4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
27 #include <sys/zfs_context.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/space_map.h>
31 #include <sys/metaslab_impl.h>
32 #include <sys/vdev_impl.h>
34 #include <sys/spa_impl.h>
35 #include <sys/zfeature.h>
37 #define WITH_DF_BLOCK_ALLOCATOR
40 * Allow allocations to switch to gang blocks quickly. We do this to
41 * avoid having to load lots of space_maps in a given txg. There are,
42 * however, some cases where we want to avoid "fast" ganging and instead
43 * we want to do an exhaustive search of all metaslabs on this device.
44 * Currently we don't allow any gang, slog, or dump device related allocations
47 #define CAN_FASTGANG(flags) \
48 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
49 METASLAB_GANG_AVOID)))
51 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
52 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
53 #define METASLAB_ACTIVE_MASK \
54 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
57 * Metaslab granularity, in bytes. This is roughly similar to what would be
58 * referred to as the "stripe size" in traditional RAID arrays. In normal
59 * operation, we will try to write this amount of data to a top-level vdev
60 * before moving on to the next one.
62 unsigned long metaslab_aliquot
= 512 << 10;
64 uint64_t metaslab_gang_bang
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
67 * The in-core space map representation is more compact than its on-disk form.
68 * The zfs_condense_pct determines how much more compact the in-core
69 * space_map representation must be before we compact it on-disk.
70 * Values should be greater than or equal to 100.
72 int zfs_condense_pct
= 200;
75 * Condensing a metaslab is not guaranteed to actually reduce the amount of
76 * space used on disk. In particular, a space map uses data in increments of
77 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
78 * same number of blocks after condensing. Since the goal of condensing is to
79 * reduce the number of IOPs required to read the space map, we only want to
80 * condense when we can be sure we will reduce the number of blocks used by the
81 * space map. Unfortunately, we cannot precisely compute whether or not this is
82 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
83 * we apply the following heuristic: do not condense a spacemap unless the
84 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
87 int zfs_metaslab_condense_block_threshold
= 4;
90 * The zfs_mg_noalloc_threshold defines which metaslab groups should
91 * be eligible for allocation. The value is defined as a percentage of
92 * free space. Metaslab groups that have more free space than
93 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
94 * a metaslab group's free space is less than or equal to the
95 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
96 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
97 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
98 * groups are allowed to accept allocations. Gang blocks are always
99 * eligible to allocate on any metaslab group. The default value of 0 means
100 * no metaslab group will be excluded based on this criterion.
102 int zfs_mg_noalloc_threshold
= 0;
105 * Metaslab groups are considered eligible for allocations if their
106 * fragmenation metric (measured as a percentage) is less than or equal to
107 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
108 * then it will be skipped unless all metaslab groups within the metaslab
109 * class have also crossed this threshold.
111 int zfs_mg_fragmentation_threshold
= 85;
114 * Allow metaslabs to keep their active state as long as their fragmentation
115 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
116 * active metaslab that exceeds this threshold will no longer keep its active
117 * status allowing better metaslabs to be selected.
119 int zfs_metaslab_fragmentation_threshold
= 70;
122 * When set will load all metaslabs when pool is first opened.
124 int metaslab_debug_load
= 0;
127 * When set will prevent metaslabs from being unloaded.
129 int metaslab_debug_unload
= 0;
132 * Minimum size which forces the dynamic allocator to change
133 * it's allocation strategy. Once the space map cannot satisfy
134 * an allocation of this size then it switches to using more
135 * aggressive strategy (i.e search by size rather than offset).
137 uint64_t metaslab_df_alloc_threshold
= SPA_MAXBLOCKSIZE
;
140 * The minimum free space, in percent, which must be available
141 * in a space map to continue allocations in a first-fit fashion.
142 * Once the space_map's free space drops below this level we dynamically
143 * switch to using best-fit allocations.
145 int metaslab_df_free_pct
= 4;
148 * Percentage of all cpus that can be used by the metaslab taskq.
150 int metaslab_load_pct
= 50;
153 * Determines how many txgs a metaslab may remain loaded without having any
154 * allocations from it. As long as a metaslab continues to be used we will
157 int metaslab_unload_delay
= TXG_SIZE
* 2;
160 * Max number of metaslabs per group to preload.
162 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
165 * Enable/disable preloading of metaslab.
167 int metaslab_preload_enabled
= B_TRUE
;
170 * Enable/disable fragmentation weighting on metaslabs.
172 int metaslab_fragmentation_factor_enabled
= B_TRUE
;
175 * Enable/disable lba weighting (i.e. outer tracks are given preference).
177 int metaslab_lba_weighting_enabled
= B_TRUE
;
180 * Enable/disable metaslab group biasing.
182 int metaslab_bias_enabled
= B_TRUE
;
184 static uint64_t metaslab_fragmentation(metaslab_t
*);
187 * ==========================================================================
189 * ==========================================================================
192 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
194 metaslab_class_t
*mc
;
196 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
206 metaslab_class_destroy(metaslab_class_t
*mc
)
208 ASSERT(mc
->mc_rotor
== NULL
);
209 ASSERT(mc
->mc_alloc
== 0);
210 ASSERT(mc
->mc_deferred
== 0);
211 ASSERT(mc
->mc_space
== 0);
212 ASSERT(mc
->mc_dspace
== 0);
214 kmem_free(mc
, sizeof (metaslab_class_t
));
218 metaslab_class_validate(metaslab_class_t
*mc
)
220 metaslab_group_t
*mg
;
224 * Must hold one of the spa_config locks.
226 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
227 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
229 if ((mg
= mc
->mc_rotor
) == NULL
)
234 ASSERT(vd
->vdev_mg
!= NULL
);
235 ASSERT3P(vd
->vdev_top
, ==, vd
);
236 ASSERT3P(mg
->mg_class
, ==, mc
);
237 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
238 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
244 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
245 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
247 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
248 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
249 atomic_add_64(&mc
->mc_space
, space_delta
);
250 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
254 metaslab_class_get_alloc(metaslab_class_t
*mc
)
256 return (mc
->mc_alloc
);
260 metaslab_class_get_deferred(metaslab_class_t
*mc
)
262 return (mc
->mc_deferred
);
266 metaslab_class_get_space(metaslab_class_t
*mc
)
268 return (mc
->mc_space
);
272 metaslab_class_get_dspace(metaslab_class_t
*mc
)
274 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
278 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
280 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
284 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
287 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
290 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
291 vdev_t
*tvd
= rvd
->vdev_child
[c
];
292 metaslab_group_t
*mg
= tvd
->vdev_mg
;
295 * Skip any holes, uninitialized top-levels, or
296 * vdevs that are not in this metalab class.
298 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
299 mg
->mg_class
!= mc
) {
303 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
304 mc_hist
[i
] += mg
->mg_histogram
[i
];
307 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
308 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
310 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
314 * Calculate the metaslab class's fragmentation metric. The metric
315 * is weighted based on the space contribution of each metaslab group.
316 * The return value will be a number between 0 and 100 (inclusive), or
317 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
318 * zfs_frag_table for more information about the metric.
321 metaslab_class_fragmentation(metaslab_class_t
*mc
)
323 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
324 uint64_t fragmentation
= 0;
327 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
329 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
330 vdev_t
*tvd
= rvd
->vdev_child
[c
];
331 metaslab_group_t
*mg
= tvd
->vdev_mg
;
334 * Skip any holes, uninitialized top-levels, or
335 * vdevs that are not in this metalab class.
337 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
338 mg
->mg_class
!= mc
) {
343 * If a metaslab group does not contain a fragmentation
344 * metric then just bail out.
346 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
347 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
348 return (ZFS_FRAG_INVALID
);
352 * Determine how much this metaslab_group is contributing
353 * to the overall pool fragmentation metric.
355 fragmentation
+= mg
->mg_fragmentation
*
356 metaslab_group_get_space(mg
);
358 fragmentation
/= metaslab_class_get_space(mc
);
360 ASSERT3U(fragmentation
, <=, 100);
361 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
362 return (fragmentation
);
366 * Calculate the amount of expandable space that is available in
367 * this metaslab class. If a device is expanded then its expandable
368 * space will be the amount of allocatable space that is currently not
369 * part of this metaslab class.
372 metaslab_class_expandable_space(metaslab_class_t
*mc
)
374 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
378 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
379 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
380 vdev_t
*tvd
= rvd
->vdev_child
[c
];
381 metaslab_group_t
*mg
= tvd
->vdev_mg
;
383 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
384 mg
->mg_class
!= mc
) {
388 space
+= tvd
->vdev_max_asize
- tvd
->vdev_asize
;
390 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
395 * ==========================================================================
397 * ==========================================================================
400 metaslab_compare(const void *x1
, const void *x2
)
402 const metaslab_t
*m1
= (const metaslab_t
*)x1
;
403 const metaslab_t
*m2
= (const metaslab_t
*)x2
;
405 int cmp
= AVL_CMP(m2
->ms_weight
, m1
->ms_weight
);
409 IMPLY(AVL_CMP(m1
->ms_start
, m2
->ms_start
) == 0, m1
== m2
);
411 return (AVL_CMP(m1
->ms_start
, m2
->ms_start
));
415 * Update the allocatable flag and the metaslab group's capacity.
416 * The allocatable flag is set to true if the capacity is below
417 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
418 * from allocatable to non-allocatable or vice versa then the metaslab
419 * group's class is updated to reflect the transition.
422 metaslab_group_alloc_update(metaslab_group_t
*mg
)
424 vdev_t
*vd
= mg
->mg_vd
;
425 metaslab_class_t
*mc
= mg
->mg_class
;
426 vdev_stat_t
*vs
= &vd
->vdev_stat
;
427 boolean_t was_allocatable
;
429 ASSERT(vd
== vd
->vdev_top
);
431 mutex_enter(&mg
->mg_lock
);
432 was_allocatable
= mg
->mg_allocatable
;
434 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
438 * A metaslab group is considered allocatable if it has plenty
439 * of free space or is not heavily fragmented. We only take
440 * fragmentation into account if the metaslab group has a valid
441 * fragmentation metric (i.e. a value between 0 and 100).
443 mg
->mg_allocatable
= (mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
444 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
445 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
448 * The mc_alloc_groups maintains a count of the number of
449 * groups in this metaslab class that are still above the
450 * zfs_mg_noalloc_threshold. This is used by the allocating
451 * threads to determine if they should avoid allocations to
452 * a given group. The allocator will avoid allocations to a group
453 * if that group has reached or is below the zfs_mg_noalloc_threshold
454 * and there are still other groups that are above the threshold.
455 * When a group transitions from allocatable to non-allocatable or
456 * vice versa we update the metaslab class to reflect that change.
457 * When the mc_alloc_groups value drops to 0 that means that all
458 * groups have reached the zfs_mg_noalloc_threshold making all groups
459 * eligible for allocations. This effectively means that all devices
460 * are balanced again.
462 if (was_allocatable
&& !mg
->mg_allocatable
)
463 mc
->mc_alloc_groups
--;
464 else if (!was_allocatable
&& mg
->mg_allocatable
)
465 mc
->mc_alloc_groups
++;
467 mutex_exit(&mg
->mg_lock
);
471 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
473 metaslab_group_t
*mg
;
475 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
476 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
477 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
478 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
481 mg
->mg_activation_count
= 0;
483 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
484 maxclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
| TASKQ_DYNAMIC
);
490 metaslab_group_destroy(metaslab_group_t
*mg
)
492 ASSERT(mg
->mg_prev
== NULL
);
493 ASSERT(mg
->mg_next
== NULL
);
495 * We may have gone below zero with the activation count
496 * either because we never activated in the first place or
497 * because we're done, and possibly removing the vdev.
499 ASSERT(mg
->mg_activation_count
<= 0);
501 taskq_destroy(mg
->mg_taskq
);
502 avl_destroy(&mg
->mg_metaslab_tree
);
503 mutex_destroy(&mg
->mg_lock
);
504 kmem_free(mg
, sizeof (metaslab_group_t
));
508 metaslab_group_activate(metaslab_group_t
*mg
)
510 metaslab_class_t
*mc
= mg
->mg_class
;
511 metaslab_group_t
*mgprev
, *mgnext
;
513 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
515 ASSERT(mc
->mc_rotor
!= mg
);
516 ASSERT(mg
->mg_prev
== NULL
);
517 ASSERT(mg
->mg_next
== NULL
);
518 ASSERT(mg
->mg_activation_count
<= 0);
520 if (++mg
->mg_activation_count
<= 0)
523 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
524 metaslab_group_alloc_update(mg
);
526 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
530 mgnext
= mgprev
->mg_next
;
531 mg
->mg_prev
= mgprev
;
532 mg
->mg_next
= mgnext
;
533 mgprev
->mg_next
= mg
;
534 mgnext
->mg_prev
= mg
;
540 metaslab_group_passivate(metaslab_group_t
*mg
)
542 metaslab_class_t
*mc
= mg
->mg_class
;
543 metaslab_group_t
*mgprev
, *mgnext
;
545 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
547 if (--mg
->mg_activation_count
!= 0) {
548 ASSERT(mc
->mc_rotor
!= mg
);
549 ASSERT(mg
->mg_prev
== NULL
);
550 ASSERT(mg
->mg_next
== NULL
);
551 ASSERT(mg
->mg_activation_count
< 0);
555 taskq_wait_outstanding(mg
->mg_taskq
, 0);
556 metaslab_group_alloc_update(mg
);
558 mgprev
= mg
->mg_prev
;
559 mgnext
= mg
->mg_next
;
564 mc
->mc_rotor
= mgnext
;
565 mgprev
->mg_next
= mgnext
;
566 mgnext
->mg_prev
= mgprev
;
574 metaslab_group_get_space(metaslab_group_t
*mg
)
576 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
580 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
583 vdev_t
*vd
= mg
->mg_vd
;
584 uint64_t ashift
= vd
->vdev_ashift
;
587 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
590 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
593 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
594 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
596 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
597 metaslab_t
*msp
= vd
->vdev_ms
[m
];
599 if (msp
->ms_sm
== NULL
)
602 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
603 mg_hist
[i
+ ashift
] +=
604 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
607 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
608 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
610 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
614 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
616 metaslab_class_t
*mc
= mg
->mg_class
;
617 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
620 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
621 if (msp
->ms_sm
== NULL
)
624 mutex_enter(&mg
->mg_lock
);
625 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
626 mg
->mg_histogram
[i
+ ashift
] +=
627 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
628 mc
->mc_histogram
[i
+ ashift
] +=
629 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
631 mutex_exit(&mg
->mg_lock
);
635 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
637 metaslab_class_t
*mc
= mg
->mg_class
;
638 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
641 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
642 if (msp
->ms_sm
== NULL
)
645 mutex_enter(&mg
->mg_lock
);
646 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
647 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
648 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
649 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
650 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
652 mg
->mg_histogram
[i
+ ashift
] -=
653 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
654 mc
->mc_histogram
[i
+ ashift
] -=
655 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
657 mutex_exit(&mg
->mg_lock
);
661 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
663 ASSERT(msp
->ms_group
== NULL
);
664 mutex_enter(&mg
->mg_lock
);
667 avl_add(&mg
->mg_metaslab_tree
, msp
);
668 mutex_exit(&mg
->mg_lock
);
670 mutex_enter(&msp
->ms_lock
);
671 metaslab_group_histogram_add(mg
, msp
);
672 mutex_exit(&msp
->ms_lock
);
676 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
678 mutex_enter(&msp
->ms_lock
);
679 metaslab_group_histogram_remove(mg
, msp
);
680 mutex_exit(&msp
->ms_lock
);
682 mutex_enter(&mg
->mg_lock
);
683 ASSERT(msp
->ms_group
== mg
);
684 avl_remove(&mg
->mg_metaslab_tree
, msp
);
685 msp
->ms_group
= NULL
;
686 mutex_exit(&mg
->mg_lock
);
690 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
693 * Although in principle the weight can be any value, in
694 * practice we do not use values in the range [1, 511].
696 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
697 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
699 mutex_enter(&mg
->mg_lock
);
700 ASSERT(msp
->ms_group
== mg
);
701 avl_remove(&mg
->mg_metaslab_tree
, msp
);
702 msp
->ms_weight
= weight
;
703 avl_add(&mg
->mg_metaslab_tree
, msp
);
704 mutex_exit(&mg
->mg_lock
);
708 * Calculate the fragmentation for a given metaslab group. We can use
709 * a simple average here since all metaslabs within the group must have
710 * the same size. The return value will be a value between 0 and 100
711 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
712 * group have a fragmentation metric.
715 metaslab_group_fragmentation(metaslab_group_t
*mg
)
717 vdev_t
*vd
= mg
->mg_vd
;
718 uint64_t fragmentation
= 0;
719 uint64_t valid_ms
= 0;
722 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
723 metaslab_t
*msp
= vd
->vdev_ms
[m
];
725 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
729 fragmentation
+= msp
->ms_fragmentation
;
732 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
733 return (ZFS_FRAG_INVALID
);
735 fragmentation
/= valid_ms
;
736 ASSERT3U(fragmentation
, <=, 100);
737 return (fragmentation
);
741 * Determine if a given metaslab group should skip allocations. A metaslab
742 * group should avoid allocations if its free capacity is less than the
743 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
744 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
745 * that can still handle allocations.
748 metaslab_group_allocatable(metaslab_group_t
*mg
)
750 vdev_t
*vd
= mg
->mg_vd
;
751 spa_t
*spa
= vd
->vdev_spa
;
752 metaslab_class_t
*mc
= mg
->mg_class
;
755 * We use two key metrics to determine if a metaslab group is
756 * considered allocatable -- free space and fragmentation. If
757 * the free space is greater than the free space threshold and
758 * the fragmentation is less than the fragmentation threshold then
759 * consider the group allocatable. There are two case when we will
760 * not consider these key metrics. The first is if the group is
761 * associated with a slog device and the second is if all groups
762 * in this metaslab class have already been consider ineligible
765 return ((mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
766 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
767 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
)) ||
768 mc
!= spa_normal_class(spa
) || mc
->mc_alloc_groups
== 0);
772 * ==========================================================================
773 * Range tree callbacks
774 * ==========================================================================
778 * Comparison function for the private size-ordered tree. Tree is sorted
779 * by size, larger sizes at the end of the tree.
782 metaslab_rangesize_compare(const void *x1
, const void *x2
)
784 const range_seg_t
*r1
= x1
;
785 const range_seg_t
*r2
= x2
;
786 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
787 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
789 int cmp
= AVL_CMP(rs_size1
, rs_size2
);
793 return (AVL_CMP(r1
->rs_start
, r2
->rs_start
));
797 * Create any block allocator specific components. The current allocators
798 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
801 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
803 metaslab_t
*msp
= arg
;
805 ASSERT3P(rt
->rt_arg
, ==, msp
);
806 ASSERT(msp
->ms_tree
== NULL
);
808 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
809 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
813 * Destroy the block allocator specific components.
816 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
818 metaslab_t
*msp
= arg
;
820 ASSERT3P(rt
->rt_arg
, ==, msp
);
821 ASSERT3P(msp
->ms_tree
, ==, rt
);
822 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
824 avl_destroy(&msp
->ms_size_tree
);
828 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
830 metaslab_t
*msp
= arg
;
832 ASSERT3P(rt
->rt_arg
, ==, msp
);
833 ASSERT3P(msp
->ms_tree
, ==, rt
);
834 VERIFY(!msp
->ms_condensing
);
835 avl_add(&msp
->ms_size_tree
, rs
);
839 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
841 metaslab_t
*msp
= arg
;
843 ASSERT3P(rt
->rt_arg
, ==, msp
);
844 ASSERT3P(msp
->ms_tree
, ==, rt
);
845 VERIFY(!msp
->ms_condensing
);
846 avl_remove(&msp
->ms_size_tree
, rs
);
850 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
852 metaslab_t
*msp
= arg
;
854 ASSERT3P(rt
->rt_arg
, ==, msp
);
855 ASSERT3P(msp
->ms_tree
, ==, rt
);
858 * Normally one would walk the tree freeing nodes along the way.
859 * Since the nodes are shared with the range trees we can avoid
860 * walking all nodes and just reinitialize the avl tree. The nodes
861 * will be freed by the range tree, so we don't want to free them here.
863 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
864 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
867 static range_tree_ops_t metaslab_rt_ops
= {
876 * ==========================================================================
877 * Metaslab block operations
878 * ==========================================================================
882 * Return the maximum contiguous segment within the metaslab.
885 metaslab_block_maxsize(metaslab_t
*msp
)
887 avl_tree_t
*t
= &msp
->ms_size_tree
;
890 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
893 return (rs
->rs_end
- rs
->rs_start
);
897 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
)
900 range_tree_t
*rt
= msp
->ms_tree
;
902 VERIFY(!msp
->ms_condensing
);
904 start
= msp
->ms_ops
->msop_alloc(msp
, size
);
905 if (start
!= -1ULL) {
906 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
908 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
909 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
910 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
911 range_tree_remove(rt
, start
, size
);
917 * ==========================================================================
918 * Common allocator routines
919 * ==========================================================================
922 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
923 defined(WITH_DF_BLOCK_ALLOCATOR) || \
924 defined(WITH_CF_BLOCK_ALLOCATOR)
926 * This is a helper function that can be used by the allocator to find
927 * a suitable block to allocate. This will search the specified AVL
928 * tree looking for a block that matches the specified criteria.
931 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
934 range_seg_t
*rs
, rsearch
;
937 rsearch
.rs_start
= *cursor
;
938 rsearch
.rs_end
= *cursor
+ size
;
940 rs
= avl_find(t
, &rsearch
, &where
);
942 rs
= avl_nearest(t
, where
, AVL_AFTER
);
945 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
947 if (offset
+ size
<= rs
->rs_end
) {
948 *cursor
= offset
+ size
;
951 rs
= AVL_NEXT(t
, rs
);
955 * If we know we've searched the whole map (*cursor == 0), give up.
956 * Otherwise, reset the cursor to the beginning and try again.
962 return (metaslab_block_picker(t
, cursor
, size
, align
));
964 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
966 #if defined(WITH_FF_BLOCK_ALLOCATOR)
968 * ==========================================================================
969 * The first-fit block allocator
970 * ==========================================================================
973 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
976 * Find the largest power of 2 block size that evenly divides the
977 * requested size. This is used to try to allocate blocks with similar
978 * alignment from the same area of the metaslab (i.e. same cursor
979 * bucket) but it does not guarantee that other allocations sizes
980 * may exist in the same region.
982 uint64_t align
= size
& -size
;
983 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
984 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
986 return (metaslab_block_picker(t
, cursor
, size
, align
));
989 static metaslab_ops_t metaslab_ff_ops
= {
993 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
994 #endif /* WITH_FF_BLOCK_ALLOCATOR */
996 #if defined(WITH_DF_BLOCK_ALLOCATOR)
998 * ==========================================================================
999 * Dynamic block allocator -
1000 * Uses the first fit allocation scheme until space get low and then
1001 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1002 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1003 * ==========================================================================
1006 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1009 * Find the largest power of 2 block size that evenly divides the
1010 * requested size. This is used to try to allocate blocks with similar
1011 * alignment from the same area of the metaslab (i.e. same cursor
1012 * bucket) but it does not guarantee that other allocations sizes
1013 * may exist in the same region.
1015 uint64_t align
= size
& -size
;
1016 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1017 range_tree_t
*rt
= msp
->ms_tree
;
1018 avl_tree_t
*t
= &rt
->rt_root
;
1019 uint64_t max_size
= metaslab_block_maxsize(msp
);
1020 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1022 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1023 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1025 if (max_size
< size
)
1029 * If we're running low on space switch to using the size
1030 * sorted AVL tree (best-fit).
1032 if (max_size
< metaslab_df_alloc_threshold
||
1033 free_pct
< metaslab_df_free_pct
) {
1034 t
= &msp
->ms_size_tree
;
1038 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1041 static metaslab_ops_t metaslab_df_ops
= {
1045 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1046 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1048 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1050 * ==========================================================================
1051 * Cursor fit block allocator -
1052 * Select the largest region in the metaslab, set the cursor to the beginning
1053 * of the range and the cursor_end to the end of the range. As allocations
1054 * are made advance the cursor. Continue allocating from the cursor until
1055 * the range is exhausted and then find a new range.
1056 * ==========================================================================
1059 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1061 range_tree_t
*rt
= msp
->ms_tree
;
1062 avl_tree_t
*t
= &msp
->ms_size_tree
;
1063 uint64_t *cursor
= &msp
->ms_lbas
[0];
1064 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1065 uint64_t offset
= 0;
1067 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1068 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1070 ASSERT3U(*cursor_end
, >=, *cursor
);
1072 if ((*cursor
+ size
) > *cursor_end
) {
1075 rs
= avl_last(&msp
->ms_size_tree
);
1076 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1079 *cursor
= rs
->rs_start
;
1080 *cursor_end
= rs
->rs_end
;
1089 static metaslab_ops_t metaslab_cf_ops
= {
1093 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1094 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1096 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1098 * ==========================================================================
1099 * New dynamic fit allocator -
1100 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1101 * contiguous blocks. If no region is found then just use the largest segment
1103 * ==========================================================================
1107 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1108 * to request from the allocator.
1110 uint64_t metaslab_ndf_clump_shift
= 4;
1113 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1115 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1117 range_seg_t
*rs
, rsearch
;
1118 uint64_t hbit
= highbit64(size
);
1119 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1120 uint64_t max_size
= metaslab_block_maxsize(msp
);
1122 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1123 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1125 if (max_size
< size
)
1128 rsearch
.rs_start
= *cursor
;
1129 rsearch
.rs_end
= *cursor
+ size
;
1131 rs
= avl_find(t
, &rsearch
, &where
);
1132 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1133 t
= &msp
->ms_size_tree
;
1135 rsearch
.rs_start
= 0;
1136 rsearch
.rs_end
= MIN(max_size
,
1137 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1138 rs
= avl_find(t
, &rsearch
, &where
);
1140 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1144 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1145 *cursor
= rs
->rs_start
+ size
;
1146 return (rs
->rs_start
);
1151 static metaslab_ops_t metaslab_ndf_ops
= {
1155 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1156 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1160 * ==========================================================================
1162 * ==========================================================================
1166 * Wait for any in-progress metaslab loads to complete.
1169 metaslab_load_wait(metaslab_t
*msp
)
1171 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1173 while (msp
->ms_loading
) {
1174 ASSERT(!msp
->ms_loaded
);
1175 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1180 metaslab_load(metaslab_t
*msp
)
1185 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1186 ASSERT(!msp
->ms_loaded
);
1187 ASSERT(!msp
->ms_loading
);
1189 msp
->ms_loading
= B_TRUE
;
1192 * If the space map has not been allocated yet, then treat
1193 * all the space in the metaslab as free and add it to the
1196 if (msp
->ms_sm
!= NULL
)
1197 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1199 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1201 msp
->ms_loaded
= (error
== 0);
1202 msp
->ms_loading
= B_FALSE
;
1204 if (msp
->ms_loaded
) {
1205 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1206 range_tree_walk(msp
->ms_defertree
[t
],
1207 range_tree_remove
, msp
->ms_tree
);
1210 cv_broadcast(&msp
->ms_load_cv
);
1215 metaslab_unload(metaslab_t
*msp
)
1217 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1218 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1219 msp
->ms_loaded
= B_FALSE
;
1220 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1224 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1227 vdev_t
*vd
= mg
->mg_vd
;
1228 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1232 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1233 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1234 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1236 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1237 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1240 * We only open space map objects that already exist. All others
1241 * will be opened when we finally allocate an object for it.
1244 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1245 ms
->ms_size
, vd
->vdev_ashift
, &ms
->ms_lock
);
1248 kmem_free(ms
, sizeof (metaslab_t
));
1252 ASSERT(ms
->ms_sm
!= NULL
);
1256 * We create the main range tree here, but we don't create the
1257 * alloctree and freetree until metaslab_sync_done(). This serves
1258 * two purposes: it allows metaslab_sync_done() to detect the
1259 * addition of new space; and for debugging, it ensures that we'd
1260 * data fault on any attempt to use this metaslab before it's ready.
1262 ms
->ms_tree
= range_tree_create(&metaslab_rt_ops
, ms
, &ms
->ms_lock
);
1263 metaslab_group_add(mg
, ms
);
1265 ms
->ms_fragmentation
= metaslab_fragmentation(ms
);
1266 ms
->ms_ops
= mg
->mg_class
->mc_ops
;
1269 * If we're opening an existing pool (txg == 0) or creating
1270 * a new one (txg == TXG_INITIAL), all space is available now.
1271 * If we're adding space to an existing pool, the new space
1272 * does not become available until after this txg has synced.
1274 if (txg
<= TXG_INITIAL
)
1275 metaslab_sync_done(ms
, 0);
1278 * If metaslab_debug_load is set and we're initializing a metaslab
1279 * that has an allocated space_map object then load the its space
1280 * map so that can verify frees.
1282 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1283 mutex_enter(&ms
->ms_lock
);
1284 VERIFY0(metaslab_load(ms
));
1285 mutex_exit(&ms
->ms_lock
);
1289 vdev_dirty(vd
, 0, NULL
, txg
);
1290 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1299 metaslab_fini(metaslab_t
*msp
)
1303 metaslab_group_t
*mg
= msp
->ms_group
;
1305 metaslab_group_remove(mg
, msp
);
1307 mutex_enter(&msp
->ms_lock
);
1309 VERIFY(msp
->ms_group
== NULL
);
1310 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1312 space_map_close(msp
->ms_sm
);
1314 metaslab_unload(msp
);
1315 range_tree_destroy(msp
->ms_tree
);
1317 for (t
= 0; t
< TXG_SIZE
; t
++) {
1318 range_tree_destroy(msp
->ms_alloctree
[t
]);
1319 range_tree_destroy(msp
->ms_freetree
[t
]);
1322 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1323 range_tree_destroy(msp
->ms_defertree
[t
]);
1326 ASSERT0(msp
->ms_deferspace
);
1328 mutex_exit(&msp
->ms_lock
);
1329 cv_destroy(&msp
->ms_load_cv
);
1330 mutex_destroy(&msp
->ms_lock
);
1332 kmem_free(msp
, sizeof (metaslab_t
));
1335 #define FRAGMENTATION_TABLE_SIZE 17
1338 * This table defines a segment size based fragmentation metric that will
1339 * allow each metaslab to derive its own fragmentation value. This is done
1340 * by calculating the space in each bucket of the spacemap histogram and
1341 * multiplying that by the fragmetation metric in this table. Doing
1342 * this for all buckets and dividing it by the total amount of free
1343 * space in this metaslab (i.e. the total free space in all buckets) gives
1344 * us the fragmentation metric. This means that a high fragmentation metric
1345 * equates to most of the free space being comprised of small segments.
1346 * Conversely, if the metric is low, then most of the free space is in
1347 * large segments. A 10% change in fragmentation equates to approximately
1348 * double the number of segments.
1350 * This table defines 0% fragmented space using 16MB segments. Testing has
1351 * shown that segments that are greater than or equal to 16MB do not suffer
1352 * from drastic performance problems. Using this value, we derive the rest
1353 * of the table. Since the fragmentation value is never stored on disk, it
1354 * is possible to change these calculations in the future.
1356 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1376 * Calclate the metaslab's fragmentation metric. A return value
1377 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1378 * not support this metric. Otherwise, the return value should be in the
1382 metaslab_fragmentation(metaslab_t
*msp
)
1384 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1385 uint64_t fragmentation
= 0;
1387 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1388 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1391 if (!feature_enabled
)
1392 return (ZFS_FRAG_INVALID
);
1395 * A null space map means that the entire metaslab is free
1396 * and thus is not fragmented.
1398 if (msp
->ms_sm
== NULL
)
1402 * If this metaslab's space_map has not been upgraded, flag it
1403 * so that we upgrade next time we encounter it.
1405 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1406 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1408 if (spa_writeable(vd
->vdev_spa
)) {
1409 uint64_t txg
= spa_syncing_txg(spa
);
1411 msp
->ms_condense_wanted
= B_TRUE
;
1412 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1413 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1414 "msp %p, vd %p", txg
, msp
, vd
);
1416 return (ZFS_FRAG_INVALID
);
1419 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1421 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1422 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1423 FRAGMENTATION_TABLE_SIZE
- 1);
1425 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1428 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1431 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1432 fragmentation
+= space
* zfs_frag_table
[idx
];
1436 fragmentation
/= total
;
1437 ASSERT3U(fragmentation
, <=, 100);
1438 return (fragmentation
);
1442 * Compute a weight -- a selection preference value -- for the given metaslab.
1443 * This is based on the amount of free space, the level of fragmentation,
1444 * the LBA range, and whether the metaslab is loaded.
1447 metaslab_weight(metaslab_t
*msp
)
1449 metaslab_group_t
*mg
= msp
->ms_group
;
1450 vdev_t
*vd
= mg
->mg_vd
;
1451 uint64_t weight
, space
;
1453 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1456 * This vdev is in the process of being removed so there is nothing
1457 * for us to do here.
1459 if (vd
->vdev_removing
) {
1460 ASSERT0(space_map_allocated(msp
->ms_sm
));
1461 ASSERT0(vd
->vdev_ms_shift
);
1466 * The baseline weight is the metaslab's free space.
1468 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1470 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1471 if (metaslab_fragmentation_factor_enabled
&&
1472 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1474 * Use the fragmentation information to inversely scale
1475 * down the baseline weight. We need to ensure that we
1476 * don't exclude this metaslab completely when it's 100%
1477 * fragmented. To avoid this we reduce the fragmented value
1480 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1483 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1484 * this metaslab again. The fragmentation metric may have
1485 * decreased the space to something smaller than
1486 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1487 * so that we can consume any remaining space.
1489 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1490 space
= SPA_MINBLOCKSIZE
;
1495 * Modern disks have uniform bit density and constant angular velocity.
1496 * Therefore, the outer recording zones are faster (higher bandwidth)
1497 * than the inner zones by the ratio of outer to inner track diameter,
1498 * which is typically around 2:1. We account for this by assigning
1499 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1500 * In effect, this means that we'll select the metaslab with the most
1501 * free bandwidth rather than simply the one with the most free space.
1503 if (!vd
->vdev_nonrot
&& metaslab_lba_weighting_enabled
) {
1504 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1505 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1509 * If this metaslab is one we're actively using, adjust its
1510 * weight to make it preferable to any inactive metaslab so
1511 * we'll polish it off. If the fragmentation on this metaslab
1512 * has exceed our threshold, then don't mark it active.
1514 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1515 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1516 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1523 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1525 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1527 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1528 metaslab_load_wait(msp
);
1529 if (!msp
->ms_loaded
) {
1530 int error
= metaslab_load(msp
);
1532 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1537 metaslab_group_sort(msp
->ms_group
, msp
,
1538 msp
->ms_weight
| activation_weight
);
1540 ASSERT(msp
->ms_loaded
);
1541 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1547 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
1550 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1551 * this metaslab again. In that case, it had better be empty,
1552 * or we would be leaving space on the table.
1554 ASSERT(size
>= SPA_MINBLOCKSIZE
|| range_tree_space(msp
->ms_tree
) == 0);
1555 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
1556 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1560 metaslab_preload(void *arg
)
1562 metaslab_t
*msp
= arg
;
1563 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1564 fstrans_cookie_t cookie
= spl_fstrans_mark();
1566 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1568 mutex_enter(&msp
->ms_lock
);
1569 metaslab_load_wait(msp
);
1570 if (!msp
->ms_loaded
)
1571 (void) metaslab_load(msp
);
1574 * Set the ms_access_txg value so that we don't unload it right away.
1576 msp
->ms_access_txg
= spa_syncing_txg(spa
) + metaslab_unload_delay
+ 1;
1577 mutex_exit(&msp
->ms_lock
);
1578 spl_fstrans_unmark(cookie
);
1582 metaslab_group_preload(metaslab_group_t
*mg
)
1584 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1586 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1589 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
1590 taskq_wait_outstanding(mg
->mg_taskq
, 0);
1594 mutex_enter(&mg
->mg_lock
);
1596 * Load the next potential metaslabs
1599 while (msp
!= NULL
) {
1600 metaslab_t
*msp_next
= AVL_NEXT(t
, msp
);
1603 * We preload only the maximum number of metaslabs specified
1604 * by metaslab_preload_limit. If a metaslab is being forced
1605 * to condense then we preload it too. This will ensure
1606 * that force condensing happens in the next txg.
1608 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
1614 * We must drop the metaslab group lock here to preserve
1615 * lock ordering with the ms_lock (when grabbing both
1616 * the mg_lock and the ms_lock, the ms_lock must be taken
1617 * first). As a result, it is possible that the ordering
1618 * of the metaslabs within the avl tree may change before
1619 * we reacquire the lock. The metaslab cannot be removed from
1620 * the tree while we're in syncing context so it is safe to
1621 * drop the mg_lock here. If the metaslabs are reordered
1622 * nothing will break -- we just may end up loading a
1623 * less than optimal one.
1625 mutex_exit(&mg
->mg_lock
);
1626 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
1627 msp
, TQ_SLEEP
) != 0);
1628 mutex_enter(&mg
->mg_lock
);
1631 mutex_exit(&mg
->mg_lock
);
1635 * Determine if the space map's on-disk footprint is past our tolerance
1636 * for inefficiency. We would like to use the following criteria to make
1639 * 1. The size of the space map object should not dramatically increase as a
1640 * result of writing out the free space range tree.
1642 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1643 * times the size than the free space range tree representation
1644 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1646 * 3. The on-disk size of the space map should actually decrease.
1648 * Checking the first condition is tricky since we don't want to walk
1649 * the entire AVL tree calculating the estimated on-disk size. Instead we
1650 * use the size-ordered range tree in the metaslab and calculate the
1651 * size required to write out the largest segment in our free tree. If the
1652 * size required to represent that segment on disk is larger than the space
1653 * map object then we avoid condensing this map.
1655 * To determine the second criterion we use a best-case estimate and assume
1656 * each segment can be represented on-disk as a single 64-bit entry. We refer
1657 * to this best-case estimate as the space map's minimal form.
1659 * Unfortunately, we cannot compute the on-disk size of the space map in this
1660 * context because we cannot accurately compute the effects of compression, etc.
1661 * Instead, we apply the heuristic described in the block comment for
1662 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1663 * is greater than a threshold number of blocks.
1666 metaslab_should_condense(metaslab_t
*msp
)
1668 space_map_t
*sm
= msp
->ms_sm
;
1670 uint64_t size
, entries
, segsz
, object_size
, optimal_size
, record_size
;
1671 dmu_object_info_t doi
;
1672 uint64_t vdev_blocksize
= 1 << msp
->ms_group
->mg_vd
->vdev_ashift
;
1674 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1675 ASSERT(msp
->ms_loaded
);
1678 * Use the ms_size_tree range tree, which is ordered by size, to
1679 * obtain the largest segment in the free tree. We always condense
1680 * metaslabs that are empty and metaslabs for which a condense
1681 * request has been made.
1683 rs
= avl_last(&msp
->ms_size_tree
);
1684 if (rs
== NULL
|| msp
->ms_condense_wanted
)
1688 * Calculate the number of 64-bit entries this segment would
1689 * require when written to disk. If this single segment would be
1690 * larger on-disk than the entire current on-disk structure, then
1691 * clearly condensing will increase the on-disk structure size.
1693 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
1694 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
1695 segsz
= entries
* sizeof (uint64_t);
1697 optimal_size
= sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
);
1698 object_size
= space_map_length(msp
->ms_sm
);
1700 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
1701 record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
1703 return (segsz
<= object_size
&&
1704 object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
1705 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
1709 * Condense the on-disk space map representation to its minimized form.
1710 * The minimized form consists of a small number of allocations followed by
1711 * the entries of the free range tree.
1714 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
1716 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1717 range_tree_t
*freetree
= msp
->ms_freetree
[txg
& TXG_MASK
];
1718 range_tree_t
*condense_tree
;
1719 space_map_t
*sm
= msp
->ms_sm
;
1722 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1723 ASSERT3U(spa_sync_pass(spa
), ==, 1);
1724 ASSERT(msp
->ms_loaded
);
1727 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
1728 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
1729 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
1730 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
1731 space_map_length(msp
->ms_sm
), avl_numnodes(&msp
->ms_tree
->rt_root
),
1732 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
1734 msp
->ms_condense_wanted
= B_FALSE
;
1737 * Create an range tree that is 100% allocated. We remove segments
1738 * that have been freed in this txg, any deferred frees that exist,
1739 * and any allocation in the future. Removing segments should be
1740 * a relatively inexpensive operation since we expect these trees to
1741 * have a small number of nodes.
1743 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
1744 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
1747 * Remove what's been freed in this txg from the condense_tree.
1748 * Since we're in sync_pass 1, we know that all the frees from
1749 * this txg are in the freetree.
1751 range_tree_walk(freetree
, range_tree_remove
, condense_tree
);
1753 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1754 range_tree_walk(msp
->ms_defertree
[t
],
1755 range_tree_remove
, condense_tree
);
1758 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1759 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
1760 range_tree_remove
, condense_tree
);
1764 * We're about to drop the metaslab's lock thus allowing
1765 * other consumers to change it's content. Set the
1766 * metaslab's ms_condensing flag to ensure that
1767 * allocations on this metaslab do not occur while we're
1768 * in the middle of committing it to disk. This is only critical
1769 * for the ms_tree as all other range trees use per txg
1770 * views of their content.
1772 msp
->ms_condensing
= B_TRUE
;
1774 mutex_exit(&msp
->ms_lock
);
1775 space_map_truncate(sm
, tx
);
1776 mutex_enter(&msp
->ms_lock
);
1779 * While we would ideally like to create a space_map representation
1780 * that consists only of allocation records, doing so can be
1781 * prohibitively expensive because the in-core free tree can be
1782 * large, and therefore computationally expensive to subtract
1783 * from the condense_tree. Instead we sync out two trees, a cheap
1784 * allocation only tree followed by the in-core free tree. While not
1785 * optimal, this is typically close to optimal, and much cheaper to
1788 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
1789 range_tree_vacate(condense_tree
, NULL
, NULL
);
1790 range_tree_destroy(condense_tree
);
1792 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
1793 msp
->ms_condensing
= B_FALSE
;
1797 * Write a metaslab to disk in the context of the specified transaction group.
1800 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
1802 metaslab_group_t
*mg
= msp
->ms_group
;
1803 vdev_t
*vd
= mg
->mg_vd
;
1804 spa_t
*spa
= vd
->vdev_spa
;
1805 objset_t
*mos
= spa_meta_objset(spa
);
1806 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
1807 range_tree_t
**freetree
= &msp
->ms_freetree
[txg
& TXG_MASK
];
1808 range_tree_t
**freed_tree
=
1809 &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1811 uint64_t object
= space_map_object(msp
->ms_sm
);
1813 ASSERT(!vd
->vdev_ishole
);
1816 * This metaslab has just been added so there's no work to do now.
1818 if (*freetree
== NULL
) {
1819 ASSERT3P(alloctree
, ==, NULL
);
1823 ASSERT3P(alloctree
, !=, NULL
);
1824 ASSERT3P(*freetree
, !=, NULL
);
1825 ASSERT3P(*freed_tree
, !=, NULL
);
1828 * Normally, we don't want to process a metaslab if there
1829 * are no allocations or frees to perform. However, if the metaslab
1830 * is being forced to condense we need to let it through.
1832 if (range_tree_space(alloctree
) == 0 &&
1833 range_tree_space(*freetree
) == 0 &&
1834 !msp
->ms_condense_wanted
)
1838 * The only state that can actually be changing concurrently with
1839 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1840 * be modifying this txg's alloctree, freetree, freed_tree, or
1841 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1842 * space_map ASSERTs. We drop it whenever we call into the DMU,
1843 * because the DMU can call down to us (e.g. via zio_free()) at
1847 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
1849 if (msp
->ms_sm
== NULL
) {
1850 uint64_t new_object
;
1852 new_object
= space_map_alloc(mos
, tx
);
1853 VERIFY3U(new_object
, !=, 0);
1855 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
1856 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
1858 ASSERT(msp
->ms_sm
!= NULL
);
1861 mutex_enter(&msp
->ms_lock
);
1864 * Note: metaslab_condense() clears the space_map's histogram.
1865 * Therefore we muse verify and remove this histogram before
1868 metaslab_group_histogram_verify(mg
);
1869 metaslab_class_histogram_verify(mg
->mg_class
);
1870 metaslab_group_histogram_remove(mg
, msp
);
1872 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
1873 metaslab_should_condense(msp
)) {
1874 metaslab_condense(msp
, txg
, tx
);
1876 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
1877 space_map_write(msp
->ms_sm
, *freetree
, SM_FREE
, tx
);
1880 if (msp
->ms_loaded
) {
1882 * When the space map is loaded, we have an accruate
1883 * histogram in the range tree. This gives us an opportunity
1884 * to bring the space map's histogram up-to-date so we clear
1885 * it first before updating it.
1887 space_map_histogram_clear(msp
->ms_sm
);
1888 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
1891 * Since the space map is not loaded we simply update the
1892 * exisiting histogram with what was freed in this txg. This
1893 * means that the on-disk histogram may not have an accurate
1894 * view of the free space but it's close enough to allow
1895 * us to make allocation decisions.
1897 space_map_histogram_add(msp
->ms_sm
, *freetree
, tx
);
1899 metaslab_group_histogram_add(mg
, msp
);
1900 metaslab_group_histogram_verify(mg
);
1901 metaslab_class_histogram_verify(mg
->mg_class
);
1904 * For sync pass 1, we avoid traversing this txg's free range tree
1905 * and instead will just swap the pointers for freetree and
1906 * freed_tree. We can safely do this since the freed_tree is
1907 * guaranteed to be empty on the initial pass.
1909 if (spa_sync_pass(spa
) == 1) {
1910 range_tree_swap(freetree
, freed_tree
);
1912 range_tree_vacate(*freetree
, range_tree_add
, *freed_tree
);
1914 range_tree_vacate(alloctree
, NULL
, NULL
);
1916 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1917 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1919 mutex_exit(&msp
->ms_lock
);
1921 if (object
!= space_map_object(msp
->ms_sm
)) {
1922 object
= space_map_object(msp
->ms_sm
);
1923 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
1924 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
1930 * Called after a transaction group has completely synced to mark
1931 * all of the metaslab's free space as usable.
1934 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1936 metaslab_group_t
*mg
= msp
->ms_group
;
1937 vdev_t
*vd
= mg
->mg_vd
;
1938 range_tree_t
**freed_tree
;
1939 range_tree_t
**defer_tree
;
1940 int64_t alloc_delta
, defer_delta
;
1943 ASSERT(!vd
->vdev_ishole
);
1945 mutex_enter(&msp
->ms_lock
);
1948 * If this metaslab is just becoming available, initialize its
1949 * alloctrees, freetrees, and defertree and add its capacity to
1952 if (msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
] == NULL
) {
1953 for (t
= 0; t
< TXG_SIZE
; t
++) {
1954 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
1955 ASSERT(msp
->ms_freetree
[t
] == NULL
);
1957 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
1959 msp
->ms_freetree
[t
] = range_tree_create(NULL
, msp
,
1963 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1964 ASSERT(msp
->ms_defertree
[t
] == NULL
);
1966 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
1970 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
1973 freed_tree
= &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1974 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
1976 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
1977 defer_delta
= range_tree_space(*freed_tree
) -
1978 range_tree_space(*defer_tree
);
1980 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1982 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1983 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1986 * If there's a metaslab_load() in progress, wait for it to complete
1987 * so that we have a consistent view of the in-core space map.
1989 metaslab_load_wait(msp
);
1992 * Move the frees from the defer_tree back to the free
1993 * range tree (if it's loaded). Swap the freed_tree and the
1994 * defer_tree -- this is safe to do because we've just emptied out
1997 range_tree_vacate(*defer_tree
,
1998 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
1999 range_tree_swap(freed_tree
, defer_tree
);
2001 space_map_update(msp
->ms_sm
);
2003 msp
->ms_deferspace
+= defer_delta
;
2004 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2005 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2006 if (msp
->ms_deferspace
!= 0) {
2008 * Keep syncing this metaslab until all deferred frees
2009 * are back in circulation.
2011 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2014 if (msp
->ms_loaded
&& msp
->ms_access_txg
< txg
) {
2015 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2016 VERIFY0(range_tree_space(
2017 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
2020 if (!metaslab_debug_unload
)
2021 metaslab_unload(msp
);
2024 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
2025 mutex_exit(&msp
->ms_lock
);
2029 metaslab_sync_reassess(metaslab_group_t
*mg
)
2031 metaslab_group_alloc_update(mg
);
2032 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2035 * Preload the next potential metaslabs
2037 metaslab_group_preload(mg
);
2041 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2043 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2044 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2045 uint64_t start
= msp
->ms_id
;
2047 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2048 return (1ULL << 63);
2051 return ((start
- offset
) << ms_shift
);
2053 return ((offset
- start
) << ms_shift
);
2058 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
2059 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2061 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2062 metaslab_t
*msp
= NULL
;
2063 uint64_t offset
= -1ULL;
2064 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2065 uint64_t activation_weight
;
2066 uint64_t target_distance
;
2069 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2070 for (i
= 0; i
< d
; i
++) {
2071 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2072 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2078 boolean_t was_active
;
2080 mutex_enter(&mg
->mg_lock
);
2081 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
2082 if (msp
->ms_weight
< asize
) {
2083 spa_dbgmsg(spa
, "%s: failed to meet weight "
2084 "requirement: vdev %llu, txg %llu, mg %p, "
2085 "msp %p, psize %llu, asize %llu, "
2086 "weight %llu", spa_name(spa
),
2087 mg
->mg_vd
->vdev_id
, txg
,
2088 mg
, msp
, psize
, asize
, msp
->ms_weight
);
2089 mutex_exit(&mg
->mg_lock
);
2094 * If the selected metaslab is condensing, skip it.
2096 if (msp
->ms_condensing
)
2099 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2100 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2103 target_distance
= min_distance
+
2104 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2107 for (i
= 0; i
< d
; i
++)
2108 if (metaslab_distance(msp
, &dva
[i
]) <
2114 mutex_exit(&mg
->mg_lock
);
2118 mutex_enter(&msp
->ms_lock
);
2121 * Ensure that the metaslab we have selected is still
2122 * capable of handling our request. It's possible that
2123 * another thread may have changed the weight while we
2124 * were blocked on the metaslab lock.
2126 if (msp
->ms_weight
< asize
|| (was_active
&&
2127 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
2128 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
2129 mutex_exit(&msp
->ms_lock
);
2133 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2134 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2135 metaslab_passivate(msp
,
2136 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2137 mutex_exit(&msp
->ms_lock
);
2141 if (metaslab_activate(msp
, activation_weight
) != 0) {
2142 mutex_exit(&msp
->ms_lock
);
2147 * If this metaslab is currently condensing then pick again as
2148 * we can't manipulate this metaslab until it's committed
2151 if (msp
->ms_condensing
) {
2152 mutex_exit(&msp
->ms_lock
);
2156 if ((offset
= metaslab_block_alloc(msp
, asize
)) != -1ULL)
2159 metaslab_passivate(msp
, metaslab_block_maxsize(msp
));
2160 mutex_exit(&msp
->ms_lock
);
2163 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2164 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2166 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, asize
);
2167 msp
->ms_access_txg
= txg
+ metaslab_unload_delay
;
2169 mutex_exit(&msp
->ms_lock
);
2175 * Allocate a block for the specified i/o.
2178 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2179 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
2181 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
2185 int zio_lock
= B_FALSE
;
2186 boolean_t allocatable
;
2187 uint64_t offset
= -1ULL;
2191 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2194 * For testing, make some blocks above a certain size be gang blocks.
2196 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
2197 return (SET_ERROR(ENOSPC
));
2200 * Start at the rotor and loop through all mgs until we find something.
2201 * Note that there's no locking on mc_rotor or mc_aliquot because
2202 * nothing actually breaks if we miss a few updates -- we just won't
2203 * allocate quite as evenly. It all balances out over time.
2205 * If we are doing ditto or log blocks, try to spread them across
2206 * consecutive vdevs. If we're forced to reuse a vdev before we've
2207 * allocated all of our ditto blocks, then try and spread them out on
2208 * that vdev as much as possible. If it turns out to not be possible,
2209 * gradually lower our standards until anything becomes acceptable.
2210 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2211 * gives us hope of containing our fault domains to something we're
2212 * able to reason about. Otherwise, any two top-level vdev failures
2213 * will guarantee the loss of data. With consecutive allocation,
2214 * only two adjacent top-level vdev failures will result in data loss.
2216 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2217 * ourselves on the same vdev as our gang block header. That
2218 * way, we can hope for locality in vdev_cache, plus it makes our
2219 * fault domains something tractable.
2222 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
2225 * It's possible the vdev we're using as the hint no
2226 * longer exists (i.e. removed). Consult the rotor when
2232 if (flags
& METASLAB_HINTBP_AVOID
&&
2233 mg
->mg_next
!= NULL
)
2238 } else if (d
!= 0) {
2239 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
2240 mg
= vd
->vdev_mg
->mg_next
;
2241 } else if (flags
& METASLAB_FASTWRITE
) {
2242 mg
= fast_mg
= mc
->mc_rotor
;
2245 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
2246 mg
->mg_vd
->vdev_pending_fastwrite
)
2248 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
2255 * If the hint put us into the wrong metaslab class, or into a
2256 * metaslab group that has been passivated, just follow the rotor.
2258 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
2265 ASSERT(mg
->mg_activation_count
== 1);
2270 * Don't allocate from faulted devices.
2273 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
2274 allocatable
= vdev_allocatable(vd
);
2275 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
2277 allocatable
= vdev_allocatable(vd
);
2281 * Determine if the selected metaslab group is eligible
2282 * for allocations. If we're ganging or have requested
2283 * an allocation for the smallest gang block size
2284 * then we don't want to avoid allocating to the this
2285 * metaslab group. If we're in this condition we should
2286 * try to allocate from any device possible so that we
2287 * don't inadvertently return ENOSPC and suspend the pool
2288 * even though space is still available.
2290 if (allocatable
&& CAN_FASTGANG(flags
) &&
2291 psize
> SPA_GANGBLOCKSIZE
)
2292 allocatable
= metaslab_group_allocatable(mg
);
2298 * Avoid writing single-copy data to a failing vdev
2299 * unless the user instructs us that it is okay.
2301 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
2302 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
2303 d
== 0 && dshift
== 3 && vd
->vdev_children
== 0) {
2308 ASSERT(mg
->mg_class
== mc
);
2310 distance
= vd
->vdev_asize
>> dshift
;
2311 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
2316 asize
= vdev_psize_to_asize(vd
, psize
);
2317 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
2319 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
2321 if (offset
!= -1ULL) {
2323 * If we've just selected this metaslab group,
2324 * figure out whether the corresponding vdev is
2325 * over- or under-used relative to the pool,
2326 * and set an allocation bias to even it out.
2328 * Bias is also used to compensate for unequally
2329 * sized vdevs so that space is allocated fairly.
2331 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
2332 vdev_stat_t
*vs
= &vd
->vdev_stat
;
2333 int64_t vs_free
= vs
->vs_space
- vs
->vs_alloc
;
2334 int64_t mc_free
= mc
->mc_space
- mc
->mc_alloc
;
2338 * Calculate how much more or less we should
2339 * try to allocate from this device during
2340 * this iteration around the rotor.
2342 * This basically introduces a zero-centered
2343 * bias towards the devices with the most
2344 * free space, while compensating for vdev
2348 * vdev V1 = 16M/128M
2349 * vdev V2 = 16M/128M
2350 * ratio(V1) = 100% ratio(V2) = 100%
2352 * vdev V1 = 16M/128M
2353 * vdev V2 = 64M/128M
2354 * ratio(V1) = 127% ratio(V2) = 72%
2356 * vdev V1 = 16M/128M
2357 * vdev V2 = 64M/512M
2358 * ratio(V1) = 40% ratio(V2) = 160%
2360 ratio
= (vs_free
* mc
->mc_alloc_groups
* 100) /
2362 mg
->mg_bias
= ((ratio
- 100) *
2363 (int64_t)mg
->mg_aliquot
) / 100;
2364 } else if (!metaslab_bias_enabled
) {
2368 if ((flags
& METASLAB_FASTWRITE
) ||
2369 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
2370 mg
->mg_aliquot
+ mg
->mg_bias
) {
2371 mc
->mc_rotor
= mg
->mg_next
;
2375 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
2376 DVA_SET_OFFSET(&dva
[d
], offset
);
2377 DVA_SET_GANG(&dva
[d
],
2378 ((flags
& METASLAB_GANG_HEADER
) ? 1 : 0));
2379 DVA_SET_ASIZE(&dva
[d
], asize
);
2381 if (flags
& METASLAB_FASTWRITE
) {
2382 atomic_add_64(&vd
->vdev_pending_fastwrite
,
2389 mc
->mc_rotor
= mg
->mg_next
;
2391 } while ((mg
= mg
->mg_next
) != rotor
);
2395 ASSERT(dshift
< 64);
2399 if (!allocatable
&& !zio_lock
) {
2405 bzero(&dva
[d
], sizeof (dva_t
));
2407 return (SET_ERROR(ENOSPC
));
2411 * Free the block represented by DVA in the context of the specified
2412 * transaction group.
2415 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
2417 uint64_t vdev
= DVA_GET_VDEV(dva
);
2418 uint64_t offset
= DVA_GET_OFFSET(dva
);
2419 uint64_t size
= DVA_GET_ASIZE(dva
);
2423 if (txg
> spa_freeze_txg(spa
))
2426 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
2427 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
2428 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
2429 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
2430 (u_longlong_t
)size
);
2434 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2436 if (DVA_GET_GANG(dva
))
2437 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2439 mutex_enter(&msp
->ms_lock
);
2442 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
2445 VERIFY(!msp
->ms_condensing
);
2446 VERIFY3U(offset
, >=, msp
->ms_start
);
2447 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
2448 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
2450 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2451 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2452 range_tree_add(msp
->ms_tree
, offset
, size
);
2454 if (range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]) == 0)
2455 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2456 range_tree_add(msp
->ms_freetree
[txg
& TXG_MASK
],
2460 mutex_exit(&msp
->ms_lock
);
2464 * Intent log support: upon opening the pool after a crash, notify the SPA
2465 * of blocks that the intent log has allocated for immediate write, but
2466 * which are still considered free by the SPA because the last transaction
2467 * group didn't commit yet.
2470 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
2472 uint64_t vdev
= DVA_GET_VDEV(dva
);
2473 uint64_t offset
= DVA_GET_OFFSET(dva
);
2474 uint64_t size
= DVA_GET_ASIZE(dva
);
2479 ASSERT(DVA_IS_VALID(dva
));
2481 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2482 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
2483 return (SET_ERROR(ENXIO
));
2485 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2487 if (DVA_GET_GANG(dva
))
2488 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2490 mutex_enter(&msp
->ms_lock
);
2492 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
2493 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
2495 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
2496 error
= SET_ERROR(ENOENT
);
2498 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
2499 mutex_exit(&msp
->ms_lock
);
2503 VERIFY(!msp
->ms_condensing
);
2504 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2505 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2506 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
2507 range_tree_remove(msp
->ms_tree
, offset
, size
);
2509 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
2510 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2511 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2512 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
2515 mutex_exit(&msp
->ms_lock
);
2521 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
2522 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
2524 dva_t
*dva
= bp
->blk_dva
;
2525 dva_t
*hintdva
= hintbp
->blk_dva
;
2528 ASSERT(bp
->blk_birth
== 0);
2529 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
2531 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2533 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
2534 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2535 return (SET_ERROR(ENOSPC
));
2538 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
2539 ASSERT(BP_GET_NDVAS(bp
) == 0);
2540 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
2542 for (d
= 0; d
< ndvas
; d
++) {
2543 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
2546 for (d
--; d
>= 0; d
--) {
2547 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
2548 bzero(&dva
[d
], sizeof (dva_t
));
2550 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2555 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
2557 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2559 BP_SET_BIRTH(bp
, txg
, 0);
2565 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
2567 const dva_t
*dva
= bp
->blk_dva
;
2568 int d
, ndvas
= BP_GET_NDVAS(bp
);
2570 ASSERT(!BP_IS_HOLE(bp
));
2571 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
2573 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
2575 for (d
= 0; d
< ndvas
; d
++)
2576 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
2578 spa_config_exit(spa
, SCL_FREE
, FTAG
);
2582 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
2584 const dva_t
*dva
= bp
->blk_dva
;
2585 int ndvas
= BP_GET_NDVAS(bp
);
2588 ASSERT(!BP_IS_HOLE(bp
));
2592 * First do a dry run to make sure all DVAs are claimable,
2593 * so we don't have to unwind from partial failures below.
2595 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
2599 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2601 for (d
= 0; d
< ndvas
; d
++)
2602 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
2605 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2607 ASSERT(error
== 0 || txg
== 0);
2613 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
2615 const dva_t
*dva
= bp
->blk_dva
;
2616 int ndvas
= BP_GET_NDVAS(bp
);
2617 uint64_t psize
= BP_GET_PSIZE(bp
);
2621 ASSERT(!BP_IS_HOLE(bp
));
2622 ASSERT(!BP_IS_EMBEDDED(bp
));
2625 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2627 for (d
= 0; d
< ndvas
; d
++) {
2628 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2630 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
2633 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2637 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
2639 const dva_t
*dva
= bp
->blk_dva
;
2640 int ndvas
= BP_GET_NDVAS(bp
);
2641 uint64_t psize
= BP_GET_PSIZE(bp
);
2645 ASSERT(!BP_IS_HOLE(bp
));
2646 ASSERT(!BP_IS_EMBEDDED(bp
));
2649 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2651 for (d
= 0; d
< ndvas
; d
++) {
2652 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2654 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
2655 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
2658 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2662 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
2666 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
2669 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2670 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2671 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
2672 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
2673 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
2674 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2675 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2678 range_tree_verify(msp
->ms_tree
, offset
, size
);
2680 for (j
= 0; j
< TXG_SIZE
; j
++)
2681 range_tree_verify(msp
->ms_freetree
[j
], offset
, size
);
2682 for (j
= 0; j
< TXG_DEFER_SIZE
; j
++)
2683 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
2685 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2688 #if defined(_KERNEL) && defined(HAVE_SPL)
2689 module_param(metaslab_aliquot
, ulong
, 0644);
2690 module_param(metaslab_debug_load
, int, 0644);
2691 module_param(metaslab_debug_unload
, int, 0644);
2692 module_param(metaslab_preload_enabled
, int, 0644);
2693 module_param(zfs_mg_noalloc_threshold
, int, 0644);
2694 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
2695 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
2696 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
2697 module_param(metaslab_lba_weighting_enabled
, int, 0644);
2698 module_param(metaslab_bias_enabled
, int, 0644);
2700 MODULE_PARM_DESC(metaslab_aliquot
,
2701 "allocation granularity (a.k.a. stripe size)");
2702 MODULE_PARM_DESC(metaslab_debug_load
,
2703 "load all metaslabs when pool is first opened");
2704 MODULE_PARM_DESC(metaslab_debug_unload
,
2705 "prevent metaslabs from being unloaded");
2706 MODULE_PARM_DESC(metaslab_preload_enabled
,
2707 "preload potential metaslabs during reassessment");
2709 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
2710 "percentage of free space for metaslab group to allow allocation");
2711 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
2712 "fragmentation for metaslab group to allow allocation");
2714 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
2715 "fragmentation for metaslab to allow allocation");
2716 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
2717 "use the fragmentation metric to prefer less fragmented metaslabs");
2718 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
2719 "prefer metaslabs with lower LBAs");
2720 MODULE_PARM_DESC(metaslab_bias_enabled
,
2721 "enable metaslab group biasing");
2722 #endif /* _KERNEL && HAVE_SPL */