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 https://opensource.org/licenses/CDDL-1.0.
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, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2015, Nexenta Systems, Inc. All rights reserved.
26 * Copyright (c) 2017, Intel Corporation.
29 #include <sys/zfs_context.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/space_map.h>
33 #include <sys/metaslab_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/vdev_draid.h>
37 #include <sys/spa_impl.h>
38 #include <sys/zfeature.h>
39 #include <sys/vdev_indirect_mapping.h>
41 #include <sys/btree.h>
43 #define WITH_DF_BLOCK_ALLOCATOR
45 #define GANG_ALLOCATION(flags) \
46 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
49 * Metaslab granularity, in bytes. This is roughly similar to what would be
50 * referred to as the "stripe size" in traditional RAID arrays. In normal
51 * operation, we will try to write this amount of data to each disk before
52 * moving on to the next top-level vdev.
54 static uint64_t metaslab_aliquot
= 1024 * 1024;
57 * For testing, make some blocks above a certain size be gang blocks.
59 uint64_t metaslab_force_ganging
= SPA_MAXBLOCKSIZE
+ 1;
62 * In pools where the log space map feature is not enabled we touch
63 * multiple metaslabs (and their respective space maps) with each
64 * transaction group. Thus, we benefit from having a small space map
65 * block size since it allows us to issue more I/O operations scattered
66 * around the disk. So a sane default for the space map block size
69 int zfs_metaslab_sm_blksz_no_log
= (1 << 14);
72 * When the log space map feature is enabled, we accumulate a lot of
73 * changes per metaslab that are flushed once in a while so we benefit
74 * from a bigger block size like 128K for the metaslab space maps.
76 int zfs_metaslab_sm_blksz_with_log
= (1 << 17);
79 * The in-core space map representation is more compact than its on-disk form.
80 * The zfs_condense_pct determines how much more compact the in-core
81 * space map representation must be before we compact it on-disk.
82 * Values should be greater than or equal to 100.
84 uint_t zfs_condense_pct
= 200;
87 * Condensing a metaslab is not guaranteed to actually reduce the amount of
88 * space used on disk. In particular, a space map uses data in increments of
89 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
90 * same number of blocks after condensing. Since the goal of condensing is to
91 * reduce the number of IOPs required to read the space map, we only want to
92 * condense when we can be sure we will reduce the number of blocks used by the
93 * space map. Unfortunately, we cannot precisely compute whether or not this is
94 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
95 * we apply the following heuristic: do not condense a spacemap unless the
96 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
99 static const int zfs_metaslab_condense_block_threshold
= 4;
102 * The zfs_mg_noalloc_threshold defines which metaslab groups should
103 * be eligible for allocation. The value is defined as a percentage of
104 * free space. Metaslab groups that have more free space than
105 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
106 * a metaslab group's free space is less than or equal to the
107 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
108 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
109 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
110 * groups are allowed to accept allocations. Gang blocks are always
111 * eligible to allocate on any metaslab group. The default value of 0 means
112 * no metaslab group will be excluded based on this criterion.
114 static uint_t zfs_mg_noalloc_threshold
= 0;
117 * Metaslab groups are considered eligible for allocations if their
118 * fragmentation metric (measured as a percentage) is less than or
119 * equal to zfs_mg_fragmentation_threshold. If a metaslab group
120 * exceeds this threshold then it will be skipped unless all metaslab
121 * groups within the metaslab class have also crossed this threshold.
123 * This tunable was introduced to avoid edge cases where we continue
124 * allocating from very fragmented disks in our pool while other, less
125 * fragmented disks, exists. On the other hand, if all disks in the
126 * pool are uniformly approaching the threshold, the threshold can
127 * be a speed bump in performance, where we keep switching the disks
128 * that we allocate from (e.g. we allocate some segments from disk A
129 * making it bypassing the threshold while freeing segments from disk
130 * B getting its fragmentation below the threshold).
132 * Empirically, we've seen that our vdev selection for allocations is
133 * good enough that fragmentation increases uniformly across all vdevs
134 * the majority of the time. Thus we set the threshold percentage high
135 * enough to avoid hitting the speed bump on pools that are being pushed
138 static uint_t zfs_mg_fragmentation_threshold
= 95;
141 * Allow metaslabs to keep their active state as long as their fragmentation
142 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
143 * active metaslab that exceeds this threshold will no longer keep its active
144 * status allowing better metaslabs to be selected.
146 static uint_t zfs_metaslab_fragmentation_threshold
= 70;
149 * When set will load all metaslabs when pool is first opened.
151 int metaslab_debug_load
= B_FALSE
;
154 * When set will prevent metaslabs from being unloaded.
156 static int metaslab_debug_unload
= B_FALSE
;
159 * Minimum size which forces the dynamic allocator to change
160 * it's allocation strategy. Once the space map cannot satisfy
161 * an allocation of this size then it switches to using more
162 * aggressive strategy (i.e search by size rather than offset).
164 uint64_t metaslab_df_alloc_threshold
= SPA_OLD_MAXBLOCKSIZE
;
167 * The minimum free space, in percent, which must be available
168 * in a space map to continue allocations in a first-fit fashion.
169 * Once the space map's free space drops below this level we dynamically
170 * switch to using best-fit allocations.
172 uint_t metaslab_df_free_pct
= 4;
175 * Maximum distance to search forward from the last offset. Without this
176 * limit, fragmented pools can see >100,000 iterations and
177 * metaslab_block_picker() becomes the performance limiting factor on
178 * high-performance storage.
180 * With the default setting of 16MB, we typically see less than 500
181 * iterations, even with very fragmented, ashift=9 pools. The maximum number
182 * of iterations possible is:
183 * metaslab_df_max_search / (2 * (1<<ashift))
184 * With the default setting of 16MB this is 16*1024 (with ashift=9) or
185 * 2048 (with ashift=12).
187 static uint_t metaslab_df_max_search
= 16 * 1024 * 1024;
190 * Forces the metaslab_block_picker function to search for at least this many
191 * segments forwards until giving up on finding a segment that the allocation
194 static const uint32_t metaslab_min_search_count
= 100;
197 * If we are not searching forward (due to metaslab_df_max_search,
198 * metaslab_df_free_pct, or metaslab_df_alloc_threshold), this tunable
199 * controls what segment is used. If it is set, we will use the largest free
200 * segment. If it is not set, we will use a segment of exactly the requested
203 static int metaslab_df_use_largest_segment
= B_FALSE
;
206 * Percentage of all cpus that can be used by the metaslab taskq.
208 int metaslab_load_pct
= 50;
211 * These tunables control how long a metaslab will remain loaded after the
212 * last allocation from it. A metaslab can't be unloaded until at least
213 * metaslab_unload_delay TXG's and metaslab_unload_delay_ms milliseconds
214 * have elapsed. However, zfs_metaslab_mem_limit may cause it to be
215 * unloaded sooner. These settings are intended to be generous -- to keep
216 * metaslabs loaded for a long time, reducing the rate of metaslab loading.
218 static uint_t metaslab_unload_delay
= 32;
219 static uint_t metaslab_unload_delay_ms
= 10 * 60 * 1000; /* ten minutes */
222 * Max number of metaslabs per group to preload.
224 uint_t metaslab_preload_limit
= 10;
227 * Enable/disable preloading of metaslab.
229 static int metaslab_preload_enabled
= B_TRUE
;
232 * Enable/disable fragmentation weighting on metaslabs.
234 static int metaslab_fragmentation_factor_enabled
= B_TRUE
;
237 * Enable/disable lba weighting (i.e. outer tracks are given preference).
239 static int metaslab_lba_weighting_enabled
= B_TRUE
;
242 * Enable/disable metaslab group biasing.
244 static int metaslab_bias_enabled
= B_TRUE
;
247 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
249 static const boolean_t zfs_remap_blkptr_enable
= B_TRUE
;
252 * Enable/disable segment-based metaslab selection.
254 static int zfs_metaslab_segment_weight_enabled
= B_TRUE
;
257 * When using segment-based metaslab selection, we will continue
258 * allocating from the active metaslab until we have exhausted
259 * zfs_metaslab_switch_threshold of its buckets.
261 static int zfs_metaslab_switch_threshold
= 2;
264 * Internal switch to enable/disable the metaslab allocation tracing
267 static const boolean_t metaslab_trace_enabled
= B_FALSE
;
270 * Maximum entries that the metaslab allocation tracing facility will keep
271 * in a given list when running in non-debug mode. We limit the number
272 * of entries in non-debug mode to prevent us from using up too much memory.
273 * The limit should be sufficiently large that we don't expect any allocation
274 * to every exceed this value. In debug mode, the system will panic if this
275 * limit is ever reached allowing for further investigation.
277 static const uint64_t metaslab_trace_max_entries
= 5000;
280 * Maximum number of metaslabs per group that can be disabled
283 static const int max_disabled_ms
= 3;
286 * Time (in seconds) to respect ms_max_size when the metaslab is not loaded.
287 * To avoid 64-bit overflow, don't set above UINT32_MAX.
289 static uint64_t zfs_metaslab_max_size_cache_sec
= 1 * 60 * 60; /* 1 hour */
292 * Maximum percentage of memory to use on storing loaded metaslabs. If loading
293 * a metaslab would take it over this percentage, the oldest selected metaslab
294 * is automatically unloaded.
296 static uint_t zfs_metaslab_mem_limit
= 25;
299 * Force the per-metaslab range trees to use 64-bit integers to store
300 * segments. Used for debugging purposes.
302 static const boolean_t zfs_metaslab_force_large_segs
= B_FALSE
;
305 * By default we only store segments over a certain size in the size-sorted
306 * metaslab trees (ms_allocatable_by_size and
307 * ms_unflushed_frees_by_size). This dramatically reduces memory usage and
308 * improves load and unload times at the cost of causing us to use slightly
309 * larger segments than we would otherwise in some cases.
311 static const uint32_t metaslab_by_size_min_shift
= 14;
314 * If not set, we will first try normal allocation. If that fails then
315 * we will do a gang allocation. If that fails then we will do a "try hard"
316 * gang allocation. If that fails then we will have a multi-layer gang
319 * If set, we will first try normal allocation. If that fails then
320 * we will do a "try hard" allocation. If that fails we will do a gang
321 * allocation. If that fails we will do a "try hard" gang allocation. If
322 * that fails then we will have a multi-layer gang block.
324 static int zfs_metaslab_try_hard_before_gang
= B_FALSE
;
327 * When not trying hard, we only consider the best zfs_metaslab_find_max_tries
328 * metaslabs. This improves performance, especially when there are many
329 * metaslabs per vdev and the allocation can't actually be satisfied (so we
330 * would otherwise iterate all the metaslabs). If there is a metaslab with a
331 * worse weight but it can actually satisfy the allocation, we won't find it
332 * until trying hard. This may happen if the worse metaslab is not loaded
333 * (and the true weight is better than we have calculated), or due to weight
334 * bucketization. E.g. we are looking for a 60K segment, and the best
335 * metaslabs all have free segments in the 32-63K bucket, but the best
336 * zfs_metaslab_find_max_tries metaslabs have ms_max_size <60KB, and a
337 * subsequent metaslab has ms_max_size >60KB (but fewer segments in this
338 * bucket, and therefore a lower weight).
340 static uint_t zfs_metaslab_find_max_tries
= 100;
342 static uint64_t metaslab_weight(metaslab_t
*, boolean_t
);
343 static void metaslab_set_fragmentation(metaslab_t
*, boolean_t
);
344 static void metaslab_free_impl(vdev_t
*, uint64_t, uint64_t, boolean_t
);
345 static void metaslab_check_free_impl(vdev_t
*, uint64_t, uint64_t);
347 static void metaslab_passivate(metaslab_t
*msp
, uint64_t weight
);
348 static uint64_t metaslab_weight_from_range_tree(metaslab_t
*msp
);
349 static void metaslab_flush_update(metaslab_t
*, dmu_tx_t
*);
350 static unsigned int metaslab_idx_func(multilist_t
*, void *);
351 static void metaslab_evict(metaslab_t
*, uint64_t);
352 static void metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
);
353 kmem_cache_t
*metaslab_alloc_trace_cache
;
355 typedef struct metaslab_stats
{
356 kstat_named_t metaslabstat_trace_over_limit
;
357 kstat_named_t metaslabstat_reload_tree
;
358 kstat_named_t metaslabstat_too_many_tries
;
359 kstat_named_t metaslabstat_try_hard
;
362 static metaslab_stats_t metaslab_stats
= {
363 { "trace_over_limit", KSTAT_DATA_UINT64
},
364 { "reload_tree", KSTAT_DATA_UINT64
},
365 { "too_many_tries", KSTAT_DATA_UINT64
},
366 { "try_hard", KSTAT_DATA_UINT64
},
369 #define METASLABSTAT_BUMP(stat) \
370 atomic_inc_64(&metaslab_stats.stat.value.ui64);
373 static kstat_t
*metaslab_ksp
;
376 metaslab_stat_init(void)
378 ASSERT(metaslab_alloc_trace_cache
== NULL
);
379 metaslab_alloc_trace_cache
= kmem_cache_create(
380 "metaslab_alloc_trace_cache", sizeof (metaslab_alloc_trace_t
),
381 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
382 metaslab_ksp
= kstat_create("zfs", 0, "metaslab_stats",
383 "misc", KSTAT_TYPE_NAMED
, sizeof (metaslab_stats
) /
384 sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
385 if (metaslab_ksp
!= NULL
) {
386 metaslab_ksp
->ks_data
= &metaslab_stats
;
387 kstat_install(metaslab_ksp
);
392 metaslab_stat_fini(void)
394 if (metaslab_ksp
!= NULL
) {
395 kstat_delete(metaslab_ksp
);
399 kmem_cache_destroy(metaslab_alloc_trace_cache
);
400 metaslab_alloc_trace_cache
= NULL
;
404 * ==========================================================================
406 * ==========================================================================
409 metaslab_class_create(spa_t
*spa
, const metaslab_ops_t
*ops
)
411 metaslab_class_t
*mc
;
413 mc
= kmem_zalloc(offsetof(metaslab_class_t
,
414 mc_allocator
[spa
->spa_alloc_count
]), KM_SLEEP
);
418 mutex_init(&mc
->mc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
419 multilist_create(&mc
->mc_metaslab_txg_list
, sizeof (metaslab_t
),
420 offsetof(metaslab_t
, ms_class_txg_node
), metaslab_idx_func
);
421 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++) {
422 metaslab_class_allocator_t
*mca
= &mc
->mc_allocator
[i
];
423 mca
->mca_rotor
= NULL
;
424 zfs_refcount_create_tracked(&mca
->mca_alloc_slots
);
431 metaslab_class_destroy(metaslab_class_t
*mc
)
433 spa_t
*spa
= mc
->mc_spa
;
435 ASSERT(mc
->mc_alloc
== 0);
436 ASSERT(mc
->mc_deferred
== 0);
437 ASSERT(mc
->mc_space
== 0);
438 ASSERT(mc
->mc_dspace
== 0);
440 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++) {
441 metaslab_class_allocator_t
*mca
= &mc
->mc_allocator
[i
];
442 ASSERT(mca
->mca_rotor
== NULL
);
443 zfs_refcount_destroy(&mca
->mca_alloc_slots
);
445 mutex_destroy(&mc
->mc_lock
);
446 multilist_destroy(&mc
->mc_metaslab_txg_list
);
447 kmem_free(mc
, offsetof(metaslab_class_t
,
448 mc_allocator
[spa
->spa_alloc_count
]));
452 metaslab_class_validate(metaslab_class_t
*mc
)
454 metaslab_group_t
*mg
;
458 * Must hold one of the spa_config locks.
460 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
461 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
463 if ((mg
= mc
->mc_allocator
[0].mca_rotor
) == NULL
)
468 ASSERT(vd
->vdev_mg
!= NULL
);
469 ASSERT3P(vd
->vdev_top
, ==, vd
);
470 ASSERT3P(mg
->mg_class
, ==, mc
);
471 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
472 } while ((mg
= mg
->mg_next
) != mc
->mc_allocator
[0].mca_rotor
);
478 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
479 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
481 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
482 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
483 atomic_add_64(&mc
->mc_space
, space_delta
);
484 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
488 metaslab_class_get_alloc(metaslab_class_t
*mc
)
490 return (mc
->mc_alloc
);
494 metaslab_class_get_deferred(metaslab_class_t
*mc
)
496 return (mc
->mc_deferred
);
500 metaslab_class_get_space(metaslab_class_t
*mc
)
502 return (mc
->mc_space
);
506 metaslab_class_get_dspace(metaslab_class_t
*mc
)
508 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
512 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
514 spa_t
*spa
= mc
->mc_spa
;
515 vdev_t
*rvd
= spa
->spa_root_vdev
;
519 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
522 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
525 mutex_enter(&mc
->mc_lock
);
526 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
527 vdev_t
*tvd
= rvd
->vdev_child
[c
];
528 metaslab_group_t
*mg
= vdev_get_mg(tvd
, mc
);
531 * Skip any holes, uninitialized top-levels, or
532 * vdevs that are not in this metalab class.
534 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
535 mg
->mg_class
!= mc
) {
539 IMPLY(mg
== mg
->mg_vd
->vdev_log_mg
,
540 mc
== spa_embedded_log_class(mg
->mg_vd
->vdev_spa
));
542 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
543 mc_hist
[i
] += mg
->mg_histogram
[i
];
546 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++) {
547 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
550 mutex_exit(&mc
->mc_lock
);
551 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
555 * Calculate the metaslab class's fragmentation metric. The metric
556 * is weighted based on the space contribution of each metaslab group.
557 * The return value will be a number between 0 and 100 (inclusive), or
558 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
559 * zfs_frag_table for more information about the metric.
562 metaslab_class_fragmentation(metaslab_class_t
*mc
)
564 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
565 uint64_t fragmentation
= 0;
567 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
569 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
570 vdev_t
*tvd
= rvd
->vdev_child
[c
];
571 metaslab_group_t
*mg
= tvd
->vdev_mg
;
574 * Skip any holes, uninitialized top-levels,
575 * or vdevs that are not in this metalab class.
577 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
578 mg
->mg_class
!= mc
) {
583 * If a metaslab group does not contain a fragmentation
584 * metric then just bail out.
586 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
587 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
588 return (ZFS_FRAG_INVALID
);
592 * Determine how much this metaslab_group is contributing
593 * to the overall pool fragmentation metric.
595 fragmentation
+= mg
->mg_fragmentation
*
596 metaslab_group_get_space(mg
);
598 fragmentation
/= metaslab_class_get_space(mc
);
600 ASSERT3U(fragmentation
, <=, 100);
601 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
602 return (fragmentation
);
606 * Calculate the amount of expandable space that is available in
607 * this metaslab class. If a device is expanded then its expandable
608 * space will be the amount of allocatable space that is currently not
609 * part of this metaslab class.
612 metaslab_class_expandable_space(metaslab_class_t
*mc
)
614 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
617 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
618 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
619 vdev_t
*tvd
= rvd
->vdev_child
[c
];
620 metaslab_group_t
*mg
= tvd
->vdev_mg
;
622 if (!vdev_is_concrete(tvd
) || tvd
->vdev_ms_shift
== 0 ||
623 mg
->mg_class
!= mc
) {
628 * Calculate if we have enough space to add additional
629 * metaslabs. We report the expandable space in terms
630 * of the metaslab size since that's the unit of expansion.
632 space
+= P2ALIGN(tvd
->vdev_max_asize
- tvd
->vdev_asize
,
633 1ULL << tvd
->vdev_ms_shift
);
635 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
640 metaslab_class_evict_old(metaslab_class_t
*mc
, uint64_t txg
)
642 multilist_t
*ml
= &mc
->mc_metaslab_txg_list
;
643 for (int i
= 0; i
< multilist_get_num_sublists(ml
); i
++) {
644 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
645 metaslab_t
*msp
= multilist_sublist_head(mls
);
646 multilist_sublist_unlock(mls
);
647 while (msp
!= NULL
) {
648 mutex_enter(&msp
->ms_lock
);
651 * If the metaslab has been removed from the list
652 * (which could happen if we were at the memory limit
653 * and it was evicted during this loop), then we can't
654 * proceed and we should restart the sublist.
656 if (!multilist_link_active(&msp
->ms_class_txg_node
)) {
657 mutex_exit(&msp
->ms_lock
);
661 mls
= multilist_sublist_lock(ml
, i
);
662 metaslab_t
*next_msp
= multilist_sublist_next(mls
, msp
);
663 multilist_sublist_unlock(mls
);
665 msp
->ms_selected_txg
+ metaslab_unload_delay
&&
666 gethrtime() > msp
->ms_selected_time
+
667 (uint64_t)MSEC2NSEC(metaslab_unload_delay_ms
)) {
668 metaslab_evict(msp
, txg
);
671 * Once we've hit a metaslab selected too
672 * recently to evict, we're done evicting for
675 mutex_exit(&msp
->ms_lock
);
678 mutex_exit(&msp
->ms_lock
);
685 metaslab_compare(const void *x1
, const void *x2
)
687 const metaslab_t
*m1
= (const metaslab_t
*)x1
;
688 const metaslab_t
*m2
= (const metaslab_t
*)x2
;
692 if (m1
->ms_allocator
!= -1 && m1
->ms_primary
)
694 else if (m1
->ms_allocator
!= -1 && !m1
->ms_primary
)
696 if (m2
->ms_allocator
!= -1 && m2
->ms_primary
)
698 else if (m2
->ms_allocator
!= -1 && !m2
->ms_primary
)
702 * Sort inactive metaslabs first, then primaries, then secondaries. When
703 * selecting a metaslab to allocate from, an allocator first tries its
704 * primary, then secondary active metaslab. If it doesn't have active
705 * metaslabs, or can't allocate from them, it searches for an inactive
706 * metaslab to activate. If it can't find a suitable one, it will steal
707 * a primary or secondary metaslab from another allocator.
714 int cmp
= TREE_CMP(m2
->ms_weight
, m1
->ms_weight
);
718 IMPLY(TREE_CMP(m1
->ms_start
, m2
->ms_start
) == 0, m1
== m2
);
720 return (TREE_CMP(m1
->ms_start
, m2
->ms_start
));
724 * ==========================================================================
726 * ==========================================================================
729 * Update the allocatable flag and the metaslab group's capacity.
730 * The allocatable flag is set to true if the capacity is below
731 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
732 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
733 * transitions from allocatable to non-allocatable or vice versa then the
734 * metaslab group's class is updated to reflect the transition.
737 metaslab_group_alloc_update(metaslab_group_t
*mg
)
739 vdev_t
*vd
= mg
->mg_vd
;
740 metaslab_class_t
*mc
= mg
->mg_class
;
741 vdev_stat_t
*vs
= &vd
->vdev_stat
;
742 boolean_t was_allocatable
;
743 boolean_t was_initialized
;
745 ASSERT(vd
== vd
->vdev_top
);
746 ASSERT3U(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_READER
), ==,
749 mutex_enter(&mg
->mg_lock
);
750 was_allocatable
= mg
->mg_allocatable
;
751 was_initialized
= mg
->mg_initialized
;
753 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
756 mutex_enter(&mc
->mc_lock
);
759 * If the metaslab group was just added then it won't
760 * have any space until we finish syncing out this txg.
761 * At that point we will consider it initialized and available
762 * for allocations. We also don't consider non-activated
763 * metaslab groups (e.g. vdevs that are in the middle of being removed)
764 * to be initialized, because they can't be used for allocation.
766 mg
->mg_initialized
= metaslab_group_initialized(mg
);
767 if (!was_initialized
&& mg
->mg_initialized
) {
769 } else if (was_initialized
&& !mg
->mg_initialized
) {
770 ASSERT3U(mc
->mc_groups
, >, 0);
773 if (mg
->mg_initialized
)
774 mg
->mg_no_free_space
= B_FALSE
;
777 * A metaslab group is considered allocatable if it has plenty
778 * of free space or is not heavily fragmented. We only take
779 * fragmentation into account if the metaslab group has a valid
780 * fragmentation metric (i.e. a value between 0 and 100).
782 mg
->mg_allocatable
= (mg
->mg_activation_count
> 0 &&
783 mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
784 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
785 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
788 * The mc_alloc_groups maintains a count of the number of
789 * groups in this metaslab class that are still above the
790 * zfs_mg_noalloc_threshold. This is used by the allocating
791 * threads to determine if they should avoid allocations to
792 * a given group. The allocator will avoid allocations to a group
793 * if that group has reached or is below the zfs_mg_noalloc_threshold
794 * and there are still other groups that are above the threshold.
795 * When a group transitions from allocatable to non-allocatable or
796 * vice versa we update the metaslab class to reflect that change.
797 * When the mc_alloc_groups value drops to 0 that means that all
798 * groups have reached the zfs_mg_noalloc_threshold making all groups
799 * eligible for allocations. This effectively means that all devices
800 * are balanced again.
802 if (was_allocatable
&& !mg
->mg_allocatable
)
803 mc
->mc_alloc_groups
--;
804 else if (!was_allocatable
&& mg
->mg_allocatable
)
805 mc
->mc_alloc_groups
++;
806 mutex_exit(&mc
->mc_lock
);
808 mutex_exit(&mg
->mg_lock
);
812 metaslab_sort_by_flushed(const void *va
, const void *vb
)
814 const metaslab_t
*a
= va
;
815 const metaslab_t
*b
= vb
;
817 int cmp
= TREE_CMP(a
->ms_unflushed_txg
, b
->ms_unflushed_txg
);
821 uint64_t a_vdev_id
= a
->ms_group
->mg_vd
->vdev_id
;
822 uint64_t b_vdev_id
= b
->ms_group
->mg_vd
->vdev_id
;
823 cmp
= TREE_CMP(a_vdev_id
, b_vdev_id
);
827 return (TREE_CMP(a
->ms_id
, b
->ms_id
));
831 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
, int allocators
)
833 metaslab_group_t
*mg
;
835 mg
= kmem_zalloc(offsetof(metaslab_group_t
,
836 mg_allocator
[allocators
]), KM_SLEEP
);
837 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
838 mutex_init(&mg
->mg_ms_disabled_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
839 cv_init(&mg
->mg_ms_disabled_cv
, NULL
, CV_DEFAULT
, NULL
);
840 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
841 sizeof (metaslab_t
), offsetof(metaslab_t
, ms_group_node
));
844 mg
->mg_activation_count
= 0;
845 mg
->mg_initialized
= B_FALSE
;
846 mg
->mg_no_free_space
= B_TRUE
;
847 mg
->mg_allocators
= allocators
;
849 for (int i
= 0; i
< allocators
; i
++) {
850 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[i
];
851 zfs_refcount_create_tracked(&mga
->mga_alloc_queue_depth
);
854 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
855 maxclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
| TASKQ_DYNAMIC
);
861 metaslab_group_destroy(metaslab_group_t
*mg
)
863 ASSERT(mg
->mg_prev
== NULL
);
864 ASSERT(mg
->mg_next
== NULL
);
866 * We may have gone below zero with the activation count
867 * either because we never activated in the first place or
868 * because we're done, and possibly removing the vdev.
870 ASSERT(mg
->mg_activation_count
<= 0);
872 taskq_destroy(mg
->mg_taskq
);
873 avl_destroy(&mg
->mg_metaslab_tree
);
874 mutex_destroy(&mg
->mg_lock
);
875 mutex_destroy(&mg
->mg_ms_disabled_lock
);
876 cv_destroy(&mg
->mg_ms_disabled_cv
);
878 for (int i
= 0; i
< mg
->mg_allocators
; i
++) {
879 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[i
];
880 zfs_refcount_destroy(&mga
->mga_alloc_queue_depth
);
882 kmem_free(mg
, offsetof(metaslab_group_t
,
883 mg_allocator
[mg
->mg_allocators
]));
887 metaslab_group_activate(metaslab_group_t
*mg
)
889 metaslab_class_t
*mc
= mg
->mg_class
;
890 spa_t
*spa
= mc
->mc_spa
;
891 metaslab_group_t
*mgprev
, *mgnext
;
893 ASSERT3U(spa_config_held(spa
, SCL_ALLOC
, RW_WRITER
), !=, 0);
895 ASSERT(mg
->mg_prev
== NULL
);
896 ASSERT(mg
->mg_next
== NULL
);
897 ASSERT(mg
->mg_activation_count
<= 0);
899 if (++mg
->mg_activation_count
<= 0)
902 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1,
903 vdev_get_ndisks(mg
->mg_vd
) - vdev_get_nparity(mg
->mg_vd
));
904 metaslab_group_alloc_update(mg
);
906 if ((mgprev
= mc
->mc_allocator
[0].mca_rotor
) == NULL
) {
910 mgnext
= mgprev
->mg_next
;
911 mg
->mg_prev
= mgprev
;
912 mg
->mg_next
= mgnext
;
913 mgprev
->mg_next
= mg
;
914 mgnext
->mg_prev
= mg
;
916 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++) {
917 mc
->mc_allocator
[i
].mca_rotor
= mg
;
923 * Passivate a metaslab group and remove it from the allocation rotor.
924 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
925 * a metaslab group. This function will momentarily drop spa_config_locks
926 * that are lower than the SCL_ALLOC lock (see comment below).
929 metaslab_group_passivate(metaslab_group_t
*mg
)
931 metaslab_class_t
*mc
= mg
->mg_class
;
932 spa_t
*spa
= mc
->mc_spa
;
933 metaslab_group_t
*mgprev
, *mgnext
;
934 int locks
= spa_config_held(spa
, SCL_ALL
, RW_WRITER
);
936 ASSERT3U(spa_config_held(spa
, SCL_ALLOC
| SCL_ZIO
, RW_WRITER
), ==,
937 (SCL_ALLOC
| SCL_ZIO
));
939 if (--mg
->mg_activation_count
!= 0) {
940 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++)
941 ASSERT(mc
->mc_allocator
[i
].mca_rotor
!= mg
);
942 ASSERT(mg
->mg_prev
== NULL
);
943 ASSERT(mg
->mg_next
== NULL
);
944 ASSERT(mg
->mg_activation_count
< 0);
949 * The spa_config_lock is an array of rwlocks, ordered as
950 * follows (from highest to lowest):
951 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
952 * SCL_ZIO > SCL_FREE > SCL_VDEV
953 * (For more information about the spa_config_lock see spa_misc.c)
954 * The higher the lock, the broader its coverage. When we passivate
955 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
956 * config locks. However, the metaslab group's taskq might be trying
957 * to preload metaslabs so we must drop the SCL_ZIO lock and any
958 * lower locks to allow the I/O to complete. At a minimum,
959 * we continue to hold the SCL_ALLOC lock, which prevents any future
960 * allocations from taking place and any changes to the vdev tree.
962 spa_config_exit(spa
, locks
& ~(SCL_ZIO
- 1), spa
);
963 taskq_wait_outstanding(mg
->mg_taskq
, 0);
964 spa_config_enter(spa
, locks
& ~(SCL_ZIO
- 1), spa
, RW_WRITER
);
965 metaslab_group_alloc_update(mg
);
966 for (int i
= 0; i
< mg
->mg_allocators
; i
++) {
967 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[i
];
968 metaslab_t
*msp
= mga
->mga_primary
;
970 mutex_enter(&msp
->ms_lock
);
971 metaslab_passivate(msp
,
972 metaslab_weight_from_range_tree(msp
));
973 mutex_exit(&msp
->ms_lock
);
975 msp
= mga
->mga_secondary
;
977 mutex_enter(&msp
->ms_lock
);
978 metaslab_passivate(msp
,
979 metaslab_weight_from_range_tree(msp
));
980 mutex_exit(&msp
->ms_lock
);
984 mgprev
= mg
->mg_prev
;
985 mgnext
= mg
->mg_next
;
990 mgprev
->mg_next
= mgnext
;
991 mgnext
->mg_prev
= mgprev
;
993 for (int i
= 0; i
< spa
->spa_alloc_count
; i
++) {
994 if (mc
->mc_allocator
[i
].mca_rotor
== mg
)
995 mc
->mc_allocator
[i
].mca_rotor
= mgnext
;
1003 metaslab_group_initialized(metaslab_group_t
*mg
)
1005 vdev_t
*vd
= mg
->mg_vd
;
1006 vdev_stat_t
*vs
= &vd
->vdev_stat
;
1008 return (vs
->vs_space
!= 0 && mg
->mg_activation_count
> 0);
1012 metaslab_group_get_space(metaslab_group_t
*mg
)
1015 * Note that the number of nodes in mg_metaslab_tree may be one less
1016 * than vdev_ms_count, due to the embedded log metaslab.
1018 mutex_enter(&mg
->mg_lock
);
1019 uint64_t ms_count
= avl_numnodes(&mg
->mg_metaslab_tree
);
1020 mutex_exit(&mg
->mg_lock
);
1021 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * ms_count
);
1025 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
1028 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1029 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
1031 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
1034 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
1037 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
1038 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
1040 mutex_enter(&mg
->mg_lock
);
1041 for (metaslab_t
*msp
= avl_first(t
);
1042 msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
1043 VERIFY3P(msp
->ms_group
, ==, mg
);
1044 /* skip if not active */
1045 if (msp
->ms_sm
== NULL
)
1048 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1049 mg_hist
[i
+ ashift
] +=
1050 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
1054 for (int i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
1055 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
1057 mutex_exit(&mg
->mg_lock
);
1059 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
1063 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
1065 metaslab_class_t
*mc
= mg
->mg_class
;
1066 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
1068 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1069 if (msp
->ms_sm
== NULL
)
1072 mutex_enter(&mg
->mg_lock
);
1073 mutex_enter(&mc
->mc_lock
);
1074 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1075 IMPLY(mg
== mg
->mg_vd
->vdev_log_mg
,
1076 mc
== spa_embedded_log_class(mg
->mg_vd
->vdev_spa
));
1077 mg
->mg_histogram
[i
+ ashift
] +=
1078 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
1079 mc
->mc_histogram
[i
+ ashift
] +=
1080 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
1082 mutex_exit(&mc
->mc_lock
);
1083 mutex_exit(&mg
->mg_lock
);
1087 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
1089 metaslab_class_t
*mc
= mg
->mg_class
;
1090 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
1092 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1093 if (msp
->ms_sm
== NULL
)
1096 mutex_enter(&mg
->mg_lock
);
1097 mutex_enter(&mc
->mc_lock
);
1098 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1099 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
1100 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
1101 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
1102 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
1103 IMPLY(mg
== mg
->mg_vd
->vdev_log_mg
,
1104 mc
== spa_embedded_log_class(mg
->mg_vd
->vdev_spa
));
1106 mg
->mg_histogram
[i
+ ashift
] -=
1107 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
1108 mc
->mc_histogram
[i
+ ashift
] -=
1109 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
1111 mutex_exit(&mc
->mc_lock
);
1112 mutex_exit(&mg
->mg_lock
);
1116 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
1118 ASSERT(msp
->ms_group
== NULL
);
1119 mutex_enter(&mg
->mg_lock
);
1122 avl_add(&mg
->mg_metaslab_tree
, msp
);
1123 mutex_exit(&mg
->mg_lock
);
1125 mutex_enter(&msp
->ms_lock
);
1126 metaslab_group_histogram_add(mg
, msp
);
1127 mutex_exit(&msp
->ms_lock
);
1131 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
1133 mutex_enter(&msp
->ms_lock
);
1134 metaslab_group_histogram_remove(mg
, msp
);
1135 mutex_exit(&msp
->ms_lock
);
1137 mutex_enter(&mg
->mg_lock
);
1138 ASSERT(msp
->ms_group
== mg
);
1139 avl_remove(&mg
->mg_metaslab_tree
, msp
);
1141 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
1142 multilist_sublist_t
*mls
=
1143 multilist_sublist_lock_obj(&mc
->mc_metaslab_txg_list
, msp
);
1144 if (multilist_link_active(&msp
->ms_class_txg_node
))
1145 multilist_sublist_remove(mls
, msp
);
1146 multilist_sublist_unlock(mls
);
1148 msp
->ms_group
= NULL
;
1149 mutex_exit(&mg
->mg_lock
);
1153 metaslab_group_sort_impl(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
1155 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1156 ASSERT(MUTEX_HELD(&mg
->mg_lock
));
1157 ASSERT(msp
->ms_group
== mg
);
1159 avl_remove(&mg
->mg_metaslab_tree
, msp
);
1160 msp
->ms_weight
= weight
;
1161 avl_add(&mg
->mg_metaslab_tree
, msp
);
1166 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
1169 * Although in principle the weight can be any value, in
1170 * practice we do not use values in the range [1, 511].
1172 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
1173 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1175 mutex_enter(&mg
->mg_lock
);
1176 metaslab_group_sort_impl(mg
, msp
, weight
);
1177 mutex_exit(&mg
->mg_lock
);
1181 * Calculate the fragmentation for a given metaslab group. We can use
1182 * a simple average here since all metaslabs within the group must have
1183 * the same size. The return value will be a value between 0 and 100
1184 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
1185 * group have a fragmentation metric.
1188 metaslab_group_fragmentation(metaslab_group_t
*mg
)
1190 vdev_t
*vd
= mg
->mg_vd
;
1191 uint64_t fragmentation
= 0;
1192 uint64_t valid_ms
= 0;
1194 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
1195 metaslab_t
*msp
= vd
->vdev_ms
[m
];
1197 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
1199 if (msp
->ms_group
!= mg
)
1203 fragmentation
+= msp
->ms_fragmentation
;
1206 if (valid_ms
<= mg
->mg_vd
->vdev_ms_count
/ 2)
1207 return (ZFS_FRAG_INVALID
);
1209 fragmentation
/= valid_ms
;
1210 ASSERT3U(fragmentation
, <=, 100);
1211 return (fragmentation
);
1215 * Determine if a given metaslab group should skip allocations. A metaslab
1216 * group should avoid allocations if its free capacity is less than the
1217 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
1218 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
1219 * that can still handle allocations. If the allocation throttle is enabled
1220 * then we skip allocations to devices that have reached their maximum
1221 * allocation queue depth unless the selected metaslab group is the only
1222 * eligible group remaining.
1225 metaslab_group_allocatable(metaslab_group_t
*mg
, metaslab_group_t
*rotor
,
1226 int flags
, uint64_t psize
, int allocator
, int d
)
1228 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1229 metaslab_class_t
*mc
= mg
->mg_class
;
1232 * We can only consider skipping this metaslab group if it's
1233 * in the normal metaslab class and there are other metaslab
1234 * groups to select from. Otherwise, we always consider it eligible
1237 if ((mc
!= spa_normal_class(spa
) &&
1238 mc
!= spa_special_class(spa
) &&
1239 mc
!= spa_dedup_class(spa
)) ||
1244 * If the metaslab group's mg_allocatable flag is set (see comments
1245 * in metaslab_group_alloc_update() for more information) and
1246 * the allocation throttle is disabled then allow allocations to this
1247 * device. However, if the allocation throttle is enabled then
1248 * check if we have reached our allocation limit (mga_alloc_queue_depth)
1249 * to determine if we should allow allocations to this metaslab group.
1250 * If all metaslab groups are no longer considered allocatable
1251 * (mc_alloc_groups == 0) or we're trying to allocate the smallest
1252 * gang block size then we allow allocations on this metaslab group
1253 * regardless of the mg_allocatable or throttle settings.
1255 if (mg
->mg_allocatable
) {
1256 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
1258 uint64_t qmax
= mga
->mga_cur_max_alloc_queue_depth
;
1260 if (!mc
->mc_alloc_throttle_enabled
)
1264 * If this metaslab group does not have any free space, then
1265 * there is no point in looking further.
1267 if (mg
->mg_no_free_space
)
1271 * Some allocations (e.g., those coming from device removal
1272 * where the * allocations are not even counted in the
1273 * metaslab * allocation queues) are allowed to bypass
1276 if (flags
& METASLAB_DONT_THROTTLE
)
1280 * Relax allocation throttling for ditto blocks. Due to
1281 * random imbalances in allocation it tends to push copies
1282 * to one vdev, that looks a bit better at the moment.
1284 qmax
= qmax
* (4 + d
) / 4;
1286 qdepth
= zfs_refcount_count(&mga
->mga_alloc_queue_depth
);
1289 * If this metaslab group is below its qmax or it's
1290 * the only allocatable metasable group, then attempt
1291 * to allocate from it.
1293 if (qdepth
< qmax
|| mc
->mc_alloc_groups
== 1)
1295 ASSERT3U(mc
->mc_alloc_groups
, >, 1);
1298 * Since this metaslab group is at or over its qmax, we
1299 * need to determine if there are metaslab groups after this
1300 * one that might be able to handle this allocation. This is
1301 * racy since we can't hold the locks for all metaslab
1302 * groups at the same time when we make this check.
1304 for (metaslab_group_t
*mgp
= mg
->mg_next
;
1305 mgp
!= rotor
; mgp
= mgp
->mg_next
) {
1306 metaslab_group_allocator_t
*mgap
=
1307 &mgp
->mg_allocator
[allocator
];
1308 qmax
= mgap
->mga_cur_max_alloc_queue_depth
;
1309 qmax
= qmax
* (4 + d
) / 4;
1311 zfs_refcount_count(&mgap
->mga_alloc_queue_depth
);
1314 * If there is another metaslab group that
1315 * might be able to handle the allocation, then
1316 * we return false so that we skip this group.
1318 if (qdepth
< qmax
&& !mgp
->mg_no_free_space
)
1323 * We didn't find another group to handle the allocation
1324 * so we can't skip this metaslab group even though
1325 * we are at or over our qmax.
1329 } else if (mc
->mc_alloc_groups
== 0 || psize
== SPA_MINBLOCKSIZE
) {
1336 * ==========================================================================
1337 * Range tree callbacks
1338 * ==========================================================================
1342 * Comparison function for the private size-ordered tree using 32-bit
1343 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1346 metaslab_rangesize32_compare(const void *x1
, const void *x2
)
1348 const range_seg32_t
*r1
= x1
;
1349 const range_seg32_t
*r2
= x2
;
1351 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
1352 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
1354 int cmp
= TREE_CMP(rs_size1
, rs_size2
);
1358 return (TREE_CMP(r1
->rs_start
, r2
->rs_start
));
1362 * Comparison function for the private size-ordered tree using 64-bit
1363 * ranges. Tree is sorted by size, larger sizes at the end of the tree.
1366 metaslab_rangesize64_compare(const void *x1
, const void *x2
)
1368 const range_seg64_t
*r1
= x1
;
1369 const range_seg64_t
*r2
= x2
;
1371 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
1372 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
1374 int cmp
= TREE_CMP(rs_size1
, rs_size2
);
1378 return (TREE_CMP(r1
->rs_start
, r2
->rs_start
));
1380 typedef struct metaslab_rt_arg
{
1381 zfs_btree_t
*mra_bt
;
1382 uint32_t mra_floor_shift
;
1383 } metaslab_rt_arg_t
;
1387 metaslab_rt_arg_t
*mra
;
1391 metaslab_size_sorted_add(void *arg
, uint64_t start
, uint64_t size
)
1393 struct mssa_arg
*mssap
= arg
;
1394 range_tree_t
*rt
= mssap
->rt
;
1395 metaslab_rt_arg_t
*mrap
= mssap
->mra
;
1396 range_seg_max_t seg
= {0};
1397 rs_set_start(&seg
, rt
, start
);
1398 rs_set_end(&seg
, rt
, start
+ size
);
1399 metaslab_rt_add(rt
, &seg
, mrap
);
1403 metaslab_size_tree_full_load(range_tree_t
*rt
)
1405 metaslab_rt_arg_t
*mrap
= rt
->rt_arg
;
1406 METASLABSTAT_BUMP(metaslabstat_reload_tree
);
1407 ASSERT0(zfs_btree_numnodes(mrap
->mra_bt
));
1408 mrap
->mra_floor_shift
= 0;
1409 struct mssa_arg arg
= {0};
1412 range_tree_walk(rt
, metaslab_size_sorted_add
, &arg
);
1416 * Create any block allocator specific components. The current allocators
1417 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
1420 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
1422 metaslab_rt_arg_t
*mrap
= arg
;
1423 zfs_btree_t
*size_tree
= mrap
->mra_bt
;
1426 int (*compare
) (const void *, const void *);
1427 switch (rt
->rt_type
) {
1429 size
= sizeof (range_seg32_t
);
1430 compare
= metaslab_rangesize32_compare
;
1433 size
= sizeof (range_seg64_t
);
1434 compare
= metaslab_rangesize64_compare
;
1437 panic("Invalid range seg type %d", rt
->rt_type
);
1439 zfs_btree_create(size_tree
, compare
, size
);
1440 mrap
->mra_floor_shift
= metaslab_by_size_min_shift
;
1444 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
1447 metaslab_rt_arg_t
*mrap
= arg
;
1448 zfs_btree_t
*size_tree
= mrap
->mra_bt
;
1450 zfs_btree_destroy(size_tree
);
1451 kmem_free(mrap
, sizeof (*mrap
));
1455 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
1457 metaslab_rt_arg_t
*mrap
= arg
;
1458 zfs_btree_t
*size_tree
= mrap
->mra_bt
;
1460 if (rs_get_end(rs
, rt
) - rs_get_start(rs
, rt
) <
1461 (1ULL << mrap
->mra_floor_shift
))
1464 zfs_btree_add(size_tree
, rs
);
1468 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
1470 metaslab_rt_arg_t
*mrap
= arg
;
1471 zfs_btree_t
*size_tree
= mrap
->mra_bt
;
1473 if (rs_get_end(rs
, rt
) - rs_get_start(rs
, rt
) < (1ULL <<
1474 mrap
->mra_floor_shift
))
1477 zfs_btree_remove(size_tree
, rs
);
1481 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
1483 metaslab_rt_arg_t
*mrap
= arg
;
1484 zfs_btree_t
*size_tree
= mrap
->mra_bt
;
1485 zfs_btree_clear(size_tree
);
1486 zfs_btree_destroy(size_tree
);
1488 metaslab_rt_create(rt
, arg
);
1491 static const range_tree_ops_t metaslab_rt_ops
= {
1492 .rtop_create
= metaslab_rt_create
,
1493 .rtop_destroy
= metaslab_rt_destroy
,
1494 .rtop_add
= metaslab_rt_add
,
1495 .rtop_remove
= metaslab_rt_remove
,
1496 .rtop_vacate
= metaslab_rt_vacate
1500 * ==========================================================================
1501 * Common allocator routines
1502 * ==========================================================================
1506 * Return the maximum contiguous segment within the metaslab.
1509 metaslab_largest_allocatable(metaslab_t
*msp
)
1511 zfs_btree_t
*t
= &msp
->ms_allocatable_by_size
;
1516 if (zfs_btree_numnodes(t
) == 0)
1517 metaslab_size_tree_full_load(msp
->ms_allocatable
);
1519 rs
= zfs_btree_last(t
, NULL
);
1523 return (rs_get_end(rs
, msp
->ms_allocatable
) - rs_get_start(rs
,
1524 msp
->ms_allocatable
));
1528 * Return the maximum contiguous segment within the unflushed frees of this
1532 metaslab_largest_unflushed_free(metaslab_t
*msp
)
1534 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1536 if (msp
->ms_unflushed_frees
== NULL
)
1539 if (zfs_btree_numnodes(&msp
->ms_unflushed_frees_by_size
) == 0)
1540 metaslab_size_tree_full_load(msp
->ms_unflushed_frees
);
1541 range_seg_t
*rs
= zfs_btree_last(&msp
->ms_unflushed_frees_by_size
,
1547 * When a range is freed from the metaslab, that range is added to
1548 * both the unflushed frees and the deferred frees. While the block
1549 * will eventually be usable, if the metaslab were loaded the range
1550 * would not be added to the ms_allocatable tree until TXG_DEFER_SIZE
1551 * txgs had passed. As a result, when attempting to estimate an upper
1552 * bound for the largest currently-usable free segment in the
1553 * metaslab, we need to not consider any ranges currently in the defer
1554 * trees. This algorithm approximates the largest available chunk in
1555 * the largest range in the unflushed_frees tree by taking the first
1556 * chunk. While this may be a poor estimate, it should only remain so
1557 * briefly and should eventually self-correct as frees are no longer
1558 * deferred. Similar logic applies to the ms_freed tree. See
1559 * metaslab_load() for more details.
1561 * There are two primary sources of inaccuracy in this estimate. Both
1562 * are tolerated for performance reasons. The first source is that we
1563 * only check the largest segment for overlaps. Smaller segments may
1564 * have more favorable overlaps with the other trees, resulting in
1565 * larger usable chunks. Second, we only look at the first chunk in
1566 * the largest segment; there may be other usable chunks in the
1567 * largest segment, but we ignore them.
1569 uint64_t rstart
= rs_get_start(rs
, msp
->ms_unflushed_frees
);
1570 uint64_t rsize
= rs_get_end(rs
, msp
->ms_unflushed_frees
) - rstart
;
1571 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1574 boolean_t found
= range_tree_find_in(msp
->ms_defer
[t
], rstart
,
1575 rsize
, &start
, &size
);
1577 if (rstart
== start
)
1579 rsize
= start
- rstart
;
1585 boolean_t found
= range_tree_find_in(msp
->ms_freed
, rstart
,
1586 rsize
, &start
, &size
);
1588 rsize
= start
- rstart
;
1593 static range_seg_t
*
1594 metaslab_block_find(zfs_btree_t
*t
, range_tree_t
*rt
, uint64_t start
,
1595 uint64_t size
, zfs_btree_index_t
*where
)
1598 range_seg_max_t rsearch
;
1600 rs_set_start(&rsearch
, rt
, start
);
1601 rs_set_end(&rsearch
, rt
, start
+ size
);
1603 rs
= zfs_btree_find(t
, &rsearch
, where
);
1605 rs
= zfs_btree_next(t
, where
, where
);
1611 #if defined(WITH_DF_BLOCK_ALLOCATOR) || \
1612 defined(WITH_CF_BLOCK_ALLOCATOR)
1615 * This is a helper function that can be used by the allocator to find a
1616 * suitable block to allocate. This will search the specified B-tree looking
1617 * for a block that matches the specified criteria.
1620 metaslab_block_picker(range_tree_t
*rt
, uint64_t *cursor
, uint64_t size
,
1621 uint64_t max_search
)
1624 *cursor
= rt
->rt_start
;
1625 zfs_btree_t
*bt
= &rt
->rt_root
;
1626 zfs_btree_index_t where
;
1627 range_seg_t
*rs
= metaslab_block_find(bt
, rt
, *cursor
, size
, &where
);
1628 uint64_t first_found
;
1629 int count_searched
= 0;
1632 first_found
= rs_get_start(rs
, rt
);
1634 while (rs
!= NULL
&& (rs_get_start(rs
, rt
) - first_found
<=
1635 max_search
|| count_searched
< metaslab_min_search_count
)) {
1636 uint64_t offset
= rs_get_start(rs
, rt
);
1637 if (offset
+ size
<= rs_get_end(rs
, rt
)) {
1638 *cursor
= offset
+ size
;
1641 rs
= zfs_btree_next(bt
, &where
, &where
);
1648 #endif /* WITH_DF/CF_BLOCK_ALLOCATOR */
1650 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1652 * ==========================================================================
1653 * Dynamic Fit (df) block allocator
1655 * Search for a free chunk of at least this size, starting from the last
1656 * offset (for this alignment of block) looking for up to
1657 * metaslab_df_max_search bytes (16MB). If a large enough free chunk is not
1658 * found within 16MB, then return a free chunk of exactly the requested size (or
1661 * If it seems like searching from the last offset will be unproductive, skip
1662 * that and just return a free chunk of exactly the requested size (or larger).
1663 * This is based on metaslab_df_alloc_threshold and metaslab_df_free_pct. This
1664 * mechanism is probably not very useful and may be removed in the future.
1666 * The behavior when not searching can be changed to return the largest free
1667 * chunk, instead of a free chunk of exactly the requested size, by setting
1668 * metaslab_df_use_largest_segment.
1669 * ==========================================================================
1672 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1675 * Find the largest power of 2 block size that evenly divides the
1676 * requested size. This is used to try to allocate blocks with similar
1677 * alignment from the same area of the metaslab (i.e. same cursor
1678 * bucket) but it does not guarantee that other allocations sizes
1679 * may exist in the same region.
1681 uint64_t align
= size
& -size
;
1682 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1683 range_tree_t
*rt
= msp
->ms_allocatable
;
1684 uint_t free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1687 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1690 * If we're running low on space, find a segment based on size,
1691 * rather than iterating based on offset.
1693 if (metaslab_largest_allocatable(msp
) < metaslab_df_alloc_threshold
||
1694 free_pct
< metaslab_df_free_pct
) {
1697 offset
= metaslab_block_picker(rt
,
1698 cursor
, size
, metaslab_df_max_search
);
1703 if (zfs_btree_numnodes(&msp
->ms_allocatable_by_size
) == 0)
1704 metaslab_size_tree_full_load(msp
->ms_allocatable
);
1706 if (metaslab_df_use_largest_segment
) {
1707 /* use largest free segment */
1708 rs
= zfs_btree_last(&msp
->ms_allocatable_by_size
, NULL
);
1710 zfs_btree_index_t where
;
1711 /* use segment of this size, or next largest */
1712 rs
= metaslab_block_find(&msp
->ms_allocatable_by_size
,
1713 rt
, msp
->ms_start
, size
, &where
);
1715 if (rs
!= NULL
&& rs_get_start(rs
, rt
) + size
<= rs_get_end(rs
,
1717 offset
= rs_get_start(rs
, rt
);
1718 *cursor
= offset
+ size
;
1725 const metaslab_ops_t zfs_metaslab_ops
= {
1728 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1730 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1732 * ==========================================================================
1733 * Cursor fit block allocator -
1734 * Select the largest region in the metaslab, set the cursor to the beginning
1735 * of the range and the cursor_end to the end of the range. As allocations
1736 * are made advance the cursor. Continue allocating from the cursor until
1737 * the range is exhausted and then find a new range.
1738 * ==========================================================================
1741 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1743 range_tree_t
*rt
= msp
->ms_allocatable
;
1744 zfs_btree_t
*t
= &msp
->ms_allocatable_by_size
;
1745 uint64_t *cursor
= &msp
->ms_lbas
[0];
1746 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1747 uint64_t offset
= 0;
1749 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1751 ASSERT3U(*cursor_end
, >=, *cursor
);
1753 if ((*cursor
+ size
) > *cursor_end
) {
1756 if (zfs_btree_numnodes(t
) == 0)
1757 metaslab_size_tree_full_load(msp
->ms_allocatable
);
1758 rs
= zfs_btree_last(t
, NULL
);
1759 if (rs
== NULL
|| (rs_get_end(rs
, rt
) - rs_get_start(rs
, rt
)) <
1763 *cursor
= rs_get_start(rs
, rt
);
1764 *cursor_end
= rs_get_end(rs
, rt
);
1773 const metaslab_ops_t zfs_metaslab_ops
= {
1776 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1778 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1780 * ==========================================================================
1781 * New dynamic fit allocator -
1782 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1783 * contiguous blocks. If no region is found then just use the largest segment
1785 * ==========================================================================
1789 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1790 * to request from the allocator.
1792 uint64_t metaslab_ndf_clump_shift
= 4;
1795 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1797 zfs_btree_t
*t
= &msp
->ms_allocatable
->rt_root
;
1798 range_tree_t
*rt
= msp
->ms_allocatable
;
1799 zfs_btree_index_t where
;
1801 range_seg_max_t rsearch
;
1802 uint64_t hbit
= highbit64(size
);
1803 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1804 uint64_t max_size
= metaslab_largest_allocatable(msp
);
1806 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1808 if (max_size
< size
)
1811 rs_set_start(&rsearch
, rt
, *cursor
);
1812 rs_set_end(&rsearch
, rt
, *cursor
+ size
);
1814 rs
= zfs_btree_find(t
, &rsearch
, &where
);
1815 if (rs
== NULL
|| (rs_get_end(rs
, rt
) - rs_get_start(rs
, rt
)) < size
) {
1816 t
= &msp
->ms_allocatable_by_size
;
1818 rs_set_start(&rsearch
, rt
, 0);
1819 rs_set_end(&rsearch
, rt
, MIN(max_size
, 1ULL << (hbit
+
1820 metaslab_ndf_clump_shift
)));
1822 rs
= zfs_btree_find(t
, &rsearch
, &where
);
1824 rs
= zfs_btree_next(t
, &where
, &where
);
1828 if ((rs_get_end(rs
, rt
) - rs_get_start(rs
, rt
)) >= size
) {
1829 *cursor
= rs_get_start(rs
, rt
) + size
;
1830 return (rs_get_start(rs
, rt
));
1835 const metaslab_ops_t zfs_metaslab_ops
= {
1838 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1842 * ==========================================================================
1844 * ==========================================================================
1848 * Wait for any in-progress metaslab loads to complete.
1851 metaslab_load_wait(metaslab_t
*msp
)
1853 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1855 while (msp
->ms_loading
) {
1856 ASSERT(!msp
->ms_loaded
);
1857 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1862 * Wait for any in-progress flushing to complete.
1865 metaslab_flush_wait(metaslab_t
*msp
)
1867 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1869 while (msp
->ms_flushing
)
1870 cv_wait(&msp
->ms_flush_cv
, &msp
->ms_lock
);
1874 metaslab_idx_func(multilist_t
*ml
, void *arg
)
1876 metaslab_t
*msp
= arg
;
1879 * ms_id values are allocated sequentially, so full 64bit
1880 * division would be a waste of time, so limit it to 32 bits.
1882 return ((unsigned int)msp
->ms_id
% multilist_get_num_sublists(ml
));
1886 metaslab_allocated_space(metaslab_t
*msp
)
1888 return (msp
->ms_allocated_space
);
1892 * Verify that the space accounting on disk matches the in-core range_trees.
1895 metaslab_verify_space(metaslab_t
*msp
, uint64_t txg
)
1897 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1898 uint64_t allocating
= 0;
1899 uint64_t sm_free_space
, msp_free_space
;
1901 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1902 ASSERT(!msp
->ms_condensing
);
1904 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
1908 * We can only verify the metaslab space when we're called
1909 * from syncing context with a loaded metaslab that has an
1910 * allocated space map. Calling this in non-syncing context
1911 * does not provide a consistent view of the metaslab since
1912 * we're performing allocations in the future.
1914 if (txg
!= spa_syncing_txg(spa
) || msp
->ms_sm
== NULL
||
1919 * Even though the smp_alloc field can get negative,
1920 * when it comes to a metaslab's space map, that should
1921 * never be the case.
1923 ASSERT3S(space_map_allocated(msp
->ms_sm
), >=, 0);
1925 ASSERT3U(space_map_allocated(msp
->ms_sm
), >=,
1926 range_tree_space(msp
->ms_unflushed_frees
));
1928 ASSERT3U(metaslab_allocated_space(msp
), ==,
1929 space_map_allocated(msp
->ms_sm
) +
1930 range_tree_space(msp
->ms_unflushed_allocs
) -
1931 range_tree_space(msp
->ms_unflushed_frees
));
1933 sm_free_space
= msp
->ms_size
- metaslab_allocated_space(msp
);
1936 * Account for future allocations since we would have
1937 * already deducted that space from the ms_allocatable.
1939 for (int t
= 0; t
< TXG_CONCURRENT_STATES
; t
++) {
1941 range_tree_space(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]);
1943 ASSERT3U(allocating
+ msp
->ms_allocated_this_txg
, ==,
1944 msp
->ms_allocating_total
);
1946 ASSERT3U(msp
->ms_deferspace
, ==,
1947 range_tree_space(msp
->ms_defer
[0]) +
1948 range_tree_space(msp
->ms_defer
[1]));
1950 msp_free_space
= range_tree_space(msp
->ms_allocatable
) + allocating
+
1951 msp
->ms_deferspace
+ range_tree_space(msp
->ms_freed
);
1953 VERIFY3U(sm_free_space
, ==, msp_free_space
);
1957 metaslab_aux_histograms_clear(metaslab_t
*msp
)
1960 * Auxiliary histograms are only cleared when resetting them,
1961 * which can only happen while the metaslab is loaded.
1963 ASSERT(msp
->ms_loaded
);
1965 memset(msp
->ms_synchist
, 0, sizeof (msp
->ms_synchist
));
1966 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++)
1967 memset(msp
->ms_deferhist
[t
], 0, sizeof (msp
->ms_deferhist
[t
]));
1971 metaslab_aux_histogram_add(uint64_t *histogram
, uint64_t shift
,
1975 * This is modeled after space_map_histogram_add(), so refer to that
1976 * function for implementation details. We want this to work like
1977 * the space map histogram, and not the range tree histogram, as we
1978 * are essentially constructing a delta that will be later subtracted
1979 * from the space map histogram.
1982 for (int i
= shift
; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++) {
1983 ASSERT3U(i
, >=, idx
+ shift
);
1984 histogram
[idx
] += rt
->rt_histogram
[i
] << (i
- idx
- shift
);
1986 if (idx
< SPACE_MAP_HISTOGRAM_SIZE
- 1) {
1987 ASSERT3U(idx
+ shift
, ==, i
);
1989 ASSERT3U(idx
, <, SPACE_MAP_HISTOGRAM_SIZE
);
1995 * Called at every sync pass that the metaslab gets synced.
1997 * The reason is that we want our auxiliary histograms to be updated
1998 * wherever the metaslab's space map histogram is updated. This way
1999 * we stay consistent on which parts of the metaslab space map's
2000 * histogram are currently not available for allocations (e.g because
2001 * they are in the defer, freed, and freeing trees).
2004 metaslab_aux_histograms_update(metaslab_t
*msp
)
2006 space_map_t
*sm
= msp
->ms_sm
;
2010 * This is similar to the metaslab's space map histogram updates
2011 * that take place in metaslab_sync(). The only difference is that
2012 * we only care about segments that haven't made it into the
2013 * ms_allocatable tree yet.
2015 if (msp
->ms_loaded
) {
2016 metaslab_aux_histograms_clear(msp
);
2018 metaslab_aux_histogram_add(msp
->ms_synchist
,
2019 sm
->sm_shift
, msp
->ms_freed
);
2021 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2022 metaslab_aux_histogram_add(msp
->ms_deferhist
[t
],
2023 sm
->sm_shift
, msp
->ms_defer
[t
]);
2027 metaslab_aux_histogram_add(msp
->ms_synchist
,
2028 sm
->sm_shift
, msp
->ms_freeing
);
2032 * Called every time we are done syncing (writing to) the metaslab,
2033 * i.e. at the end of each sync pass.
2034 * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist]
2037 metaslab_aux_histograms_update_done(metaslab_t
*msp
, boolean_t defer_allowed
)
2039 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2040 space_map_t
*sm
= msp
->ms_sm
;
2044 * We came here from metaslab_init() when creating/opening a
2045 * pool, looking at a metaslab that hasn't had any allocations
2052 * This is similar to the actions that we take for the ms_freed
2053 * and ms_defer trees in metaslab_sync_done().
2055 uint64_t hist_index
= spa_syncing_txg(spa
) % TXG_DEFER_SIZE
;
2056 if (defer_allowed
) {
2057 memcpy(msp
->ms_deferhist
[hist_index
], msp
->ms_synchist
,
2058 sizeof (msp
->ms_synchist
));
2060 memset(msp
->ms_deferhist
[hist_index
], 0,
2061 sizeof (msp
->ms_deferhist
[hist_index
]));
2063 memset(msp
->ms_synchist
, 0, sizeof (msp
->ms_synchist
));
2067 * Ensure that the metaslab's weight and fragmentation are consistent
2068 * with the contents of the histogram (either the range tree's histogram
2069 * or the space map's depending whether the metaslab is loaded).
2072 metaslab_verify_weight_and_frag(metaslab_t
*msp
)
2074 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2076 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
2080 * We can end up here from vdev_remove_complete(), in which case we
2081 * cannot do these assertions because we hold spa config locks and
2082 * thus we are not allowed to read from the DMU.
2084 * We check if the metaslab group has been removed and if that's
2085 * the case we return immediately as that would mean that we are
2086 * here from the aforementioned code path.
2088 if (msp
->ms_group
== NULL
)
2092 * Devices being removed always return a weight of 0 and leave
2093 * fragmentation and ms_max_size as is - there is nothing for
2094 * us to verify here.
2096 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
2097 if (vd
->vdev_removing
)
2101 * If the metaslab is dirty it probably means that we've done
2102 * some allocations or frees that have changed our histograms
2103 * and thus the weight.
2105 for (int t
= 0; t
< TXG_SIZE
; t
++) {
2106 if (txg_list_member(&vd
->vdev_ms_list
, msp
, t
))
2111 * This verification checks that our in-memory state is consistent
2112 * with what's on disk. If the pool is read-only then there aren't
2113 * any changes and we just have the initially-loaded state.
2115 if (!spa_writeable(msp
->ms_group
->mg_vd
->vdev_spa
))
2118 /* some extra verification for in-core tree if you can */
2119 if (msp
->ms_loaded
) {
2120 range_tree_stat_verify(msp
->ms_allocatable
);
2121 VERIFY(space_map_histogram_verify(msp
->ms_sm
,
2122 msp
->ms_allocatable
));
2125 uint64_t weight
= msp
->ms_weight
;
2126 uint64_t was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2127 boolean_t space_based
= WEIGHT_IS_SPACEBASED(msp
->ms_weight
);
2128 uint64_t frag
= msp
->ms_fragmentation
;
2129 uint64_t max_segsize
= msp
->ms_max_size
;
2132 msp
->ms_fragmentation
= 0;
2135 * This function is used for verification purposes and thus should
2136 * not introduce any side-effects/mutations on the system's state.
2138 * Regardless of whether metaslab_weight() thinks this metaslab
2139 * should be active or not, we want to ensure that the actual weight
2140 * (and therefore the value of ms_weight) would be the same if it
2141 * was to be recalculated at this point.
2143 * In addition we set the nodirty flag so metaslab_weight() does
2144 * not dirty the metaslab for future TXGs (e.g. when trying to
2145 * force condensing to upgrade the metaslab spacemaps).
2147 msp
->ms_weight
= metaslab_weight(msp
, B_TRUE
) | was_active
;
2149 VERIFY3U(max_segsize
, ==, msp
->ms_max_size
);
2152 * If the weight type changed then there is no point in doing
2153 * verification. Revert fields to their original values.
2155 if ((space_based
&& !WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) ||
2156 (!space_based
&& WEIGHT_IS_SPACEBASED(msp
->ms_weight
))) {
2157 msp
->ms_fragmentation
= frag
;
2158 msp
->ms_weight
= weight
;
2162 VERIFY3U(msp
->ms_fragmentation
, ==, frag
);
2163 VERIFY3U(msp
->ms_weight
, ==, weight
);
2167 * If we're over the zfs_metaslab_mem_limit, select the loaded metaslab from
2168 * this class that was used longest ago, and attempt to unload it. We don't
2169 * want to spend too much time in this loop to prevent performance
2170 * degradation, and we expect that most of the time this operation will
2171 * succeed. Between that and the normal unloading processing during txg sync,
2172 * we expect this to keep the metaslab memory usage under control.
2175 metaslab_potentially_evict(metaslab_class_t
*mc
)
2178 uint64_t allmem
= arc_all_memory();
2179 uint64_t inuse
= spl_kmem_cache_inuse(zfs_btree_leaf_cache
);
2180 uint64_t size
= spl_kmem_cache_entry_size(zfs_btree_leaf_cache
);
2182 for (; allmem
* zfs_metaslab_mem_limit
/ 100 < inuse
* size
&&
2183 tries
< multilist_get_num_sublists(&mc
->mc_metaslab_txg_list
) * 2;
2185 unsigned int idx
= multilist_get_random_index(
2186 &mc
->mc_metaslab_txg_list
);
2187 multilist_sublist_t
*mls
=
2188 multilist_sublist_lock(&mc
->mc_metaslab_txg_list
, idx
);
2189 metaslab_t
*msp
= multilist_sublist_head(mls
);
2190 multilist_sublist_unlock(mls
);
2191 while (msp
!= NULL
&& allmem
* zfs_metaslab_mem_limit
/ 100 <
2193 VERIFY3P(mls
, ==, multilist_sublist_lock(
2194 &mc
->mc_metaslab_txg_list
, idx
));
2196 metaslab_idx_func(&mc
->mc_metaslab_txg_list
, msp
));
2198 if (!multilist_link_active(&msp
->ms_class_txg_node
)) {
2199 multilist_sublist_unlock(mls
);
2202 metaslab_t
*next_msp
= multilist_sublist_next(mls
, msp
);
2203 multilist_sublist_unlock(mls
);
2205 * If the metaslab is currently loading there are two
2206 * cases. If it's the metaslab we're evicting, we
2207 * can't continue on or we'll panic when we attempt to
2208 * recursively lock the mutex. If it's another
2209 * metaslab that's loading, it can be safely skipped,
2210 * since we know it's very new and therefore not a
2211 * good eviction candidate. We check later once the
2212 * lock is held that the metaslab is fully loaded
2213 * before actually unloading it.
2215 if (msp
->ms_loading
) {
2218 spl_kmem_cache_inuse(zfs_btree_leaf_cache
);
2222 * We can't unload metaslabs with no spacemap because
2223 * they're not ready to be unloaded yet. We can't
2224 * unload metaslabs with outstanding allocations
2225 * because doing so could cause the metaslab's weight
2226 * to decrease while it's unloaded, which violates an
2227 * invariant that we use to prevent unnecessary
2228 * loading. We also don't unload metaslabs that are
2229 * currently active because they are high-weight
2230 * metaslabs that are likely to be used in the near
2233 mutex_enter(&msp
->ms_lock
);
2234 if (msp
->ms_allocator
== -1 && msp
->ms_sm
!= NULL
&&
2235 msp
->ms_allocating_total
== 0) {
2236 metaslab_unload(msp
);
2238 mutex_exit(&msp
->ms_lock
);
2240 inuse
= spl_kmem_cache_inuse(zfs_btree_leaf_cache
);
2244 (void) mc
, (void) zfs_metaslab_mem_limit
;
2249 metaslab_load_impl(metaslab_t
*msp
)
2253 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2254 ASSERT(msp
->ms_loading
);
2255 ASSERT(!msp
->ms_condensing
);
2258 * We temporarily drop the lock to unblock other operations while we
2259 * are reading the space map. Therefore, metaslab_sync() and
2260 * metaslab_sync_done() can run at the same time as we do.
2262 * If we are using the log space maps, metaslab_sync() can't write to
2263 * the metaslab's space map while we are loading as we only write to
2264 * it when we are flushing the metaslab, and that can't happen while
2265 * we are loading it.
2267 * If we are not using log space maps though, metaslab_sync() can
2268 * append to the space map while we are loading. Therefore we load
2269 * only entries that existed when we started the load. Additionally,
2270 * metaslab_sync_done() has to wait for the load to complete because
2271 * there are potential races like metaslab_load() loading parts of the
2272 * space map that are currently being appended by metaslab_sync(). If
2273 * we didn't, the ms_allocatable would have entries that
2274 * metaslab_sync_done() would try to re-add later.
2276 * That's why before dropping the lock we remember the synced length
2277 * of the metaslab and read up to that point of the space map,
2278 * ignoring entries appended by metaslab_sync() that happen after we
2281 uint64_t length
= msp
->ms_synced_length
;
2282 mutex_exit(&msp
->ms_lock
);
2284 hrtime_t load_start
= gethrtime();
2285 metaslab_rt_arg_t
*mrap
;
2286 if (msp
->ms_allocatable
->rt_arg
== NULL
) {
2287 mrap
= kmem_zalloc(sizeof (*mrap
), KM_SLEEP
);
2289 mrap
= msp
->ms_allocatable
->rt_arg
;
2290 msp
->ms_allocatable
->rt_ops
= NULL
;
2291 msp
->ms_allocatable
->rt_arg
= NULL
;
2293 mrap
->mra_bt
= &msp
->ms_allocatable_by_size
;
2294 mrap
->mra_floor_shift
= metaslab_by_size_min_shift
;
2296 if (msp
->ms_sm
!= NULL
) {
2297 error
= space_map_load_length(msp
->ms_sm
, msp
->ms_allocatable
,
2300 /* Now, populate the size-sorted tree. */
2301 metaslab_rt_create(msp
->ms_allocatable
, mrap
);
2302 msp
->ms_allocatable
->rt_ops
= &metaslab_rt_ops
;
2303 msp
->ms_allocatable
->rt_arg
= mrap
;
2305 struct mssa_arg arg
= {0};
2306 arg
.rt
= msp
->ms_allocatable
;
2308 range_tree_walk(msp
->ms_allocatable
, metaslab_size_sorted_add
,
2312 * Add the size-sorted tree first, since we don't need to load
2313 * the metaslab from the spacemap.
2315 metaslab_rt_create(msp
->ms_allocatable
, mrap
);
2316 msp
->ms_allocatable
->rt_ops
= &metaslab_rt_ops
;
2317 msp
->ms_allocatable
->rt_arg
= mrap
;
2319 * The space map has not been allocated yet, so treat
2320 * all the space in the metaslab as free and add it to the
2321 * ms_allocatable tree.
2323 range_tree_add(msp
->ms_allocatable
,
2324 msp
->ms_start
, msp
->ms_size
);
2328 * If the ms_sm doesn't exist, this means that this
2329 * metaslab hasn't gone through metaslab_sync() and
2330 * thus has never been dirtied. So we shouldn't
2331 * expect any unflushed allocs or frees from previous
2334 ASSERT(range_tree_is_empty(msp
->ms_unflushed_allocs
));
2335 ASSERT(range_tree_is_empty(msp
->ms_unflushed_frees
));
2340 * We need to grab the ms_sync_lock to prevent metaslab_sync() from
2341 * changing the ms_sm (or log_sm) and the metaslab's range trees
2342 * while we are about to use them and populate the ms_allocatable.
2343 * The ms_lock is insufficient for this because metaslab_sync() doesn't
2344 * hold the ms_lock while writing the ms_checkpointing tree to disk.
2346 mutex_enter(&msp
->ms_sync_lock
);
2347 mutex_enter(&msp
->ms_lock
);
2349 ASSERT(!msp
->ms_condensing
);
2350 ASSERT(!msp
->ms_flushing
);
2353 mutex_exit(&msp
->ms_sync_lock
);
2357 ASSERT3P(msp
->ms_group
, !=, NULL
);
2358 msp
->ms_loaded
= B_TRUE
;
2361 * Apply all the unflushed changes to ms_allocatable right
2362 * away so any manipulations we do below have a clear view
2363 * of what is allocated and what is free.
2365 range_tree_walk(msp
->ms_unflushed_allocs
,
2366 range_tree_remove
, msp
->ms_allocatable
);
2367 range_tree_walk(msp
->ms_unflushed_frees
,
2368 range_tree_add
, msp
->ms_allocatable
);
2370 ASSERT3P(msp
->ms_group
, !=, NULL
);
2371 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2372 if (spa_syncing_log_sm(spa
) != NULL
) {
2373 ASSERT(spa_feature_is_enabled(spa
,
2374 SPA_FEATURE_LOG_SPACEMAP
));
2377 * If we use a log space map we add all the segments
2378 * that are in ms_unflushed_frees so they are available
2381 * ms_allocatable needs to contain all free segments
2382 * that are ready for allocations (thus not segments
2383 * from ms_freeing, ms_freed, and the ms_defer trees).
2384 * But if we grab the lock in this code path at a sync
2385 * pass later that 1, then it also contains the
2386 * segments of ms_freed (they were added to it earlier
2387 * in this path through ms_unflushed_frees). So we
2388 * need to remove all the segments that exist in
2389 * ms_freed from ms_allocatable as they will be added
2390 * later in metaslab_sync_done().
2392 * When there's no log space map, the ms_allocatable
2393 * correctly doesn't contain any segments that exist
2394 * in ms_freed [see ms_synced_length].
2396 range_tree_walk(msp
->ms_freed
,
2397 range_tree_remove
, msp
->ms_allocatable
);
2401 * If we are not using the log space map, ms_allocatable
2402 * contains the segments that exist in the ms_defer trees
2403 * [see ms_synced_length]. Thus we need to remove them
2404 * from ms_allocatable as they will be added again in
2405 * metaslab_sync_done().
2407 * If we are using the log space map, ms_allocatable still
2408 * contains the segments that exist in the ms_defer trees.
2409 * Not because it read them through the ms_sm though. But
2410 * because these segments are part of ms_unflushed_frees
2411 * whose segments we add to ms_allocatable earlier in this
2414 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2415 range_tree_walk(msp
->ms_defer
[t
],
2416 range_tree_remove
, msp
->ms_allocatable
);
2420 * Call metaslab_recalculate_weight_and_sort() now that the
2421 * metaslab is loaded so we get the metaslab's real weight.
2423 * Unless this metaslab was created with older software and
2424 * has not yet been converted to use segment-based weight, we
2425 * expect the new weight to be better or equal to the weight
2426 * that the metaslab had while it was not loaded. This is
2427 * because the old weight does not take into account the
2428 * consolidation of adjacent segments between TXGs. [see
2429 * comment for ms_synchist and ms_deferhist[] for more info]
2431 uint64_t weight
= msp
->ms_weight
;
2432 uint64_t max_size
= msp
->ms_max_size
;
2433 metaslab_recalculate_weight_and_sort(msp
);
2434 if (!WEIGHT_IS_SPACEBASED(weight
))
2435 ASSERT3U(weight
, <=, msp
->ms_weight
);
2436 msp
->ms_max_size
= metaslab_largest_allocatable(msp
);
2437 ASSERT3U(max_size
, <=, msp
->ms_max_size
);
2438 hrtime_t load_end
= gethrtime();
2439 msp
->ms_load_time
= load_end
;
2440 zfs_dbgmsg("metaslab_load: txg %llu, spa %s, vdev_id %llu, "
2441 "ms_id %llu, smp_length %llu, "
2442 "unflushed_allocs %llu, unflushed_frees %llu, "
2443 "freed %llu, defer %llu + %llu, unloaded time %llu ms, "
2444 "loading_time %lld ms, ms_max_size %llu, "
2445 "max size error %lld, "
2446 "old_weight %llx, new_weight %llx",
2447 (u_longlong_t
)spa_syncing_txg(spa
), spa_name(spa
),
2448 (u_longlong_t
)msp
->ms_group
->mg_vd
->vdev_id
,
2449 (u_longlong_t
)msp
->ms_id
,
2450 (u_longlong_t
)space_map_length(msp
->ms_sm
),
2451 (u_longlong_t
)range_tree_space(msp
->ms_unflushed_allocs
),
2452 (u_longlong_t
)range_tree_space(msp
->ms_unflushed_frees
),
2453 (u_longlong_t
)range_tree_space(msp
->ms_freed
),
2454 (u_longlong_t
)range_tree_space(msp
->ms_defer
[0]),
2455 (u_longlong_t
)range_tree_space(msp
->ms_defer
[1]),
2456 (longlong_t
)((load_start
- msp
->ms_unload_time
) / 1000000),
2457 (longlong_t
)((load_end
- load_start
) / 1000000),
2458 (u_longlong_t
)msp
->ms_max_size
,
2459 (u_longlong_t
)msp
->ms_max_size
- max_size
,
2460 (u_longlong_t
)weight
, (u_longlong_t
)msp
->ms_weight
);
2462 metaslab_verify_space(msp
, spa_syncing_txg(spa
));
2463 mutex_exit(&msp
->ms_sync_lock
);
2468 metaslab_load(metaslab_t
*msp
)
2470 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2473 * There may be another thread loading the same metaslab, if that's
2474 * the case just wait until the other thread is done and return.
2476 metaslab_load_wait(msp
);
2479 VERIFY(!msp
->ms_loading
);
2480 ASSERT(!msp
->ms_condensing
);
2483 * We set the loading flag BEFORE potentially dropping the lock to
2484 * wait for an ongoing flush (see ms_flushing below). This way other
2485 * threads know that there is already a thread that is loading this
2488 msp
->ms_loading
= B_TRUE
;
2491 * Wait for any in-progress flushing to finish as we drop the ms_lock
2492 * both here (during space_map_load()) and in metaslab_flush() (when
2493 * we flush our changes to the ms_sm).
2495 if (msp
->ms_flushing
)
2496 metaslab_flush_wait(msp
);
2499 * In the possibility that we were waiting for the metaslab to be
2500 * flushed (where we temporarily dropped the ms_lock), ensure that
2501 * no one else loaded the metaslab somehow.
2503 ASSERT(!msp
->ms_loaded
);
2506 * If we're loading a metaslab in the normal class, consider evicting
2507 * another one to keep our memory usage under the limit defined by the
2508 * zfs_metaslab_mem_limit tunable.
2510 if (spa_normal_class(msp
->ms_group
->mg_class
->mc_spa
) ==
2511 msp
->ms_group
->mg_class
) {
2512 metaslab_potentially_evict(msp
->ms_group
->mg_class
);
2515 int error
= metaslab_load_impl(msp
);
2517 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2518 msp
->ms_loading
= B_FALSE
;
2519 cv_broadcast(&msp
->ms_load_cv
);
2525 metaslab_unload(metaslab_t
*msp
)
2527 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2530 * This can happen if a metaslab is selected for eviction (in
2531 * metaslab_potentially_evict) and then unloaded during spa_sync (via
2532 * metaslab_class_evict_old).
2534 if (!msp
->ms_loaded
)
2537 range_tree_vacate(msp
->ms_allocatable
, NULL
, NULL
);
2538 msp
->ms_loaded
= B_FALSE
;
2539 msp
->ms_unload_time
= gethrtime();
2541 msp
->ms_activation_weight
= 0;
2542 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
2544 if (msp
->ms_group
!= NULL
) {
2545 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
2546 multilist_sublist_t
*mls
=
2547 multilist_sublist_lock_obj(&mc
->mc_metaslab_txg_list
, msp
);
2548 if (multilist_link_active(&msp
->ms_class_txg_node
))
2549 multilist_sublist_remove(mls
, msp
);
2550 multilist_sublist_unlock(mls
);
2552 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2553 zfs_dbgmsg("metaslab_unload: txg %llu, spa %s, vdev_id %llu, "
2554 "ms_id %llu, weight %llx, "
2555 "selected txg %llu (%llu ms ago), alloc_txg %llu, "
2556 "loaded %llu ms ago, max_size %llu",
2557 (u_longlong_t
)spa_syncing_txg(spa
), spa_name(spa
),
2558 (u_longlong_t
)msp
->ms_group
->mg_vd
->vdev_id
,
2559 (u_longlong_t
)msp
->ms_id
,
2560 (u_longlong_t
)msp
->ms_weight
,
2561 (u_longlong_t
)msp
->ms_selected_txg
,
2562 (u_longlong_t
)(msp
->ms_unload_time
-
2563 msp
->ms_selected_time
) / 1000 / 1000,
2564 (u_longlong_t
)msp
->ms_alloc_txg
,
2565 (u_longlong_t
)(msp
->ms_unload_time
-
2566 msp
->ms_load_time
) / 1000 / 1000,
2567 (u_longlong_t
)msp
->ms_max_size
);
2571 * We explicitly recalculate the metaslab's weight based on its space
2572 * map (as it is now not loaded). We want unload metaslabs to always
2573 * have their weights calculated from the space map histograms, while
2574 * loaded ones have it calculated from their in-core range tree
2575 * [see metaslab_load()]. This way, the weight reflects the information
2576 * available in-core, whether it is loaded or not.
2578 * If ms_group == NULL means that we came here from metaslab_fini(),
2579 * at which point it doesn't make sense for us to do the recalculation
2582 if (msp
->ms_group
!= NULL
)
2583 metaslab_recalculate_weight_and_sort(msp
);
2587 * We want to optimize the memory use of the per-metaslab range
2588 * trees. To do this, we store the segments in the range trees in
2589 * units of sectors, zero-indexing from the start of the metaslab. If
2590 * the vdev_ms_shift - the vdev_ashift is less than 32, we can store
2591 * the ranges using two uint32_ts, rather than two uint64_ts.
2594 metaslab_calculate_range_tree_type(vdev_t
*vdev
, metaslab_t
*msp
,
2595 uint64_t *start
, uint64_t *shift
)
2597 if (vdev
->vdev_ms_shift
- vdev
->vdev_ashift
< 32 &&
2598 !zfs_metaslab_force_large_segs
) {
2599 *shift
= vdev
->vdev_ashift
;
2600 *start
= msp
->ms_start
;
2601 return (RANGE_SEG32
);
2605 return (RANGE_SEG64
);
2610 metaslab_set_selected_txg(metaslab_t
*msp
, uint64_t txg
)
2612 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2613 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
2614 multilist_sublist_t
*mls
=
2615 multilist_sublist_lock_obj(&mc
->mc_metaslab_txg_list
, msp
);
2616 if (multilist_link_active(&msp
->ms_class_txg_node
))
2617 multilist_sublist_remove(mls
, msp
);
2618 msp
->ms_selected_txg
= txg
;
2619 msp
->ms_selected_time
= gethrtime();
2620 multilist_sublist_insert_tail(mls
, msp
);
2621 multilist_sublist_unlock(mls
);
2625 metaslab_space_update(vdev_t
*vd
, metaslab_class_t
*mc
, int64_t alloc_delta
,
2626 int64_t defer_delta
, int64_t space_delta
)
2628 vdev_space_update(vd
, alloc_delta
, defer_delta
, space_delta
);
2630 ASSERT3P(vd
->vdev_spa
->spa_root_vdev
, ==, vd
->vdev_parent
);
2631 ASSERT(vd
->vdev_ms_count
!= 0);
2633 metaslab_class_space_update(mc
, alloc_delta
, defer_delta
, space_delta
,
2634 vdev_deflated_space(vd
, space_delta
));
2638 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
,
2639 uint64_t txg
, metaslab_t
**msp
)
2641 vdev_t
*vd
= mg
->mg_vd
;
2642 spa_t
*spa
= vd
->vdev_spa
;
2643 objset_t
*mos
= spa
->spa_meta_objset
;
2647 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
2648 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
2649 mutex_init(&ms
->ms_sync_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
2650 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
2651 cv_init(&ms
->ms_flush_cv
, NULL
, CV_DEFAULT
, NULL
);
2652 multilist_link_init(&ms
->ms_class_txg_node
);
2655 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
2656 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
2657 ms
->ms_allocator
= -1;
2658 ms
->ms_new
= B_TRUE
;
2660 vdev_ops_t
*ops
= vd
->vdev_ops
;
2661 if (ops
->vdev_op_metaslab_init
!= NULL
)
2662 ops
->vdev_op_metaslab_init(vd
, &ms
->ms_start
, &ms
->ms_size
);
2665 * We only open space map objects that already exist. All others
2666 * will be opened when we finally allocate an object for it. For
2667 * readonly pools there is no need to open the space map object.
2670 * When called from vdev_expand(), we can't call into the DMU as
2671 * we are holding the spa_config_lock as a writer and we would
2672 * deadlock [see relevant comment in vdev_metaslab_init()]. in
2673 * that case, the object parameter is zero though, so we won't
2674 * call into the DMU.
2676 if (object
!= 0 && !(spa
->spa_mode
== SPA_MODE_READ
&&
2677 !spa
->spa_read_spacemaps
)) {
2678 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
2679 ms
->ms_size
, vd
->vdev_ashift
);
2682 kmem_free(ms
, sizeof (metaslab_t
));
2686 ASSERT(ms
->ms_sm
!= NULL
);
2687 ms
->ms_allocated_space
= space_map_allocated(ms
->ms_sm
);
2690 uint64_t shift
, start
;
2691 range_seg_type_t type
=
2692 metaslab_calculate_range_tree_type(vd
, ms
, &start
, &shift
);
2694 ms
->ms_allocatable
= range_tree_create(NULL
, type
, NULL
, start
, shift
);
2695 for (int t
= 0; t
< TXG_SIZE
; t
++) {
2696 ms
->ms_allocating
[t
] = range_tree_create(NULL
, type
,
2697 NULL
, start
, shift
);
2699 ms
->ms_freeing
= range_tree_create(NULL
, type
, NULL
, start
, shift
);
2700 ms
->ms_freed
= range_tree_create(NULL
, type
, NULL
, start
, shift
);
2701 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2702 ms
->ms_defer
[t
] = range_tree_create(NULL
, type
, NULL
,
2705 ms
->ms_checkpointing
=
2706 range_tree_create(NULL
, type
, NULL
, start
, shift
);
2707 ms
->ms_unflushed_allocs
=
2708 range_tree_create(NULL
, type
, NULL
, start
, shift
);
2710 metaslab_rt_arg_t
*mrap
= kmem_zalloc(sizeof (*mrap
), KM_SLEEP
);
2711 mrap
->mra_bt
= &ms
->ms_unflushed_frees_by_size
;
2712 mrap
->mra_floor_shift
= metaslab_by_size_min_shift
;
2713 ms
->ms_unflushed_frees
= range_tree_create(&metaslab_rt_ops
,
2714 type
, mrap
, start
, shift
);
2716 ms
->ms_trim
= range_tree_create(NULL
, type
, NULL
, start
, shift
);
2718 metaslab_group_add(mg
, ms
);
2719 metaslab_set_fragmentation(ms
, B_FALSE
);
2722 * If we're opening an existing pool (txg == 0) or creating
2723 * a new one (txg == TXG_INITIAL), all space is available now.
2724 * If we're adding space to an existing pool, the new space
2725 * does not become available until after this txg has synced.
2726 * The metaslab's weight will also be initialized when we sync
2727 * out this txg. This ensures that we don't attempt to allocate
2728 * from it before we have initialized it completely.
2730 if (txg
<= TXG_INITIAL
) {
2731 metaslab_sync_done(ms
, 0);
2732 metaslab_space_update(vd
, mg
->mg_class
,
2733 metaslab_allocated_space(ms
), 0, 0);
2737 vdev_dirty(vd
, 0, NULL
, txg
);
2738 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
2747 metaslab_fini_flush_data(metaslab_t
*msp
)
2749 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2751 if (metaslab_unflushed_txg(msp
) == 0) {
2752 ASSERT3P(avl_find(&spa
->spa_metaslabs_by_flushed
, msp
, NULL
),
2756 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
));
2758 mutex_enter(&spa
->spa_flushed_ms_lock
);
2759 avl_remove(&spa
->spa_metaslabs_by_flushed
, msp
);
2760 mutex_exit(&spa
->spa_flushed_ms_lock
);
2762 spa_log_sm_decrement_mscount(spa
, metaslab_unflushed_txg(msp
));
2763 spa_log_summary_decrement_mscount(spa
, metaslab_unflushed_txg(msp
),
2764 metaslab_unflushed_dirty(msp
));
2768 metaslab_unflushed_changes_memused(metaslab_t
*ms
)
2770 return ((range_tree_numsegs(ms
->ms_unflushed_allocs
) +
2771 range_tree_numsegs(ms
->ms_unflushed_frees
)) *
2772 ms
->ms_unflushed_allocs
->rt_root
.bt_elem_size
);
2776 metaslab_fini(metaslab_t
*msp
)
2778 metaslab_group_t
*mg
= msp
->ms_group
;
2779 vdev_t
*vd
= mg
->mg_vd
;
2780 spa_t
*spa
= vd
->vdev_spa
;
2782 metaslab_fini_flush_data(msp
);
2784 metaslab_group_remove(mg
, msp
);
2786 mutex_enter(&msp
->ms_lock
);
2787 VERIFY(msp
->ms_group
== NULL
);
2790 * If this metaslab hasn't been through metaslab_sync_done() yet its
2791 * space hasn't been accounted for in its vdev and doesn't need to be
2795 metaslab_space_update(vd
, mg
->mg_class
,
2796 -metaslab_allocated_space(msp
), 0, -msp
->ms_size
);
2799 space_map_close(msp
->ms_sm
);
2802 metaslab_unload(msp
);
2804 range_tree_destroy(msp
->ms_allocatable
);
2805 range_tree_destroy(msp
->ms_freeing
);
2806 range_tree_destroy(msp
->ms_freed
);
2808 ASSERT3U(spa
->spa_unflushed_stats
.sus_memused
, >=,
2809 metaslab_unflushed_changes_memused(msp
));
2810 spa
->spa_unflushed_stats
.sus_memused
-=
2811 metaslab_unflushed_changes_memused(msp
);
2812 range_tree_vacate(msp
->ms_unflushed_allocs
, NULL
, NULL
);
2813 range_tree_destroy(msp
->ms_unflushed_allocs
);
2814 range_tree_destroy(msp
->ms_checkpointing
);
2815 range_tree_vacate(msp
->ms_unflushed_frees
, NULL
, NULL
);
2816 range_tree_destroy(msp
->ms_unflushed_frees
);
2818 for (int t
= 0; t
< TXG_SIZE
; t
++) {
2819 range_tree_destroy(msp
->ms_allocating
[t
]);
2821 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
2822 range_tree_destroy(msp
->ms_defer
[t
]);
2824 ASSERT0(msp
->ms_deferspace
);
2826 for (int t
= 0; t
< TXG_SIZE
; t
++)
2827 ASSERT(!txg_list_member(&vd
->vdev_ms_list
, msp
, t
));
2829 range_tree_vacate(msp
->ms_trim
, NULL
, NULL
);
2830 range_tree_destroy(msp
->ms_trim
);
2832 mutex_exit(&msp
->ms_lock
);
2833 cv_destroy(&msp
->ms_load_cv
);
2834 cv_destroy(&msp
->ms_flush_cv
);
2835 mutex_destroy(&msp
->ms_lock
);
2836 mutex_destroy(&msp
->ms_sync_lock
);
2837 ASSERT3U(msp
->ms_allocator
, ==, -1);
2839 kmem_free(msp
, sizeof (metaslab_t
));
2842 #define FRAGMENTATION_TABLE_SIZE 17
2845 * This table defines a segment size based fragmentation metric that will
2846 * allow each metaslab to derive its own fragmentation value. This is done
2847 * by calculating the space in each bucket of the spacemap histogram and
2848 * multiplying that by the fragmentation metric in this table. Doing
2849 * this for all buckets and dividing it by the total amount of free
2850 * space in this metaslab (i.e. the total free space in all buckets) gives
2851 * us the fragmentation metric. This means that a high fragmentation metric
2852 * equates to most of the free space being comprised of small segments.
2853 * Conversely, if the metric is low, then most of the free space is in
2854 * large segments. A 10% change in fragmentation equates to approximately
2855 * double the number of segments.
2857 * This table defines 0% fragmented space using 16MB segments. Testing has
2858 * shown that segments that are greater than or equal to 16MB do not suffer
2859 * from drastic performance problems. Using this value, we derive the rest
2860 * of the table. Since the fragmentation value is never stored on disk, it
2861 * is possible to change these calculations in the future.
2863 static const int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
2883 * Calculate the metaslab's fragmentation metric and set ms_fragmentation.
2884 * Setting this value to ZFS_FRAG_INVALID means that the metaslab has not
2885 * been upgraded and does not support this metric. Otherwise, the return
2886 * value should be in the range [0, 100].
2889 metaslab_set_fragmentation(metaslab_t
*msp
, boolean_t nodirty
)
2891 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
2892 uint64_t fragmentation
= 0;
2894 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
2895 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
2897 if (!feature_enabled
) {
2898 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
2903 * A null space map means that the entire metaslab is free
2904 * and thus is not fragmented.
2906 if (msp
->ms_sm
== NULL
) {
2907 msp
->ms_fragmentation
= 0;
2912 * If this metaslab's space map has not been upgraded, flag it
2913 * so that we upgrade next time we encounter it.
2915 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
2916 uint64_t txg
= spa_syncing_txg(spa
);
2917 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
2920 * If we've reached the final dirty txg, then we must
2921 * be shutting down the pool. We don't want to dirty
2922 * any data past this point so skip setting the condense
2923 * flag. We can retry this action the next time the pool
2924 * is imported. We also skip marking this metaslab for
2925 * condensing if the caller has explicitly set nodirty.
2928 spa_writeable(spa
) && txg
< spa_final_dirty_txg(spa
)) {
2929 msp
->ms_condense_wanted
= B_TRUE
;
2930 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2931 zfs_dbgmsg("txg %llu, requesting force condense: "
2932 "ms_id %llu, vdev_id %llu", (u_longlong_t
)txg
,
2933 (u_longlong_t
)msp
->ms_id
,
2934 (u_longlong_t
)vd
->vdev_id
);
2936 msp
->ms_fragmentation
= ZFS_FRAG_INVALID
;
2940 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
2942 uint8_t shift
= msp
->ms_sm
->sm_shift
;
2944 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
2945 FRAGMENTATION_TABLE_SIZE
- 1);
2947 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
2950 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
2953 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
2954 fragmentation
+= space
* zfs_frag_table
[idx
];
2958 fragmentation
/= total
;
2959 ASSERT3U(fragmentation
, <=, 100);
2961 msp
->ms_fragmentation
= fragmentation
;
2965 * Compute a weight -- a selection preference value -- for the given metaslab.
2966 * This is based on the amount of free space, the level of fragmentation,
2967 * the LBA range, and whether the metaslab is loaded.
2970 metaslab_space_weight(metaslab_t
*msp
)
2972 metaslab_group_t
*mg
= msp
->ms_group
;
2973 vdev_t
*vd
= mg
->mg_vd
;
2974 uint64_t weight
, space
;
2976 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
2979 * The baseline weight is the metaslab's free space.
2981 space
= msp
->ms_size
- metaslab_allocated_space(msp
);
2983 if (metaslab_fragmentation_factor_enabled
&&
2984 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
2986 * Use the fragmentation information to inversely scale
2987 * down the baseline weight. We need to ensure that we
2988 * don't exclude this metaslab completely when it's 100%
2989 * fragmented. To avoid this we reduce the fragmented value
2992 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
2995 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
2996 * this metaslab again. The fragmentation metric may have
2997 * decreased the space to something smaller than
2998 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
2999 * so that we can consume any remaining space.
3001 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
3002 space
= SPA_MINBLOCKSIZE
;
3007 * Modern disks have uniform bit density and constant angular velocity.
3008 * Therefore, the outer recording zones are faster (higher bandwidth)
3009 * than the inner zones by the ratio of outer to inner track diameter,
3010 * which is typically around 2:1. We account for this by assigning
3011 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
3012 * In effect, this means that we'll select the metaslab with the most
3013 * free bandwidth rather than simply the one with the most free space.
3015 if (!vd
->vdev_nonrot
&& metaslab_lba_weighting_enabled
) {
3016 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
3017 ASSERT(weight
>= space
&& weight
<= 2 * space
);
3021 * If this metaslab is one we're actively using, adjust its
3022 * weight to make it preferable to any inactive metaslab so
3023 * we'll polish it off. If the fragmentation on this metaslab
3024 * has exceed our threshold, then don't mark it active.
3026 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
3027 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
3028 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
3031 WEIGHT_SET_SPACEBASED(weight
);
3036 * Return the weight of the specified metaslab, according to the segment-based
3037 * weighting algorithm. The metaslab must be loaded. This function can
3038 * be called within a sync pass since it relies only on the metaslab's
3039 * range tree which is always accurate when the metaslab is loaded.
3042 metaslab_weight_from_range_tree(metaslab_t
*msp
)
3044 uint64_t weight
= 0;
3045 uint32_t segments
= 0;
3047 ASSERT(msp
->ms_loaded
);
3049 for (int i
= RANGE_TREE_HISTOGRAM_SIZE
- 1; i
>= SPA_MINBLOCKSHIFT
;
3051 uint8_t shift
= msp
->ms_group
->mg_vd
->vdev_ashift
;
3052 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
3055 segments
+= msp
->ms_allocatable
->rt_histogram
[i
];
3058 * The range tree provides more precision than the space map
3059 * and must be downgraded so that all values fit within the
3060 * space map's histogram. This allows us to compare loaded
3061 * vs. unloaded metaslabs to determine which metaslab is
3062 * considered "best".
3067 if (segments
!= 0) {
3068 WEIGHT_SET_COUNT(weight
, segments
);
3069 WEIGHT_SET_INDEX(weight
, i
);
3070 WEIGHT_SET_ACTIVE(weight
, 0);
3078 * Calculate the weight based on the on-disk histogram. Should be applied
3079 * only to unloaded metaslabs (i.e no incoming allocations) in-order to
3080 * give results consistent with the on-disk state
3083 metaslab_weight_from_spacemap(metaslab_t
*msp
)
3085 space_map_t
*sm
= msp
->ms_sm
;
3086 ASSERT(!msp
->ms_loaded
);
3088 ASSERT3U(space_map_object(sm
), !=, 0);
3089 ASSERT3U(sm
->sm_dbuf
->db_size
, ==, sizeof (space_map_phys_t
));
3092 * Create a joint histogram from all the segments that have made
3093 * it to the metaslab's space map histogram, that are not yet
3094 * available for allocation because they are still in the freeing
3095 * pipeline (e.g. freeing, freed, and defer trees). Then subtract
3096 * these segments from the space map's histogram to get a more
3099 uint64_t deferspace_histogram
[SPACE_MAP_HISTOGRAM_SIZE
] = {0};
3100 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
3101 deferspace_histogram
[i
] += msp
->ms_synchist
[i
];
3102 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
3103 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
3104 deferspace_histogram
[i
] += msp
->ms_deferhist
[t
][i
];
3108 uint64_t weight
= 0;
3109 for (int i
= SPACE_MAP_HISTOGRAM_SIZE
- 1; i
>= 0; i
--) {
3110 ASSERT3U(sm
->sm_phys
->smp_histogram
[i
], >=,
3111 deferspace_histogram
[i
]);
3113 sm
->sm_phys
->smp_histogram
[i
] - deferspace_histogram
[i
];
3115 WEIGHT_SET_COUNT(weight
, count
);
3116 WEIGHT_SET_INDEX(weight
, i
+ sm
->sm_shift
);
3117 WEIGHT_SET_ACTIVE(weight
, 0);
3125 * Compute a segment-based weight for the specified metaslab. The weight
3126 * is determined by highest bucket in the histogram. The information
3127 * for the highest bucket is encoded into the weight value.
3130 metaslab_segment_weight(metaslab_t
*msp
)
3132 metaslab_group_t
*mg
= msp
->ms_group
;
3133 uint64_t weight
= 0;
3134 uint8_t shift
= mg
->mg_vd
->vdev_ashift
;
3136 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3139 * The metaslab is completely free.
3141 if (metaslab_allocated_space(msp
) == 0) {
3142 int idx
= highbit64(msp
->ms_size
) - 1;
3143 int max_idx
= SPACE_MAP_HISTOGRAM_SIZE
+ shift
- 1;
3145 if (idx
< max_idx
) {
3146 WEIGHT_SET_COUNT(weight
, 1ULL);
3147 WEIGHT_SET_INDEX(weight
, idx
);
3149 WEIGHT_SET_COUNT(weight
, 1ULL << (idx
- max_idx
));
3150 WEIGHT_SET_INDEX(weight
, max_idx
);
3152 WEIGHT_SET_ACTIVE(weight
, 0);
3153 ASSERT(!WEIGHT_IS_SPACEBASED(weight
));
3157 ASSERT3U(msp
->ms_sm
->sm_dbuf
->db_size
, ==, sizeof (space_map_phys_t
));
3160 * If the metaslab is fully allocated then just make the weight 0.
3162 if (metaslab_allocated_space(msp
) == msp
->ms_size
)
3165 * If the metaslab is already loaded, then use the range tree to
3166 * determine the weight. Otherwise, we rely on the space map information
3167 * to generate the weight.
3169 if (msp
->ms_loaded
) {
3170 weight
= metaslab_weight_from_range_tree(msp
);
3172 weight
= metaslab_weight_from_spacemap(msp
);
3176 * If the metaslab was active the last time we calculated its weight
3177 * then keep it active. We want to consume the entire region that
3178 * is associated with this weight.
3180 if (msp
->ms_activation_weight
!= 0 && weight
!= 0)
3181 WEIGHT_SET_ACTIVE(weight
, WEIGHT_GET_ACTIVE(msp
->ms_weight
));
3186 * Determine if we should attempt to allocate from this metaslab. If the
3187 * metaslab is loaded, then we can determine if the desired allocation
3188 * can be satisfied by looking at the size of the maximum free segment
3189 * on that metaslab. Otherwise, we make our decision based on the metaslab's
3190 * weight. For segment-based weighting we can determine the maximum
3191 * allocation based on the index encoded in its value. For space-based
3192 * weights we rely on the entire weight (excluding the weight-type bit).
3195 metaslab_should_allocate(metaslab_t
*msp
, uint64_t asize
, boolean_t try_hard
)
3198 * If the metaslab is loaded, ms_max_size is definitive and we can use
3199 * the fast check. If it's not, the ms_max_size is a lower bound (once
3200 * set), and we should use the fast check as long as we're not in
3201 * try_hard and it's been less than zfs_metaslab_max_size_cache_sec
3202 * seconds since the metaslab was unloaded.
3204 if (msp
->ms_loaded
||
3205 (msp
->ms_max_size
!= 0 && !try_hard
&& gethrtime() <
3206 msp
->ms_unload_time
+ SEC2NSEC(zfs_metaslab_max_size_cache_sec
)))
3207 return (msp
->ms_max_size
>= asize
);
3209 boolean_t should_allocate
;
3210 if (!WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
3212 * The metaslab segment weight indicates segments in the
3213 * range [2^i, 2^(i+1)), where i is the index in the weight.
3214 * Since the asize might be in the middle of the range, we
3215 * should attempt the allocation if asize < 2^(i+1).
3217 should_allocate
= (asize
<
3218 1ULL << (WEIGHT_GET_INDEX(msp
->ms_weight
) + 1));
3220 should_allocate
= (asize
<=
3221 (msp
->ms_weight
& ~METASLAB_WEIGHT_TYPE
));
3224 return (should_allocate
);
3228 metaslab_weight(metaslab_t
*msp
, boolean_t nodirty
)
3230 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
3231 spa_t
*spa
= vd
->vdev_spa
;
3234 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3236 metaslab_set_fragmentation(msp
, nodirty
);
3239 * Update the maximum size. If the metaslab is loaded, this will
3240 * ensure that we get an accurate maximum size if newly freed space
3241 * has been added back into the free tree. If the metaslab is
3242 * unloaded, we check if there's a larger free segment in the
3243 * unflushed frees. This is a lower bound on the largest allocatable
3244 * segment size. Coalescing of adjacent entries may reveal larger
3245 * allocatable segments, but we aren't aware of those until loading
3246 * the space map into a range tree.
3248 if (msp
->ms_loaded
) {
3249 msp
->ms_max_size
= metaslab_largest_allocatable(msp
);
3251 msp
->ms_max_size
= MAX(msp
->ms_max_size
,
3252 metaslab_largest_unflushed_free(msp
));
3256 * Segment-based weighting requires space map histogram support.
3258 if (zfs_metaslab_segment_weight_enabled
&&
3259 spa_feature_is_enabled(spa
, SPA_FEATURE_SPACEMAP_HISTOGRAM
) &&
3260 (msp
->ms_sm
== NULL
|| msp
->ms_sm
->sm_dbuf
->db_size
==
3261 sizeof (space_map_phys_t
))) {
3262 weight
= metaslab_segment_weight(msp
);
3264 weight
= metaslab_space_weight(msp
);
3270 metaslab_recalculate_weight_and_sort(metaslab_t
*msp
)
3272 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3274 /* note: we preserve the mask (e.g. indication of primary, etc..) */
3275 uint64_t was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
3276 metaslab_group_sort(msp
->ms_group
, msp
,
3277 metaslab_weight(msp
, B_FALSE
) | was_active
);
3281 metaslab_activate_allocator(metaslab_group_t
*mg
, metaslab_t
*msp
,
3282 int allocator
, uint64_t activation_weight
)
3284 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
3285 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3288 * If we're activating for the claim code, we don't want to actually
3289 * set the metaslab up for a specific allocator.
3291 if (activation_weight
== METASLAB_WEIGHT_CLAIM
) {
3292 ASSERT0(msp
->ms_activation_weight
);
3293 msp
->ms_activation_weight
= msp
->ms_weight
;
3294 metaslab_group_sort(mg
, msp
, msp
->ms_weight
|
3299 metaslab_t
**mspp
= (activation_weight
== METASLAB_WEIGHT_PRIMARY
?
3300 &mga
->mga_primary
: &mga
->mga_secondary
);
3302 mutex_enter(&mg
->mg_lock
);
3303 if (*mspp
!= NULL
) {
3304 mutex_exit(&mg
->mg_lock
);
3309 ASSERT3S(msp
->ms_allocator
, ==, -1);
3310 msp
->ms_allocator
= allocator
;
3311 msp
->ms_primary
= (activation_weight
== METASLAB_WEIGHT_PRIMARY
);
3313 ASSERT0(msp
->ms_activation_weight
);
3314 msp
->ms_activation_weight
= msp
->ms_weight
;
3315 metaslab_group_sort_impl(mg
, msp
,
3316 msp
->ms_weight
| activation_weight
);
3317 mutex_exit(&mg
->mg_lock
);
3323 metaslab_activate(metaslab_t
*msp
, int allocator
, uint64_t activation_weight
)
3325 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3328 * The current metaslab is already activated for us so there
3329 * is nothing to do. Already activated though, doesn't mean
3330 * that this metaslab is activated for our allocator nor our
3331 * requested activation weight. The metaslab could have started
3332 * as an active one for our allocator but changed allocators
3333 * while we were waiting to grab its ms_lock or we stole it
3334 * [see find_valid_metaslab()]. This means that there is a
3335 * possibility of passivating a metaslab of another allocator
3336 * or from a different activation mask, from this thread.
3338 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) != 0) {
3339 ASSERT(msp
->ms_loaded
);
3343 int error
= metaslab_load(msp
);
3345 metaslab_group_sort(msp
->ms_group
, msp
, 0);
3350 * When entering metaslab_load() we may have dropped the
3351 * ms_lock because we were loading this metaslab, or we
3352 * were waiting for another thread to load it for us. In
3353 * that scenario, we recheck the weight of the metaslab
3354 * to see if it was activated by another thread.
3356 * If the metaslab was activated for another allocator or
3357 * it was activated with a different activation weight (e.g.
3358 * we wanted to make it a primary but it was activated as
3359 * secondary) we return error (EBUSY).
3361 * If the metaslab was activated for the same allocator
3362 * and requested activation mask, skip activating it.
3364 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) != 0) {
3365 if (msp
->ms_allocator
!= allocator
)
3368 if ((msp
->ms_weight
& activation_weight
) == 0)
3369 return (SET_ERROR(EBUSY
));
3371 EQUIV((activation_weight
== METASLAB_WEIGHT_PRIMARY
),
3377 * If the metaslab has literally 0 space, it will have weight 0. In
3378 * that case, don't bother activating it. This can happen if the
3379 * metaslab had space during find_valid_metaslab, but another thread
3380 * loaded it and used all that space while we were waiting to grab the
3383 if (msp
->ms_weight
== 0) {
3384 ASSERT0(range_tree_space(msp
->ms_allocatable
));
3385 return (SET_ERROR(ENOSPC
));
3388 if ((error
= metaslab_activate_allocator(msp
->ms_group
, msp
,
3389 allocator
, activation_weight
)) != 0) {
3393 ASSERT(msp
->ms_loaded
);
3394 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
3400 metaslab_passivate_allocator(metaslab_group_t
*mg
, metaslab_t
*msp
,
3403 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3404 ASSERT(msp
->ms_loaded
);
3406 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
) {
3407 metaslab_group_sort(mg
, msp
, weight
);
3411 mutex_enter(&mg
->mg_lock
);
3412 ASSERT3P(msp
->ms_group
, ==, mg
);
3413 ASSERT3S(0, <=, msp
->ms_allocator
);
3414 ASSERT3U(msp
->ms_allocator
, <, mg
->mg_allocators
);
3416 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[msp
->ms_allocator
];
3417 if (msp
->ms_primary
) {
3418 ASSERT3P(mga
->mga_primary
, ==, msp
);
3419 ASSERT(msp
->ms_weight
& METASLAB_WEIGHT_PRIMARY
);
3420 mga
->mga_primary
= NULL
;
3422 ASSERT3P(mga
->mga_secondary
, ==, msp
);
3423 ASSERT(msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
);
3424 mga
->mga_secondary
= NULL
;
3426 msp
->ms_allocator
= -1;
3427 metaslab_group_sort_impl(mg
, msp
, weight
);
3428 mutex_exit(&mg
->mg_lock
);
3432 metaslab_passivate(metaslab_t
*msp
, uint64_t weight
)
3434 uint64_t size __maybe_unused
= weight
& ~METASLAB_WEIGHT_TYPE
;
3437 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
3438 * this metaslab again. In that case, it had better be empty,
3439 * or we would be leaving space on the table.
3441 ASSERT(!WEIGHT_IS_SPACEBASED(msp
->ms_weight
) ||
3442 size
>= SPA_MINBLOCKSIZE
||
3443 range_tree_space(msp
->ms_allocatable
) == 0);
3444 ASSERT0(weight
& METASLAB_ACTIVE_MASK
);
3446 ASSERT(msp
->ms_activation_weight
!= 0);
3447 msp
->ms_activation_weight
= 0;
3448 metaslab_passivate_allocator(msp
->ms_group
, msp
, weight
);
3449 ASSERT0(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
3453 * Segment-based metaslabs are activated once and remain active until
3454 * we either fail an allocation attempt (similar to space-based metaslabs)
3455 * or have exhausted the free space in zfs_metaslab_switch_threshold
3456 * buckets since the metaslab was activated. This function checks to see
3457 * if we've exhausted the zfs_metaslab_switch_threshold buckets in the
3458 * metaslab and passivates it proactively. This will allow us to select a
3459 * metaslab with a larger contiguous region, if any, remaining within this
3460 * metaslab group. If we're in sync pass > 1, then we continue using this
3461 * metaslab so that we don't dirty more block and cause more sync passes.
3464 metaslab_segment_may_passivate(metaslab_t
*msp
)
3466 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
3468 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
) || spa_sync_pass(spa
) > 1)
3472 * Since we are in the middle of a sync pass, the most accurate
3473 * information that is accessible to us is the in-core range tree
3474 * histogram; calculate the new weight based on that information.
3476 uint64_t weight
= metaslab_weight_from_range_tree(msp
);
3477 int activation_idx
= WEIGHT_GET_INDEX(msp
->ms_activation_weight
);
3478 int current_idx
= WEIGHT_GET_INDEX(weight
);
3480 if (current_idx
<= activation_idx
- zfs_metaslab_switch_threshold
)
3481 metaslab_passivate(msp
, weight
);
3485 metaslab_preload(void *arg
)
3487 metaslab_t
*msp
= arg
;
3488 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
3489 spa_t
*spa
= mc
->mc_spa
;
3490 fstrans_cookie_t cookie
= spl_fstrans_mark();
3492 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
3494 mutex_enter(&msp
->ms_lock
);
3495 (void) metaslab_load(msp
);
3496 metaslab_set_selected_txg(msp
, spa_syncing_txg(spa
));
3497 mutex_exit(&msp
->ms_lock
);
3498 spl_fstrans_unmark(cookie
);
3502 metaslab_group_preload(metaslab_group_t
*mg
)
3504 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
3506 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
3509 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
3510 taskq_wait_outstanding(mg
->mg_taskq
, 0);
3514 mutex_enter(&mg
->mg_lock
);
3517 * Load the next potential metaslabs
3519 for (msp
= avl_first(t
); msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
3520 ASSERT3P(msp
->ms_group
, ==, mg
);
3523 * We preload only the maximum number of metaslabs specified
3524 * by metaslab_preload_limit. If a metaslab is being forced
3525 * to condense then we preload it too. This will ensure
3526 * that force condensing happens in the next txg.
3528 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
3532 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
3533 msp
, TQ_SLEEP
) != TASKQID_INVALID
);
3535 mutex_exit(&mg
->mg_lock
);
3539 * Determine if the space map's on-disk footprint is past our tolerance for
3540 * inefficiency. We would like to use the following criteria to make our
3543 * 1. Do not condense if the size of the space map object would dramatically
3544 * increase as a result of writing out the free space range tree.
3546 * 2. Condense if the on on-disk space map representation is at least
3547 * zfs_condense_pct/100 times the size of the optimal representation
3548 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, optimal = 1.1MB).
3550 * 3. Do not condense if the on-disk size of the space map does not actually
3553 * Unfortunately, we cannot compute the on-disk size of the space map in this
3554 * context because we cannot accurately compute the effects of compression, etc.
3555 * Instead, we apply the heuristic described in the block comment for
3556 * zfs_metaslab_condense_block_threshold - we only condense if the space used
3557 * is greater than a threshold number of blocks.
3560 metaslab_should_condense(metaslab_t
*msp
)
3562 space_map_t
*sm
= msp
->ms_sm
;
3563 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
3564 uint64_t vdev_blocksize
= 1ULL << vd
->vdev_ashift
;
3566 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3567 ASSERT(msp
->ms_loaded
);
3569 ASSERT3U(spa_sync_pass(vd
->vdev_spa
), ==, 1);
3572 * We always condense metaslabs that are empty and metaslabs for
3573 * which a condense request has been made.
3575 if (range_tree_numsegs(msp
->ms_allocatable
) == 0 ||
3576 msp
->ms_condense_wanted
)
3579 uint64_t record_size
= MAX(sm
->sm_blksz
, vdev_blocksize
);
3580 uint64_t object_size
= space_map_length(sm
);
3581 uint64_t optimal_size
= space_map_estimate_optimal_size(sm
,
3582 msp
->ms_allocatable
, SM_NO_VDEVID
);
3584 return (object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
3585 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
3589 * Condense the on-disk space map representation to its minimized form.
3590 * The minimized form consists of a small number of allocations followed
3591 * by the entries of the free range tree (ms_allocatable). The condensed
3592 * spacemap contains all the entries of previous TXGs (including those in
3593 * the pool-wide log spacemaps; thus this is effectively a superset of
3594 * metaslab_flush()), but this TXG's entries still need to be written.
3597 metaslab_condense(metaslab_t
*msp
, dmu_tx_t
*tx
)
3599 range_tree_t
*condense_tree
;
3600 space_map_t
*sm
= msp
->ms_sm
;
3601 uint64_t txg
= dmu_tx_get_txg(tx
);
3602 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
3604 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3605 ASSERT(msp
->ms_loaded
);
3606 ASSERT(msp
->ms_sm
!= NULL
);
3609 * In order to condense the space map, we need to change it so it
3610 * only describes which segments are currently allocated and free.
3612 * All the current free space resides in the ms_allocatable, all
3613 * the ms_defer trees, and all the ms_allocating trees. We ignore
3614 * ms_freed because it is empty because we're in sync pass 1. We
3615 * ignore ms_freeing because these changes are not yet reflected
3616 * in the spacemap (they will be written later this txg).
3618 * So to truncate the space map to represent all the entries of
3619 * previous TXGs we do the following:
3621 * 1] We create a range tree (condense tree) that is 100% empty.
3622 * 2] We add to it all segments found in the ms_defer trees
3623 * as those segments are marked as free in the original space
3624 * map. We do the same with the ms_allocating trees for the same
3625 * reason. Adding these segments should be a relatively
3626 * inexpensive operation since we expect these trees to have a
3627 * small number of nodes.
3628 * 3] We vacate any unflushed allocs, since they are not frees we
3629 * need to add to the condense tree. Then we vacate any
3630 * unflushed frees as they should already be part of ms_allocatable.
3631 * 4] At this point, we would ideally like to add all segments
3632 * in the ms_allocatable tree from the condense tree. This way
3633 * we would write all the entries of the condense tree as the
3634 * condensed space map, which would only contain freed
3635 * segments with everything else assumed to be allocated.
3637 * Doing so can be prohibitively expensive as ms_allocatable can
3638 * be large, and therefore computationally expensive to add to
3639 * the condense_tree. Instead we first sync out an entry marking
3640 * everything as allocated, then the condense_tree and then the
3641 * ms_allocatable, in the condensed space map. While this is not
3642 * optimal, it is typically close to optimal and more importantly
3643 * much cheaper to compute.
3645 * 5] Finally, as both of the unflushed trees were written to our
3646 * new and condensed metaslab space map, we basically flushed
3647 * all the unflushed changes to disk, thus we call
3648 * metaslab_flush_update().
3650 ASSERT3U(spa_sync_pass(spa
), ==, 1);
3651 ASSERT(range_tree_is_empty(msp
->ms_freed
)); /* since it is pass 1 */
3653 zfs_dbgmsg("condensing: txg %llu, msp[%llu] %px, vdev id %llu, "
3654 "spa %s, smp size %llu, segments %llu, forcing condense=%s",
3655 (u_longlong_t
)txg
, (u_longlong_t
)msp
->ms_id
, msp
,
3656 (u_longlong_t
)msp
->ms_group
->mg_vd
->vdev_id
,
3657 spa
->spa_name
, (u_longlong_t
)space_map_length(msp
->ms_sm
),
3658 (u_longlong_t
)range_tree_numsegs(msp
->ms_allocatable
),
3659 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
3661 msp
->ms_condense_wanted
= B_FALSE
;
3663 range_seg_type_t type
;
3664 uint64_t shift
, start
;
3665 type
= metaslab_calculate_range_tree_type(msp
->ms_group
->mg_vd
, msp
,
3668 condense_tree
= range_tree_create(NULL
, type
, NULL
, start
, shift
);
3670 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
3671 range_tree_walk(msp
->ms_defer
[t
],
3672 range_tree_add
, condense_tree
);
3675 for (int t
= 0; t
< TXG_CONCURRENT_STATES
; t
++) {
3676 range_tree_walk(msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
],
3677 range_tree_add
, condense_tree
);
3680 ASSERT3U(spa
->spa_unflushed_stats
.sus_memused
, >=,
3681 metaslab_unflushed_changes_memused(msp
));
3682 spa
->spa_unflushed_stats
.sus_memused
-=
3683 metaslab_unflushed_changes_memused(msp
);
3684 range_tree_vacate(msp
->ms_unflushed_allocs
, NULL
, NULL
);
3685 range_tree_vacate(msp
->ms_unflushed_frees
, NULL
, NULL
);
3688 * We're about to drop the metaslab's lock thus allowing other
3689 * consumers to change it's content. Set the metaslab's ms_condensing
3690 * flag to ensure that allocations on this metaslab do not occur
3691 * while we're in the middle of committing it to disk. This is only
3692 * critical for ms_allocatable as all other range trees use per TXG
3693 * views of their content.
3695 msp
->ms_condensing
= B_TRUE
;
3697 mutex_exit(&msp
->ms_lock
);
3698 uint64_t object
= space_map_object(msp
->ms_sm
);
3699 space_map_truncate(sm
,
3700 spa_feature_is_enabled(spa
, SPA_FEATURE_LOG_SPACEMAP
) ?
3701 zfs_metaslab_sm_blksz_with_log
: zfs_metaslab_sm_blksz_no_log
, tx
);
3704 * space_map_truncate() may have reallocated the spacemap object.
3705 * If so, update the vdev_ms_array.
3707 if (space_map_object(msp
->ms_sm
) != object
) {
3708 object
= space_map_object(msp
->ms_sm
);
3709 dmu_write(spa
->spa_meta_objset
,
3710 msp
->ms_group
->mg_vd
->vdev_ms_array
, sizeof (uint64_t) *
3711 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
3716 * When the log space map feature is enabled, each space map will
3717 * always have ALLOCS followed by FREES for each sync pass. This is
3718 * typically true even when the log space map feature is disabled,
3719 * except from the case where a metaslab goes through metaslab_sync()
3720 * and gets condensed. In that case the metaslab's space map will have
3721 * ALLOCS followed by FREES (due to condensing) followed by ALLOCS
3722 * followed by FREES (due to space_map_write() in metaslab_sync()) for
3725 range_tree_t
*tmp_tree
= range_tree_create(NULL
, type
, NULL
, start
,
3727 range_tree_add(tmp_tree
, msp
->ms_start
, msp
->ms_size
);
3728 space_map_write(sm
, tmp_tree
, SM_ALLOC
, SM_NO_VDEVID
, tx
);
3729 space_map_write(sm
, msp
->ms_allocatable
, SM_FREE
, SM_NO_VDEVID
, tx
);
3730 space_map_write(sm
, condense_tree
, SM_FREE
, SM_NO_VDEVID
, tx
);
3732 range_tree_vacate(condense_tree
, NULL
, NULL
);
3733 range_tree_destroy(condense_tree
);
3734 range_tree_vacate(tmp_tree
, NULL
, NULL
);
3735 range_tree_destroy(tmp_tree
);
3736 mutex_enter(&msp
->ms_lock
);
3738 msp
->ms_condensing
= B_FALSE
;
3739 metaslab_flush_update(msp
, tx
);
3743 metaslab_unflushed_add(metaslab_t
*msp
, dmu_tx_t
*tx
)
3745 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
3746 ASSERT(spa_syncing_log_sm(spa
) != NULL
);
3747 ASSERT(msp
->ms_sm
!= NULL
);
3748 ASSERT(range_tree_is_empty(msp
->ms_unflushed_allocs
));
3749 ASSERT(range_tree_is_empty(msp
->ms_unflushed_frees
));
3751 mutex_enter(&spa
->spa_flushed_ms_lock
);
3752 metaslab_set_unflushed_txg(msp
, spa_syncing_txg(spa
), tx
);
3753 metaslab_set_unflushed_dirty(msp
, B_TRUE
);
3754 avl_add(&spa
->spa_metaslabs_by_flushed
, msp
);
3755 mutex_exit(&spa
->spa_flushed_ms_lock
);
3757 spa_log_sm_increment_current_mscount(spa
);
3758 spa_log_summary_add_flushed_metaslab(spa
, B_TRUE
);
3762 metaslab_unflushed_bump(metaslab_t
*msp
, dmu_tx_t
*tx
, boolean_t dirty
)
3764 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
3765 ASSERT(spa_syncing_log_sm(spa
) != NULL
);
3766 ASSERT(msp
->ms_sm
!= NULL
);
3767 ASSERT(metaslab_unflushed_txg(msp
) != 0);
3768 ASSERT3P(avl_find(&spa
->spa_metaslabs_by_flushed
, msp
, NULL
), ==, msp
);
3769 ASSERT(range_tree_is_empty(msp
->ms_unflushed_allocs
));
3770 ASSERT(range_tree_is_empty(msp
->ms_unflushed_frees
));
3772 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(spa
));
3774 /* update metaslab's position in our flushing tree */
3775 uint64_t ms_prev_flushed_txg
= metaslab_unflushed_txg(msp
);
3776 boolean_t ms_prev_flushed_dirty
= metaslab_unflushed_dirty(msp
);
3777 mutex_enter(&spa
->spa_flushed_ms_lock
);
3778 avl_remove(&spa
->spa_metaslabs_by_flushed
, msp
);
3779 metaslab_set_unflushed_txg(msp
, spa_syncing_txg(spa
), tx
);
3780 metaslab_set_unflushed_dirty(msp
, dirty
);
3781 avl_add(&spa
->spa_metaslabs_by_flushed
, msp
);
3782 mutex_exit(&spa
->spa_flushed_ms_lock
);
3784 /* update metaslab counts of spa_log_sm_t nodes */
3785 spa_log_sm_decrement_mscount(spa
, ms_prev_flushed_txg
);
3786 spa_log_sm_increment_current_mscount(spa
);
3788 /* update log space map summary */
3789 spa_log_summary_decrement_mscount(spa
, ms_prev_flushed_txg
,
3790 ms_prev_flushed_dirty
);
3791 spa_log_summary_add_flushed_metaslab(spa
, dirty
);
3793 /* cleanup obsolete logs if any */
3794 spa_cleanup_old_sm_logs(spa
, tx
);
3798 * Called when the metaslab has been flushed (its own spacemap now reflects
3799 * all the contents of the pool-wide spacemap log). Updates the metaslab's
3800 * metadata and any pool-wide related log space map data (e.g. summary,
3801 * obsolete logs, etc..) to reflect that.
3804 metaslab_flush_update(metaslab_t
*msp
, dmu_tx_t
*tx
)
3806 metaslab_group_t
*mg
= msp
->ms_group
;
3807 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
3809 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3811 ASSERT3U(spa_sync_pass(spa
), ==, 1);
3814 * Just because a metaslab got flushed, that doesn't mean that
3815 * it will pass through metaslab_sync_done(). Thus, make sure to
3816 * update ms_synced_length here in case it doesn't.
3818 msp
->ms_synced_length
= space_map_length(msp
->ms_sm
);
3821 * We may end up here from metaslab_condense() without the
3822 * feature being active. In that case this is a no-op.
3824 if (!spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
) ||
3825 metaslab_unflushed_txg(msp
) == 0)
3828 metaslab_unflushed_bump(msp
, tx
, B_FALSE
);
3832 metaslab_flush(metaslab_t
*msp
, dmu_tx_t
*tx
)
3834 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
3836 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
3837 ASSERT3U(spa_sync_pass(spa
), ==, 1);
3838 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
));
3840 ASSERT(msp
->ms_sm
!= NULL
);
3841 ASSERT(metaslab_unflushed_txg(msp
) != 0);
3842 ASSERT(avl_find(&spa
->spa_metaslabs_by_flushed
, msp
, NULL
) != NULL
);
3845 * There is nothing wrong with flushing the same metaslab twice, as
3846 * this codepath should work on that case. However, the current
3847 * flushing scheme makes sure to avoid this situation as we would be
3848 * making all these calls without having anything meaningful to write
3849 * to disk. We assert this behavior here.
3851 ASSERT3U(metaslab_unflushed_txg(msp
), <, dmu_tx_get_txg(tx
));
3854 * We can not flush while loading, because then we would
3855 * not load the ms_unflushed_{allocs,frees}.
3857 if (msp
->ms_loading
)
3860 metaslab_verify_space(msp
, dmu_tx_get_txg(tx
));
3861 metaslab_verify_weight_and_frag(msp
);
3864 * Metaslab condensing is effectively flushing. Therefore if the
3865 * metaslab can be condensed we can just condense it instead of
3868 * Note that metaslab_condense() does call metaslab_flush_update()
3869 * so we can just return immediately after condensing. We also
3870 * don't need to care about setting ms_flushing or broadcasting
3871 * ms_flush_cv, even if we temporarily drop the ms_lock in
3872 * metaslab_condense(), as the metaslab is already loaded.
3874 if (msp
->ms_loaded
&& metaslab_should_condense(msp
)) {
3875 metaslab_group_t
*mg
= msp
->ms_group
;
3878 * For all histogram operations below refer to the
3879 * comments of metaslab_sync() where we follow a
3880 * similar procedure.
3882 metaslab_group_histogram_verify(mg
);
3883 metaslab_class_histogram_verify(mg
->mg_class
);
3884 metaslab_group_histogram_remove(mg
, msp
);
3886 metaslab_condense(msp
, tx
);
3888 space_map_histogram_clear(msp
->ms_sm
);
3889 space_map_histogram_add(msp
->ms_sm
, msp
->ms_allocatable
, tx
);
3890 ASSERT(range_tree_is_empty(msp
->ms_freed
));
3891 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
3892 space_map_histogram_add(msp
->ms_sm
,
3893 msp
->ms_defer
[t
], tx
);
3895 metaslab_aux_histograms_update(msp
);
3897 metaslab_group_histogram_add(mg
, msp
);
3898 metaslab_group_histogram_verify(mg
);
3899 metaslab_class_histogram_verify(mg
->mg_class
);
3901 metaslab_verify_space(msp
, dmu_tx_get_txg(tx
));
3904 * Since we recreated the histogram (and potentially
3905 * the ms_sm too while condensing) ensure that the
3906 * weight is updated too because we are not guaranteed
3907 * that this metaslab is dirty and will go through
3908 * metaslab_sync_done().
3910 metaslab_recalculate_weight_and_sort(msp
);
3914 msp
->ms_flushing
= B_TRUE
;
3915 uint64_t sm_len_before
= space_map_length(msp
->ms_sm
);
3917 mutex_exit(&msp
->ms_lock
);
3918 space_map_write(msp
->ms_sm
, msp
->ms_unflushed_allocs
, SM_ALLOC
,
3920 space_map_write(msp
->ms_sm
, msp
->ms_unflushed_frees
, SM_FREE
,
3922 mutex_enter(&msp
->ms_lock
);
3924 uint64_t sm_len_after
= space_map_length(msp
->ms_sm
);
3925 if (zfs_flags
& ZFS_DEBUG_LOG_SPACEMAP
) {
3926 zfs_dbgmsg("flushing: txg %llu, spa %s, vdev_id %llu, "
3927 "ms_id %llu, unflushed_allocs %llu, unflushed_frees %llu, "
3928 "appended %llu bytes", (u_longlong_t
)dmu_tx_get_txg(tx
),
3930 (u_longlong_t
)msp
->ms_group
->mg_vd
->vdev_id
,
3931 (u_longlong_t
)msp
->ms_id
,
3932 (u_longlong_t
)range_tree_space(msp
->ms_unflushed_allocs
),
3933 (u_longlong_t
)range_tree_space(msp
->ms_unflushed_frees
),
3934 (u_longlong_t
)(sm_len_after
- sm_len_before
));
3937 ASSERT3U(spa
->spa_unflushed_stats
.sus_memused
, >=,
3938 metaslab_unflushed_changes_memused(msp
));
3939 spa
->spa_unflushed_stats
.sus_memused
-=
3940 metaslab_unflushed_changes_memused(msp
);
3941 range_tree_vacate(msp
->ms_unflushed_allocs
, NULL
, NULL
);
3942 range_tree_vacate(msp
->ms_unflushed_frees
, NULL
, NULL
);
3944 metaslab_verify_space(msp
, dmu_tx_get_txg(tx
));
3945 metaslab_verify_weight_and_frag(msp
);
3947 metaslab_flush_update(msp
, tx
);
3949 metaslab_verify_space(msp
, dmu_tx_get_txg(tx
));
3950 metaslab_verify_weight_and_frag(msp
);
3952 msp
->ms_flushing
= B_FALSE
;
3953 cv_broadcast(&msp
->ms_flush_cv
);
3958 * Write a metaslab to disk in the context of the specified transaction group.
3961 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
3963 metaslab_group_t
*mg
= msp
->ms_group
;
3964 vdev_t
*vd
= mg
->mg_vd
;
3965 spa_t
*spa
= vd
->vdev_spa
;
3966 objset_t
*mos
= spa_meta_objset(spa
);
3967 range_tree_t
*alloctree
= msp
->ms_allocating
[txg
& TXG_MASK
];
3970 ASSERT(!vd
->vdev_ishole
);
3973 * This metaslab has just been added so there's no work to do now.
3976 ASSERT0(range_tree_space(alloctree
));
3977 ASSERT0(range_tree_space(msp
->ms_freeing
));
3978 ASSERT0(range_tree_space(msp
->ms_freed
));
3979 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
3980 ASSERT0(range_tree_space(msp
->ms_trim
));
3985 * Normally, we don't want to process a metaslab if there are no
3986 * allocations or frees to perform. However, if the metaslab is being
3987 * forced to condense, it's loaded and we're not beyond the final
3988 * dirty txg, we need to let it through. Not condensing beyond the
3989 * final dirty txg prevents an issue where metaslabs that need to be
3990 * condensed but were loaded for other reasons could cause a panic
3991 * here. By only checking the txg in that branch of the conditional,
3992 * we preserve the utility of the VERIFY statements in all other
3995 if (range_tree_is_empty(alloctree
) &&
3996 range_tree_is_empty(msp
->ms_freeing
) &&
3997 range_tree_is_empty(msp
->ms_checkpointing
) &&
3998 !(msp
->ms_loaded
&& msp
->ms_condense_wanted
&&
3999 txg
<= spa_final_dirty_txg(spa
)))
4003 VERIFY3U(txg
, <=, spa_final_dirty_txg(spa
));
4006 * The only state that can actually be changing concurrently
4007 * with metaslab_sync() is the metaslab's ms_allocatable. No
4008 * other thread can be modifying this txg's alloc, freeing,
4009 * freed, or space_map_phys_t. We drop ms_lock whenever we
4010 * could call into the DMU, because the DMU can call down to
4011 * us (e.g. via zio_free()) at any time.
4013 * The spa_vdev_remove_thread() can be reading metaslab state
4014 * concurrently, and it is locked out by the ms_sync_lock.
4015 * Note that the ms_lock is insufficient for this, because it
4016 * is dropped by space_map_write().
4018 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
4021 * Generate a log space map if one doesn't exist already.
4023 spa_generate_syncing_log_sm(spa
, tx
);
4025 if (msp
->ms_sm
== NULL
) {
4026 uint64_t new_object
= space_map_alloc(mos
,
4027 spa_feature_is_enabled(spa
, SPA_FEATURE_LOG_SPACEMAP
) ?
4028 zfs_metaslab_sm_blksz_with_log
:
4029 zfs_metaslab_sm_blksz_no_log
, tx
);
4030 VERIFY3U(new_object
, !=, 0);
4032 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
4033 msp
->ms_id
, sizeof (uint64_t), &new_object
, tx
);
4035 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
4036 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
));
4037 ASSERT(msp
->ms_sm
!= NULL
);
4039 ASSERT(range_tree_is_empty(msp
->ms_unflushed_allocs
));
4040 ASSERT(range_tree_is_empty(msp
->ms_unflushed_frees
));
4041 ASSERT0(metaslab_allocated_space(msp
));
4044 if (!range_tree_is_empty(msp
->ms_checkpointing
) &&
4045 vd
->vdev_checkpoint_sm
== NULL
) {
4046 ASSERT(spa_has_checkpoint(spa
));
4048 uint64_t new_object
= space_map_alloc(mos
,
4049 zfs_vdev_standard_sm_blksz
, tx
);
4050 VERIFY3U(new_object
, !=, 0);
4052 VERIFY0(space_map_open(&vd
->vdev_checkpoint_sm
,
4053 mos
, new_object
, 0, vd
->vdev_asize
, vd
->vdev_ashift
));
4054 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
4057 * We save the space map object as an entry in vdev_top_zap
4058 * so it can be retrieved when the pool is reopened after an
4059 * export or through zdb.
4061 VERIFY0(zap_add(vd
->vdev_spa
->spa_meta_objset
,
4062 vd
->vdev_top_zap
, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM
,
4063 sizeof (new_object
), 1, &new_object
, tx
));
4066 mutex_enter(&msp
->ms_sync_lock
);
4067 mutex_enter(&msp
->ms_lock
);
4070 * Note: metaslab_condense() clears the space map's histogram.
4071 * Therefore we must verify and remove this histogram before
4074 metaslab_group_histogram_verify(mg
);
4075 metaslab_class_histogram_verify(mg
->mg_class
);
4076 metaslab_group_histogram_remove(mg
, msp
);
4078 if (spa
->spa_sync_pass
== 1 && msp
->ms_loaded
&&
4079 metaslab_should_condense(msp
))
4080 metaslab_condense(msp
, tx
);
4083 * We'll be going to disk to sync our space accounting, thus we
4084 * drop the ms_lock during that time so allocations coming from
4085 * open-context (ZIL) for future TXGs do not block.
4087 mutex_exit(&msp
->ms_lock
);
4088 space_map_t
*log_sm
= spa_syncing_log_sm(spa
);
4089 if (log_sm
!= NULL
) {
4090 ASSERT(spa_feature_is_enabled(spa
, SPA_FEATURE_LOG_SPACEMAP
));
4091 if (metaslab_unflushed_txg(msp
) == 0)
4092 metaslab_unflushed_add(msp
, tx
);
4093 else if (!metaslab_unflushed_dirty(msp
))
4094 metaslab_unflushed_bump(msp
, tx
, B_TRUE
);
4096 space_map_write(log_sm
, alloctree
, SM_ALLOC
,
4098 space_map_write(log_sm
, msp
->ms_freeing
, SM_FREE
,
4100 mutex_enter(&msp
->ms_lock
);
4102 ASSERT3U(spa
->spa_unflushed_stats
.sus_memused
, >=,
4103 metaslab_unflushed_changes_memused(msp
));
4104 spa
->spa_unflushed_stats
.sus_memused
-=
4105 metaslab_unflushed_changes_memused(msp
);
4106 range_tree_remove_xor_add(alloctree
,
4107 msp
->ms_unflushed_frees
, msp
->ms_unflushed_allocs
);
4108 range_tree_remove_xor_add(msp
->ms_freeing
,
4109 msp
->ms_unflushed_allocs
, msp
->ms_unflushed_frees
);
4110 spa
->spa_unflushed_stats
.sus_memused
+=
4111 metaslab_unflushed_changes_memused(msp
);
4113 ASSERT(!spa_feature_is_enabled(spa
, SPA_FEATURE_LOG_SPACEMAP
));
4115 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
,
4117 space_map_write(msp
->ms_sm
, msp
->ms_freeing
, SM_FREE
,
4119 mutex_enter(&msp
->ms_lock
);
4122 msp
->ms_allocated_space
+= range_tree_space(alloctree
);
4123 ASSERT3U(msp
->ms_allocated_space
, >=,
4124 range_tree_space(msp
->ms_freeing
));
4125 msp
->ms_allocated_space
-= range_tree_space(msp
->ms_freeing
);
4127 if (!range_tree_is_empty(msp
->ms_checkpointing
)) {
4128 ASSERT(spa_has_checkpoint(spa
));
4129 ASSERT3P(vd
->vdev_checkpoint_sm
, !=, NULL
);
4132 * Since we are doing writes to disk and the ms_checkpointing
4133 * tree won't be changing during that time, we drop the
4134 * ms_lock while writing to the checkpoint space map, for the
4135 * same reason mentioned above.
4137 mutex_exit(&msp
->ms_lock
);
4138 space_map_write(vd
->vdev_checkpoint_sm
,
4139 msp
->ms_checkpointing
, SM_FREE
, SM_NO_VDEVID
, tx
);
4140 mutex_enter(&msp
->ms_lock
);
4142 spa
->spa_checkpoint_info
.sci_dspace
+=
4143 range_tree_space(msp
->ms_checkpointing
);
4144 vd
->vdev_stat
.vs_checkpoint_space
+=
4145 range_tree_space(msp
->ms_checkpointing
);
4146 ASSERT3U(vd
->vdev_stat
.vs_checkpoint_space
, ==,
4147 -space_map_allocated(vd
->vdev_checkpoint_sm
));
4149 range_tree_vacate(msp
->ms_checkpointing
, NULL
, NULL
);
4152 if (msp
->ms_loaded
) {
4154 * When the space map is loaded, we have an accurate
4155 * histogram in the range tree. This gives us an opportunity
4156 * to bring the space map's histogram up-to-date so we clear
4157 * it first before updating it.
4159 space_map_histogram_clear(msp
->ms_sm
);
4160 space_map_histogram_add(msp
->ms_sm
, msp
->ms_allocatable
, tx
);
4163 * Since we've cleared the histogram we need to add back
4164 * any free space that has already been processed, plus
4165 * any deferred space. This allows the on-disk histogram
4166 * to accurately reflect all free space even if some space
4167 * is not yet available for allocation (i.e. deferred).
4169 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freed
, tx
);
4172 * Add back any deferred free space that has not been
4173 * added back into the in-core free tree yet. This will
4174 * ensure that we don't end up with a space map histogram
4175 * that is completely empty unless the metaslab is fully
4178 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
4179 space_map_histogram_add(msp
->ms_sm
,
4180 msp
->ms_defer
[t
], tx
);
4185 * Always add the free space from this sync pass to the space
4186 * map histogram. We want to make sure that the on-disk histogram
4187 * accounts for all free space. If the space map is not loaded,
4188 * then we will lose some accuracy but will correct it the next
4189 * time we load the space map.
4191 space_map_histogram_add(msp
->ms_sm
, msp
->ms_freeing
, tx
);
4192 metaslab_aux_histograms_update(msp
);
4194 metaslab_group_histogram_add(mg
, msp
);
4195 metaslab_group_histogram_verify(mg
);
4196 metaslab_class_histogram_verify(mg
->mg_class
);
4199 * For sync pass 1, we avoid traversing this txg's free range tree
4200 * and instead will just swap the pointers for freeing and freed.
4201 * We can safely do this since the freed_tree is guaranteed to be
4202 * empty on the initial pass.
4204 * Keep in mind that even if we are currently using a log spacemap
4205 * we want current frees to end up in the ms_allocatable (but not
4206 * get appended to the ms_sm) so their ranges can be reused as usual.
4208 if (spa_sync_pass(spa
) == 1) {
4209 range_tree_swap(&msp
->ms_freeing
, &msp
->ms_freed
);
4210 ASSERT0(msp
->ms_allocated_this_txg
);
4212 range_tree_vacate(msp
->ms_freeing
,
4213 range_tree_add
, msp
->ms_freed
);
4215 msp
->ms_allocated_this_txg
+= range_tree_space(alloctree
);
4216 range_tree_vacate(alloctree
, NULL
, NULL
);
4218 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
4219 ASSERT0(range_tree_space(msp
->ms_allocating
[TXG_CLEAN(txg
)
4221 ASSERT0(range_tree_space(msp
->ms_freeing
));
4222 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
4224 mutex_exit(&msp
->ms_lock
);
4227 * Verify that the space map object ID has been recorded in the
4231 VERIFY0(dmu_read(mos
, vd
->vdev_ms_array
,
4232 msp
->ms_id
* sizeof (uint64_t), sizeof (uint64_t), &object
, 0));
4233 VERIFY3U(object
, ==, space_map_object(msp
->ms_sm
));
4235 mutex_exit(&msp
->ms_sync_lock
);
4240 metaslab_evict(metaslab_t
*msp
, uint64_t txg
)
4242 if (!msp
->ms_loaded
|| msp
->ms_disabled
!= 0)
4245 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
4246 VERIFY0(range_tree_space(
4247 msp
->ms_allocating
[(txg
+ t
) & TXG_MASK
]));
4249 if (msp
->ms_allocator
!= -1)
4250 metaslab_passivate(msp
, msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
4252 if (!metaslab_debug_unload
)
4253 metaslab_unload(msp
);
4257 * Called after a transaction group has completely synced to mark
4258 * all of the metaslab's free space as usable.
4261 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
4263 metaslab_group_t
*mg
= msp
->ms_group
;
4264 vdev_t
*vd
= mg
->mg_vd
;
4265 spa_t
*spa
= vd
->vdev_spa
;
4266 range_tree_t
**defer_tree
;
4267 int64_t alloc_delta
, defer_delta
;
4268 boolean_t defer_allowed
= B_TRUE
;
4270 ASSERT(!vd
->vdev_ishole
);
4272 mutex_enter(&msp
->ms_lock
);
4275 /* this is a new metaslab, add its capacity to the vdev */
4276 metaslab_space_update(vd
, mg
->mg_class
, 0, 0, msp
->ms_size
);
4278 /* there should be no allocations nor frees at this point */
4279 VERIFY0(msp
->ms_allocated_this_txg
);
4280 VERIFY0(range_tree_space(msp
->ms_freed
));
4283 ASSERT0(range_tree_space(msp
->ms_freeing
));
4284 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
4286 defer_tree
= &msp
->ms_defer
[txg
% TXG_DEFER_SIZE
];
4288 uint64_t free_space
= metaslab_class_get_space(spa_normal_class(spa
)) -
4289 metaslab_class_get_alloc(spa_normal_class(spa
));
4290 if (free_space
<= spa_get_slop_space(spa
) || vd
->vdev_removing
) {
4291 defer_allowed
= B_FALSE
;
4295 alloc_delta
= msp
->ms_allocated_this_txg
-
4296 range_tree_space(msp
->ms_freed
);
4298 if (defer_allowed
) {
4299 defer_delta
= range_tree_space(msp
->ms_freed
) -
4300 range_tree_space(*defer_tree
);
4302 defer_delta
-= range_tree_space(*defer_tree
);
4304 metaslab_space_update(vd
, mg
->mg_class
, alloc_delta
+ defer_delta
,
4307 if (spa_syncing_log_sm(spa
) == NULL
) {
4309 * If there's a metaslab_load() in progress and we don't have
4310 * a log space map, it means that we probably wrote to the
4311 * metaslab's space map. If this is the case, we need to
4312 * make sure that we wait for the load to complete so that we
4313 * have a consistent view at the in-core side of the metaslab.
4315 metaslab_load_wait(msp
);
4317 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
));
4321 * When auto-trimming is enabled, free ranges which are added to
4322 * ms_allocatable are also be added to ms_trim. The ms_trim tree is
4323 * periodically consumed by the vdev_autotrim_thread() which issues
4324 * trims for all ranges and then vacates the tree. The ms_trim tree
4325 * can be discarded at any time with the sole consequence of recent
4326 * frees not being trimmed.
4328 if (spa_get_autotrim(spa
) == SPA_AUTOTRIM_ON
) {
4329 range_tree_walk(*defer_tree
, range_tree_add
, msp
->ms_trim
);
4330 if (!defer_allowed
) {
4331 range_tree_walk(msp
->ms_freed
, range_tree_add
,
4335 range_tree_vacate(msp
->ms_trim
, NULL
, NULL
);
4339 * Move the frees from the defer_tree back to the free
4340 * range tree (if it's loaded). Swap the freed_tree and
4341 * the defer_tree -- this is safe to do because we've
4342 * just emptied out the defer_tree.
4344 range_tree_vacate(*defer_tree
,
4345 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_allocatable
);
4346 if (defer_allowed
) {
4347 range_tree_swap(&msp
->ms_freed
, defer_tree
);
4349 range_tree_vacate(msp
->ms_freed
,
4350 msp
->ms_loaded
? range_tree_add
: NULL
,
4351 msp
->ms_allocatable
);
4354 msp
->ms_synced_length
= space_map_length(msp
->ms_sm
);
4356 msp
->ms_deferspace
+= defer_delta
;
4357 ASSERT3S(msp
->ms_deferspace
, >=, 0);
4358 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
4359 if (msp
->ms_deferspace
!= 0) {
4361 * Keep syncing this metaslab until all deferred frees
4362 * are back in circulation.
4364 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
4366 metaslab_aux_histograms_update_done(msp
, defer_allowed
);
4369 msp
->ms_new
= B_FALSE
;
4370 mutex_enter(&mg
->mg_lock
);
4372 mutex_exit(&mg
->mg_lock
);
4376 * Re-sort metaslab within its group now that we've adjusted
4377 * its allocatable space.
4379 metaslab_recalculate_weight_and_sort(msp
);
4381 ASSERT0(range_tree_space(msp
->ms_allocating
[txg
& TXG_MASK
]));
4382 ASSERT0(range_tree_space(msp
->ms_freeing
));
4383 ASSERT0(range_tree_space(msp
->ms_freed
));
4384 ASSERT0(range_tree_space(msp
->ms_checkpointing
));
4385 msp
->ms_allocating_total
-= msp
->ms_allocated_this_txg
;
4386 msp
->ms_allocated_this_txg
= 0;
4387 mutex_exit(&msp
->ms_lock
);
4391 metaslab_sync_reassess(metaslab_group_t
*mg
)
4393 spa_t
*spa
= mg
->mg_class
->mc_spa
;
4395 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
4396 metaslab_group_alloc_update(mg
);
4397 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
4400 * Preload the next potential metaslabs but only on active
4401 * metaslab groups. We can get into a state where the metaslab
4402 * is no longer active since we dirty metaslabs as we remove a
4403 * a device, thus potentially making the metaslab group eligible
4406 if (mg
->mg_activation_count
> 0) {
4407 metaslab_group_preload(mg
);
4409 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
4413 * When writing a ditto block (i.e. more than one DVA for a given BP) on
4414 * the same vdev as an existing DVA of this BP, then try to allocate it
4415 * on a different metaslab than existing DVAs (i.e. a unique metaslab).
4418 metaslab_is_unique(metaslab_t
*msp
, dva_t
*dva
)
4422 if (DVA_GET_ASIZE(dva
) == 0)
4425 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
4428 dva_ms_id
= DVA_GET_OFFSET(dva
) >> msp
->ms_group
->mg_vd
->vdev_ms_shift
;
4430 return (msp
->ms_id
!= dva_ms_id
);
4434 * ==========================================================================
4435 * Metaslab allocation tracing facility
4436 * ==========================================================================
4440 * Add an allocation trace element to the allocation tracing list.
4443 metaslab_trace_add(zio_alloc_list_t
*zal
, metaslab_group_t
*mg
,
4444 metaslab_t
*msp
, uint64_t psize
, uint32_t dva_id
, uint64_t offset
,
4447 metaslab_alloc_trace_t
*mat
;
4449 if (!metaslab_trace_enabled
)
4453 * When the tracing list reaches its maximum we remove
4454 * the second element in the list before adding a new one.
4455 * By removing the second element we preserve the original
4456 * entry as a clue to what allocations steps have already been
4459 if (zal
->zal_size
== metaslab_trace_max_entries
) {
4460 metaslab_alloc_trace_t
*mat_next
;
4462 panic("too many entries in allocation list");
4464 METASLABSTAT_BUMP(metaslabstat_trace_over_limit
);
4466 mat_next
= list_next(&zal
->zal_list
, list_head(&zal
->zal_list
));
4467 list_remove(&zal
->zal_list
, mat_next
);
4468 kmem_cache_free(metaslab_alloc_trace_cache
, mat_next
);
4471 mat
= kmem_cache_alloc(metaslab_alloc_trace_cache
, KM_SLEEP
);
4472 list_link_init(&mat
->mat_list_node
);
4475 mat
->mat_size
= psize
;
4476 mat
->mat_dva_id
= dva_id
;
4477 mat
->mat_offset
= offset
;
4478 mat
->mat_weight
= 0;
4479 mat
->mat_allocator
= allocator
;
4482 mat
->mat_weight
= msp
->ms_weight
;
4485 * The list is part of the zio so locking is not required. Only
4486 * a single thread will perform allocations for a given zio.
4488 list_insert_tail(&zal
->zal_list
, mat
);
4491 ASSERT3U(zal
->zal_size
, <=, metaslab_trace_max_entries
);
4495 metaslab_trace_init(zio_alloc_list_t
*zal
)
4497 list_create(&zal
->zal_list
, sizeof (metaslab_alloc_trace_t
),
4498 offsetof(metaslab_alloc_trace_t
, mat_list_node
));
4503 metaslab_trace_fini(zio_alloc_list_t
*zal
)
4505 metaslab_alloc_trace_t
*mat
;
4507 while ((mat
= list_remove_head(&zal
->zal_list
)) != NULL
)
4508 kmem_cache_free(metaslab_alloc_trace_cache
, mat
);
4509 list_destroy(&zal
->zal_list
);
4514 * ==========================================================================
4515 * Metaslab block operations
4516 * ==========================================================================
4520 metaslab_group_alloc_increment(spa_t
*spa
, uint64_t vdev
, const void *tag
,
4521 int flags
, int allocator
)
4523 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
4524 (flags
& METASLAB_DONT_THROTTLE
))
4527 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
4528 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
4531 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
4532 (void) zfs_refcount_add(&mga
->mga_alloc_queue_depth
, tag
);
4536 metaslab_group_increment_qdepth(metaslab_group_t
*mg
, int allocator
)
4538 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
4539 metaslab_class_allocator_t
*mca
=
4540 &mg
->mg_class
->mc_allocator
[allocator
];
4541 uint64_t max
= mg
->mg_max_alloc_queue_depth
;
4542 uint64_t cur
= mga
->mga_cur_max_alloc_queue_depth
;
4544 if (atomic_cas_64(&mga
->mga_cur_max_alloc_queue_depth
,
4545 cur
, cur
+ 1) == cur
) {
4546 atomic_inc_64(&mca
->mca_alloc_max_slots
);
4549 cur
= mga
->mga_cur_max_alloc_queue_depth
;
4554 metaslab_group_alloc_decrement(spa_t
*spa
, uint64_t vdev
, const void *tag
,
4555 int flags
, int allocator
, boolean_t io_complete
)
4557 if (!(flags
& METASLAB_ASYNC_ALLOC
) ||
4558 (flags
& METASLAB_DONT_THROTTLE
))
4561 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
4562 if (!mg
->mg_class
->mc_alloc_throttle_enabled
)
4565 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
4566 (void) zfs_refcount_remove(&mga
->mga_alloc_queue_depth
, tag
);
4568 metaslab_group_increment_qdepth(mg
, allocator
);
4572 metaslab_group_alloc_verify(spa_t
*spa
, const blkptr_t
*bp
, const void *tag
,
4576 const dva_t
*dva
= bp
->blk_dva
;
4577 int ndvas
= BP_GET_NDVAS(bp
);
4579 for (int d
= 0; d
< ndvas
; d
++) {
4580 uint64_t vdev
= DVA_GET_VDEV(&dva
[d
]);
4581 metaslab_group_t
*mg
= vdev_lookup_top(spa
, vdev
)->vdev_mg
;
4582 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
4583 VERIFY(zfs_refcount_not_held(&mga
->mga_alloc_queue_depth
, tag
));
4589 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
, uint64_t txg
)
4592 range_tree_t
*rt
= msp
->ms_allocatable
;
4593 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
4595 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
4596 VERIFY(!msp
->ms_condensing
);
4597 VERIFY0(msp
->ms_disabled
);
4599 start
= mc
->mc_ops
->msop_alloc(msp
, size
);
4600 if (start
!= -1ULL) {
4601 metaslab_group_t
*mg
= msp
->ms_group
;
4602 vdev_t
*vd
= mg
->mg_vd
;
4604 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
4605 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
4606 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
4607 range_tree_remove(rt
, start
, size
);
4608 range_tree_clear(msp
->ms_trim
, start
, size
);
4610 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
4611 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
4613 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
], start
, size
);
4614 msp
->ms_allocating_total
+= size
;
4616 /* Track the last successful allocation */
4617 msp
->ms_alloc_txg
= txg
;
4618 metaslab_verify_space(msp
, txg
);
4622 * Now that we've attempted the allocation we need to update the
4623 * metaslab's maximum block size since it may have changed.
4625 msp
->ms_max_size
= metaslab_largest_allocatable(msp
);
4630 * Find the metaslab with the highest weight that is less than what we've
4631 * already tried. In the common case, this means that we will examine each
4632 * metaslab at most once. Note that concurrent callers could reorder metaslabs
4633 * by activation/passivation once we have dropped the mg_lock. If a metaslab is
4634 * activated by another thread, and we fail to allocate from the metaslab we
4635 * have selected, we may not try the newly-activated metaslab, and instead
4636 * activate another metaslab. This is not optimal, but generally does not cause
4637 * any problems (a possible exception being if every metaslab is completely full
4638 * except for the newly-activated metaslab which we fail to examine).
4641 find_valid_metaslab(metaslab_group_t
*mg
, uint64_t activation_weight
,
4642 dva_t
*dva
, int d
, boolean_t want_unique
, uint64_t asize
, int allocator
,
4643 boolean_t try_hard
, zio_alloc_list_t
*zal
, metaslab_t
*search
,
4644 boolean_t
*was_active
)
4647 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
4648 metaslab_t
*msp
= avl_find(t
, search
, &idx
);
4650 msp
= avl_nearest(t
, idx
, AVL_AFTER
);
4653 for (; msp
!= NULL
; msp
= AVL_NEXT(t
, msp
)) {
4656 if (!try_hard
&& tries
> zfs_metaslab_find_max_tries
) {
4657 METASLABSTAT_BUMP(metaslabstat_too_many_tries
);
4662 if (!metaslab_should_allocate(msp
, asize
, try_hard
)) {
4663 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
4664 TRACE_TOO_SMALL
, allocator
);
4669 * If the selected metaslab is condensing or disabled,
4672 if (msp
->ms_condensing
|| msp
->ms_disabled
> 0)
4675 *was_active
= msp
->ms_allocator
!= -1;
4677 * If we're activating as primary, this is our first allocation
4678 * from this disk, so we don't need to check how close we are.
4679 * If the metaslab under consideration was already active,
4680 * we're getting desperate enough to steal another allocator's
4681 * metaslab, so we still don't care about distances.
4683 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
|| *was_active
)
4686 for (i
= 0; i
< d
; i
++) {
4688 !metaslab_is_unique(msp
, &dva
[i
]))
4689 break; /* try another metaslab */
4696 search
->ms_weight
= msp
->ms_weight
;
4697 search
->ms_start
= msp
->ms_start
+ 1;
4698 search
->ms_allocator
= msp
->ms_allocator
;
4699 search
->ms_primary
= msp
->ms_primary
;
4705 metaslab_active_mask_verify(metaslab_t
*msp
)
4707 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
4709 if ((zfs_flags
& ZFS_DEBUG_METASLAB_VERIFY
) == 0)
4712 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0)
4715 if (msp
->ms_weight
& METASLAB_WEIGHT_PRIMARY
) {
4716 VERIFY0(msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
);
4717 VERIFY0(msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
);
4718 VERIFY3S(msp
->ms_allocator
, !=, -1);
4719 VERIFY(msp
->ms_primary
);
4723 if (msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) {
4724 VERIFY0(msp
->ms_weight
& METASLAB_WEIGHT_PRIMARY
);
4725 VERIFY0(msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
);
4726 VERIFY3S(msp
->ms_allocator
, !=, -1);
4727 VERIFY(!msp
->ms_primary
);
4731 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
) {
4732 VERIFY0(msp
->ms_weight
& METASLAB_WEIGHT_PRIMARY
);
4733 VERIFY0(msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
);
4734 VERIFY3S(msp
->ms_allocator
, ==, -1);
4740 metaslab_group_alloc_normal(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
4741 uint64_t asize
, uint64_t txg
, boolean_t want_unique
, dva_t
*dva
, int d
,
4742 int allocator
, boolean_t try_hard
)
4744 metaslab_t
*msp
= NULL
;
4745 uint64_t offset
= -1ULL;
4747 uint64_t activation_weight
= METASLAB_WEIGHT_PRIMARY
;
4748 for (int i
= 0; i
< d
; i
++) {
4749 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
&&
4750 DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
4751 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
4752 } else if (activation_weight
== METASLAB_WEIGHT_SECONDARY
&&
4753 DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
4754 activation_weight
= METASLAB_WEIGHT_CLAIM
;
4760 * If we don't have enough metaslabs active to fill the entire array, we
4761 * just use the 0th slot.
4763 if (mg
->mg_ms_ready
< mg
->mg_allocators
* 3)
4765 metaslab_group_allocator_t
*mga
= &mg
->mg_allocator
[allocator
];
4767 ASSERT3U(mg
->mg_vd
->vdev_ms_count
, >=, 2);
4769 metaslab_t
*search
= kmem_alloc(sizeof (*search
), KM_SLEEP
);
4770 search
->ms_weight
= UINT64_MAX
;
4771 search
->ms_start
= 0;
4773 * At the end of the metaslab tree are the already-active metaslabs,
4774 * first the primaries, then the secondaries. When we resume searching
4775 * through the tree, we need to consider ms_allocator and ms_primary so
4776 * we start in the location right after where we left off, and don't
4777 * accidentally loop forever considering the same metaslabs.
4779 search
->ms_allocator
= -1;
4780 search
->ms_primary
= B_TRUE
;
4782 boolean_t was_active
= B_FALSE
;
4784 mutex_enter(&mg
->mg_lock
);
4786 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
&&
4787 mga
->mga_primary
!= NULL
) {
4788 msp
= mga
->mga_primary
;
4791 * Even though we don't hold the ms_lock for the
4792 * primary metaslab, those fields should not
4793 * change while we hold the mg_lock. Thus it is
4794 * safe to make assertions on them.
4796 ASSERT(msp
->ms_primary
);
4797 ASSERT3S(msp
->ms_allocator
, ==, allocator
);
4798 ASSERT(msp
->ms_loaded
);
4800 was_active
= B_TRUE
;
4801 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
4802 } else if (activation_weight
== METASLAB_WEIGHT_SECONDARY
&&
4803 mga
->mga_secondary
!= NULL
) {
4804 msp
= mga
->mga_secondary
;
4807 * See comment above about the similar assertions
4808 * for the primary metaslab.
4810 ASSERT(!msp
->ms_primary
);
4811 ASSERT3S(msp
->ms_allocator
, ==, allocator
);
4812 ASSERT(msp
->ms_loaded
);
4814 was_active
= B_TRUE
;
4815 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
4817 msp
= find_valid_metaslab(mg
, activation_weight
, dva
, d
,
4818 want_unique
, asize
, allocator
, try_hard
, zal
,
4819 search
, &was_active
);
4822 mutex_exit(&mg
->mg_lock
);
4824 kmem_free(search
, sizeof (*search
));
4827 mutex_enter(&msp
->ms_lock
);
4829 metaslab_active_mask_verify(msp
);
4832 * This code is disabled out because of issues with
4833 * tracepoints in non-gpl kernel modules.
4836 DTRACE_PROBE3(ms__activation__attempt
,
4837 metaslab_t
*, msp
, uint64_t, activation_weight
,
4838 boolean_t
, was_active
);
4842 * Ensure that the metaslab we have selected is still
4843 * capable of handling our request. It's possible that
4844 * another thread may have changed the weight while we
4845 * were blocked on the metaslab lock. We check the
4846 * active status first to see if we need to set_selected_txg
4849 if (was_active
&& !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
)) {
4850 ASSERT3S(msp
->ms_allocator
, ==, -1);
4851 mutex_exit(&msp
->ms_lock
);
4856 * If the metaslab was activated for another allocator
4857 * while we were waiting in the ms_lock above, or it's
4858 * a primary and we're seeking a secondary (or vice versa),
4859 * we go back and select a new metaslab.
4861 if (!was_active
&& (msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
4862 (msp
->ms_allocator
!= -1) &&
4863 (msp
->ms_allocator
!= allocator
|| ((activation_weight
==
4864 METASLAB_WEIGHT_PRIMARY
) != msp
->ms_primary
))) {
4865 ASSERT(msp
->ms_loaded
);
4866 ASSERT((msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
) ||
4867 msp
->ms_allocator
!= -1);
4868 mutex_exit(&msp
->ms_lock
);
4873 * This metaslab was used for claiming regions allocated
4874 * by the ZIL during pool import. Once these regions are
4875 * claimed we don't need to keep the CLAIM bit set
4876 * anymore. Passivate this metaslab to zero its activation
4879 if (msp
->ms_weight
& METASLAB_WEIGHT_CLAIM
&&
4880 activation_weight
!= METASLAB_WEIGHT_CLAIM
) {
4881 ASSERT(msp
->ms_loaded
);
4882 ASSERT3S(msp
->ms_allocator
, ==, -1);
4883 metaslab_passivate(msp
, msp
->ms_weight
&
4884 ~METASLAB_WEIGHT_CLAIM
);
4885 mutex_exit(&msp
->ms_lock
);
4889 metaslab_set_selected_txg(msp
, txg
);
4891 int activation_error
=
4892 metaslab_activate(msp
, allocator
, activation_weight
);
4893 metaslab_active_mask_verify(msp
);
4896 * If the metaslab was activated by another thread for
4897 * another allocator or activation_weight (EBUSY), or it
4898 * failed because another metaslab was assigned as primary
4899 * for this allocator (EEXIST) we continue using this
4900 * metaslab for our allocation, rather than going on to a
4901 * worse metaslab (we waited for that metaslab to be loaded
4904 * If the activation failed due to an I/O error or ENOSPC we
4905 * skip to the next metaslab.
4907 boolean_t activated
;
4908 if (activation_error
== 0) {
4910 } else if (activation_error
== EBUSY
||
4911 activation_error
== EEXIST
) {
4912 activated
= B_FALSE
;
4914 mutex_exit(&msp
->ms_lock
);
4917 ASSERT(msp
->ms_loaded
);
4920 * Now that we have the lock, recheck to see if we should
4921 * continue to use this metaslab for this allocation. The
4922 * the metaslab is now loaded so metaslab_should_allocate()
4923 * can accurately determine if the allocation attempt should
4926 if (!metaslab_should_allocate(msp
, asize
, try_hard
)) {
4927 /* Passivate this metaslab and select a new one. */
4928 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
4929 TRACE_TOO_SMALL
, allocator
);
4934 * If this metaslab is currently condensing then pick again
4935 * as we can't manipulate this metaslab until it's committed
4936 * to disk. If this metaslab is being initialized, we shouldn't
4937 * allocate from it since the allocated region might be
4938 * overwritten after allocation.
4940 if (msp
->ms_condensing
) {
4941 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
4942 TRACE_CONDENSING
, allocator
);
4944 metaslab_passivate(msp
, msp
->ms_weight
&
4945 ~METASLAB_ACTIVE_MASK
);
4947 mutex_exit(&msp
->ms_lock
);
4949 } else if (msp
->ms_disabled
> 0) {
4950 metaslab_trace_add(zal
, mg
, msp
, asize
, d
,
4951 TRACE_DISABLED
, allocator
);
4953 metaslab_passivate(msp
, msp
->ms_weight
&
4954 ~METASLAB_ACTIVE_MASK
);
4956 mutex_exit(&msp
->ms_lock
);
4960 offset
= metaslab_block_alloc(msp
, asize
, txg
);
4961 metaslab_trace_add(zal
, mg
, msp
, asize
, d
, offset
, allocator
);
4963 if (offset
!= -1ULL) {
4964 /* Proactively passivate the metaslab, if needed */
4966 metaslab_segment_may_passivate(msp
);
4970 ASSERT(msp
->ms_loaded
);
4973 * This code is disabled out because of issues with
4974 * tracepoints in non-gpl kernel modules.
4977 DTRACE_PROBE2(ms__alloc__failure
, metaslab_t
*, msp
,
4982 * We were unable to allocate from this metaslab so determine
4983 * a new weight for this metaslab. Now that we have loaded
4984 * the metaslab we can provide a better hint to the metaslab
4987 * For space-based metaslabs, we use the maximum block size.
4988 * This information is only available when the metaslab
4989 * is loaded and is more accurate than the generic free
4990 * space weight that was calculated by metaslab_weight().
4991 * This information allows us to quickly compare the maximum
4992 * available allocation in the metaslab to the allocation
4993 * size being requested.
4995 * For segment-based metaslabs, determine the new weight
4996 * based on the highest bucket in the range tree. We
4997 * explicitly use the loaded segment weight (i.e. the range
4998 * tree histogram) since it contains the space that is
4999 * currently available for allocation and is accurate
5000 * even within a sync pass.
5003 if (WEIGHT_IS_SPACEBASED(msp
->ms_weight
)) {
5004 weight
= metaslab_largest_allocatable(msp
);
5005 WEIGHT_SET_SPACEBASED(weight
);
5007 weight
= metaslab_weight_from_range_tree(msp
);
5011 metaslab_passivate(msp
, weight
);
5014 * For the case where we use the metaslab that is
5015 * active for another allocator we want to make
5016 * sure that we retain the activation mask.
5018 * Note that we could attempt to use something like
5019 * metaslab_recalculate_weight_and_sort() that
5020 * retains the activation mask here. That function
5021 * uses metaslab_weight() to set the weight though
5022 * which is not as accurate as the calculations
5025 weight
|= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
5026 metaslab_group_sort(mg
, msp
, weight
);
5028 metaslab_active_mask_verify(msp
);
5031 * We have just failed an allocation attempt, check
5032 * that metaslab_should_allocate() agrees. Otherwise,
5033 * we may end up in an infinite loop retrying the same
5036 ASSERT(!metaslab_should_allocate(msp
, asize
, try_hard
));
5038 mutex_exit(&msp
->ms_lock
);
5040 mutex_exit(&msp
->ms_lock
);
5041 kmem_free(search
, sizeof (*search
));
5046 metaslab_group_alloc(metaslab_group_t
*mg
, zio_alloc_list_t
*zal
,
5047 uint64_t asize
, uint64_t txg
, boolean_t want_unique
, dva_t
*dva
, int d
,
5048 int allocator
, boolean_t try_hard
)
5051 ASSERT(mg
->mg_initialized
);
5053 offset
= metaslab_group_alloc_normal(mg
, zal
, asize
, txg
, want_unique
,
5054 dva
, d
, allocator
, try_hard
);
5056 mutex_enter(&mg
->mg_lock
);
5057 if (offset
== -1ULL) {
5058 mg
->mg_failed_allocations
++;
5059 metaslab_trace_add(zal
, mg
, NULL
, asize
, d
,
5060 TRACE_GROUP_FAILURE
, allocator
);
5061 if (asize
== SPA_GANGBLOCKSIZE
) {
5063 * This metaslab group was unable to allocate
5064 * the minimum gang block size so it must be out of
5065 * space. We must notify the allocation throttle
5066 * to start skipping allocation attempts to this
5067 * metaslab group until more space becomes available.
5068 * Note: this failure cannot be caused by the
5069 * allocation throttle since the allocation throttle
5070 * is only responsible for skipping devices and
5071 * not failing block allocations.
5073 mg
->mg_no_free_space
= B_TRUE
;
5076 mg
->mg_allocations
++;
5077 mutex_exit(&mg
->mg_lock
);
5082 * Allocate a block for the specified i/o.
5085 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
5086 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
,
5087 zio_alloc_list_t
*zal
, int allocator
)
5089 metaslab_class_allocator_t
*mca
= &mc
->mc_allocator
[allocator
];
5090 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
5092 boolean_t try_hard
= B_FALSE
;
5094 ASSERT(!DVA_IS_VALID(&dva
[d
]));
5097 * For testing, make some blocks above a certain size be gang blocks.
5098 * This will result in more split blocks when using device removal,
5099 * and a large number of split blocks coupled with ztest-induced
5100 * damage can result in extremely long reconstruction times. This
5101 * will also test spilling from special to normal.
5103 if (psize
>= metaslab_force_ganging
&& (random_in_range(100) < 3)) {
5104 metaslab_trace_add(zal
, NULL
, NULL
, psize
, d
, TRACE_FORCE_GANG
,
5106 return (SET_ERROR(ENOSPC
));
5110 * Start at the rotor and loop through all mgs until we find something.
5111 * Note that there's no locking on mca_rotor or mca_aliquot because
5112 * nothing actually breaks if we miss a few updates -- we just won't
5113 * allocate quite as evenly. It all balances out over time.
5115 * If we are doing ditto or log blocks, try to spread them across
5116 * consecutive vdevs. If we're forced to reuse a vdev before we've
5117 * allocated all of our ditto blocks, then try and spread them out on
5118 * that vdev as much as possible. If it turns out to not be possible,
5119 * gradually lower our standards until anything becomes acceptable.
5120 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
5121 * gives us hope of containing our fault domains to something we're
5122 * able to reason about. Otherwise, any two top-level vdev failures
5123 * will guarantee the loss of data. With consecutive allocation,
5124 * only two adjacent top-level vdev failures will result in data loss.
5126 * If we are doing gang blocks (hintdva is non-NULL), try to keep
5127 * ourselves on the same vdev as our gang block header. That
5128 * way, we can hope for locality in vdev_cache, plus it makes our
5129 * fault domains something tractable.
5132 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
5135 * It's possible the vdev we're using as the hint no
5136 * longer exists or its mg has been closed (e.g. by
5137 * device removal). Consult the rotor when
5140 if (vd
!= NULL
&& vd
->vdev_mg
!= NULL
) {
5141 mg
= vdev_get_mg(vd
, mc
);
5143 if (flags
& METASLAB_HINTBP_AVOID
)
5146 mg
= mca
->mca_rotor
;
5148 } else if (d
!= 0) {
5149 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
5150 mg
= vd
->vdev_mg
->mg_next
;
5151 } else if (flags
& METASLAB_FASTWRITE
) {
5152 mg
= fast_mg
= mca
->mca_rotor
;
5155 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
5156 mg
->mg_vd
->vdev_pending_fastwrite
)
5158 } while ((fast_mg
= fast_mg
->mg_next
) != mca
->mca_rotor
);
5161 ASSERT(mca
->mca_rotor
!= NULL
);
5162 mg
= mca
->mca_rotor
;
5166 * If the hint put us into the wrong metaslab class, or into a
5167 * metaslab group that has been passivated, just follow the rotor.
5169 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
5170 mg
= mca
->mca_rotor
;
5175 boolean_t allocatable
;
5177 ASSERT(mg
->mg_activation_count
== 1);
5181 * Don't allocate from faulted devices.
5184 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
5185 allocatable
= vdev_allocatable(vd
);
5186 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
5188 allocatable
= vdev_allocatable(vd
);
5192 * Determine if the selected metaslab group is eligible
5193 * for allocations. If we're ganging then don't allow
5194 * this metaslab group to skip allocations since that would
5195 * inadvertently return ENOSPC and suspend the pool
5196 * even though space is still available.
5198 if (allocatable
&& !GANG_ALLOCATION(flags
) && !try_hard
) {
5199 allocatable
= metaslab_group_allocatable(mg
, rotor
,
5200 flags
, psize
, allocator
, d
);
5204 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
5205 TRACE_NOT_ALLOCATABLE
, allocator
);
5209 ASSERT(mg
->mg_initialized
);
5212 * Avoid writing single-copy data to an unhealthy,
5213 * non-redundant vdev, unless we've already tried all
5216 if (vd
->vdev_state
< VDEV_STATE_HEALTHY
&&
5217 d
== 0 && !try_hard
&& vd
->vdev_children
== 0) {
5218 metaslab_trace_add(zal
, mg
, NULL
, psize
, d
,
5219 TRACE_VDEV_ERROR
, allocator
);
5223 ASSERT(mg
->mg_class
== mc
);
5225 uint64_t asize
= vdev_psize_to_asize(vd
, psize
);
5226 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
5229 * If we don't need to try hard, then require that the
5230 * block be on a different metaslab from any other DVAs
5231 * in this BP (unique=true). If we are trying hard, then
5232 * allow any metaslab to be used (unique=false).
5234 uint64_t offset
= metaslab_group_alloc(mg
, zal
, asize
, txg
,
5235 !try_hard
, dva
, d
, allocator
, try_hard
);
5237 if (offset
!= -1ULL) {
5239 * If we've just selected this metaslab group,
5240 * figure out whether the corresponding vdev is
5241 * over- or under-used relative to the pool,
5242 * and set an allocation bias to even it out.
5244 * Bias is also used to compensate for unequally
5245 * sized vdevs so that space is allocated fairly.
5247 if (mca
->mca_aliquot
== 0 && metaslab_bias_enabled
) {
5248 vdev_stat_t
*vs
= &vd
->vdev_stat
;
5249 int64_t vs_free
= vs
->vs_space
- vs
->vs_alloc
;
5250 int64_t mc_free
= mc
->mc_space
- mc
->mc_alloc
;
5254 * Calculate how much more or less we should
5255 * try to allocate from this device during
5256 * this iteration around the rotor.
5258 * This basically introduces a zero-centered
5259 * bias towards the devices with the most
5260 * free space, while compensating for vdev
5264 * vdev V1 = 16M/128M
5265 * vdev V2 = 16M/128M
5266 * ratio(V1) = 100% ratio(V2) = 100%
5268 * vdev V1 = 16M/128M
5269 * vdev V2 = 64M/128M
5270 * ratio(V1) = 127% ratio(V2) = 72%
5272 * vdev V1 = 16M/128M
5273 * vdev V2 = 64M/512M
5274 * ratio(V1) = 40% ratio(V2) = 160%
5276 ratio
= (vs_free
* mc
->mc_alloc_groups
* 100) /
5278 mg
->mg_bias
= ((ratio
- 100) *
5279 (int64_t)mg
->mg_aliquot
) / 100;
5280 } else if (!metaslab_bias_enabled
) {
5284 if ((flags
& METASLAB_FASTWRITE
) ||
5285 atomic_add_64_nv(&mca
->mca_aliquot
, asize
) >=
5286 mg
->mg_aliquot
+ mg
->mg_bias
) {
5287 mca
->mca_rotor
= mg
->mg_next
;
5288 mca
->mca_aliquot
= 0;
5291 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
5292 DVA_SET_OFFSET(&dva
[d
], offset
);
5293 DVA_SET_GANG(&dva
[d
],
5294 ((flags
& METASLAB_GANG_HEADER
) ? 1 : 0));
5295 DVA_SET_ASIZE(&dva
[d
], asize
);
5297 if (flags
& METASLAB_FASTWRITE
) {
5298 atomic_add_64(&vd
->vdev_pending_fastwrite
,
5305 mca
->mca_rotor
= mg
->mg_next
;
5306 mca
->mca_aliquot
= 0;
5307 } while ((mg
= mg
->mg_next
) != rotor
);
5310 * If we haven't tried hard, perhaps do so now.
5312 if (!try_hard
&& (zfs_metaslab_try_hard_before_gang
||
5313 GANG_ALLOCATION(flags
) || (flags
& METASLAB_ZIL
) != 0 ||
5314 psize
<= 1 << spa
->spa_min_ashift
)) {
5315 METASLABSTAT_BUMP(metaslabstat_try_hard
);
5320 memset(&dva
[d
], 0, sizeof (dva_t
));
5322 metaslab_trace_add(zal
, rotor
, NULL
, psize
, d
, TRACE_ENOSPC
, allocator
);
5323 return (SET_ERROR(ENOSPC
));
5327 metaslab_free_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t asize
,
5328 boolean_t checkpoint
)
5331 spa_t
*spa
= vd
->vdev_spa
;
5333 ASSERT(vdev_is_concrete(vd
));
5334 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
5335 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
5337 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
5339 VERIFY(!msp
->ms_condensing
);
5340 VERIFY3U(offset
, >=, msp
->ms_start
);
5341 VERIFY3U(offset
+ asize
, <=, msp
->ms_start
+ msp
->ms_size
);
5342 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
5343 VERIFY0(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
));
5345 metaslab_check_free_impl(vd
, offset
, asize
);
5347 mutex_enter(&msp
->ms_lock
);
5348 if (range_tree_is_empty(msp
->ms_freeing
) &&
5349 range_tree_is_empty(msp
->ms_checkpointing
)) {
5350 vdev_dirty(vd
, VDD_METASLAB
, msp
, spa_syncing_txg(spa
));
5354 ASSERT(spa_has_checkpoint(spa
));
5355 range_tree_add(msp
->ms_checkpointing
, offset
, asize
);
5357 range_tree_add(msp
->ms_freeing
, offset
, asize
);
5359 mutex_exit(&msp
->ms_lock
);
5363 metaslab_free_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
5364 uint64_t size
, void *arg
)
5366 (void) inner_offset
;
5367 boolean_t
*checkpoint
= arg
;
5369 ASSERT3P(checkpoint
, !=, NULL
);
5371 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
)
5372 vdev_indirect_mark_obsolete(vd
, offset
, size
);
5374 metaslab_free_impl(vd
, offset
, size
, *checkpoint
);
5378 metaslab_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
5379 boolean_t checkpoint
)
5381 spa_t
*spa
= vd
->vdev_spa
;
5383 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
5385 if (spa_syncing_txg(spa
) > spa_freeze_txg(spa
))
5388 if (spa
->spa_vdev_removal
!= NULL
&&
5389 spa
->spa_vdev_removal
->svr_vdev_id
== vd
->vdev_id
&&
5390 vdev_is_concrete(vd
)) {
5392 * Note: we check if the vdev is concrete because when
5393 * we complete the removal, we first change the vdev to be
5394 * an indirect vdev (in open context), and then (in syncing
5395 * context) clear spa_vdev_removal.
5397 free_from_removing_vdev(vd
, offset
, size
);
5398 } else if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
5399 vdev_indirect_mark_obsolete(vd
, offset
, size
);
5400 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
5401 metaslab_free_impl_cb
, &checkpoint
);
5403 metaslab_free_concrete(vd
, offset
, size
, checkpoint
);
5407 typedef struct remap_blkptr_cb_arg
{
5409 spa_remap_cb_t rbca_cb
;
5410 vdev_t
*rbca_remap_vd
;
5411 uint64_t rbca_remap_offset
;
5413 } remap_blkptr_cb_arg_t
;
5416 remap_blkptr_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
5417 uint64_t size
, void *arg
)
5419 remap_blkptr_cb_arg_t
*rbca
= arg
;
5420 blkptr_t
*bp
= rbca
->rbca_bp
;
5422 /* We can not remap split blocks. */
5423 if (size
!= DVA_GET_ASIZE(&bp
->blk_dva
[0]))
5425 ASSERT0(inner_offset
);
5427 if (rbca
->rbca_cb
!= NULL
) {
5429 * At this point we know that we are not handling split
5430 * blocks and we invoke the callback on the previous
5431 * vdev which must be indirect.
5433 ASSERT3P(rbca
->rbca_remap_vd
->vdev_ops
, ==, &vdev_indirect_ops
);
5435 rbca
->rbca_cb(rbca
->rbca_remap_vd
->vdev_id
,
5436 rbca
->rbca_remap_offset
, size
, rbca
->rbca_cb_arg
);
5438 /* set up remap_blkptr_cb_arg for the next call */
5439 rbca
->rbca_remap_vd
= vd
;
5440 rbca
->rbca_remap_offset
= offset
;
5444 * The phys birth time is that of dva[0]. This ensures that we know
5445 * when each dva was written, so that resilver can determine which
5446 * blocks need to be scrubbed (i.e. those written during the time
5447 * the vdev was offline). It also ensures that the key used in
5448 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
5449 * we didn't change the phys_birth, a lookup in the ARC for a
5450 * remapped BP could find the data that was previously stored at
5451 * this vdev + offset.
5453 vdev_t
*oldvd
= vdev_lookup_top(vd
->vdev_spa
,
5454 DVA_GET_VDEV(&bp
->blk_dva
[0]));
5455 vdev_indirect_births_t
*vib
= oldvd
->vdev_indirect_births
;
5456 bp
->blk_phys_birth
= vdev_indirect_births_physbirth(vib
,
5457 DVA_GET_OFFSET(&bp
->blk_dva
[0]), DVA_GET_ASIZE(&bp
->blk_dva
[0]));
5459 DVA_SET_VDEV(&bp
->blk_dva
[0], vd
->vdev_id
);
5460 DVA_SET_OFFSET(&bp
->blk_dva
[0], offset
);
5464 * If the block pointer contains any indirect DVAs, modify them to refer to
5465 * concrete DVAs. Note that this will sometimes not be possible, leaving
5466 * the indirect DVA in place. This happens if the indirect DVA spans multiple
5467 * segments in the mapping (i.e. it is a "split block").
5469 * If the BP was remapped, calls the callback on the original dva (note the
5470 * callback can be called multiple times if the original indirect DVA refers
5471 * to another indirect DVA, etc).
5473 * Returns TRUE if the BP was remapped.
5476 spa_remap_blkptr(spa_t
*spa
, blkptr_t
*bp
, spa_remap_cb_t callback
, void *arg
)
5478 remap_blkptr_cb_arg_t rbca
;
5480 if (!zfs_remap_blkptr_enable
)
5483 if (!spa_feature_is_enabled(spa
, SPA_FEATURE_OBSOLETE_COUNTS
))
5487 * Dedup BP's can not be remapped, because ddt_phys_select() depends
5488 * on DVA[0] being the same in the BP as in the DDT (dedup table).
5490 if (BP_GET_DEDUP(bp
))
5494 * Gang blocks can not be remapped, because
5495 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
5496 * the BP used to read the gang block header (GBH) being the same
5497 * as the DVA[0] that we allocated for the GBH.
5503 * Embedded BP's have no DVA to remap.
5505 if (BP_GET_NDVAS(bp
) < 1)
5509 * Note: we only remap dva[0]. If we remapped other dvas, we
5510 * would no longer know what their phys birth txg is.
5512 dva_t
*dva
= &bp
->blk_dva
[0];
5514 uint64_t offset
= DVA_GET_OFFSET(dva
);
5515 uint64_t size
= DVA_GET_ASIZE(dva
);
5516 vdev_t
*vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(dva
));
5518 if (vd
->vdev_ops
->vdev_op_remap
== NULL
)
5522 rbca
.rbca_cb
= callback
;
5523 rbca
.rbca_remap_vd
= vd
;
5524 rbca
.rbca_remap_offset
= offset
;
5525 rbca
.rbca_cb_arg
= arg
;
5528 * remap_blkptr_cb() will be called in order for each level of
5529 * indirection, until a concrete vdev is reached or a split block is
5530 * encountered. old_vd and old_offset are updated within the callback
5531 * as we go from the one indirect vdev to the next one (either concrete
5532 * or indirect again) in that order.
5534 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
, remap_blkptr_cb
, &rbca
);
5536 /* Check if the DVA wasn't remapped because it is a split block */
5537 if (DVA_GET_VDEV(&rbca
.rbca_bp
->blk_dva
[0]) == vd
->vdev_id
)
5544 * Undo the allocation of a DVA which happened in the given transaction group.
5547 metaslab_unalloc_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
5551 uint64_t vdev
= DVA_GET_VDEV(dva
);
5552 uint64_t offset
= DVA_GET_OFFSET(dva
);
5553 uint64_t size
= DVA_GET_ASIZE(dva
);
5555 ASSERT(DVA_IS_VALID(dva
));
5556 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
5558 if (txg
> spa_freeze_txg(spa
))
5561 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
5562 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
5563 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
5564 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
5565 (u_longlong_t
)size
);
5569 ASSERT(!vd
->vdev_removing
);
5570 ASSERT(vdev_is_concrete(vd
));
5571 ASSERT0(vd
->vdev_indirect_config
.vic_mapping_object
);
5572 ASSERT3P(vd
->vdev_indirect_mapping
, ==, NULL
);
5574 if (DVA_GET_GANG(dva
))
5575 size
= vdev_gang_header_asize(vd
);
5577 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
5579 mutex_enter(&msp
->ms_lock
);
5580 range_tree_remove(msp
->ms_allocating
[txg
& TXG_MASK
],
5582 msp
->ms_allocating_total
-= size
;
5584 VERIFY(!msp
->ms_condensing
);
5585 VERIFY3U(offset
, >=, msp
->ms_start
);
5586 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
5587 VERIFY3U(range_tree_space(msp
->ms_allocatable
) + size
, <=,
5589 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
5590 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
5591 range_tree_add(msp
->ms_allocatable
, offset
, size
);
5592 mutex_exit(&msp
->ms_lock
);
5596 * Free the block represented by the given DVA.
5599 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, boolean_t checkpoint
)
5601 uint64_t vdev
= DVA_GET_VDEV(dva
);
5602 uint64_t offset
= DVA_GET_OFFSET(dva
);
5603 uint64_t size
= DVA_GET_ASIZE(dva
);
5604 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
5606 ASSERT(DVA_IS_VALID(dva
));
5607 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
5609 if (DVA_GET_GANG(dva
)) {
5610 size
= vdev_gang_header_asize(vd
);
5613 metaslab_free_impl(vd
, offset
, size
, checkpoint
);
5617 * Reserve some allocation slots. The reservation system must be called
5618 * before we call into the allocator. If there aren't any available slots
5619 * then the I/O will be throttled until an I/O completes and its slots are
5620 * freed up. The function returns true if it was successful in placing
5624 metaslab_class_throttle_reserve(metaslab_class_t
*mc
, int slots
, int allocator
,
5625 zio_t
*zio
, int flags
)
5627 metaslab_class_allocator_t
*mca
= &mc
->mc_allocator
[allocator
];
5628 uint64_t max
= mca
->mca_alloc_max_slots
;
5630 ASSERT(mc
->mc_alloc_throttle_enabled
);
5631 if (GANG_ALLOCATION(flags
) || (flags
& METASLAB_MUST_RESERVE
) ||
5632 zfs_refcount_count(&mca
->mca_alloc_slots
) + slots
<= max
) {
5634 * The potential race between _count() and _add() is covered
5635 * by the allocator lock in most cases, or irrelevant due to
5636 * GANG_ALLOCATION() or METASLAB_MUST_RESERVE set in others.
5637 * But even if we assume some other non-existing scenario, the
5638 * worst that can happen is few more I/Os get to allocation
5639 * earlier, that is not a problem.
5641 * We reserve the slots individually so that we can unreserve
5642 * them individually when an I/O completes.
5644 for (int d
= 0; d
< slots
; d
++)
5645 zfs_refcount_add(&mca
->mca_alloc_slots
, zio
);
5646 zio
->io_flags
|= ZIO_FLAG_IO_ALLOCATING
;
5653 metaslab_class_throttle_unreserve(metaslab_class_t
*mc
, int slots
,
5654 int allocator
, zio_t
*zio
)
5656 metaslab_class_allocator_t
*mca
= &mc
->mc_allocator
[allocator
];
5658 ASSERT(mc
->mc_alloc_throttle_enabled
);
5659 for (int d
= 0; d
< slots
; d
++)
5660 zfs_refcount_remove(&mca
->mca_alloc_slots
, zio
);
5664 metaslab_claim_concrete(vdev_t
*vd
, uint64_t offset
, uint64_t size
,
5668 spa_t
*spa
= vd
->vdev_spa
;
5671 if (offset
>> vd
->vdev_ms_shift
>= vd
->vdev_ms_count
)
5672 return (SET_ERROR(ENXIO
));
5674 ASSERT3P(vd
->vdev_ms
, !=, NULL
);
5675 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
5677 mutex_enter(&msp
->ms_lock
);
5679 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
) {
5680 error
= metaslab_activate(msp
, 0, METASLAB_WEIGHT_CLAIM
);
5681 if (error
== EBUSY
) {
5682 ASSERT(msp
->ms_loaded
);
5683 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
5689 !range_tree_contains(msp
->ms_allocatable
, offset
, size
))
5690 error
= SET_ERROR(ENOENT
);
5692 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
5693 mutex_exit(&msp
->ms_lock
);
5697 VERIFY(!msp
->ms_condensing
);
5698 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
5699 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
5700 VERIFY3U(range_tree_space(msp
->ms_allocatable
) - size
, <=,
5702 range_tree_remove(msp
->ms_allocatable
, offset
, size
);
5703 range_tree_clear(msp
->ms_trim
, offset
, size
);
5705 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(8) */
5706 metaslab_class_t
*mc
= msp
->ms_group
->mg_class
;
5707 multilist_sublist_t
*mls
=
5708 multilist_sublist_lock_obj(&mc
->mc_metaslab_txg_list
, msp
);
5709 if (!multilist_link_active(&msp
->ms_class_txg_node
)) {
5710 msp
->ms_selected_txg
= txg
;
5711 multilist_sublist_insert_head(mls
, msp
);
5713 multilist_sublist_unlock(mls
);
5715 if (range_tree_is_empty(msp
->ms_allocating
[txg
& TXG_MASK
]))
5716 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
5717 range_tree_add(msp
->ms_allocating
[txg
& TXG_MASK
],
5719 msp
->ms_allocating_total
+= size
;
5722 mutex_exit(&msp
->ms_lock
);
5727 typedef struct metaslab_claim_cb_arg_t
{
5730 } metaslab_claim_cb_arg_t
;
5733 metaslab_claim_impl_cb(uint64_t inner_offset
, vdev_t
*vd
, uint64_t offset
,
5734 uint64_t size
, void *arg
)
5736 (void) inner_offset
;
5737 metaslab_claim_cb_arg_t
*mcca_arg
= arg
;
5739 if (mcca_arg
->mcca_error
== 0) {
5740 mcca_arg
->mcca_error
= metaslab_claim_concrete(vd
, offset
,
5741 size
, mcca_arg
->mcca_txg
);
5746 metaslab_claim_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
, uint64_t txg
)
5748 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
5749 metaslab_claim_cb_arg_t arg
;
5752 * Only zdb(8) can claim on indirect vdevs. This is used
5753 * to detect leaks of mapped space (that are not accounted
5754 * for in the obsolete counts, spacemap, or bpobj).
5756 ASSERT(!spa_writeable(vd
->vdev_spa
));
5760 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
5761 metaslab_claim_impl_cb
, &arg
);
5763 if (arg
.mcca_error
== 0) {
5764 arg
.mcca_error
= metaslab_claim_concrete(vd
,
5767 return (arg
.mcca_error
);
5769 return (metaslab_claim_concrete(vd
, offset
, size
, txg
));
5774 * Intent log support: upon opening the pool after a crash, notify the SPA
5775 * of blocks that the intent log has allocated for immediate write, but
5776 * which are still considered free by the SPA because the last transaction
5777 * group didn't commit yet.
5780 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
5782 uint64_t vdev
= DVA_GET_VDEV(dva
);
5783 uint64_t offset
= DVA_GET_OFFSET(dva
);
5784 uint64_t size
= DVA_GET_ASIZE(dva
);
5787 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
) {
5788 return (SET_ERROR(ENXIO
));
5791 ASSERT(DVA_IS_VALID(dva
));
5793 if (DVA_GET_GANG(dva
))
5794 size
= vdev_gang_header_asize(vd
);
5796 return (metaslab_claim_impl(vd
, offset
, size
, txg
));
5800 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
5801 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
,
5802 zio_alloc_list_t
*zal
, zio_t
*zio
, int allocator
)
5804 dva_t
*dva
= bp
->blk_dva
;
5805 dva_t
*hintdva
= (hintbp
!= NULL
) ? hintbp
->blk_dva
: NULL
;
5808 ASSERT(bp
->blk_birth
== 0);
5809 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
5811 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
5813 if (mc
->mc_allocator
[allocator
].mca_rotor
== NULL
) {
5814 /* no vdevs in this class */
5815 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
5816 return (SET_ERROR(ENOSPC
));
5819 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
5820 ASSERT(BP_GET_NDVAS(bp
) == 0);
5821 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
5822 ASSERT3P(zal
, !=, NULL
);
5824 for (int d
= 0; d
< ndvas
; d
++) {
5825 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
5826 txg
, flags
, zal
, allocator
);
5828 for (d
--; d
>= 0; d
--) {
5829 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
5830 metaslab_group_alloc_decrement(spa
,
5831 DVA_GET_VDEV(&dva
[d
]), zio
, flags
,
5832 allocator
, B_FALSE
);
5833 memset(&dva
[d
], 0, sizeof (dva_t
));
5835 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
5839 * Update the metaslab group's queue depth
5840 * based on the newly allocated dva.
5842 metaslab_group_alloc_increment(spa
,
5843 DVA_GET_VDEV(&dva
[d
]), zio
, flags
, allocator
);
5847 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
5849 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
5851 BP_SET_BIRTH(bp
, txg
, 0);
5857 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
5859 const dva_t
*dva
= bp
->blk_dva
;
5860 int ndvas
= BP_GET_NDVAS(bp
);
5862 ASSERT(!BP_IS_HOLE(bp
));
5863 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
5866 * If we have a checkpoint for the pool we need to make sure that
5867 * the blocks that we free that are part of the checkpoint won't be
5868 * reused until the checkpoint is discarded or we revert to it.
5870 * The checkpoint flag is passed down the metaslab_free code path
5871 * and is set whenever we want to add a block to the checkpoint's
5872 * accounting. That is, we "checkpoint" blocks that existed at the
5873 * time the checkpoint was created and are therefore referenced by
5874 * the checkpointed uberblock.
5876 * Note that, we don't checkpoint any blocks if the current
5877 * syncing txg <= spa_checkpoint_txg. We want these frees to sync
5878 * normally as they will be referenced by the checkpointed uberblock.
5880 boolean_t checkpoint
= B_FALSE
;
5881 if (bp
->blk_birth
<= spa
->spa_checkpoint_txg
&&
5882 spa_syncing_txg(spa
) > spa
->spa_checkpoint_txg
) {
5884 * At this point, if the block is part of the checkpoint
5885 * there is no way it was created in the current txg.
5888 ASSERT3U(spa_syncing_txg(spa
), ==, txg
);
5889 checkpoint
= B_TRUE
;
5892 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
5894 for (int d
= 0; d
< ndvas
; d
++) {
5896 metaslab_unalloc_dva(spa
, &dva
[d
], txg
);
5898 ASSERT3U(txg
, ==, spa_syncing_txg(spa
));
5899 metaslab_free_dva(spa
, &dva
[d
], checkpoint
);
5903 spa_config_exit(spa
, SCL_FREE
, FTAG
);
5907 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
5909 const dva_t
*dva
= bp
->blk_dva
;
5910 int ndvas
= BP_GET_NDVAS(bp
);
5913 ASSERT(!BP_IS_HOLE(bp
));
5917 * First do a dry run to make sure all DVAs are claimable,
5918 * so we don't have to unwind from partial failures below.
5920 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
5924 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
5926 for (int d
= 0; d
< ndvas
; d
++) {
5927 error
= metaslab_claim_dva(spa
, &dva
[d
], txg
);
5932 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
5934 ASSERT(error
== 0 || txg
== 0);
5940 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
5942 const dva_t
*dva
= bp
->blk_dva
;
5943 int ndvas
= BP_GET_NDVAS(bp
);
5944 uint64_t psize
= BP_GET_PSIZE(bp
);
5948 ASSERT(!BP_IS_HOLE(bp
));
5949 ASSERT(!BP_IS_EMBEDDED(bp
));
5952 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
5954 for (d
= 0; d
< ndvas
; d
++) {
5955 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
5957 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
5960 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
5964 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
5966 const dva_t
*dva
= bp
->blk_dva
;
5967 int ndvas
= BP_GET_NDVAS(bp
);
5968 uint64_t psize
= BP_GET_PSIZE(bp
);
5972 ASSERT(!BP_IS_HOLE(bp
));
5973 ASSERT(!BP_IS_EMBEDDED(bp
));
5976 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
5978 for (d
= 0; d
< ndvas
; d
++) {
5979 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
5981 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
5982 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
5985 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
5989 metaslab_check_free_impl_cb(uint64_t inner
, vdev_t
*vd
, uint64_t offset
,
5990 uint64_t size
, void *arg
)
5992 (void) inner
, (void) arg
;
5994 if (vd
->vdev_ops
== &vdev_indirect_ops
)
5997 metaslab_check_free_impl(vd
, offset
, size
);
6001 metaslab_check_free_impl(vdev_t
*vd
, uint64_t offset
, uint64_t size
)
6004 spa_t
*spa __maybe_unused
= vd
->vdev_spa
;
6006 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
6009 if (vd
->vdev_ops
->vdev_op_remap
!= NULL
) {
6010 vd
->vdev_ops
->vdev_op_remap(vd
, offset
, size
,
6011 metaslab_check_free_impl_cb
, NULL
);
6015 ASSERT(vdev_is_concrete(vd
));
6016 ASSERT3U(offset
>> vd
->vdev_ms_shift
, <, vd
->vdev_ms_count
);
6017 ASSERT3U(spa_config_held(spa
, SCL_ALL
, RW_READER
), !=, 0);
6019 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
6021 mutex_enter(&msp
->ms_lock
);
6022 if (msp
->ms_loaded
) {
6023 range_tree_verify_not_present(msp
->ms_allocatable
,
6028 * Check all segments that currently exist in the freeing pipeline.
6030 * It would intuitively make sense to also check the current allocating
6031 * tree since metaslab_unalloc_dva() exists for extents that are
6032 * allocated and freed in the same sync pass within the same txg.
6033 * Unfortunately there are places (e.g. the ZIL) where we allocate a
6034 * segment but then we free part of it within the same txg
6035 * [see zil_sync()]. Thus, we don't call range_tree_verify() in the
6036 * current allocating tree.
6038 range_tree_verify_not_present(msp
->ms_freeing
, offset
, size
);
6039 range_tree_verify_not_present(msp
->ms_checkpointing
, offset
, size
);
6040 range_tree_verify_not_present(msp
->ms_freed
, offset
, size
);
6041 for (int j
= 0; j
< TXG_DEFER_SIZE
; j
++)
6042 range_tree_verify_not_present(msp
->ms_defer
[j
], offset
, size
);
6043 range_tree_verify_not_present(msp
->ms_trim
, offset
, size
);
6044 mutex_exit(&msp
->ms_lock
);
6048 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
6050 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
6053 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
6054 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
6055 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
6056 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
6057 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
6058 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
6060 if (DVA_GET_GANG(&bp
->blk_dva
[i
]))
6061 size
= vdev_gang_header_asize(vd
);
6063 ASSERT3P(vd
, !=, NULL
);
6065 metaslab_check_free_impl(vd
, offset
, size
);
6067 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
6071 metaslab_group_disable_wait(metaslab_group_t
*mg
)
6073 ASSERT(MUTEX_HELD(&mg
->mg_ms_disabled_lock
));
6074 while (mg
->mg_disabled_updating
) {
6075 cv_wait(&mg
->mg_ms_disabled_cv
, &mg
->mg_ms_disabled_lock
);
6080 metaslab_group_disabled_increment(metaslab_group_t
*mg
)
6082 ASSERT(MUTEX_HELD(&mg
->mg_ms_disabled_lock
));
6083 ASSERT(mg
->mg_disabled_updating
);
6085 while (mg
->mg_ms_disabled
>= max_disabled_ms
) {
6086 cv_wait(&mg
->mg_ms_disabled_cv
, &mg
->mg_ms_disabled_lock
);
6088 mg
->mg_ms_disabled
++;
6089 ASSERT3U(mg
->mg_ms_disabled
, <=, max_disabled_ms
);
6093 * Mark the metaslab as disabled to prevent any allocations on this metaslab.
6094 * We must also track how many metaslabs are currently disabled within a
6095 * metaslab group and limit them to prevent allocation failures from
6096 * occurring because all metaslabs are disabled.
6099 metaslab_disable(metaslab_t
*msp
)
6101 ASSERT(!MUTEX_HELD(&msp
->ms_lock
));
6102 metaslab_group_t
*mg
= msp
->ms_group
;
6104 mutex_enter(&mg
->mg_ms_disabled_lock
);
6107 * To keep an accurate count of how many threads have disabled
6108 * a specific metaslab group, we only allow one thread to mark
6109 * the metaslab group at a time. This ensures that the value of
6110 * ms_disabled will be accurate when we decide to mark a metaslab
6111 * group as disabled. To do this we force all other threads
6112 * to wait till the metaslab's mg_disabled_updating flag is no
6115 metaslab_group_disable_wait(mg
);
6116 mg
->mg_disabled_updating
= B_TRUE
;
6117 if (msp
->ms_disabled
== 0) {
6118 metaslab_group_disabled_increment(mg
);
6120 mutex_enter(&msp
->ms_lock
);
6122 mutex_exit(&msp
->ms_lock
);
6124 mg
->mg_disabled_updating
= B_FALSE
;
6125 cv_broadcast(&mg
->mg_ms_disabled_cv
);
6126 mutex_exit(&mg
->mg_ms_disabled_lock
);
6130 metaslab_enable(metaslab_t
*msp
, boolean_t sync
, boolean_t unload
)
6132 metaslab_group_t
*mg
= msp
->ms_group
;
6133 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
6136 * Wait for the outstanding IO to be synced to prevent newly
6137 * allocated blocks from being overwritten. This used by
6138 * initialize and TRIM which are modifying unallocated space.
6141 txg_wait_synced(spa_get_dsl(spa
), 0);
6143 mutex_enter(&mg
->mg_ms_disabled_lock
);
6144 mutex_enter(&msp
->ms_lock
);
6145 if (--msp
->ms_disabled
== 0) {
6146 mg
->mg_ms_disabled
--;
6147 cv_broadcast(&mg
->mg_ms_disabled_cv
);
6149 metaslab_unload(msp
);
6151 mutex_exit(&msp
->ms_lock
);
6152 mutex_exit(&mg
->mg_ms_disabled_lock
);
6156 metaslab_set_unflushed_dirty(metaslab_t
*ms
, boolean_t dirty
)
6158 ms
->ms_unflushed_dirty
= dirty
;
6162 metaslab_update_ondisk_flush_data(metaslab_t
*ms
, dmu_tx_t
*tx
)
6164 vdev_t
*vd
= ms
->ms_group
->mg_vd
;
6165 spa_t
*spa
= vd
->vdev_spa
;
6166 objset_t
*mos
= spa_meta_objset(spa
);
6168 ASSERT(spa_feature_is_active(spa
, SPA_FEATURE_LOG_SPACEMAP
));
6170 metaslab_unflushed_phys_t entry
= {
6171 .msp_unflushed_txg
= metaslab_unflushed_txg(ms
),
6173 uint64_t entry_size
= sizeof (entry
);
6174 uint64_t entry_offset
= ms
->ms_id
* entry_size
;
6176 uint64_t object
= 0;
6177 int err
= zap_lookup(mos
, vd
->vdev_top_zap
,
6178 VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS
, sizeof (uint64_t), 1,
6180 if (err
== ENOENT
) {
6181 object
= dmu_object_alloc(mos
, DMU_OTN_UINT64_METADATA
,
6182 SPA_OLD_MAXBLOCKSIZE
, DMU_OT_NONE
, 0, tx
);
6183 VERIFY0(zap_add(mos
, vd
->vdev_top_zap
,
6184 VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS
, sizeof (uint64_t), 1,
6190 dmu_write(spa_meta_objset(spa
), object
, entry_offset
, entry_size
,
6195 metaslab_set_unflushed_txg(metaslab_t
*ms
, uint64_t txg
, dmu_tx_t
*tx
)
6197 ms
->ms_unflushed_txg
= txg
;
6198 metaslab_update_ondisk_flush_data(ms
, tx
);
6202 metaslab_unflushed_dirty(metaslab_t
*ms
)
6204 return (ms
->ms_unflushed_dirty
);
6208 metaslab_unflushed_txg(metaslab_t
*ms
)
6210 return (ms
->ms_unflushed_txg
);
6213 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, aliquot
, U64
, ZMOD_RW
,
6214 "Allocation granularity (a.k.a. stripe size)");
6216 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, debug_load
, INT
, ZMOD_RW
,
6217 "Load all metaslabs when pool is first opened");
6219 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, debug_unload
, INT
, ZMOD_RW
,
6220 "Prevent metaslabs from being unloaded");
6222 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, preload_enabled
, INT
, ZMOD_RW
,
6223 "Preload potential metaslabs during reassessment");
6225 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, unload_delay
, UINT
, ZMOD_RW
,
6226 "Delay in txgs after metaslab was last used before unloading");
6228 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, unload_delay_ms
, UINT
, ZMOD_RW
,
6229 "Delay in milliseconds after metaslab was last used before unloading");
6232 ZFS_MODULE_PARAM(zfs_mg
, zfs_mg_
, noalloc_threshold
, UINT
, ZMOD_RW
,
6233 "Percentage of metaslab group size that should be free to make it "
6234 "eligible for allocation");
6236 ZFS_MODULE_PARAM(zfs_mg
, zfs_mg_
, fragmentation_threshold
, UINT
, ZMOD_RW
,
6237 "Percentage of metaslab group size that should be considered eligible "
6238 "for allocations unless all metaslab groups within the metaslab class "
6239 "have also crossed this threshold");
6241 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, fragmentation_factor_enabled
, INT
,
6243 "Use the fragmentation metric to prefer less fragmented metaslabs");
6246 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, fragmentation_threshold
, UINT
,
6247 ZMOD_RW
, "Fragmentation for metaslab to allow allocation");
6249 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, lba_weighting_enabled
, INT
, ZMOD_RW
,
6250 "Prefer metaslabs with lower LBAs");
6252 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, bias_enabled
, INT
, ZMOD_RW
,
6253 "Enable metaslab group biasing");
6255 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, segment_weight_enabled
, INT
,
6256 ZMOD_RW
, "Enable segment-based metaslab selection");
6258 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, switch_threshold
, INT
, ZMOD_RW
,
6259 "Segment-based metaslab selection maximum buckets before switching");
6261 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, force_ganging
, U64
, ZMOD_RW
,
6262 "Blocks larger than this size are forced to be gang blocks");
6264 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, df_max_search
, UINT
, ZMOD_RW
,
6265 "Max distance (bytes) to search forward before using size tree");
6267 ZFS_MODULE_PARAM(zfs_metaslab
, metaslab_
, df_use_largest_segment
, INT
, ZMOD_RW
,
6268 "When looking in size tree, use largest segment instead of exact fit");
6270 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, max_size_cache_sec
, U64
,
6271 ZMOD_RW
, "How long to trust the cached max chunk size of a metaslab");
6273 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, mem_limit
, UINT
, ZMOD_RW
,
6274 "Percentage of memory that can be used to store metaslab range trees");
6276 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, try_hard_before_gang
, INT
,
6277 ZMOD_RW
, "Try hard to allocate before ganging");
6279 ZFS_MODULE_PARAM(zfs_metaslab
, zfs_metaslab_
, find_max_tries
, UINT
, ZMOD_RW
,
6280 "Normally only consider this many of the best metaslabs in each vdev");