg
->mg_next
== NULL
);
771 ASSERT(mg
->mg_activation_count
< 0);
776 * The spa_config_lock is an array of rwlocks, ordered as
777 * follows (from highest to lowest):
778 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
779 * SCL_ZIO > SCL_FREE > SCL_VDEV
780 * (For more information about the spa_config_lock see spa_misc.c)
781 * The higher the lock, the broader its coverage. When we passivate
782 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
783 * config locks. However, the metaslab group's taskq might be trying
784 * to preload metaslabs so we must drop the SCL_ZIO lock and any
785 * lower locks to allow the I/O to complete. At a minimum,
786 * we continue to hold the SCL_ALLOC lock, which prevents any future
787 * allocations from taking place and any changes to the vdev tree.
789 spa_config_exit(spa
, locks
& ~(SCL_ZIO
- 1), spa
);
790 taskq_wait_outstanding(mg
->mg_taskq
, 0);
791 spa_config_enter(spa
, locks
& ~(SCL_ZIO
- 1), spa
, RW_WRITER
);
792 metaslab_group_alloc_update(mg
);
793 for (int i
= 0; i
< mg
->mg_allocators
; i
++) {
794 metaslab_t
*msp
= mg
->mg_primaries
[i
];
796 mutex_enter(&msp
->ms_lock
);
797 metaslab_passivate(msp
,
798 metaslab_weight_from_range_tree(msp
));
799 mutex_exit(&msp
->ms_lock
);
801 msp
= mg
->mg_secondaries
[i
];
803 mutex_enter(&msp
->ms_lock
);
804 metaslab_passivate(msp
,
805 metaslab_weight_from_range_tree(msp
));
806 mutex_exit(&msp
->ms_lock
);
810 mgprev
= mg
->mg_prev
;
811 mgnext
= mg
->mg_next
;
816 mc
->mc_rotor
= mgnext
;
817 mgprev
->mg_next
= mgnext
;
818 mgnext
->mg_prev
= mgprev
;
826 metaslab_group_initialized(metaslab_group_t
*mg
)
828 vdev_t
*vd
= mg
->mg_vd
;
829 vdev_stat_t
*vs
= &vd
->vdev_stat
;
831 return (vs
->vs_space
!= 0 && mg
->mg_activation_count
> 0);
835 metaslab_group_get_space(metaslab_group_t
*mg
)
837 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
841 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
844 vdev_t
*vd
= mg
->mg_vd
;
845 uint64_t ashift
= vd
->vdev_ashift
;
848 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
851 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
854 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
855 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
857 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
858 metaslab_t
*msp
= vd
->vdev_ms
[m
];
860 /* skip if not active or not a member */
861 if (msp
->ms_sm
== NULL
|| msp
->ms_group
!= mg
)
864 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
865 mg_hist
[i
+ ashift
] +=
866 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
869 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
870 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
872 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
876 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
878 metaslab_class_t
*mc
= mg
->mg_class
;
879 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
881 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
882 if (msp
->ms_sm
== NULL
)
885 mutex_enter(&mg
->mg_lock
);
886 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
887 mg
->mg_histogram
[i
+ ashift
] +=
888 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
889 mc
->mc_histogram
[i
+ ashift
] +=
890 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
892 mutex_exit(&mg
->mg_lock
);
896 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
898 metaslab_class_t
*mc
= mg
->mg_class
;
899 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
901 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
902 if (msp
->ms_sm
== NULL
)
905 mutex_enter(&mg
->mg_lock
);
906 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
907 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
908 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
909 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
910 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
912 mg
->mg_histogram
[i
+ ashift
] -=
913 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
914 mc
->mc_histogram
[i
+ ashift
] -=
915 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
917 mutex_exit(&mg
->mg_lock
);
921 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
923 ASSERT(msp
->ms_group
== NULL
);
924 mutex_enter(&mg
->mg_lock
);
927 avl_add(&mg
->mg_metaslab_tree
, msp
);
928 mutex_exit(&mg
->mg_lock
);
930 mutex_enter(&msp
->ms_lock
);
931 metaslab_group_histogram_add(mg
, msp
);
932 mutex_exit(&msp
->ms_lock
);
936 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
938 mutex_enter(&msp
->ms_lock
);
939 metaslab_group_histogram_remove(mg
, msp
);
940 mutex_exit(&msp
->ms_lock
);
942 mutex_enter(&mg
->mg_lock
);
943 ASSERT(msp
->ms_group
== mg
);
944 avl_remove(&mg
->mg_metaslab_tree
, msp
);
945 msp
->ms_group
= NULL
;
946 mutex_exit(&mg
->mg_lock
);
950 metaslab_group_sort_impl(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
952 ASSERT(MUTEX_HELD(&mg
->mg_lock
));
953 ASSERT(msp
->ms_group
== mg
);
954 avl_remove(&mg
->mg_metaslab_tree
, msp
);
955 msp
->ms_weight
= weight
;
956 avl_add(&mg
->mg_metaslab_tree
, msp
);
961 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
964 * Although in principle the weight can be any value, in
965 * practice we do not use values in the range [1, 511].
967 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
968 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
970 mutex_enter(&mg
->mg_lock
);
971 metaslab_group_sort_impl(mg
, msp
, weight
);
972 mutex_exit(&mg
->mg_lock
);
976 * Calculate the fragmentation for a given metaslab group. We can use
977 * a simple average here since all metaslabs within the group must have
978 * the same size. The return value will be a value between 0 and 100
979 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
980 * group have a fragmentation metric.
983 metaslab_group_fragmentation(metaslab_group_t
*mg
)
985 vdev_t
*vd
= mg
->mg_vd
;
986 uint64_t fragmentation
= 0;
987 uint64_t valid_ms
= 0;
989 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
990 metaslab_t
*msp
= vd
->vdev_ms
[m
];
992 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
994 if (msp
->ms_group
!= mg
)
998 fragmentation
+= msp
->ms_fragmentation
;
1001 if (valid_ms
<= mg
->mg_vd
->vdev_ms_count
/ 2)
1002 return (ZFS_FRAG_INVALID
);
1004 fragmentation
/= valid_ms
;
1005 ASSERT3U(fragmentation
, <=, 100);
1006 return (fragmentation
);
1010 * Determine if a given metaslab group should skip allocations. A metaslab
1011 * group should avoid allocations if its free capacity is less than the
1012 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
1013 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
1014 * that can still handle allocations. If the allocation throttle is enabled
1015 * then we skip allocations to devices that have reached their maximum
1016 * allocation queue depth unless the selected metaslab group is the only
1017 * eligible group remaining.
1020 metaslab_group_allocatable(metaslab_group_t
*mg
, metaslab_group_t
*rotor
,
1021 uint64_t psize
, int allocator
, int d
)
1023 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1024 metaslab_class_t
*mc
= mg
->mg_class
;
1027 * We can only consider skipping this metaslab group if it's
1028 * in the normal metaslab class and there are other metaslab
1029 * groups to select from. Otherwise, we always consider it eligible
1032 if ((mc
!= spa_normal_class(spa
) &&
1033 mc
!= spa_special_class(spa
) &&
1034 mc
!= spa_dedup_class(spa
)) ||
1039 * If the metaslab group's mg_allocatable flag is set (see comments
1040 * in metaslab_group_alloc_update() for more information) and
1041 * the allocation throttle is disabled then allow allocations to this
1042 * device. However, if the allocation throttle is enabled then
1043 * check if we have reached our allocation limit (mg_alloc_queue_depth)
1044 * to determine if we should allow allocations to this metaslab group.
1045 * If all metaslab groups are no longer considered allocatable
1046 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1047 * gang block size then we allow allocations on this metaslab group
1048 * regardless of the mg_allocatable or throttle settings.
1050 if (mg
->mg_allocatable
) {
1051 metaslab_group_t
*mgp
;
1053 uint64_t qmax
= mg
->mg_cur_max_alloc_queue_depth
[allocator
];
1055 if (!mc
->mc_alloc_throttle_enabled
)
1059 * If this metaslab group does not have any free space, then
1060 * there is no point in looking further.
1062 if (mg
->mg_no_free_space
)
1066 * Relax allocation throttling for ditto blocks. Due to
1067 * random imbalances in allocation it tends to push copies
1068 * to one vdev, that looks a bit better at the moment.
1070 qmax
= qmax
* (4 + d
) / 4;
1072 qdepth
= zfs_refcount_count(
1073 &mg
->mg_alloc_queue_depth
[allocator
]);
1076 * If this metaslab group is below its qmax or it's
1077 * the only allocatable metasable group, then attempt
1078 * to allocate from it.
1080 if (qdepth
< qmax
|| mc
->mc_alloc_groups
== 1)
1082 ASSERT3U(mc
->mc_alloc_groups
, >, 1);
1085 * Since this metaslab group is at or over its qmax, we
1086 * need to determine if there are metaslab groups after this
1087 * one that might be able to handle this allocation. This is
1088 * racy since we can't hold the locks for all metaslab
1089 * groups at the same time when we make this check.
1091 for (mgp
= mg
->mg_next
; mgp
!= rotor
; mgp
= mgp
->mg_next
) {
1092 qmax
= mgp
->mg_cur_max_alloc_queue_depth
[allocator
];
1093 qmax
= qmax
* (4 + d
) / 4;
1094 qdepth
= zfs_refcount_count(
1095 &mgp
->mg_alloc_queue_depth
[allocator
]);
1098 * If there is another metaslab group that
1099 * might be able to handle the allocation, then
1100 * we return false so that we skip this group.
1102 if (qdepth
< qmax
&& !mgp
->mg_no_free_space
)
1107 * We didn't find another group to handle the allocation
1108 * so we can't skip this metaslab group even though
1109 * we are at or over our qmax.
1113 } else if (mc
->mc_alloc_groups
== 0 || psize
== SPA_MINBLOCKSIZE
) {
1120 * ==========================================================================
1121 * Range tree callbacks
1122 * ==========================================================================
1126 * Comparison function for the private size-ordered tree. Tree is sorted
1127 * by size, larger sizes at the end of the tree.
1130 metaslab_rangesize_compare(const void *x1
, const void *x2
)
1132 const range_seg_t
*r1
= x1
;
1133 const range_seg_t
*r2
= x2
;
1134 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
1135 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
1137 int cmp
= AVL_CMP(rs_size1
, rs_size2
);
1141 return (AVL_CMP(r1
->rs_start
, r2
->rs_start
));
1145 * ==========================================================================
1146 * Common allocator routines
1147 * ==========================================================================
1151 * Return the maximum contiguous segment within the metaslab.
1154 metaslab_block_maxsize(metaslab_t
*msp
)
1156 avl_tree_t
*t
= &msp
->ms_allocatable_by_size
;
1159 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
1162 return (rs
->rs_end
- rs
->rs_start
);
1165 static range_seg_t
*
1166 metaslab_block_find(avl_tree_t
*t
, uint64_t start
, uint64_t size
)
1168 range_seg_t
*rs
, rsearch
;
1171 rsearch
.rs_start
= start
;
1172 rsearch
.rs_end
= start
+ size
;
1174 rs
= avl_find(t
, &rsearch
, &where
);
1176 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1182 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
1183 defined(WITH_DF_BLOCK_ALLOCATOR) || \
1184 defined(WITH_CF_BLOCK_ALLOCATOR)
1186 * This is a helper function that can be used by the allocator to find
1187 * a suitable block to allocate. This will search the specified AVL
1188 * tree looking for a block that matches the specified criteria.
1191 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
1194 range_seg_t
*rs
= metaslab_block_find(t
, *cursor
, size
);
1196 while (rs
!= NULL
) {
1197 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
1199 if (offset
+ size
<= rs
->rs_end
) {
1200 *cursor
= offset
+ size
;
1203 rs
= AVL_NEXT(t
, rs
);
1207 * If we know we've searched the whole map (*cursor == 0), give up.
1208 * Otherwise, reset the cursor to the beginning and try again.
1214 return (metaslab_block_picker(t
, cursor
, size
, align
));
1216 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
1218 #if defined(WITH_FF_BLOCK_ALLOCATOR)
1220 * ==========================================================================
1221 * The first-fit block allocator
1222 * ==========================================================================
1225 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
1228 * Find the largest power of 2 block size that evenly divides the
1229 * requested size. This is used to try to allocate blocks with similar
1230 * alignment from the same area of the metaslab (i.e. same cursor
1231 * bucket) but it does not guarantee that other allocations sizes
1232 * may exist in the same region.
1234 uint64_t align
= size
& -size
;
1235 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1236 avl_tree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1238 return (metaslab_block_picker(t
, cursor
, size
, align
));
1241 static metaslab_ops_t metaslab_ff_ops
= {
1245 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
1246 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1248 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1250 * ==========================================================================
1251 * Dynamic block allocator -
1252 * Uses the first fit allocation scheme until space get low and then
1253 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1254 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1255 * ==========================================================================
1258 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1261 * Find the largest power of 2 block size that evenly divides the
1262 * requested size. This is used to try to allocate blocks with similar
1263 * alignment from the same area of the metaslab (i.e. same cursor
1264 * bucket) but it does not guarantee that other allocations sizes
1265 * may exist in the same region.
1267 uint64_t align
= size
& -size
;
1268 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1269 range_tree_t
*rt
= msp
->ms_allocatable
;
1270 avl_tree_t
*t
= &rt
->rt_root
;
1271 uint64_t max_size
= metaslab_block_maxsize(msp
);
1272 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1274 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1275 ASSERT3U(avl_numnodes(t
), ==,
1276 avl_numnodes(&msp
->ms_allocatable_by_size
));
1278 if (max_size
< size
)
1282 * If we're running low on space switch to using the size
1283 * sorted AVL tree (best-fit).
1285 if (max_size
< metaslab_df_alloc_threshold
||
1286 free_pct
< metaslab_df_free_pct
) {
1287 t
= &msp
->ms_allocatable_by_size
;
1291 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1294 static metaslab_ops_t metaslab_df_ops
= {
1298 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1299 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1301 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1303 * ==========================================================================
1304 * Cursor fit block allocator -
1305 * Select the largest region in the metaslab, set the cursor to the beginning
1306 * of the range and the cursor_end to the end of the range. As allocations
1307 * are made advance the cursor. Continue allocating from the cursor until
1308 * the range is exhausted and then find a new range.
1309 * ==========================================================================
1312 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1314 range_tree_t
*rt
= msp
->ms_allocatable
;
1315 avl_tree_t
*t
= &msp
->ms_allocatable_by_size
;
1316 uint64_t *cursor
= &msp
->ms_lbas
[0];
1317 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1318 uint64_t offset
= 0;
1320 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1321 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1323 ASSERT3U(*cursor_end
, >=, *cursor
);
1325 if ((*cursor
+ size
) > *cursor_end
) {
1328 rs
= avl_last(&msp
->ms_allocatable_by_size
);
1329 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1332 *cursor
= rs
->rs_start
;
1333 *cursor_end
= rs
->rs_end
;
1342 static metaslab_ops_t metaslab_cf_ops
= {
1346 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1347 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1349 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1351 * ==========================================================================
1352 * New dynamic fit allocator -
1353 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1354 * contiguous blocks. If no region is found then just use the largest segment
1356 * ==========================================================================
1360 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1361 * to request from the allocator.
1363 uint64_t metaslab_ndf_clump_shift
= 4;
1366 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1368 avl_tree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1370 range_seg_t
*rs
, rsearch
;
1371 uint64_t hbit
= highbit64(size
);
1372 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1373 uint64_t max_size
= metaslab_block_maxsize(msp
);
1375 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1376 ASSERT3U(avl_numnodes(t
), ==,
1377 avl_numnodes(&msp
->ms_allocatable_by_size
));
1379 if (max_size
< size
)
1382 rsearch
.rs_start
= *cursor
;
1383 rsearch
.rs_end
= *cursor
+ size
;
1385 rs
= avl_find(t
, &rsearch
, &where
);
1386 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1387 t
= &msp
->ms_allocatable_by_size
;
1389 rsearch
.rs_start
= 0;
1390 rsearch
.rs_end
= MIN(max_size
,
1391 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1392 rs
= avl_find(t
, &rsearch
, &where
);
1394 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1398 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1399 *cursor
= rs
->rs_start
+ size
;
1400 return (rs
->rs_start
);
1405 static metaslab_ops_t metaslab_ndf_ops
= {
1409 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1410 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1414 * ==========================================================================
1416 * ==========================================================================
1420 * Wait for any in-progress metaslab loads to complete.
1423 metaslab_load_wait(metaslab_t
*msp
)
1425 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1427 while (msp
->ms_loading
) {
1428 ASSERT(!msp
->ms_loaded
);
1429 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1434 metaslab_load_impl(metaslab_t
*msp
)
1438 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1439 ASSERT(msp
->ms_loading
);
1440 ASSERT(!msp
->ms_condensing
);
1443 * We temporarily drop the lock to unblock other operations while we
1444 * are reading the space map. Therefore, metaslab_sync() and
1445 * metaslab_sync_done() can run at the same time as we do.
1447 * metaslab_sync() can append to the space map while we are loading.
1448 * Therefore we load only entries that existed when we started the
1449 * load. Additionally, metaslab_sync_done() has to wait for the load
1450 * to complete because there are potential races like metaslab_load()
1451 * loading parts of the space map that are currently being appended
1452 * by metaslab_sync(). If we didn't, the ms_allocatable would have
1453 * entries that metaslab_sync_done() would try to re-add later.
1455 * That's why before dropping the lock we remember the synced length
1456 * of the metaslab and read up to that point of the space map,
1457 * ignoring entries appended by metaslab_sync() that happen after we
1460 uint64_t length
= msp
->ms_synced_length
;
1461 mutex_exit(&msp
->ms_lock
);
1463 if (msp
->ms_sm
!= NULL
) {
1464 error
= space_map_load_length(msp
->ms_sm
, msp
->ms_allocatable
,
1468 * The space map has not been allocated yet, so treat
1469 * all the space in the metaslab as free and add it to the
1470 * ms_allocatable tree.
1472 range_tree_add(msp
->ms_allocatable
,
1473 msp
->ms_start
, msp
->ms_size
);
1477 * We need to grab the ms_sync_lock to prevent metaslab_sync() from
1478 * changing the ms_sm and the metaslab's range trees while we are
1479 * about to use them and populate the ms_allocatable. The ms_lock
1480 * is insufficient for this because metaslab_sync() doesn't hold
1481 * the ms_lock while writing the ms_checkpointing tree to disk.
1483 mutex_enter(&msp
->ms_sync_lock
);
1484 mutex_enter(&msp
->ms_lock
);
1485 ASSERT(!msp
->ms_condensing
);
1490 ASSERT3P(msp
->ms_group
, !=, NULL
);
1491 msp
->ms_loaded
= B_TRUE
;
1494 * The ms_allocatable contains the segments that exist in the
1495 * ms_defer trees [see ms_synced_length]. Thus we need to remove
1496 * them from ms_allocatable as they will be added again in
1497 * metaslab_sync_done().
1499 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1500 range_tree_walk(msp
->ms_defer
[t
],
1501 range_tree_remove
, msp
->ms_allocatable
);
1504 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
1506 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1507 metaslab_verify_space(msp
, spa_syncing_txg(spa
));
1508 mutex_exit(&msp
->ms_sync_lock
);
1514 metaslab_load(metaslab_t
*msp
)
1516 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1519 * There may be another thread loading the same metaslab, if that's
1520 * the case just wait until the other thread is done and return.
1522 metaslab_load_wait(msp
);
1525 VERIFY(!msp
->ms_loading
);
1526 ASSERT(!msp
->ms_condensing
);
1528 msp
->ms_loading
= B_TRUE
;
1529 int error
= metaslab_load_impl(msp
);
1530 msp
->ms_loading
= B_FALSE
;
1531 cv_broadcast(&msp
->ms_load_cv
);
1537 metaslab_unload(metaslab_t
*msp
)
1539 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1540 range_tree_vacate(msp
->ms_allocatable
, NULL
, NULL
);
1541 msp
->ms_loaded
= B_FALSE
;
1542 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1543 msp
->ms_max_size
= 0;
1547 metaslab_space_update(vdev_t
*vd
, metaslab_class_t
*mc
, int64_t alloc_delta
,
1548 int64_t defer_delta
, int64_t space_delta
)
1550 vdev_space_update(vd
, alloc_delta
, defer_delta
, space_delta
);
1552 ASSERT3P(vd
->vdev_spa
->spa_root_vdev
, ==, vd
->vdev_parent
);
1553 ASSERT(vd
->vdev_ms_count
!= 0);
1555 metaslab_class_space_update(mc
, alloc_delta
, defer_delta
, space_delta
,
1556 vdev_deflated_space(vd
, space_delta
));
1560 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1563 vdev_t
*vd
= mg
->mg_vd
;
1564 spa_t
*spa
= vd
->vdev_spa
;
1565 objset_t
*mos
= spa
->spa_meta_objset
;
1569 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1570 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1571 mutex_init(&ms
->ms_sync_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1572 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1575 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1576 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1577 ms
->ms_allocator
= -1;
1578 ms
->ms_new
= B_TRUE
;
1581 * We only open space map objects that already exist. All others
1582 * will be opened when we finally allocate an object for it.
1585 * When called from vdev_expand(), we can't call into the DMU as
1586 * we are holding the spa_config_lock as a writer and we would
1587 * deadlock [see relevant comment in vdev_metaslab_init()]. in
1588 * that case, the object parameter is zero though, so we won't
1589 * call into the DMU.
1592 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1593 ms
->ms_size
, vd
->vdev_ashift
);
1596 kmem_free(ms
, sizeof (metaslab_t
));
1600 ASSERT(ms
->ms_sm
!= NULL
);
1601 ms
->ms_allocated_space
= space_map_allocated(ms
->ms_sm
);
1605 * We create the ms_allocatable here, but we don't create the
1606 * other range trees until metaslab_sync_done(). This serves
1607 * two purposes: it allows metaslab_sync_done() to detect the
1608 * addition of new space; and for debugging, it ensures that
1609 * we'd data fault on any attempt to use this metaslab before
1612 ms
->ms_allocatable
= range_tree_create_impl(&rt_avl_ops
,
1613 &ms
->ms_allocatable_by_size
, metaslab_rangesize_compare
, 0);
1614 metaslab_group_add(mg
, ms
);
1616 metaslab_set_fragmentation(ms
);
1619 * If we're opening an existing pool (txg == 0) or creating
1620 * a new one (txg == TXG_INITIAL), all space is available now.
1621 * If we're adding space to an existing pool, the new space
1622 * does not become available until after this txg has synced.
1623 * The metaslab's weight will also be initialized when we sync
1624 * out this txg. This ensures that we don't attempt to allocate
1625 * from it before we have initialized it completely.
1627 if (txg
<= TXG_INITIAL
) {
1628 metaslab_sync_done(ms
, 0);
1629 metaslab_space_update(vd
, mg
->mg_class
,
1630 metaslab_allocated_space(ms
), 0, 0);
1634 * If metaslab_debug_load is set and we're initializing a metaslab
1635 * that has an allocated space map object then load the space map
1636 * so that we can verify frees.
1638 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1639 mutex_enter(&ms
->ms_lock
);
1640 VERIFY0(metaslab_load(ms
));
1641 mutex_exit(&ms
->ms_lock
);
1645 vdev_dirty(vd
, 0, NULL
, txg
);
1646 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1655 metaslab_fini(metaslab_t
*msp
)
1657 metaslab_group_t
*mg
= msp
->ms_group
;
1658 vdev_t
*vd
= mg
->mg_vd
;
1660 metaslab_group_remove(mg
, msp
);
1662 mutex_enter(&msp
->ms_lock
);
1663 VERIFY(msp
->ms_group
== NULL
);
1664 metaslab_space_update(vd
, mg
->mg_class
,
1665 -metaslab_allocated_space(msp
), 0, -msp
->ms_size
);
1667 space_map_close(msp
->ms_sm
);
1669 metaslab_unload(msp
);
1671 range_tree_destroy(msp
->ms_allocatable
);
1672 range_tree_destroy(msp
->ms_freeing
);
1673 range_tree_destroy(msp
->ms_freed
);
1675 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1676 range_tree_destroy(msp
->ms_allocating
[t
]);
1679 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1680 range_tree_destroy(msp
->ms_defer
[t
]);
1682 ASSERT0(msp
->ms_deferspace
);
1684 range_tree_destroy(msp
->ms_checkpointing
);
1686 mutex_exit(&msp
->ms_lock
);
1687 cv_destroy(&msp
->ms_load_cv
);
1688 mutex_destroy(&msp
->ms_lock
);
1689 mutex_destroy(&msp
->ms_sync_lock
);
1690 ASSERT3U(msp
->ms_allocator
, ==, -1);
1692 kmem_free(msp
, sizeof (metaslab_t
));
1695 #define FRAGMENTATION_TABLE_SIZE 17
1698 * This table defines a segment size based fragmentation metric that will
1699 * allow each metaslab to derive its own fragmentation value. This is done
1700 * by calculating the space in each bucket of the spacemap histogram and
1701 * multiplying that by the fragmetation metric in this table. Doing
1702 * this for all buckets and dividing it by the total amount of free
1703 * space in this metaslab (i.e. the total free space in all buckets) gives
1704 * us the fragmentation metric. This means that a high fragmentation metric
1705 * equates to most of the free space being comprised of small segments.
1706 * Conversely, if the metric is low, then most of the free space is in
1707 * large segments. A 10% change in fragmentation equates to approximately
1708 * double the number of segments.
1710 * This table defines 0% fragmented space using 16MB segments. Testing has
1711 * shown that segments that are greater than or equal to 16MB do not suffer
1712 * from drastic performance problems. Using this value, we derive the rest
1713 * of the table. Since the fragmentation value is never stored on disk, it
1714 * is possible to change these calculations in the future.
1716 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1736 * Calculate the metaslab's fragmentation metric and set ms_fragmentation.
1737 * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not
1738 * been upgraded and does not support this metric. Otherwise, the return
1739 * value should be in the range [0, 100].
1742 metaslab_set_fragmentation(metaslab_t
*msp
)
1744 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1745 uint64_t fragmentation
= 0;
1747 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1748 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1750 if (!feature_enabled
) {
1751 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1756 * A null space map means that the entire metaslab is free
1757 * and thus is not fragmented.
1759 if (msp
->ms_sm
== NULL
) {
1760 msp
->ms_fragmentation
= 0;
1765 * If this metaslab's space map has not been upgraded, flag it
1766 * so that we upgrade next time we encounter it.
1768 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1769 uint64_t txg
= spa_syncing_txg(spa
);
1770 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1773 * If we've reached the final dirty txg, then we must
1774 * be shutting down the pool. We don't want to dirty
1775 * any data past this point so skip setting the condense
1776 * flag. We can retry this action the next time the pool
1779 if (spa_writeable(spa
) && txg
< spa_final_dirty_txg(spa
)) {
1780 msp
->ms_condense_wanted
= B_TRUE
;
1781 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1782 zfs_dbgmsg("txg %llu, requesting force condense: "
1783 "ms_id %llu, vdev_id %llu", txg
, msp
->ms_id
,
1786 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
1790 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1792 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1794 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1795 FRAGMENTATION_TABLE_SIZE
- 1);
1797 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1800 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1803 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1804 fragmentation
+= space
* zfs_frag_table
[idx
];
1808 fragmentation
/= total
;
1809 ASSERT3U(fragmentation
, <=, 100);
1811 msp
->ms_fragmentation
= fragmentation
;
1815 * Compute a weight -- a selection preference value -- for the given metaslab.
1816 * This is based on the amount of free space, the level of fragmentation,
1817 * the LBA range, and whether the metaslab is loaded.
1820 metaslab_space_weight(metaslab_t
*msp
)
1822 metaslab_group_t
*mg
= msp
->ms_group
;
1823 vdev_t
*vd
= mg
->mg_vd
;
1824 uint64_t weight
, space
;
1826 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1827 ASSERT(!vd
->vdev_removing
);
1830 * The baseline weight is the metaslab's free space.
1832 space
= msp
->ms_size
- metaslab_allocated_space(msp
);
1834 if (metaslab_fragmentation_factor_enabled
&&
1835 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1837 * Use the fragmentation information to inversely scale
1838 * down the baseline weight. We need to ensure that we
1839 * don't exclude this metaslab completely when it's 100%
1840 * fragmented. To avoid this we reduce the fragmented value
1843 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1846 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1847 * this metaslab again. The fragmentation metric may have
1848 * decreased the space to something smaller than
1849 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1850 * so that we can consume any remaining space.
1852 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1853 space
= SPA_MINBLOCKSIZE
;
1858 * Modern disks have uniform bit density and constant angular velocity.
1859 * Therefore, the outer recording zones are faster (higher bandwidth)
1860 * than the inner zones by the ratio of outer to inner track diameter,
1861 * which is typically around 2:1. We account for this by assigning
1862 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1863 * In effect, this means that we'll select the metaslab with the most
1864 * free bandwidth rather than simply the one with the most free space.
1866 if (!vd
->vdev_nonrot
&& metaslab_lba_weighting_enabled
) {
1867 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1868 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1872 * If this metaslab is one we're actively using, adjust its
1873 * weight to make it preferable to any inactive metaslab so
1874 * we'll polish it off. If the fragmentation on this metaslab
1875 * has exceed our threshold, then don't mark it active.
1877 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1878 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1879 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1882 WEIGHT_SET_SPACEBASED(weight
);
1887 * Return the weight of the specified metaslab, according to the segment-based
1888 * weighting algorithm. The metaslab must be loaded. This function can
1889 * be called within a sync pass since it relies only on the metaslab's
1890 * range tree which is always accurate when the metaslab is loaded.
1893 metaslab_weight_from_range_tree(metaslab_t
*msp
)
1895 uint64_t weight
= 0;
1896 uint32_t segments
= 0;
1898 ASSERT(msp
->ms_loaded
);
1900 for (int i
= RANGE_TREE_HISTOGRAM_SIZE
- 1; i
>= SPA_MINBLOCKSHIFT
;
1902 uint8_t shift
= msp
->ms_group
->mg_vd
->vdev_ashift
;
1903 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1906 segments
+= msp
->ms_allocatable
->rt_histogram
[i
];
1909 * The range tree provides more precision than the space map
1910 * and must be downgraded so that all values fit within the
1911 * space map's histogram. This allows us to compare loaded
1912 * vs. unloaded metaslabs to determine which metaslab is
1913 * considered "best".
1918 if (segments
!= 0) {
1919 WEIGHT_SET_COUNT(weight
, segments
);
1920 WEIGHT_SET_INDEX(weight
, i
);
1921 WEIGHT_SET_ACTIVE(weight
, 0);
1929 * Calculate the weight based on the on-disk histogram. This should only
1930 * be called after a sync pass has completely finished since the on-disk
1931 * information is updated in metaslab_sync().
1934 metaslab_weight_from_spacemap(metaslab_t
*msp
)
1936 uint64_t weight
= 0;
1938 for (int i
= SPACE_MAP_HISTOGRAM_SIZE
- 1; i
>= 0; i
--) {
1939 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] != 0) {
1940 WEIGHT_SET_COUNT(weight
,
1941 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
1942 WEIGHT_SET_INDEX(weight
, i
+
1943 msp
->ms_sm
->sm_shift
);
1944 WEIGHT_SET_ACTIVE(weight
, 0);
1952 * Compute a segment-based weight for the specified metaslab. The weight
1953 * is determined by highest bucket in the histogram. The information
1954 * for the highest bucket is encoded into the weight value.
1957 metaslab_segment_weight(metaslab_t
*msp
)
1959 metaslab_group_t
*mg
= msp
->ms_group
;
1960 uint64_t weight
= 0;
1961 uint8_t shift
= mg
->mg_vd
->vdev_ashift
;
1963 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1966 * The metaslab is completely free.
1968 if (metaslab_allocated_space(msp
) == 0) {
1969 int idx
= highbit64(msp
->ms_size
) - 1;
1970 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
1972 if (idx
< max_idx
) {
1973 WEIGHT_SET_COUNT(weight
, 1ULL);
1974 WEIGHT_SET_INDEX(weight
, idx
);
1976 WEIGHT_SET_COUNT(weight
, 1ULL << (idx
- max_idx
));
1977 WEIGHT_SET_INDEX(weight
, max_idx
);
1979 WEIGHT_SET_ACTIVE(weight
, 0);
1980 ASSERT(!WEIGHT_IS_SPACEBASED(weight
));
1985 ASSERT3U(msp
->ms_sm
->sm_dbuf
->db_size
, ==, sizeof (space_map_phys_t
));
1988 * If the metaslab is fully allocated then just make the weight 0.
1990 if (metaslab_allocated_space(msp
) == msp
->ms_size
)
1993 * If the metaslab is already loaded, then use the range tree to
1994 * determine the weight. Otherwise, we rely on the space map information
1995 * to generate the weight.
1997 if (msp
->ms_loaded
) {
1998 weight
= metaslab_weight_from_range_tree(msp
);
2000 weight
= metaslab_weight_from_spacemap(msp
);
2004 * If the metaslab was active the last time we calculated its weight
2005 * then keep it active. We want to consume the entire region that
2006 * is associated with this weight.
2008 if (msp
->ms_activation_weight
!= 0 && weight
!= 0)
2009 WEIGHT_SET_ACTIVE(weight
, WEIGHT_GET_ACTIVE(msp
->ms_weight
));
2014 * Determine if we should attempt to allocate from this metaslab. If the
2015 * metaslab has a maximum size then we can quickly determine if the desired
2016 * allocation size can be satisfied. Otherwise, if we're using segment-based
2017 * weighting then we can determine the maximum allocation that this metaslab
2018 * can accommodate based on the index encoded in the weight. If we're using
2019 * space-based weights then rely on the entire weight (excluding the weight
2023 metaslab_should_allocate(metaslab_t
*msp
, uint64_t asize
)
2025 boolean_t should_allocate
;
2027 if (msp
->ms_max_size
!= 0)
2028 return (msp
->ms_max_size
>= asize
);
2030 if (!WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
2032 * The metaslab segment weight indicates segments in the
2033 * range [2^i, 2^(i+1)), where i is the index in the weight.
2034 * Since the asize might be in the middle of the range, we
2035 * should attempt the allocation if asize < 2^(i+1).
2037 should_allocate
= (asize
<
2038 1ULL << (WEIGHT_GET_INDEX(msp
->ms_weight
) + 1));
2040 should_allocate
= (asize
<=
2041 (msp
->ms_weight
& ~METASLAB_WEIGHT_TYPE
));
2043 return (should_allocate
);
2046 metaslab_weight(metaslab_t
*msp
)
2048 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
2049 spa_t
*spa
= vd
->vdev_spa
;
2052 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2055 * If this vdev is in the process of being removed, there is nothing
2056 * for us to do here.
2058 if (vd
->vdev_removing
)
2061 metaslab_set_fragmentation(msp
);
2064 * Update the maximum size if the metaslab is loaded. This will
2065 * ensure that we get an accurate maximum size if newly freed space
2066 * has been added back into the free tree.
2069 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
2071 ASSERT0(msp
->ms_max_size
);
2074 * Segment-based weighting requires space map histogram support.
2076 if (zfs_metaslab_segment_weight_enabled
&&
2077 spa_feature_is_enabled(spa
, SPA_FEATURE_SPACEMAP_HISTOGRAM
) &&
2078 (msp
->ms_sm
== NULL
|| msp
->ms_sm
->sm_dbuf
->db_size
==
2079 sizeof (space_map_phys_t
))) {
2080 weight
= metaslab_segment_weight(msp
);
2082 weight
= metaslab_space_weight(msp
);
2088 metaslab_activate_allocator(metaslab_group_t
*mg
, metaslab_t
*msp
,
2089 int allocator
, uint64_t activation_weight
)
2092 * If we're activating for the claim code, we don't want to actually
2093 * set the metaslab up for a specific allocator.
2095 if (activation_weight
== METASLAB_WEIGHT_CLAIM
)
2097 metaslab_t
**arr
= (activation_weight
== METASLAB_WEIGHT_PRIMARY
?
2098 mg
->mg_primaries
: mg
->mg_secondaries
);
2100 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2101 mutex_enter(&mg
->mg_lock
);
2102 if (arr
[allocator
] != NULL
) {
2103 mutex_exit(&mg
->mg_lock
);
2107 arr
[allocator
] = msp
;
2108 ASSERT3S(msp
->ms_allocator
, ==, -1);
2109 msp
->ms_allocator
= allocator
;
2110 msp
->ms_primary
= (activation_weight
== METASLAB_WEIGHT_PRIMARY
);
2111 mutex_exit(&mg
->mg_lock
);
2117 metaslab_activate(metaslab_t
*msp
, int allocator
, uint64_t activation_weight
)
2119 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2121 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
2122 int error
= metaslab_load(msp
);
2124 metaslab_group_sort(msp
->ms_group
, msp
, 0);
2127 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) != 0) {
2129 * The metaslab was activated for another allocator
2130 * while we were waiting, we should reselect.
2132 return (SET_ERROR(EBUSY
));
2134 if ((error
= metaslab_activate_allocator(msp
->ms_group
, msp
,
2135 allocator
, activation_weight
)) != 0) {
2139 msp
->ms_activation_weight
= msp
->ms_weight
;
2140 metaslab_group_sort(msp
->ms_group
, msp
,
2141 msp
->ms_weight
| activation_weight
);
2143 ASSERT(msp
->ms_loaded
);
2144 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
2150 metaslab_passivate_allocator(metaslab_group_t
*mg
, metaslab_t
*msp
,
2153 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2154 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
) {
2155 metaslab_group_sort(mg
, msp
, weight
);
2159 mutex_enter(&mg
->mg_lock
);
2160 ASSERT3P(msp
->ms_group
, ==, mg
);
2161 if (msp
->ms_primary
) {
2162 ASSERT3U(0, <=, msp
->ms_allocator
);
2163 ASSERT3U(msp
->ms_allocator
, <, mg
->mg_allocators
);
2164 ASSERT3P(mg
->mg_primaries
[msp
->ms_allocator
], ==, msp
);
2165 ASSERT(msp
->ms_weight
& METASLAB_WEIGHT_PRIMARY
);
2166 mg
->mg_primaries
[msp
->ms_allocator
] = NULL
;
2168 ASSERT(msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
);
2169 ASSERT3P(mg
->mg_secondaries
[msp
->ms_allocator
], ==, msp
);
2170 mg
->mg_secondaries
[msp
->ms_allocator
] = NULL
;
2172 msp
->ms_allocator
= -1;
2173 metaslab_group_sort_impl(mg
, msp
, weight
);
2174 mutex_exit(&mg
->mg_lock
);
2178 metaslab_passivate(metaslab_t
*msp
, uint64_t weight
)
2180 ASSERTV(uint64_t size
= weight
& ~METASLAB_WEIGHT_TYPE
);
2183 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
2184 * this metaslab again. In that case, it had better be empty,
2185 * or we would be leaving space on the table.
2187 ASSERT(!WEIGHT_IS_SPACEBASED(msp
->ms_weight
) ||
2188 size
>= SPA_MINBLOCKSIZE
||
2189 range_tree_space(msp
->ms_allocatable
) == 0);
2190 ASSERT0(weight
& METASLAB_ACTIVE_MASK
);
2192 msp
->ms_activation_weight
= 0;
2193 metaslab_passivate_allocator(msp
->ms_group
, msp
, weight
);
2194 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
2198 * Segment-based metaslabs are activated once and remain active until
2199 * we either fail an allocation attempt (similar to space-based metaslabs)
2200 * or have exhausted the free space in zfs_metaslab_switch_threshold
2201 * buckets since the metaslab was activated. This function checks to see
2202 * if we've exhaused the zfs_metaslab_switch_threshold buckets in the
2203 * metaslab and passivates it proactively. This will allow us to select a
2204 * metaslab with a larger contiguous region, if any, remaining within this
2205 * metaslab group. If we're in sync pass > 1, then we continue using this
2206 * metaslab so that we don't dirty more block and cause more sync passes.
2209 metaslab_segment_may_passivate(metaslab_t
*msp
)
2211 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2213 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
) || spa_sync_pass(spa
) > 1)
2217 * Since we are in the middle of a sync pass, the most accurate
2218 * information that is accessible to us is the in-core range tree
2219 * histogram; calculate the new weight based on that information.
2221 uint64_t weight
= metaslab_weight_from_range_tree(msp
);
2222 int activation_idx
= WEIGHT_GET_INDEX(msp
->ms_activation_weight
);
2223 int current_idx
= WEIGHT_GET_INDEX(weight
);
2225 if (current_idx
<= activation_idx
- zfs_metaslab_switch_threshold
)
2226 metaslab_passivate(msp
, weight
);
2230 metaslab_preload(void *arg
)
2232 metaslab_t
*msp
= arg
;
2233 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2234 fstrans_cookie_t cookie
= spl_fstrans_mark();
2236 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
2238 mutex_enter(&msp
->ms_lock
);
2239 (void) metaslab_load(msp
);
2240 msp
->ms_selected_txg
= spa_syncing_txg(spa
);
2241 mutex_exit(&msp
->ms_lock
);
2242 spl_fstrans_unmark(cookie
);
2246 metaslab_group_preload(metaslab_group_t
*mg
)
2248 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2250 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2253 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
2254 taskq_wait_outstanding(mg
->mg_taskq
, 0);
2258 mutex_enter(&mg
->mg_lock
);
2261 * Load the next potential metaslabs
2263 for (msp
= avl_first(t
); msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
2264 ASSERT3P(msp
->ms_group
, ==, mg
);
2267 * We preload only the maximum number of metaslabs specified
2268 * by metaslab_preload_limit. If a metaslab is being forced
2269 * to condense then we preload it too. This will ensure
2270 * that force condensing happens in the next txg.
2272 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
2276 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
2277 msp
, TQ_SLEEP
) != TASKQID_INVALID
);
2279 mutex_exit(&mg
->mg_lock
);
2283 * Determine if the space map's on-disk footprint is past our tolerance
2284 * for inefficiency. We would like to use the following criteria to make
2287 * 1. The size of the space map object should not dramatically increase as a
2288 * result of writing out the free space range tree.
2290 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2291 * times the size than the free space range tree representation
2292 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1MB).
2294 * 3. The on-disk size of the space map should actually decrease.
2296 * Unfortunately, we cannot compute the on-disk size of the space map in this
2297 * context because we cannot accurately compute the effects of compression, etc.
2298 * Instead, we apply the heuristic described in the block comment for
2299 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2300 * is greater than a threshold number of blocks.
2303 metaslab_should_condense(metaslab_t
*msp
)
2305 space_map_t
*sm
= msp
->ms_sm
;
2306 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
2307 uint64_t vdev_blocksize
= 1 << vd
->vdev_ashift
;
2308 uint64_t current_txg
= spa_syncing_txg(vd
->vdev_spa
);
2310 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2311 ASSERT(msp
->ms_loaded
);
2314 * Allocations and frees in early passes are generally more space
2315 * efficient (in terms of blocks described in space map entries)
2316 * than the ones in later passes (e.g. we don't compress after
2317 * sync pass 5) and condensing a metaslab multiple times in a txg
2318 * could degrade performance.
2320 * Thus we prefer condensing each metaslab at most once every txg at
2321 * the earliest sync pass possible. If a metaslab is eligible for
2322 * condensing again after being considered for condensing within the
2323 * same txg, it will hopefully be dirty in the next txg where it will
2324 * be condensed at an earlier pass.
2326 if (msp
->ms_condense_checked_txg
== current_txg
)
2328 msp
->ms_condense_checked_txg
= current_txg
;
2331 * We always condense metaslabs that are empty and metaslabs for
2332 * which a condense request has been made.
2334 if (avl_is_empty(&msp
->ms_allocatable_by_size
) ||
2335 msp
->ms_condense_wanted
)
2338 uint64_t object_size
= space_map_length(msp
->ms_sm
);
2339 uint64_t optimal_size
= space_map_estimate_optimal_size(sm
,
2340 msp
->ms_allocatable
, SM_NO_VDEVID
);
2342 dmu_object_info_t doi
;
2343 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
2344 uint64_t record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
2346 return (object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
2347 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
2351 * Condense the on-disk space map representation to its minimized form.
2352 * The minimized form consists of a small number of allocations followed by
2353 * the entries of the free range tree.
2356 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
2358 range_tree_t
*condense_tree
;
2359 space_map_t
*sm
= msp
->ms_sm
;
2361 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2362 ASSERT(msp
->ms_loaded
);
2365 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2366 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
2367 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
2368 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
2369 space_map_length(msp
->ms_sm
),
2370 avl_numnodes(&msp
->ms_allocatable
->rt_root
),
2371 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
2373 msp
->ms_condense_wanted
= B_FALSE
;
2376 * Create an range tree that is 100% allocated. We remove segments
2377 * that have been freed in this txg, any deferred frees that exist,
2378 * and any allocation in the future. Removing segments should be
2379 * a relatively inexpensive operation since we expect these trees to
2380 * have a small number of nodes.
2382 condense_tree
= range_tree_create(NULL
, NULL
);
2383 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
2385 range_tree_walk(msp
->ms_freeing
, range_tree_remove
, condense_tree
);
2386 range_tree_walk(msp
->ms_freed
, range_tree_remove
, condense_tree
);
2388 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2389 range_tree_walk(msp
->ms_defer
[t
],
2390 range_tree_remove
, condense_tree
);
2393 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2394 range_tree_walk(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
],
2395 range_tree_remove
, condense_tree
);
2399 * We're about to drop the metaslab's lock thus allowing
2400 * other consumers to change it's content. Set the
2401 * metaslab's ms_condensing flag to ensure that
2402 * allocations on this metaslab do not occur while we're
2403 * in the middle of committing it to disk. This is only critical
2404 * for ms_allocatable as all other range trees use per txg
2405 * views of their content.
2407 msp
->ms_condensing
= B_TRUE
;
2409 mutex_exit(&msp
->ms_lock
);
2410 space_map_truncate(sm
, zfs_metaslab_sm_blksz
, tx
);
2413 * While we would ideally like to create a space map representation
2414 * that consists only of allocation records, doing so can be
2415 * prohibitively expensive because the in-core free tree can be
2416 * large, and therefore computationally expensive to subtract
2417 * from the condense_tree. Instead we sync out two trees, a cheap
2418 * allocation only tree followed by the in-core free tree. While not
2419 * optimal, this is typically close to optimal, and much cheaper to
2422 space_map_write(sm
, condense_tree
, SM_ALLOC
, SM_NO_VDEVID
, tx
);
2423 range_tree_vacate(condense_tree
, NULL
, NULL
);
2424 range_tree_destroy(condense_tree
);
2426 space_map_write(sm
, msp
->ms_allocatable
, SM_FREE
, SM_NO_VDEVID
, tx
);
2427 mutex_enter(&msp
->ms_lock
);
2428 msp
->ms_condensing
= B_FALSE
;
2432 * Write a metaslab to disk in the context of the specified transaction group.
2435 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
2437 metaslab_group_t
*mg
= msp
->ms_group
;
2438 vdev_t
*vd
= mg
->mg_vd
;
2439 spa_t
*spa
= vd
->vdev_spa
;
2440 objset_t
*mos
= spa_meta_objset(spa
);
2441 range_tree_t
*alloctree
= msp
->ms_allocating
[txg
& TXG_MASK
];
2443 uint64_t object
= space_map_object(msp
->ms_sm
);
2445 ASSERT(!vd
->vdev_ishole
);
2448 * This metaslab has just been added so there's no work to do now.
2450 if (msp
->ms_freeing
== NULL
) {
2451 ASSERT3P(alloctree
, ==, NULL
);
2455 ASSERT3P(alloctree
, !=, NULL
);
2456 ASSERT3P(msp
->ms_freeing
, !=, NULL
);
2457 ASSERT3P(msp
->ms_freed
, !=, NULL
);
2458 ASSERT3P(msp
->ms_checkpointing
, !=, NULL
);
2461 * Normally, we don't want to process a metaslab if there are no
2462 * allocations or frees to perform. However, if the metaslab is being
2463 * forced to condense and it's loaded, we need to let it through.
2465 if (range_tree_is_empty(alloctree
) &&
2466 range_tree_is_empty(msp
->ms_freeing
) &&
2467 range_tree_is_empty(msp
->ms_checkpointing
) &&
2468 !(msp
->ms_loaded
&& msp
->ms_condense_wanted
))
2472 VERIFY(txg
<= spa_final_dirty_txg(spa
));
2475 * The only state that can actually be changing concurrently
2476 * with metaslab_sync() is the metaslab's ms_allocatable. No
2477 * other thread can be modifying this txg's alloc, freeing,
2478 * freed, or space_map_phys_t. We drop ms_lock whenever we
2479 * could call into the DMU, because the DMU can call down to
2480 * us (e.g. via zio_free()) at any time.
2482 * The spa_vdev_remove_thread() can be reading metaslab state
2483 * concurrently, and it is locked out by the ms_sync_lock.
2484 * Note that the ms_lock is insufficient for this, because it
2485 * is dropped by space_map_write().
2487 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
2489 if (msp
->ms_sm
== NULL
) {
2490 uint64_t new_object
;
2492 new_object
= space_map_alloc(mos
, zfs_metaslab_sm_blksz
, tx
);
2493 VERIFY3U(new_object
, !=, 0);
2495 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
2496 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
));
2498 ASSERT(msp
->ms_sm
!= NULL
);
2499 ASSERT0(metaslab_allocated_space(msp
));
2502 if (!range_tree_is_empty(msp
->ms_checkpointing
) &&
2503 vd
->vdev_checkpoint_sm
== NULL
) {
2504 ASSERT(spa_has_checkpoint(spa
));
2506 uint64_t new_object
= space_map_alloc(mos
,
2507 vdev_standard_sm_blksz
, tx
);
2508 VERIFY3U(new_object
, !=, 0);
2510 VERIFY0(space_map_open(&vd
->vdev_checkpoint_sm
,
2511 mos
, new_object
, 0, vd
->vdev_asize
, vd
->vdev_ashift
));
2512 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
2515 * We save the space map object as an entry in vdev_top_zap
2516 * so it can be retrieved when the pool is reopened after an
2517 * export or through zdb.
2519 VERIFY0(zap_add(vd
->vdev_spa
->spa_meta_objset
,
2520 vd
->vdev_top_zap
, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM
,
2521 sizeof (new_object
), 1, &new_object
, tx
));
2524 mutex_enter(&msp
->ms_sync_lock
);
2525 mutex_enter(&msp
->ms_lock
);
2528 * Note: metaslab_condense() clears the space map's histogram.
2529 * Therefore we must verify and remove this histogram before
2532 metaslab_group_histogram_verify(mg
);
2533 metaslab_class_histogram_verify(mg
->mg_class
);
2534 metaslab_group_histogram_remove(mg
, msp
);
2536 if (msp
->ms_loaded
&& metaslab_should_condense(msp
)) {
2537 metaslab_condense(msp
, txg
, tx
);
2539 mutex_exit(&msp
->ms_lock
);
2540 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
,
2542 space_map_write(msp
->ms_sm
, msp
->ms_freeing
, SM_FREE
,
2544 mutex_enter(&msp
->ms_lock
);
2547 msp
->ms_allocated_space
+= range_tree_space(alloctree
);
2548 ASSERT3U(msp
->ms_allocated_space
, >=,
2549 range_tree_space(msp
->ms_freeing
));
2550 msp
->ms_allocated_space
-= range_tree_space(msp
->ms_freeing
);
2552 if (!range_tree_is_empty(msp
->ms_checkpointing
)) {
2553 ASSERT(spa_has_checkpoint(spa
));
2554 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
2557 * Since we are doing writes to disk and the ms_checkpointing
2558 * tree won't be changing during that time, we drop the
2559 * ms_lock while writing to the checkpoint space map.
2561 mutex_exit(&msp
->ms_lock
);
2562 space_map_write(vd
->vdev_checkpoint_sm
,
2563 msp
->ms_checkpointing
, SM_FREE
, SM_NO_VDEVID
, tx
);
2564 mutex_enter(&msp
->ms_lock
);
2566 spa
->spa_checkpoint_info
.sci_dspace
+=
2567 range_tree_space(msp
->ms_checkpointing
);
2568 vd
->vdev_stat
.vs_checkpoint_space
+=
2569 range_tree_space(msp
->ms_checkpointing
);
2570 ASSERT3U(vd
->vdev_stat
.vs_checkpoint_space
, ==,
2571 -space_map_allocated(vd
->vdev_checkpoint_sm
));
2573 range_tree_vacate(msp
->ms_checkpointing
, NULL
, NULL
);
2576 if (msp
->ms_loaded
) {
2578 * When the space map is loaded, we have an accurate
2579 * histogram in the range tree. This gives us an opportunity
2580 * to bring the space map's histogram up-to-date so we clear
2581 * it first before updating it.
2583 space_map_histogram_clear(msp
->ms_sm
);
2584 space_map_histogram_add(msp
->ms_sm
, msp
->ms_allocatable
, tx
);
2587 * Since we've cleared the histogram we need to add back
2588 * any free space that has already been processed, plus
2589 * any deferred space. This allows the on-disk histogram
2590 * to accurately reflect all free space even if some space
2591 * is not yet available for allocation (i.e. deferred).
2593 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freed
, tx
);
2596 * Add back any deferred free space that has not been
2597 * added back into the in-core free tree yet. This will
2598 * ensure that we don't end up with a space map histogram
2599 * that is completely empty unless the metaslab is fully
2602 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2603 space_map_histogram_add(msp
->ms_sm
,
2604 msp
->ms_defer
[t
], tx
);
2609 * Always add the free space from this sync pass to the space
2610 * map histogram. We want to make sure that the on-disk histogram
2611 * accounts for all free space. If the space map is not loaded,
2612 * then we will lose some accuracy but will correct it the next
2613 * time we load the space map.
2615 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freeing
, tx
);
2617 metaslab_group_histogram_add(mg
, msp
);
2618 metaslab_group_histogram_verify(mg
);
2619 metaslab_class_histogram_verify(mg
->mg_class
);
2622 * For sync pass 1, we avoid traversing this txg's free range tree
2623 * and instead will just swap the pointers for freeing and freed.
2624 * We can safely do this since the freed_tree is guaranteed to be
2625 * empty on the initial pass.
2627 if (spa_sync_pass(spa
) == 1) {
2628 range_tree_swap(&msp
->ms_freeing
, &msp
->ms_freed
);
2629 ASSERT0(msp
->ms_allocated_this_txg
);
2631 range_tree_vacate(msp
->ms_freeing
,
2632 range_tree_add
, msp
->ms_freed
);
2634 msp
->ms_allocated_this_txg
+= range_tree_space(alloctree
);
2635 range_tree_vacate(alloctree
, NULL
, NULL
);
2637 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
2638 ASSERT0(range_tree_space(msp
->ms_allocating
[TXG_CLEAN(txg
)
2640 ASSERT0(range_tree_space(msp
->ms_freeing
));
2641 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2643 mutex_exit(&msp
->ms_lock
);
2645 if (object
!= space_map_object(msp
->ms_sm
)) {
2646 object
= space_map_object(msp
->ms_sm
);
2647 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
2648 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
2650 mutex_exit(&msp
->ms_sync_lock
);
2655 * Called after a transaction group has completely synced to mark
2656 * all of the metaslab's free space as usable.
2659 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
2661 metaslab_group_t
*mg
= msp
->ms_group
;
2662 vdev_t
*vd
= mg
->mg_vd
;
2663 spa_t
*spa
= vd
->vdev_spa
;
2664 range_tree_t
**defer_tree
;
2665 int64_t alloc_delta
, defer_delta
;
2666 boolean_t defer_allowed
= B_TRUE
;
2668 ASSERT(!vd
->vdev_ishole
);
2670 mutex_enter(&msp
->ms_lock
);
2673 * If this metaslab is just becoming available, initialize its
2674 * range trees and add its capacity to the vdev.
2676 if (msp
->ms_freed
== NULL
) {
2677 for (int t
= 0; t
< TXG_SIZE
; t
++) {
2678 ASSERT(msp
->ms_allocating
[t
] == NULL
);
2680 msp
->ms_allocating
[t
] = range_tree_create(NULL
, NULL
);
2683 ASSERT3P(msp
->ms_freeing
, ==, NULL
);
2684 msp
->ms_freeing
= range_tree_create(NULL
, NULL
);
2686 ASSERT3P(msp
->ms_freed
, ==, NULL
);
2687 msp
->ms_freed
= range_tree_create(NULL
, NULL
);
2689 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2690 ASSERT(msp
->ms_defer
[t
] == NULL
);
2692 msp
->ms_defer
[t
] = range_tree_create(NULL
, NULL
);
2695 ASSERT3P(msp
->ms_checkpointing
, ==, NULL
);
2696 msp
->ms_checkpointing
= range_tree_create(NULL
, NULL
);
2698 metaslab_space_update(vd
, mg
->mg_class
, 0, 0, msp
->ms_size
);
2700 ASSERT0(range_tree_space(msp
->ms_freeing
));
2701 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2703 defer_tree
= &msp
->ms_defer
[txg
% TXG_DEFER_SIZE
];
2705 uint64_t free_space
= metaslab_class_get_space(spa_normal_class(spa
)) -
2706 metaslab_class_get_alloc(spa_normal_class(spa
));
2707 if (free_space
<= spa_get_slop_space(spa
) || vd
->vdev_removing
) {
2708 defer_allowed
= B_FALSE
;
2712 alloc_delta
= msp
->ms_allocated_this_txg
-
2713 range_tree_space(msp
->ms_freed
);
2714 if (defer_allowed
) {
2715 defer_delta
= range_tree_space(msp
->ms_freed
) -
2716 range_tree_space(*defer_tree
);
2718 defer_delta
-= range_tree_space(*defer_tree
);
2721 metaslab_space_update(vd
, mg
->mg_class
, alloc_delta
+ defer_delta
,
2725 * If there's a metaslab_load() in progress, wait for it to complete
2726 * so that we have a consistent view of the in-core space map.
2728 metaslab_load_wait(msp
);
2731 * Move the frees from the defer_tree back to the free
2732 * range tree (if it's loaded). Swap the freed_tree and
2733 * the defer_tree -- this is safe to do because we've
2734 * just emptied out the defer_tree.
2736 range_tree_vacate(*defer_tree
,
2737 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_allocatable
);
2738 if (defer_allowed
) {
2739 range_tree_swap(&msp
->ms_freed
, defer_tree
);
2741 range_tree_vacate(msp
->ms_freed
,
2742 msp
->ms_loaded
? range_tree_add
: NULL
,
2743 msp
->ms_allocatable
);
2746 msp
->ms_synced_length
= space_map_length(msp
->ms_sm
);
2748 msp
->ms_deferspace
+= defer_delta
;
2749 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2750 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2751 if (msp
->ms_deferspace
!= 0) {
2753 * Keep syncing this metaslab until all deferred frees
2754 * are back in circulation.
2756 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2760 msp
->ms_new
= B_FALSE
;
2761 mutex_enter(&mg
->mg_lock
);
2763 mutex_exit(&mg
->mg_lock
);
2766 * Calculate the new weights before unloading any metaslabs.
2767 * This will give us the most accurate weighting.
2769 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
) |
2770 (msp
->ms_weight
& METASLAB_ACTIVE_MASK
));
2773 * If the metaslab is loaded and we've not tried to load or allocate
2774 * from it in 'metaslab_unload_delay' txgs, then unload it.
2776 if (msp
->ms_loaded
&&
2777 msp
->ms_initializing
== 0 &&
2778 msp
->ms_selected_txg
+ metaslab_unload_delay
< txg
) {
2780 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2781 VERIFY0(range_tree_space(
2782 msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]));
2784 if (msp
->ms_allocator
!= -1) {
2785 metaslab_passivate(msp
, msp
->ms_weight
&
2786 ~METASLAB_ACTIVE_MASK
);
2789 if (!metaslab_debug_unload
)
2790 metaslab_unload(msp
);
2793 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
2794 ASSERT0(range_tree_space(msp
->ms_freeing
));
2795 ASSERT0(range_tree_space(msp
->ms_freed
));
2796 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
2798 msp
->ms_allocated_this_txg
= 0;
2799 mutex_exit(&msp
->ms_lock
);
2803 metaslab_sync_reassess(metaslab_group_t
*mg
)
2805 spa_t
*spa
= mg
->mg_class
->mc_spa
;
2807 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2808 metaslab_group_alloc_update(mg
);
2809 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2812 * Preload the next potential metaslabs but only on active
2813 * metaslab groups. We can get into a state where the metaslab
2814 * is no longer active since we dirty metaslabs as we remove a
2815 * a device, thus potentially making the metaslab group eligible
2818 if (mg
->mg_activation_count
> 0) {
2819 metaslab_group_preload(mg
);
2821 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2825 * When writing a ditto block (i.e. more than one DVA for a given BP) on
2826 * the same vdev as an existing DVA of this BP, then try to allocate it
2827 * on a different metaslab than existing DVAs (i.e. a unique metaslab).
2830 metaslab_is_unique(metaslab_t
*msp
, dva_t
*dva
)
2834 if (DVA_GET_ASIZE(dva
) == 0)
2837 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2840 dva_ms_id
= DVA_GET_OFFSET(dva
) >> msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2842 return (msp
->ms_id
!= dva_ms_id
);
2846 * ==========================================================================
2847 * Metaslab allocation tracing facility
2848 * ==========================================================================
2850 #ifdef _METASLAB_TRACING
2851 kstat_t
*metaslab_trace_ksp
;
2852 kstat_named_t metaslab_trace_over_limit
;
2855 metaslab_alloc_trace_init(void)
2857 ASSERT(metaslab_alloc_trace_cache
== NULL
);
2858 metaslab_alloc_trace_cache
= kmem_cache_create(
2859 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t
),
2860 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
2861 metaslab_trace_ksp
= kstat_create("zfs", 0, "metaslab_trace_stats",
2862 "misc", KSTAT_TYPE_NAMED
, 1, KSTAT_FLAG_VIRTUAL
);
2863 if (metaslab_trace_ksp
!= NULL
) {
2864 metaslab_trace_ksp
->ks_data
= &metaslab_trace_over_limit
;
2865 kstat_named_init(&metaslab_trace_over_limit
,
2866 "metaslab_trace_over_limit", KSTAT_DATA_UINT64
);
2867 kstat_install(metaslab_trace_ksp
);
2872 metaslab_alloc_trace_fini(void)
2874 if (metaslab_trace_ksp
!= NULL
) {
2875 kstat_delete(metaslab_trace_ksp
);
2876 metaslab_trace_ksp
= NULL
;
2878 kmem_cache_destroy(metaslab_alloc_trace_cache
);
2879 metaslab_alloc_trace_cache
= NULL
;
2883 * Add an allocation trace element to the allocation tracing list.
2886 metaslab_trace_add(zio_alloc_list_t
*zal
, metaslab_group_t
*mg
,
2887 metaslab_t
*msp
, uint64_t psize
, uint32_t dva_id
, uint64_t offset
,
2890 metaslab_alloc_trace_t
*mat
;
2892 if (!metaslab_trace_enabled
)
2896 * When the tracing list reaches its maximum we remove
2897 * the second element in the list before adding a new one.
2898 * By removing the second element we preserve the original
2899 * entry as a clue to what allocations steps have already been
2902 if (zal
->zal_size
== metaslab_trace_max_entries
) {
2903 metaslab_alloc_trace_t
*mat_next
;
2905 panic("too many entries in allocation list");
2907 atomic_inc_64(&metaslab_trace_over_limit
.value
.ui64
);
2909 mat_next
= list_next(&zal
->zal_list
, list_head(&zal
->zal_list
));
2910 list_remove(&zal
->zal_list
, mat_next
);
2911 kmem_cache_free(metaslab_alloc_trace_cache
, mat_next
);
2914 mat
= kmem_cache_alloc(metaslab_alloc_trace_cache
, KM_SLEEP
);
2915 list_link_init(&mat
->mat_list_node
);
2918 mat
->mat_size
= psize
;
2919 mat
->mat_dva_id
= dva_id
;
2920 mat
->mat_offset
= offset
;
2921 mat
->mat_weight
= 0;
2922 mat
->mat_allocator
= allocator
;
2925 mat
->mat_weight
= msp
->ms_weight
;
2928 * The list is part of the zio so locking is not required. Only
2929 * a single thread will perform allocations for a given zio.
2931 list_insert_tail(&zal
->zal_list
, mat
);
2934 ASSERT3U(zal
->zal_size
, <=, metaslab_trace_max_entries
);
2938 metaslab_trace_init(zio_alloc_list_t
*zal
)
2940 list_create(&zal
->zal_list
, sizeof (metaslab_alloc_trace_t
),
2941 offsetof(metaslab_alloc_trace_t
, mat_list_node
));
2946 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2948 metaslab_alloc_trace_t
*mat
;
2950 while ((mat
= list_remove_head(&zal
->zal_list
)) != NULL
)
2951 kmem_cache_free(metaslab_alloc_trace_cache
, mat
);
2952 list_destroy(&zal
->zal_list
);
2957 #define metaslab_trace_add(zal, mg, msp, psize, id, off, alloc)
2960 metaslab_alloc_trace_init(void)
2965 metaslab_alloc_trace_fini(void)
2970 metaslab_trace_init(zio_alloc_list_t
*zal
)
2975 metaslab_trace_fini(zio_alloc_list_t
*zal
)
2979 #endif /* _METASLAB_TRACING */
2982 * ==========================================================================
2983 * Metaslab block operations
2984 * ==========================================================================
2988 metaslab_group_alloc_increment(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
,
2991 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
2992 (flags
& METASLAB_DONT_THROTTLE
))
2995 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
2996 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
2999 (void) zfs_refcount_add(&mg
->mg_alloc_queue_depth
[allocator
], tag
);
3003 metaslab_group_increment_qdepth(metaslab_group_t
*mg
, int allocator
)
3005 uint64_t max
= mg
->mg_max_alloc_queue_depth
;
3006 uint64_t cur
= mg
->mg_cur_max_alloc_queue_depth
[allocator
];
3008 if (atomic_cas_64(&mg
->mg_cur_max_alloc_queue_depth
[allocator
],
3009 cur
, cur
+ 1) == cur
) {
3011 &mg
->mg_class
->mc_alloc_max_slots
[allocator
]);
3014 cur
= mg
->mg_cur_max_alloc_queue_depth
[allocator
];
3019 metaslab_group_alloc_decrement(spa_t
*spa
, uint64_t vdev
, void *tag
, int flags
,
3020 int allocator
, boolean_t io_complete
)
3022 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
3023 (flags
& METASLAB_DONT_THROTTLE
))
3026 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
3027 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
3030 (void) zfs_refcount_remove(&mg
->mg_alloc_queue_depth
[allocator
], tag
);
3032 metaslab_group_increment_qdepth(mg
, allocator
);
3036 metaslab_group_alloc_verify(spa_t
*spa
, const blkptr_t
*bp
, void *tag
,
3040 const dva_t
*dva
= bp
->blk_dva
;
3041 int ndvas
= BP_GET_NDVAS(bp
);
3043 for (int d
= 0; d
< ndvas
; d
++) {
3044 uint64_t vdev
= DVA_GET_VDEV(&dva
[d
]);
3045 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
3046 VERIFY(zfs_refcount_not_held(
3047 &mg
->mg_alloc_queue_depth
[allocator
], tag
));
3053 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
, uint64_t txg
)
3056 range_tree_t
*rt
= msp
->ms_allocatable
;
3057 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
3059 VERIFY(!msp
->ms_condensing
);
3060 VERIFY0(msp
->ms_initializing
);
3062 start
= mc
->mc_ops
->msop_alloc(msp
, size
);
3063 if (start
!= -1ULL) {
3064 metaslab_group_t
*mg
= msp
->ms_group
;
3065 vdev_t
*vd
= mg
->mg_vd
;
3067 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
3068 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3069 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
3070 range_tree_remove(rt
, start
, size
);
3072 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
3073 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
3075 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
], start
, size
);
3077 /* Track the last successful allocation */
3078 msp
->ms_alloc_txg
= txg
;
3079 metaslab_verify_space(msp
, txg
);
3083 * Now that we've attempted the allocation we need to update the
3084 * metaslab's maximum block size since it may have changed.
3086 msp
->ms_max_size
= metaslab_block_maxsize(msp
);
3091 * Find the metaslab with the highest weight that is less than what we've
3092 * already tried. In the common case, this means that we will examine each
3093 * metaslab at most once. Note that concurrent callers could reorder metaslabs
3094 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
3095 * activated by another thread, and we fail to allocate from the metaslab we
3096 * have selected, we may not try the newly-activated metaslab, and instead
3097 * activate another metaslab. This is not optimal, but generally does not cause
3098 * any problems (a possible exception being if every metaslab is completely full
3099 * except for the the newly-activated metaslab which we fail to examine).
3102 find_valid_metaslab(metaslab_group_t
*mg
, uint64_t activation_weight
,
3103 dva_t
*dva
, int d
, boolean_t want_unique
, uint64_t asize
, int allocator
,
3104 zio_alloc_list_t
*zal
, metaslab_t
*search
, boolean_t
*was_active
)
3107 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
3108 metaslab_t
*msp
= avl_find(t
, search
, &idx
);
3110 msp
= avl_nearest(t
, idx
, AVL_AFTER
);
3112 for (; msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
3114 if (!metaslab_should_allocate(msp
, asize
)) {
3115 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3116 TRACE_TOO_SMALL
, allocator
);
3121 * If the selected metaslab is condensing or being
3122 * initialized, skip it.
3124 if (msp
->ms_condensing
|| msp
->ms_initializing
> 0)
3127 *was_active
= msp
->ms_allocator
!= -1;
3129 * If we're activating as primary, this is our first allocation
3130 * from this disk, so we don't need to check how close we are.
3131 * If the metaslab under consideration was already active,
3132 * we're getting desperate enough to steal another allocator's
3133 * metaslab, so we still don't care about distances.
3135 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
|| *was_active
)
3138 for (i
= 0; i
< d
; i
++) {
3140 !metaslab_is_unique(msp
, &dva
[i
]))
3141 break; /* try another metaslab */
3148 search
->ms_weight
= msp
->ms_weight
;
3149 search
->ms_start
= msp
->ms_start
+ 1;
3150 search
->ms_allocator
= msp
->ms_allocator
;
3151 search
->ms_primary
= msp
->ms_primary
;
3158 metaslab_group_alloc_normal(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
3159 uint64_t asize
, uint64_t txg
, boolean_t want_unique
, dva_t
*dva
,
3160 int d
, int allocator
)
3162 metaslab_t
*msp
= NULL
;
3163 uint64_t offset
= -1ULL;
3164 uint64_t activation_weight
;
3166 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
3167 for (int i
= 0; i
< d
; i
++) {
3168 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
&&
3169 DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
3170 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
3171 } else if (activation_weight
== METASLAB_WEIGHT_SECONDARY
&&
3172 DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
3173 activation_weight
= METASLAB_WEIGHT_CLAIM
;
3179 * If we don't have enough metaslabs active to fill the entire array, we
3180 * just use the 0th slot.
3182 if (mg
->mg_ms_ready
< mg
->mg_allocators
* 3)
3185 ASSERT3U(mg
->mg_vd
->vdev_ms_count
, >=, 2);
3187 metaslab_t
*search
= kmem_alloc(sizeof (*search
), KM_SLEEP
);
3188 search
->ms_weight
= UINT64_MAX
;
3189 search
->ms_start
= 0;
3191 * At the end of the metaslab tree are the already-active metaslabs,
3192 * first the primaries, then the secondaries. When we resume searching
3193 * through the tree, we need to consider ms_allocator and ms_primary so
3194 * we start in the location right after where we left off, and don't
3195 * accidentally loop forever considering the same metaslabs.
3197 search
->ms_allocator
= -1;
3198 search
->ms_primary
= B_TRUE
;
3200 boolean_t was_active
= B_FALSE
;
3202 mutex_enter(&mg
->mg_lock
);
3204 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
&&
3205 mg
->mg_primaries
[allocator
] != NULL
) {
3206 msp
= mg
->mg_primaries
[allocator
];
3207 was_active
= B_TRUE
;
3208 } else if (activation_weight
== METASLAB_WEIGHT_SECONDARY
&&
3209 mg
->mg_secondaries
[allocator
] != NULL
) {
3210 msp
= mg
->mg_secondaries
[allocator
];
3211 was_active
= B_TRUE
;
3213 msp
= find_valid_metaslab(mg
, activation_weight
, dva
, d
,
3214 want_unique
, asize
, allocator
, zal
, search
,
3218 mutex_exit(&mg
->mg_lock
);
3220 kmem_free(search
, sizeof (*search
));
3224 mutex_enter(&msp
->ms_lock
);
3226 * Ensure that the metaslab we have selected is still
3227 * capable of handling our request. It's possible that
3228 * another thread may have changed the weight while we
3229 * were blocked on the metaslab lock. We check the
3230 * active status first to see if we need to reselect
3233 if (was_active
&& !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
)) {
3234 mutex_exit(&msp
->ms_lock
);
3239 * If the metaslab is freshly activated for an allocator that
3240 * isn't the one we're allocating from, or if it's a primary and
3241 * we're seeking a secondary (or vice versa), we go back and
3242 * select a new metaslab.
3244 if (!was_active
&& (msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
3245 (msp
->ms_allocator
!= -1) &&
3246 (msp
->ms_allocator
!= allocator
|| ((activation_weight
==
3247 METASLAB_WEIGHT_PRIMARY
) != msp
->ms_primary
))) {
3248 mutex_exit(&msp
->ms_lock
);
3252 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
&&
3253 activation_weight
!= METASLAB_WEIGHT_CLAIM
) {
3254 metaslab_passivate(msp
, msp
->ms_weight
&
3255 ~METASLAB_WEIGHT_CLAIM
);
3256 mutex_exit(&msp
->ms_lock
);
3260 if (metaslab_activate(msp
, allocator
, activation_weight
) != 0) {
3261 mutex_exit(&msp
->ms_lock
);
3265 msp
->ms_selected_txg
= txg
;
3268 * Now that we have the lock, recheck to see if we should
3269 * continue to use this metaslab for this allocation. The
3270 * the metaslab is now loaded so metaslab_should_allocate() can
3271 * accurately determine if the allocation attempt should
3274 if (!metaslab_should_allocate(msp
, asize
)) {
3275 /* Passivate this metaslab and select a new one. */
3276 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3277 TRACE_TOO_SMALL
, allocator
);
3283 * If this metaslab is currently condensing then pick again as
3284 * we can't manipulate this metaslab until it's committed
3285 * to disk. If this metaslab is being initialized, we shouldn't
3286 * allocate from it since the allocated region might be
3287 * overwritten after allocation.
3289 if (msp
->ms_condensing
) {
3290 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3291 TRACE_CONDENSING
, allocator
);
3292 metaslab_passivate(msp
, msp
->ms_weight
&
3293 ~METASLAB_ACTIVE_MASK
);
3294 mutex_exit(&msp
->ms_lock
);
3296 } else if (msp
->ms_initializing
> 0) {
3297 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
3298 TRACE_INITIALIZING
, allocator
);
3299 metaslab_passivate(msp
, msp
->ms_weight
&
3300 ~METASLAB_ACTIVE_MASK
);
3301 mutex_exit(&msp
->ms_lock
);
3305 offset
= metaslab_block_alloc(msp
, asize
, txg
);
3306 metaslab_trace_add(zal
, mg
, msp
, asize
, d
, offset
, allocator
);
3308 if (offset
!= -1ULL) {
3309 /* Proactively passivate the metaslab, if needed */
3310 metaslab_segment_may_passivate(msp
);
3314 ASSERT(msp
->ms_loaded
);
3317 * We were unable to allocate from this metaslab so determine
3318 * a new weight for this metaslab. Now that we have loaded
3319 * the metaslab we can provide a better hint to the metaslab
3322 * For space-based metaslabs, we use the maximum block size.
3323 * This information is only available when the metaslab
3324 * is loaded and is more accurate than the generic free
3325 * space weight that was calculated by metaslab_weight().
3326 * This information allows us to quickly compare the maximum
3327 * available allocation in the metaslab to the allocation
3328 * size being requested.
3330 * For segment-based metaslabs, determine the new weight
3331 * based on the highest bucket in the range tree. We
3332 * explicitly use the loaded segment weight (i.e. the range
3333 * tree histogram) since it contains the space that is
3334 * currently available for allocation and is accurate
3335 * even within a sync pass.
3337 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
3338 uint64_t weight
= metaslab_block_maxsize(msp
);
3339 WEIGHT_SET_SPACEBASED(weight
);
3340 metaslab_passivate(msp
, weight
);
3342 metaslab_passivate(msp
,
3343 metaslab_weight_from_range_tree(msp
));
3347 * We have just failed an allocation attempt, check
3348 * that metaslab_should_allocate() agrees. Otherwise,
3349 * we may end up in an infinite loop retrying the same
3352 ASSERT(!metaslab_should_allocate(msp
, asize
));
3354 mutex_exit(&msp
->ms_lock
);
3356 mutex_exit(&msp
->ms_lock
);
3357 kmem_free(search
, sizeof (*search
));
3362 metaslab_group_alloc(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
3363 uint64_t asize
, uint64_t txg
, boolean_t want_unique
, dva_t
*dva
,
3364 int d
, int allocator
)
3367 ASSERT(mg
->mg_initialized
);
3369 offset
= metaslab_group_alloc_normal(mg
, zal
, asize
, txg
, want_unique
,
3372 mutex_enter(&mg
->mg_lock
);
3373 if (offset
== -1ULL) {
3374 mg
->mg_failed_allocations
++;
3375 metaslab_trace_add(zal
, mg
, NULL
, asize
, d
,
3376 TRACE_GROUP_FAILURE
, allocator
);
3377 if (asize
== SPA_GANGBLOCKSIZE
) {
3379 * This metaslab group was unable to allocate
3380 * the minimum gang block size so it must be out of
3381 * space. We must notify the allocation throttle
3382 * to start skipping allocation attempts to this
3383 * metaslab group until more space becomes available.
3384 * Note: this failure cannot be caused by the
3385 * allocation throttle since the allocation throttle
3386 * is only responsible for skipping devices and
3387 * not failing block allocations.
3389 mg
->mg_no_free_space
= B_TRUE
;
3392 mg
->mg_allocations
++;
3393 mutex_exit(&mg
->mg_lock
);
3398 * Allocate a block for the specified i/o.
3401 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
3402 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
,
3403 zio_alloc_list_t
*zal
, int allocator
)
3405 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
3407 boolean_t try_hard
= B_FALSE
;
3409 ASSERT(!DVA_IS_VALID(&dva
[d
]));
3412 * For testing, make some blocks above a certain size be gang blocks.
3413 * This will result in more split blocks when using device removal,
3414 * and a large number of split blocks coupled with ztest-induced
3415 * damage can result in extremely long reconstruction times. This
3416 * will also test spilling from special to normal.
3418 if (psize
>= metaslab_force_ganging
&& (spa_get_random(100) < 3)) {
3419 metaslab_trace_add(zal
, NULL
, NULL
, psize
, d
, TRACE_FORCE_GANG
,
3421 return (SET_ERROR(ENOSPC
));
3425 * Start at the rotor and loop through all mgs until we find something.
3426 * Note that there's no locking on mc_rotor or mc_aliquot because
3427 * nothing actually breaks if we miss a few updates -- we just won't
3428 * allocate quite as evenly. It all balances out over time.
3430 * If we are doing ditto or log blocks, try to spread them across
3431 * consecutive vdevs. If we're forced to reuse a vdev before we've
3432 * allocated all of our ditto blocks, then try and spread them out on
3433 * that vdev as much as possible. If it turns out to not be possible,
3434 * gradually lower our standards until anything becomes acceptable.
3435 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3436 * gives us hope of containing our fault domains to something we're
3437 * able to reason about. Otherwise, any two top-level vdev failures
3438 * will guarantee the loss of data. With consecutive allocation,
3439 * only two adjacent top-level vdev failures will result in data loss.
3441 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3442 * ourselves on the same vdev as our gang block header. That
3443 * way, we can hope for locality in vdev_cache, plus it makes our
3444 * fault domains something tractable.
3447 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
3450 * It's possible the vdev we're using as the hint no
3451 * longer exists or its mg has been closed (e.g. by
3452 * device removal). Consult the rotor when
3455 if (vd
!= NULL
&& vd
->vdev_mg
!= NULL
) {
3458 if (flags
& METASLAB_HINTBP_AVOID
&&
3459 mg
->mg_next
!= NULL
)
3464 } else if (d
!= 0) {
3465 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
3466 mg
= vd
->vdev_mg
->mg_next
;
3467 } else if (flags
& METASLAB_FASTWRITE
) {
3468 mg
= fast_mg
= mc
->mc_rotor
;
3471 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
3472 mg
->mg_vd
->vdev_pending_fastwrite
)
3474 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
3477 ASSERT(mc
->mc_rotor
!= NULL
);
3482 * If the hint put us into the wrong metaslab class, or into a
3483 * metaslab group that has been passivated, just follow the rotor.
3485 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
3491 boolean_t allocatable
;
3493 ASSERT(mg
->mg_activation_count
== 1);
3497 * Don't allocate from faulted devices.
3500 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
3501 allocatable
= vdev_allocatable(vd
);
3502 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
3504 allocatable
= vdev_allocatable(vd
);
3508 * Determine if the selected metaslab group is eligible
3509 * for allocations. If we're ganging then don't allow
3510 * this metaslab group to skip allocations since that would
3511 * inadvertently return ENOSPC and suspend the pool
3512 * even though space is still available.
3514 if (allocatable
&& !GANG_ALLOCATION(flags
) && !try_hard
) {
3515 allocatable
= metaslab_group_allocatable(mg
, rotor
,
3516 psize
, allocator
, d
);
3520 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3521 TRACE_NOT_ALLOCATABLE
, allocator
);
3525 ASSERT(mg
->mg_initialized
);
3528 * Avoid writing single-copy data to a failing,
3529 * non-redundant vdev, unless we've already tried all
3532 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
3533 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
3534 d
== 0 && !try_hard
&& vd
->vdev_children
== 0) {
3535 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
3536 TRACE_VDEV_ERROR
, allocator
);
3540 ASSERT(mg
->mg_class
== mc
);
3542 uint64_t asize
= vdev_psize_to_asize(vd
, psize
);
3543 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
3546 * If we don't need to try hard, then require that the
3547 * block be on an different metaslab from any other DVAs
3548 * in this BP (unique=true). If we are trying hard, then
3549 * allow any metaslab to be used (unique=false).
3551 uint64_t offset
= metaslab_group_alloc(mg
, zal
, asize
, txg
,
3552 !try_hard
, dva
, d
, allocator
);
3554 if (offset
!= -1ULL) {
3556 * If we've just selected this metaslab group,
3557 * figure out whether the corresponding vdev is
3558 * over- or under-used relative to the pool,
3559 * and set an allocation bias to even it out.
3561 * Bias is also used to compensate for unequally
3562 * sized vdevs so that space is allocated fairly.
3564 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
3565 vdev_stat_t
*vs
= &vd
->vdev_stat
;
3566 int64_t vs_free
= vs
->vs_space
- vs
->vs_alloc
;
3567 int64_t mc_free
= mc
->mc_space
- mc
->mc_alloc
;
3571 * Calculate how much more or less we should
3572 * try to allocate from this device during
3573 * this iteration around the rotor.
3575 * This basically introduces a zero-centered
3576 * bias towards the devices with the most
3577 * free space, while compensating for vdev
3581 * vdev V1 = 16M/128M
3582 * vdev V2 = 16M/128M
3583 * ratio(V1) = 100% ratio(V2) = 100%
3585 * vdev V1 = 16M/128M
3586 * vdev V2 = 64M/128M
3587 * ratio(V1) = 127% ratio(V2) = 72%
3589 * vdev V1 = 16M/128M
3590 * vdev V2 = 64M/512M
3591 * ratio(V1) = 40% ratio(V2) = 160%
3593 ratio
= (vs_free
* mc
->mc_alloc_groups
* 100) /
3595 mg
->mg_bias
= ((ratio
- 100) *
3596 (int64_t)mg
->mg_aliquot
) / 100;
3597 } else if (!metaslab_bias_enabled
) {
3601 if ((flags
& METASLAB_FASTWRITE
) ||
3602 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
3603 mg
->mg_aliquot
+ mg
->mg_bias
) {
3604 mc
->mc_rotor
= mg
->mg_next
;
3608 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
3609 DVA_SET_OFFSET(&dva
[d
], offset
);
3610 DVA_SET_GANG(&dva
[d
],
3611 ((flags
& METASLAB_GANG_HEADER
) ? 1 : 0));
3612 DVA_SET_ASIZE(&dva
[d
], asize
);
3614 if (flags
& METASLAB_FASTWRITE
) {
3615 atomic_add_64(&vd
->vdev_pending_fastwrite
,
3622 mc
->mc_rotor
= mg
->mg_next
;
3624 } while ((mg
= mg
->mg_next
) != rotor
);
3627 * If we haven't tried hard, do so now.
3634 bzero(&dva
[d
], sizeof (dva_t
));
3636 metaslab_trace_add(zal
, rotor
, NULL
, psize
, d
, TRACE_ENOSPC
, allocator
);
3637 return (SET_ERROR(ENOSPC
));
3641 metaslab_free_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t asize
,
3642 boolean_t checkpoint
)
3645 spa_t
*spa
= vd
->vdev_spa
;
3647 ASSERT(vdev_is_concrete(vd
));
3648 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3649 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
3651 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3653 VERIFY(!msp
->ms_condensing
);
3654 VERIFY3U(offset
, >=, msp
->ms_start
);
3655 VERIFY3U(offset
+ asize
, <=, msp
->ms_start
+ msp
->ms_size
);
3656 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3657 VERIFY0(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
));
3659 metaslab_check_free_impl(vd
, offset
, asize
);
3661 mutex_enter(&msp
->ms_lock
);
3662 if (range_tree_is_empty(msp
->ms_freeing
) &&
3663 range_tree_is_empty(msp
->ms_checkpointing
)) {
3664 vdev_dirty(vd
, VDD_METASLAB
, msp
, spa_syncing_txg(spa
));
3668 ASSERT(spa_has_checkpoint(spa
));
3669 range_tree_add(msp
->ms_checkpointing
, offset
, asize
);
3671 range_tree_add(msp
->ms_freeing
, offset
, asize
);
3673 mutex_exit(&msp
->ms_lock
);
3678 metaslab_free_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3679 uint64_t size
, void *arg
)
3681 boolean_t
*checkpoint
= arg
;
3683 ASSERT3P(checkpoint
, !=, NULL
);
3685 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
)
3686 vdev_indirect_mark_obsolete(vd
, offset
, size
);
3688 metaslab_free_impl(vd
, offset
, size
, *checkpoint
);
3692 metaslab_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
3693 boolean_t checkpoint
)
3695 spa_t
*spa
= vd
->vdev_spa
;
3697 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3699 if (spa_syncing_txg(spa
) > spa_freeze_txg(spa
))
3702 if (spa
->spa_vdev_removal
!= NULL
&&
3703 spa
->spa_vdev_removal
->svr_vdev_id
== vd
->vdev_id
&&
3704 vdev_is_concrete(vd
)) {
3706 * Note: we check if the vdev is concrete because when
3707 * we complete the removal, we first change the vdev to be
3708 * an indirect vdev (in open context), and then (in syncing
3709 * context) clear spa_vdev_removal.
3711 free_from_removing_vdev(vd
, offset
, size
);
3712 } else if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
3713 vdev_indirect_mark_obsolete(vd
, offset
, size
);
3714 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
3715 metaslab_free_impl_cb
, &checkpoint
);
3717 metaslab_free_concrete(vd
, offset
, size
, checkpoint
);
3721 typedef struct remap_blkptr_cb_arg
{
3723 spa_remap_cb_t rbca_cb
;
3724 vdev_t
*rbca_remap_vd
;
3725 uint64_t rbca_remap_offset
;
3727 } remap_blkptr_cb_arg_t
;
3730 remap_blkptr_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
3731 uint64_t size
, void *arg
)
3733 remap_blkptr_cb_arg_t
*rbca
= arg
;
3734 blkptr_t
*bp
= rbca
->rbca_bp
;
3736 /* We can not remap split blocks. */
3737 if (size
!= DVA_GET_ASIZE(&bp
->blk_dva
[0]))
3739 ASSERT0(inner_offset
);
3741 if (rbca
->rbca_cb
!= NULL
) {
3743 * At this point we know that we are not handling split
3744 * blocks and we invoke the callback on the previous
3745 * vdev which must be indirect.
3747 ASSERT3P(rbca
->rbca_remap_vd
->vdev_ops
, ==, &vdev_indirect_ops
);
3749 rbca
->rbca_cb(rbca
->rbca_remap_vd
->vdev_id
,
3750 rbca
->rbca_remap_offset
, size
, rbca
->rbca_cb_arg
);
3752 /* set up remap_blkptr_cb_arg for the next call */
3753 rbca
->rbca_remap_vd
= vd
;
3754 rbca
->rbca_remap_offset
= offset
;
3758 * The phys birth time is that of dva[0]. This ensures that we know
3759 * when each dva was written, so that resilver can determine which
3760 * blocks need to be scrubbed (i.e. those written during the time
3761 * the vdev was offline). It also ensures that the key used in
3762 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
3763 * we didn't change the phys_birth, a lookup in the ARC for a
3764 * remapped BP could find the data that was previously stored at
3765 * this vdev + offset.
3767 vdev_t
*oldvd
= vdev_lookup_top(vd
->vdev_spa
,
3768 DVA_GET_VDEV(&bp
->blk_dva
[0]));
3769 vdev_indirect_births_t
*vib
= oldvd
->vdev_indirect_births
;
3770 bp
->blk_phys_birth
= vdev_indirect_births_physbirth(vib
,
3771 DVA_GET_OFFSET(&bp
->blk_dva
[0]), DVA_GET_ASIZE(&bp
->blk_dva
[0]));
3773 DVA_SET_VDEV(&bp
->blk_dva
[0], vd
->vdev_id
);
3774 DVA_SET_OFFSET(&bp
->blk_dva
[0], offset
);
3778 * If the block pointer contains any indirect DVAs, modify them to refer to
3779 * concrete DVAs. Note that this will sometimes not be possible, leaving
3780 * the indirect DVA in place. This happens if the indirect DVA spans multiple
3781 * segments in the mapping (i.e. it is a "split block").
3783 * If the BP was remapped, calls the callback on the original dva (note the
3784 * callback can be called multiple times if the original indirect DVA refers
3785 * to another indirect DVA, etc).
3787 * Returns TRUE if the BP was remapped.
3790 spa_remap_blkptr(spa_t
*spa
, blkptr_t
*bp
, spa_remap_cb_t callback
, void *arg
)
3792 remap_blkptr_cb_arg_t rbca
;
3794 if (!zfs_remap_blkptr_enable
)
3797 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
))
3801 * Dedup BP's can not be remapped, because ddt_phys_select() depends
3802 * on DVA[0] being the same in the BP as in the DDT (dedup table).
3804 if (BP_GET_DEDUP(bp
))
3808 * Gang blocks can not be remapped, because
3809 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
3810 * the BP used to read the gang block header (GBH) being the same
3811 * as the DVA[0] that we allocated for the GBH.
3817 * Embedded BP's have no DVA to remap.
3819 if (BP_GET_NDVAS(bp
) < 1)
3823 * Note: we only remap dva[0]. If we remapped other dvas, we
3824 * would no longer know what their phys birth txg is.
3826 dva_t
*dva
= &bp
->blk_dva
[0];
3828 uint64_t offset
= DVA_GET_OFFSET(dva
);
3829 uint64_t size
= DVA_GET_ASIZE(dva
);
3830 vdev_t
*vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
3832 if (vd
->vdev_ops
->vdev_op_remap
== NULL
)
3836 rbca
.rbca_cb
= callback
;
3837 rbca
.rbca_remap_vd
= vd
;
3838 rbca
.rbca_remap_offset
= offset
;
3839 rbca
.rbca_cb_arg
= arg
;
3842 * remap_blkptr_cb() will be called in order for each level of
3843 * indirection, until a concrete vdev is reached or a split block is
3844 * encountered. old_vd and old_offset are updated within the callback
3845 * as we go from the one indirect vdev to the next one (either concrete
3846 * or indirect again) in that order.
3848 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
, remap_blkptr_cb
, &rbca
);
3850 /* Check if the DVA wasn't remapped because it is a split block */
3851 if (DVA_GET_VDEV(&rbca
.rbca_bp
->blk_dva
[0]) == vd
->vdev_id
)
3858 * Undo the allocation of a DVA which happened in the given transaction group.
3861 metaslab_unalloc_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
3865 uint64_t vdev
= DVA_GET_VDEV(dva
);
3866 uint64_t offset
= DVA_GET_OFFSET(dva
);
3867 uint64_t size
= DVA_GET_ASIZE(dva
);
3869 ASSERT(DVA_IS_VALID(dva
));
3870 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3872 if (txg
> spa_freeze_txg(spa
))
3875 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
3876 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
3877 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
3878 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
3879 (u_longlong_t
)size
);
3883 ASSERT(!vd
->vdev_removing
);
3884 ASSERT(vdev_is_concrete(vd
));
3885 ASSERT0(vd
->vdev_indirect_config
.vic_mapping_object
);
3886 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
3888 if (DVA_GET_GANG(dva
))
3889 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3891 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3893 mutex_enter(&msp
->ms_lock
);
3894 range_tree_remove(msp
->ms_allocating
[txg
& TXG_MASK
],
3897 VERIFY(!msp
->ms_condensing
);
3898 VERIFY3U(offset
, >=, msp
->ms_start
);
3899 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
3900 VERIFY3U(range_tree_space(msp
->ms_allocatable
) + size
, <=,
3902 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
3903 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
3904 range_tree_add(msp
->ms_allocatable
, offset
, size
);
3905 mutex_exit(&msp
->ms_lock
);
3909 * Free the block represented by the given DVA.
3912 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, boolean_t checkpoint
)
3914 uint64_t vdev
= DVA_GET_VDEV(dva
);
3915 uint64_t offset
= DVA_GET_OFFSET(dva
);
3916 uint64_t size
= DVA_GET_ASIZE(dva
);
3917 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
3919 ASSERT(DVA_IS_VALID(dva
));
3920 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
3922 if (DVA_GET_GANG(dva
)) {
3923 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
3926 metaslab_free_impl(vd
, offset
, size
, checkpoint
);
3930 * Reserve some allocation slots. The reservation system must be called
3931 * before we call into the allocator. If there aren't any available slots
3932 * then the I/O will be throttled until an I/O completes and its slots are
3933 * freed up. The function returns true if it was successful in placing
3937 metaslab_class_throttle_reserve(metaslab_class_t
*mc
, int slots
, int allocator
,
3938 zio_t
*zio
, int flags
)
3940 uint64_t available_slots
= 0;
3941 boolean_t slot_reserved
= B_FALSE
;
3942 uint64_t max
= mc
->mc_alloc_max_slots
[allocator
];
3944 ASSERT(mc
->mc_alloc_throttle_enabled
);
3945 mutex_enter(&mc
->mc_lock
);
3947 uint64_t reserved_slots
=
3948 zfs_refcount_count(&mc
->mc_alloc_slots
[allocator
]);
3949 if (reserved_slots
< max
)
3950 available_slots
= max
- reserved_slots
;
3952 if (slots
<= available_slots
|| GANG_ALLOCATION(flags
) ||
3953 flags
& METASLAB_MUST_RESERVE
) {
3955 * We reserve the slots individually so that we can unreserve
3956 * them individually when an I/O completes.
3958 for (int d
= 0; d
< slots
; d
++) {
3960 zfs_refcount_add(&mc
->mc_alloc_slots
[allocator
],
3963 zio
->io_flags
|= ZIO_FLAG_IO_ALLOCATING
;
3964 slot_reserved
= B_TRUE
;
3967 mutex_exit(&mc
->mc_lock
);
3968 return (slot_reserved
);
3972 metaslab_class_throttle_unreserve(metaslab_class_t
*mc
, int slots
,
3973 int allocator
, zio_t
*zio
)
3975 ASSERT(mc
->mc_alloc_throttle_enabled
);
3976 mutex_enter(&mc
->mc_lock
);
3977 for (int d
= 0; d
< slots
; d
++) {
3978 (void) zfs_refcount_remove(&mc
->mc_alloc_slots
[allocator
],
3981 mutex_exit(&mc
->mc_lock
);
3985 metaslab_claim_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
3989 spa_t
*spa
= vd
->vdev_spa
;
3992 if (offset
>> vd
->vdev_ms_shift
>= vd
->vdev_ms_count
)
3993 return (SET_ERROR(ENXIO
));
3995 ASSERT3P(vd
->vdev_ms
, !=, NULL
);
3996 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
3998 mutex_enter(&msp
->ms_lock
);
4000 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
) {
4001 error
= metaslab_activate(msp
, 0, METASLAB_WEIGHT_CLAIM
);
4002 if (error
== EBUSY
) {
4003 ASSERT(msp
->ms_loaded
);
4004 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
4010 !range_tree_contains(msp
->ms_allocatable
, offset
, size
))
4011 error
= SET_ERROR(ENOENT
);
4013 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
4014 mutex_exit(&msp
->ms_lock
);
4018 VERIFY(!msp
->ms_condensing
);
4019 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
4020 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
4021 VERIFY3U(range_tree_space(msp
->ms_allocatable
) - size
, <=,
4023 range_tree_remove(msp
->ms_allocatable
, offset
, size
);
4025 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
4026 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
4027 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
4028 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
],
4032 mutex_exit(&msp
->ms_lock
);
4037 typedef struct metaslab_claim_cb_arg_t
{
4040 } metaslab_claim_cb_arg_t
;
4044 metaslab_claim_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
4045 uint64_t size
, void *arg
)
4047 metaslab_claim_cb_arg_t
*mcca_arg
= arg
;
4049 if (mcca_arg
->mcca_error
== 0) {
4050 mcca_arg
->mcca_error
= metaslab_claim_concrete(vd
, offset
,
4051 size
, mcca_arg
->mcca_txg
);
4056 metaslab_claim_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
, uint64_t txg
)
4058 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
4059 metaslab_claim_cb_arg_t arg
;
4062 * Only zdb(1M) can claim on indirect vdevs. This is used
4063 * to detect leaks of mapped space (that are not accounted
4064 * for in the obsolete counts, spacemap, or bpobj).
4066 ASSERT(!spa_writeable(vd
->vdev_spa
));
4070 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
4071 metaslab_claim_impl_cb
, &arg
);
4073 if (arg
.mcca_error
== 0) {
4074 arg
.mcca_error
= metaslab_claim_concrete(vd
,
4077 return (arg
.mcca_error
);
4079 return (metaslab_claim_concrete(vd
, offset
, size
, txg
));
4084 * Intent log support: upon opening the pool after a crash, notify the SPA
4085 * of blocks that the intent log has allocated for immediate write, but
4086 * which are still considered free by the SPA because the last transaction
4087 * group didn't commit yet.
4090 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
4092 uint64_t vdev
= DVA_GET_VDEV(dva
);
4093 uint64_t offset
= DVA_GET_OFFSET(dva
);
4094 uint64_t size
= DVA_GET_ASIZE(dva
);
4097 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
) {
4098 return (SET_ERROR(ENXIO
));
4101 ASSERT(DVA_IS_VALID(dva
));
4103 if (DVA_GET_GANG(dva
))
4104 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
4106 return (metaslab_claim_impl(vd
, offset
, size
, txg
));
4110 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
4111 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
,
4112 zio_alloc_list_t
*zal
, zio_t
*zio
, int allocator
)
4114 dva_t
*dva
= bp
->blk_dva
;
4115 dva_t
*hintdva
= hintbp
->blk_dva
;
4118 ASSERT(bp
->blk_birth
== 0);
4119 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
4121 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
4123 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
4124 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4125 return (SET_ERROR(ENOSPC
));
4128 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
4129 ASSERT(BP_GET_NDVAS(bp
) == 0);
4130 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
4131 ASSERT3P(zal
, !=, NULL
);
4133 for (int d
= 0; d
< ndvas
; d
++) {
4134 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
4135 txg
, flags
, zal
, allocator
);
4137 for (d
--; d
>= 0; d
--) {
4138 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
4139 metaslab_group_alloc_decrement(spa
,
4140 DVA_GET_VDEV(&dva
[d
]), zio
, flags
,
4141 allocator
, B_FALSE
);
4142 bzero(&dva
[d
], sizeof (dva_t
));
4144 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4148 * Update the metaslab group's queue depth
4149 * based on the newly allocated dva.
4151 metaslab_group_alloc_increment(spa
,
4152 DVA_GET_VDEV(&dva
[d
]), zio
, flags
, allocator
);
4157 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
4159 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4161 BP_SET_BIRTH(bp
, txg
, 0);
4167 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
4169 const dva_t
*dva
= bp
->blk_dva
;
4170 int ndvas
= BP_GET_NDVAS(bp
);
4172 ASSERT(!BP_IS_HOLE(bp
));
4173 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
4176 * If we have a checkpoint for the pool we need to make sure that
4177 * the blocks that we free that are part of the checkpoint won't be
4178 * reused until the checkpoint is discarded or we revert to it.
4180 * The checkpoint flag is passed down the metaslab_free code path
4181 * and is set whenever we want to add a block to the checkpoint's
4182 * accounting. That is, we "checkpoint" blocks that existed at the
4183 * time the checkpoint was created and are therefore referenced by
4184 * the checkpointed uberblock.
4186 * Note that, we don't checkpoint any blocks if the current
4187 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
4188 * normally as they will be referenced by the checkpointed uberblock.
4190 boolean_t checkpoint
= B_FALSE
;
4191 if (bp
->blk_birth
<= spa
->spa_checkpoint_txg
&&
4192 spa_syncing_txg(spa
) > spa
->spa_checkpoint_txg
) {
4194 * At this point, if the block is part of the checkpoint
4195 * there is no way it was created in the current txg.
4198 ASSERT3U(spa_syncing_txg(spa
), ==, txg
);
4199 checkpoint
= B_TRUE
;
4202 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
4204 for (int d
= 0; d
< ndvas
; d
++) {
4206 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
4208 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
4209 metaslab_free_dva(spa
, &dva
[d
], checkpoint
);
4213 spa_config_exit(spa
, SCL_FREE
, FTAG
);
4217 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
4219 const dva_t
*dva
= bp
->blk_dva
;
4220 int ndvas
= BP_GET_NDVAS(bp
);
4223 ASSERT(!BP_IS_HOLE(bp
));
4227 * First do a dry run to make sure all DVAs are claimable,
4228 * so we don't have to unwind from partial failures below.
4230 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
4234 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
4236 for (int d
= 0; d
< ndvas
; d
++) {
4237 error
= metaslab_claim_dva(spa
, &dva
[d
], txg
);
4242 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4244 ASSERT(error
== 0 || txg
== 0);
4250 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
4252 const dva_t
*dva
= bp
->blk_dva
;
4253 int ndvas
= BP_GET_NDVAS(bp
);
4254 uint64_t psize
= BP_GET_PSIZE(bp
);
4258 ASSERT(!BP_IS_HOLE(bp
));
4259 ASSERT(!BP_IS_EMBEDDED(bp
));
4262 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4264 for (d
= 0; d
< ndvas
; d
++) {
4265 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
4267 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
4270 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4274 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
4276 const dva_t
*dva
= bp
->blk_dva
;
4277 int ndvas
= BP_GET_NDVAS(bp
);
4278 uint64_t psize
= BP_GET_PSIZE(bp
);
4282 ASSERT(!BP_IS_HOLE(bp
));
4283 ASSERT(!BP_IS_EMBEDDED(bp
));
4286 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4288 for (d
= 0; d
< ndvas
; d
++) {
4289 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
4291 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
4292 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
4295 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4300 metaslab_check_free_impl_cb(uint64_t inner
, vdev_t
*vd
, uint64_t offset
,
4301 uint64_t size
, void *arg
)
4303 if (vd
->vdev_ops
== &vdev_indirect_ops
)
4306 metaslab_check_free_impl(vd
, offset
, size
);
4310 metaslab_check_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
4313 ASSERTV(spa_t
*spa
= vd
->vdev_spa
);
4315 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
4318 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
4319 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
4320 metaslab_check_free_impl_cb
, NULL
);
4324 ASSERT(vdev_is_concrete(vd
));
4325 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
4326 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
4328 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
4330 mutex_enter(&msp
->ms_lock
);
4331 if (msp
->ms_loaded
) {
4332 range_tree_verify_not_present(msp
->ms_allocatable
,
4336 range_tree_verify_not_present(msp
->ms_freeing
, offset
, size
);
4337 range_tree_verify_not_present(msp
->ms_checkpointing
, offset
, size
);
4338 range_tree_verify_not_present(msp
->ms_freed
, offset
, size
);
4339 for (int j
= 0; j
< TXG_DEFER_SIZE
; j
++)
4340 range_tree_verify_not_present(msp
->ms_defer
[j
], offset
, size
);
4341 mutex_exit(&msp
->ms_lock
);
4345 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
4347 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
4350 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
4351 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
4352 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
4353 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
4354 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
4355 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
4357 if (DVA_GET_GANG(&bp
->blk_dva
[i
]))
4358 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
4360 ASSERT3P(vd
, !=, NULL
);
4362 metaslab_check_free_impl(vd
, offset
, size
);
4364 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
4367 #if defined(_KERNEL)
4369 module_param(metaslab_aliquot
, ulong
, 0644);
4370 MODULE_PARM_DESC(metaslab_aliquot
,
4371 "allocation granularity (a.k.a. stripe size)");
4373 module_param(metaslab_debug_load
, int, 0644);
4374 MODULE_PARM_DESC(metaslab_debug_load
,
4375 "load all metaslabs when pool is first opened");
4377 module_param(metaslab_debug_unload
, int, 0644);
4378 MODULE_PARM_DESC(metaslab_debug_unload
,
4379 "prevent metaslabs from being unloaded");
4381 module_param(metaslab_preload_enabled
, int, 0644);
4382 MODULE_PARM_DESC(metaslab_preload_enabled
,
4383 "preload potential metaslabs during reassessment");
4385 module_param(zfs_mg_noalloc_threshold
, int, 0644);
4386 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
4387 "percentage of free space for metaslab group to allow allocation");
4389 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
4390 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
4391 "fragmentation for metaslab group to allow allocation");
4393 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
4394 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
4395 "fragmentation for metaslab to allow allocation");
4397 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
4398 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
4399 "use the fragmentation metric to prefer less fragmented metaslabs");
4401 module_param(metaslab_lba_weighting_enabled
, int, 0644);
4402 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
4403 "prefer metaslabs with lower LBAs");
4405 module_param(metaslab_bias_enabled
, int, 0644);
4406 MODULE_PARM_DESC(metaslab_bias_enabled
,
4407 "enable metaslab group biasing");
4409 module_param(zfs_metaslab_segment_weight_enabled
, int, 0644);
4410 MODULE_PARM_DESC(zfs_metaslab_segment_weight_enabled
,
4411 "enable segment-based metaslab selection");
4413 module_param(zfs_metaslab_switch_threshold
, int, 0644);
4414 MODULE_PARM_DESC(zfs_metaslab_switch_threshold
,
4415 "segment-based metaslab selection maximum buckets before switching");
4417 module_param(metaslab_force_ganging
, ulong
, 0644);
4418 MODULE_PARM_DESC(metaslab_force_ganging
,
4419 "blocks larger than this size are forced to be gang blocks");