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
= x1
;
403 const metaslab_t
*m2
= x2
;
405 if (m1
->ms_weight
< m2
->ms_weight
)
407 if (m1
->ms_weight
> m2
->ms_weight
)
411 * If the weights are identical, use the offset to force uniqueness.
413 if (m1
->ms_start
< m2
->ms_start
)
415 if (m1
->ms_start
> m2
->ms_start
)
418 ASSERT3P(m1
, ==, m2
);
424 * Update the allocatable flag and the metaslab group's capacity.
425 * The allocatable flag is set to true if the capacity is below
426 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
427 * from allocatable to non-allocatable or vice versa then the metaslab
428 * group's class is updated to reflect the transition.
431 metaslab_group_alloc_update(metaslab_group_t
*mg
)
433 vdev_t
*vd
= mg
->mg_vd
;
434 metaslab_class_t
*mc
= mg
->mg_class
;
435 vdev_stat_t
*vs
= &vd
->vdev_stat
;
436 boolean_t was_allocatable
;
438 ASSERT(vd
== vd
->vdev_top
);
440 mutex_enter(&mg
->mg_lock
);
441 was_allocatable
= mg
->mg_allocatable
;
443 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
447 * A metaslab group is considered allocatable if it has plenty
448 * of free space or is not heavily fragmented. We only take
449 * fragmentation into account if the metaslab group has a valid
450 * fragmentation metric (i.e. a value between 0 and 100).
452 mg
->mg_allocatable
= (mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
453 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
454 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
457 * The mc_alloc_groups maintains a count of the number of
458 * groups in this metaslab class that are still above the
459 * zfs_mg_noalloc_threshold. This is used by the allocating
460 * threads to determine if they should avoid allocations to
461 * a given group. The allocator will avoid allocations to a group
462 * if that group has reached or is below the zfs_mg_noalloc_threshold
463 * and there are still other groups that are above the threshold.
464 * When a group transitions from allocatable to non-allocatable or
465 * vice versa we update the metaslab class to reflect that change.
466 * When the mc_alloc_groups value drops to 0 that means that all
467 * groups have reached the zfs_mg_noalloc_threshold making all groups
468 * eligible for allocations. This effectively means that all devices
469 * are balanced again.
471 if (was_allocatable
&& !mg
->mg_allocatable
)
472 mc
->mc_alloc_groups
--;
473 else if (!was_allocatable
&& mg
->mg_allocatable
)
474 mc
->mc_alloc_groups
++;
476 mutex_exit(&mg
->mg_lock
);
480 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
482 metaslab_group_t
*mg
;
484 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
485 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
486 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
487 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
490 mg
->mg_activation_count
= 0;
492 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
493 maxclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
| TASKQ_DYNAMIC
);
499 metaslab_group_destroy(metaslab_group_t
*mg
)
501 ASSERT(mg
->mg_prev
== NULL
);
502 ASSERT(mg
->mg_next
== NULL
);
504 * We may have gone below zero with the activation count
505 * either because we never activated in the first place or
506 * because we're done, and possibly removing the vdev.
508 ASSERT(mg
->mg_activation_count
<= 0);
510 taskq_destroy(mg
->mg_taskq
);
511 avl_destroy(&mg
->mg_metaslab_tree
);
512 mutex_destroy(&mg
->mg_lock
);
513 kmem_free(mg
, sizeof (metaslab_group_t
));
517 metaslab_group_activate(metaslab_group_t
*mg
)
519 metaslab_class_t
*mc
= mg
->mg_class
;
520 metaslab_group_t
*mgprev
, *mgnext
;
522 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
524 ASSERT(mc
->mc_rotor
!= mg
);
525 ASSERT(mg
->mg_prev
== NULL
);
526 ASSERT(mg
->mg_next
== NULL
);
527 ASSERT(mg
->mg_activation_count
<= 0);
529 if (++mg
->mg_activation_count
<= 0)
532 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
533 metaslab_group_alloc_update(mg
);
535 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
539 mgnext
= mgprev
->mg_next
;
540 mg
->mg_prev
= mgprev
;
541 mg
->mg_next
= mgnext
;
542 mgprev
->mg_next
= mg
;
543 mgnext
->mg_prev
= mg
;
549 metaslab_group_passivate(metaslab_group_t
*mg
)
551 metaslab_class_t
*mc
= mg
->mg_class
;
552 metaslab_group_t
*mgprev
, *mgnext
;
554 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
556 if (--mg
->mg_activation_count
!= 0) {
557 ASSERT(mc
->mc_rotor
!= mg
);
558 ASSERT(mg
->mg_prev
== NULL
);
559 ASSERT(mg
->mg_next
== NULL
);
560 ASSERT(mg
->mg_activation_count
< 0);
564 taskq_wait_outstanding(mg
->mg_taskq
, 0);
565 metaslab_group_alloc_update(mg
);
567 mgprev
= mg
->mg_prev
;
568 mgnext
= mg
->mg_next
;
573 mc
->mc_rotor
= mgnext
;
574 mgprev
->mg_next
= mgnext
;
575 mgnext
->mg_prev
= mgprev
;
583 metaslab_group_get_space(metaslab_group_t
*mg
)
585 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
589 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
592 vdev_t
*vd
= mg
->mg_vd
;
593 uint64_t ashift
= vd
->vdev_ashift
;
596 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
599 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
602 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
603 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
605 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
606 metaslab_t
*msp
= vd
->vdev_ms
[m
];
608 if (msp
->ms_sm
== NULL
)
611 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
612 mg_hist
[i
+ ashift
] +=
613 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
616 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
617 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
619 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
623 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
625 metaslab_class_t
*mc
= mg
->mg_class
;
626 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
629 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
630 if (msp
->ms_sm
== NULL
)
633 mutex_enter(&mg
->mg_lock
);
634 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
635 mg
->mg_histogram
[i
+ ashift
] +=
636 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
637 mc
->mc_histogram
[i
+ ashift
] +=
638 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
640 mutex_exit(&mg
->mg_lock
);
644 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
646 metaslab_class_t
*mc
= mg
->mg_class
;
647 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
650 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
651 if (msp
->ms_sm
== NULL
)
654 mutex_enter(&mg
->mg_lock
);
655 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
656 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
657 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
658 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
659 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
661 mg
->mg_histogram
[i
+ ashift
] -=
662 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
663 mc
->mc_histogram
[i
+ ashift
] -=
664 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
666 mutex_exit(&mg
->mg_lock
);
670 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
672 ASSERT(msp
->ms_group
== NULL
);
673 mutex_enter(&mg
->mg_lock
);
676 avl_add(&mg
->mg_metaslab_tree
, msp
);
677 mutex_exit(&mg
->mg_lock
);
679 mutex_enter(&msp
->ms_lock
);
680 metaslab_group_histogram_add(mg
, msp
);
681 mutex_exit(&msp
->ms_lock
);
685 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
687 mutex_enter(&msp
->ms_lock
);
688 metaslab_group_histogram_remove(mg
, msp
);
689 mutex_exit(&msp
->ms_lock
);
691 mutex_enter(&mg
->mg_lock
);
692 ASSERT(msp
->ms_group
== mg
);
693 avl_remove(&mg
->mg_metaslab_tree
, msp
);
694 msp
->ms_group
= NULL
;
695 mutex_exit(&mg
->mg_lock
);
699 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
702 * Although in principle the weight can be any value, in
703 * practice we do not use values in the range [1, 511].
705 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
706 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
708 mutex_enter(&mg
->mg_lock
);
709 ASSERT(msp
->ms_group
== mg
);
710 avl_remove(&mg
->mg_metaslab_tree
, msp
);
711 msp
->ms_weight
= weight
;
712 avl_add(&mg
->mg_metaslab_tree
, msp
);
713 mutex_exit(&mg
->mg_lock
);
717 * Calculate the fragmentation for a given metaslab group. We can use
718 * a simple average here since all metaslabs within the group must have
719 * the same size. The return value will be a value between 0 and 100
720 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
721 * group have a fragmentation metric.
724 metaslab_group_fragmentation(metaslab_group_t
*mg
)
726 vdev_t
*vd
= mg
->mg_vd
;
727 uint64_t fragmentation
= 0;
728 uint64_t valid_ms
= 0;
731 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
732 metaslab_t
*msp
= vd
->vdev_ms
[m
];
734 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
738 fragmentation
+= msp
->ms_fragmentation
;
741 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
742 return (ZFS_FRAG_INVALID
);
744 fragmentation
/= valid_ms
;
745 ASSERT3U(fragmentation
, <=, 100);
746 return (fragmentation
);
750 * Determine if a given metaslab group should skip allocations. A metaslab
751 * group should avoid allocations if its free capacity is less than the
752 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
753 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
754 * that can still handle allocations.
757 metaslab_group_allocatable(metaslab_group_t
*mg
)
759 vdev_t
*vd
= mg
->mg_vd
;
760 spa_t
*spa
= vd
->vdev_spa
;
761 metaslab_class_t
*mc
= mg
->mg_class
;
764 * We use two key metrics to determine if a metaslab group is
765 * considered allocatable -- free space and fragmentation. If
766 * the free space is greater than the free space threshold and
767 * the fragmentation is less than the fragmentation threshold then
768 * consider the group allocatable. There are two case when we will
769 * not consider these key metrics. The first is if the group is
770 * associated with a slog device and the second is if all groups
771 * in this metaslab class have already been consider ineligible
774 return ((mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
775 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
776 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
)) ||
777 mc
!= spa_normal_class(spa
) || mc
->mc_alloc_groups
== 0);
781 * ==========================================================================
782 * Range tree callbacks
783 * ==========================================================================
787 * Comparison function for the private size-ordered tree. Tree is sorted
788 * by size, larger sizes at the end of the tree.
791 metaslab_rangesize_compare(const void *x1
, const void *x2
)
793 const range_seg_t
*r1
= x1
;
794 const range_seg_t
*r2
= x2
;
795 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
796 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
798 if (rs_size1
< rs_size2
)
800 if (rs_size1
> rs_size2
)
803 if (r1
->rs_start
< r2
->rs_start
)
806 if (r1
->rs_start
> r2
->rs_start
)
813 * Create any block allocator specific components. The current allocators
814 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
817 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
819 metaslab_t
*msp
= arg
;
821 ASSERT3P(rt
->rt_arg
, ==, msp
);
822 ASSERT(msp
->ms_tree
== NULL
);
824 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
825 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
829 * Destroy the block allocator specific components.
832 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
834 metaslab_t
*msp
= arg
;
836 ASSERT3P(rt
->rt_arg
, ==, msp
);
837 ASSERT3P(msp
->ms_tree
, ==, rt
);
838 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
840 avl_destroy(&msp
->ms_size_tree
);
844 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
846 metaslab_t
*msp
= arg
;
848 ASSERT3P(rt
->rt_arg
, ==, msp
);
849 ASSERT3P(msp
->ms_tree
, ==, rt
);
850 VERIFY(!msp
->ms_condensing
);
851 avl_add(&msp
->ms_size_tree
, rs
);
855 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
857 metaslab_t
*msp
= arg
;
859 ASSERT3P(rt
->rt_arg
, ==, msp
);
860 ASSERT3P(msp
->ms_tree
, ==, rt
);
861 VERIFY(!msp
->ms_condensing
);
862 avl_remove(&msp
->ms_size_tree
, rs
);
866 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
868 metaslab_t
*msp
= arg
;
870 ASSERT3P(rt
->rt_arg
, ==, msp
);
871 ASSERT3P(msp
->ms_tree
, ==, rt
);
874 * Normally one would walk the tree freeing nodes along the way.
875 * Since the nodes are shared with the range trees we can avoid
876 * walking all nodes and just reinitialize the avl tree. The nodes
877 * will be freed by the range tree, so we don't want to free them here.
879 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
880 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
883 static range_tree_ops_t metaslab_rt_ops
= {
892 * ==========================================================================
893 * Metaslab block operations
894 * ==========================================================================
898 * Return the maximum contiguous segment within the metaslab.
901 metaslab_block_maxsize(metaslab_t
*msp
)
903 avl_tree_t
*t
= &msp
->ms_size_tree
;
906 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
909 return (rs
->rs_end
- rs
->rs_start
);
913 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
)
916 range_tree_t
*rt
= msp
->ms_tree
;
918 VERIFY(!msp
->ms_condensing
);
920 start
= msp
->ms_ops
->msop_alloc(msp
, size
);
921 if (start
!= -1ULL) {
922 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
924 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
925 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
926 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
927 range_tree_remove(rt
, start
, size
);
933 * ==========================================================================
934 * Common allocator routines
935 * ==========================================================================
938 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
939 defined(WITH_DF_BLOCK_ALLOCATOR) || \
940 defined(WITH_CF_BLOCK_ALLOCATOR)
942 * This is a helper function that can be used by the allocator to find
943 * a suitable block to allocate. This will search the specified AVL
944 * tree looking for a block that matches the specified criteria.
947 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
950 range_seg_t
*rs
, rsearch
;
953 rsearch
.rs_start
= *cursor
;
954 rsearch
.rs_end
= *cursor
+ size
;
956 rs
= avl_find(t
, &rsearch
, &where
);
958 rs
= avl_nearest(t
, where
, AVL_AFTER
);
961 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
963 if (offset
+ size
<= rs
->rs_end
) {
964 *cursor
= offset
+ size
;
967 rs
= AVL_NEXT(t
, rs
);
971 * If we know we've searched the whole map (*cursor == 0), give up.
972 * Otherwise, reset the cursor to the beginning and try again.
978 return (metaslab_block_picker(t
, cursor
, size
, align
));
980 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
982 #if defined(WITH_FF_BLOCK_ALLOCATOR)
984 * ==========================================================================
985 * The first-fit block allocator
986 * ==========================================================================
989 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
992 * Find the largest power of 2 block size that evenly divides the
993 * requested size. This is used to try to allocate blocks with similar
994 * alignment from the same area of the metaslab (i.e. same cursor
995 * bucket) but it does not guarantee that other allocations sizes
996 * may exist in the same region.
998 uint64_t align
= size
& -size
;
999 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1000 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1002 return (metaslab_block_picker(t
, cursor
, size
, align
));
1005 static metaslab_ops_t metaslab_ff_ops
= {
1009 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
1010 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1012 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1014 * ==========================================================================
1015 * Dynamic block allocator -
1016 * Uses the first fit allocation scheme until space get low and then
1017 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1018 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1019 * ==========================================================================
1022 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1025 * Find the largest power of 2 block size that evenly divides the
1026 * requested size. This is used to try to allocate blocks with similar
1027 * alignment from the same area of the metaslab (i.e. same cursor
1028 * bucket) but it does not guarantee that other allocations sizes
1029 * may exist in the same region.
1031 uint64_t align
= size
& -size
;
1032 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1033 range_tree_t
*rt
= msp
->ms_tree
;
1034 avl_tree_t
*t
= &rt
->rt_root
;
1035 uint64_t max_size
= metaslab_block_maxsize(msp
);
1036 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1038 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1039 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1041 if (max_size
< size
)
1045 * If we're running low on space switch to using the size
1046 * sorted AVL tree (best-fit).
1048 if (max_size
< metaslab_df_alloc_threshold
||
1049 free_pct
< metaslab_df_free_pct
) {
1050 t
= &msp
->ms_size_tree
;
1054 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1057 static metaslab_ops_t metaslab_df_ops
= {
1061 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1062 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1064 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1066 * ==========================================================================
1067 * Cursor fit block allocator -
1068 * Select the largest region in the metaslab, set the cursor to the beginning
1069 * of the range and the cursor_end to the end of the range. As allocations
1070 * are made advance the cursor. Continue allocating from the cursor until
1071 * the range is exhausted and then find a new range.
1072 * ==========================================================================
1075 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1077 range_tree_t
*rt
= msp
->ms_tree
;
1078 avl_tree_t
*t
= &msp
->ms_size_tree
;
1079 uint64_t *cursor
= &msp
->ms_lbas
[0];
1080 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1081 uint64_t offset
= 0;
1083 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1084 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1086 ASSERT3U(*cursor_end
, >=, *cursor
);
1088 if ((*cursor
+ size
) > *cursor_end
) {
1091 rs
= avl_last(&msp
->ms_size_tree
);
1092 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1095 *cursor
= rs
->rs_start
;
1096 *cursor_end
= rs
->rs_end
;
1105 static metaslab_ops_t metaslab_cf_ops
= {
1109 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1110 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1112 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1114 * ==========================================================================
1115 * New dynamic fit allocator -
1116 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1117 * contiguous blocks. If no region is found then just use the largest segment
1119 * ==========================================================================
1123 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1124 * to request from the allocator.
1126 uint64_t metaslab_ndf_clump_shift
= 4;
1129 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1131 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1133 range_seg_t
*rs
, rsearch
;
1134 uint64_t hbit
= highbit64(size
);
1135 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1136 uint64_t max_size
= metaslab_block_maxsize(msp
);
1138 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1139 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1141 if (max_size
< size
)
1144 rsearch
.rs_start
= *cursor
;
1145 rsearch
.rs_end
= *cursor
+ size
;
1147 rs
= avl_find(t
, &rsearch
, &where
);
1148 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1149 t
= &msp
->ms_size_tree
;
1151 rsearch
.rs_start
= 0;
1152 rsearch
.rs_end
= MIN(max_size
,
1153 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1154 rs
= avl_find(t
, &rsearch
, &where
);
1156 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1160 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1161 *cursor
= rs
->rs_start
+ size
;
1162 return (rs
->rs_start
);
1167 static metaslab_ops_t metaslab_ndf_ops
= {
1171 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1172 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1176 * ==========================================================================
1178 * ==========================================================================
1182 * Wait for any in-progress metaslab loads to complete.
1185 metaslab_load_wait(metaslab_t
*msp
)
1187 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1189 while (msp
->ms_loading
) {
1190 ASSERT(!msp
->ms_loaded
);
1191 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1196 metaslab_load(metaslab_t
*msp
)
1201 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1202 ASSERT(!msp
->ms_loaded
);
1203 ASSERT(!msp
->ms_loading
);
1205 msp
->ms_loading
= B_TRUE
;
1208 * If the space map has not been allocated yet, then treat
1209 * all the space in the metaslab as free and add it to the
1212 if (msp
->ms_sm
!= NULL
)
1213 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1215 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1217 msp
->ms_loaded
= (error
== 0);
1218 msp
->ms_loading
= B_FALSE
;
1220 if (msp
->ms_loaded
) {
1221 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1222 range_tree_walk(msp
->ms_defertree
[t
],
1223 range_tree_remove
, msp
->ms_tree
);
1226 cv_broadcast(&msp
->ms_load_cv
);
1231 metaslab_unload(metaslab_t
*msp
)
1233 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1234 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1235 msp
->ms_loaded
= B_FALSE
;
1236 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1240 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1243 vdev_t
*vd
= mg
->mg_vd
;
1244 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1248 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1249 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1250 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1252 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1253 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1256 * We only open space map objects that already exist. All others
1257 * will be opened when we finally allocate an object for it.
1260 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1261 ms
->ms_size
, vd
->vdev_ashift
, &ms
->ms_lock
);
1264 kmem_free(ms
, sizeof (metaslab_t
));
1268 ASSERT(ms
->ms_sm
!= NULL
);
1272 * We create the main range tree here, but we don't create the
1273 * alloctree and freetree until metaslab_sync_done(). This serves
1274 * two purposes: it allows metaslab_sync_done() to detect the
1275 * addition of new space; and for debugging, it ensures that we'd
1276 * data fault on any attempt to use this metaslab before it's ready.
1278 ms
->ms_tree
= range_tree_create(&metaslab_rt_ops
, ms
, &ms
->ms_lock
);
1279 metaslab_group_add(mg
, ms
);
1281 ms
->ms_fragmentation
= metaslab_fragmentation(ms
);
1282 ms
->ms_ops
= mg
->mg_class
->mc_ops
;
1285 * If we're opening an existing pool (txg == 0) or creating
1286 * a new one (txg == TXG_INITIAL), all space is available now.
1287 * If we're adding space to an existing pool, the new space
1288 * does not become available until after this txg has synced.
1290 if (txg
<= TXG_INITIAL
)
1291 metaslab_sync_done(ms
, 0);
1294 * If metaslab_debug_load is set and we're initializing a metaslab
1295 * that has an allocated space_map object then load the its space
1296 * map so that can verify frees.
1298 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1299 mutex_enter(&ms
->ms_lock
);
1300 VERIFY0(metaslab_load(ms
));
1301 mutex_exit(&ms
->ms_lock
);
1305 vdev_dirty(vd
, 0, NULL
, txg
);
1306 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1315 metaslab_fini(metaslab_t
*msp
)
1319 metaslab_group_t
*mg
= msp
->ms_group
;
1321 metaslab_group_remove(mg
, msp
);
1323 mutex_enter(&msp
->ms_lock
);
1325 VERIFY(msp
->ms_group
== NULL
);
1326 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1328 space_map_close(msp
->ms_sm
);
1330 metaslab_unload(msp
);
1331 range_tree_destroy(msp
->ms_tree
);
1333 for (t
= 0; t
< TXG_SIZE
; t
++) {
1334 range_tree_destroy(msp
->ms_alloctree
[t
]);
1335 range_tree_destroy(msp
->ms_freetree
[t
]);
1338 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1339 range_tree_destroy(msp
->ms_defertree
[t
]);
1342 ASSERT0(msp
->ms_deferspace
);
1344 mutex_exit(&msp
->ms_lock
);
1345 cv_destroy(&msp
->ms_load_cv
);
1346 mutex_destroy(&msp
->ms_lock
);
1348 kmem_free(msp
, sizeof (metaslab_t
));
1351 #define FRAGMENTATION_TABLE_SIZE 17
1354 * This table defines a segment size based fragmentation metric that will
1355 * allow each metaslab to derive its own fragmentation value. This is done
1356 * by calculating the space in each bucket of the spacemap histogram and
1357 * multiplying that by the fragmetation metric in this table. Doing
1358 * this for all buckets and dividing it by the total amount of free
1359 * space in this metaslab (i.e. the total free space in all buckets) gives
1360 * us the fragmentation metric. This means that a high fragmentation metric
1361 * equates to most of the free space being comprised of small segments.
1362 * Conversely, if the metric is low, then most of the free space is in
1363 * large segments. A 10% change in fragmentation equates to approximately
1364 * double the number of segments.
1366 * This table defines 0% fragmented space using 16MB segments. Testing has
1367 * shown that segments that are greater than or equal to 16MB do not suffer
1368 * from drastic performance problems. Using this value, we derive the rest
1369 * of the table. Since the fragmentation value is never stored on disk, it
1370 * is possible to change these calculations in the future.
1372 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1392 * Calclate the metaslab's fragmentation metric. A return value
1393 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1394 * not support this metric. Otherwise, the return value should be in the
1398 metaslab_fragmentation(metaslab_t
*msp
)
1400 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1401 uint64_t fragmentation
= 0;
1403 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1404 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1407 if (!feature_enabled
)
1408 return (ZFS_FRAG_INVALID
);
1411 * A null space map means that the entire metaslab is free
1412 * and thus is not fragmented.
1414 if (msp
->ms_sm
== NULL
)
1418 * If this metaslab's space_map has not been upgraded, flag it
1419 * so that we upgrade next time we encounter it.
1421 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1422 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1424 if (spa_writeable(vd
->vdev_spa
)) {
1425 uint64_t txg
= spa_syncing_txg(spa
);
1427 msp
->ms_condense_wanted
= B_TRUE
;
1428 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1429 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1430 "msp %p, vd %p", txg
, msp
, vd
);
1432 return (ZFS_FRAG_INVALID
);
1435 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1437 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1438 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1439 FRAGMENTATION_TABLE_SIZE
- 1);
1441 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1444 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1447 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1448 fragmentation
+= space
* zfs_frag_table
[idx
];
1452 fragmentation
/= total
;
1453 ASSERT3U(fragmentation
, <=, 100);
1454 return (fragmentation
);
1458 * Compute a weight -- a selection preference value -- for the given metaslab.
1459 * This is based on the amount of free space, the level of fragmentation,
1460 * the LBA range, and whether the metaslab is loaded.
1463 metaslab_weight(metaslab_t
*msp
)
1465 metaslab_group_t
*mg
= msp
->ms_group
;
1466 vdev_t
*vd
= mg
->mg_vd
;
1467 uint64_t weight
, space
;
1469 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1472 * This vdev is in the process of being removed so there is nothing
1473 * for us to do here.
1475 if (vd
->vdev_removing
) {
1476 ASSERT0(space_map_allocated(msp
->ms_sm
));
1477 ASSERT0(vd
->vdev_ms_shift
);
1482 * The baseline weight is the metaslab's free space.
1484 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1486 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1487 if (metaslab_fragmentation_factor_enabled
&&
1488 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1490 * Use the fragmentation information to inversely scale
1491 * down the baseline weight. We need to ensure that we
1492 * don't exclude this metaslab completely when it's 100%
1493 * fragmented. To avoid this we reduce the fragmented value
1496 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1499 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1500 * this metaslab again. The fragmentation metric may have
1501 * decreased the space to something smaller than
1502 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1503 * so that we can consume any remaining space.
1505 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1506 space
= SPA_MINBLOCKSIZE
;
1511 * Modern disks have uniform bit density and constant angular velocity.
1512 * Therefore, the outer recording zones are faster (higher bandwidth)
1513 * than the inner zones by the ratio of outer to inner track diameter,
1514 * which is typically around 2:1. We account for this by assigning
1515 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1516 * In effect, this means that we'll select the metaslab with the most
1517 * free bandwidth rather than simply the one with the most free space.
1519 if (!vd
->vdev_nonrot
&& metaslab_lba_weighting_enabled
) {
1520 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1521 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1525 * If this metaslab is one we're actively using, adjust its
1526 * weight to make it preferable to any inactive metaslab so
1527 * we'll polish it off. If the fragmentation on this metaslab
1528 * has exceed our threshold, then don't mark it active.
1530 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1531 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1532 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1539 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1541 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1543 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1544 metaslab_load_wait(msp
);
1545 if (!msp
->ms_loaded
) {
1546 int error
= metaslab_load(msp
);
1548 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1553 metaslab_group_sort(msp
->ms_group
, msp
,
1554 msp
->ms_weight
| activation_weight
);
1556 ASSERT(msp
->ms_loaded
);
1557 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1563 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
1566 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1567 * this metaslab again. In that case, it had better be empty,
1568 * or we would be leaving space on the table.
1570 ASSERT(size
>= SPA_MINBLOCKSIZE
|| range_tree_space(msp
->ms_tree
) == 0);
1571 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
1572 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1576 metaslab_preload(void *arg
)
1578 metaslab_t
*msp
= arg
;
1579 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1580 fstrans_cookie_t cookie
= spl_fstrans_mark();
1582 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1584 mutex_enter(&msp
->ms_lock
);
1585 metaslab_load_wait(msp
);
1586 if (!msp
->ms_loaded
)
1587 (void) metaslab_load(msp
);
1590 * Set the ms_access_txg value so that we don't unload it right away.
1592 msp
->ms_access_txg
= spa_syncing_txg(spa
) + metaslab_unload_delay
+ 1;
1593 mutex_exit(&msp
->ms_lock
);
1594 spl_fstrans_unmark(cookie
);
1598 metaslab_group_preload(metaslab_group_t
*mg
)
1600 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1602 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1605 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
1606 taskq_wait_outstanding(mg
->mg_taskq
, 0);
1610 mutex_enter(&mg
->mg_lock
);
1612 * Load the next potential metaslabs
1615 while (msp
!= NULL
) {
1616 metaslab_t
*msp_next
= AVL_NEXT(t
, msp
);
1619 * We preload only the maximum number of metaslabs specified
1620 * by metaslab_preload_limit. If a metaslab is being forced
1621 * to condense then we preload it too. This will ensure
1622 * that force condensing happens in the next txg.
1624 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
1630 * We must drop the metaslab group lock here to preserve
1631 * lock ordering with the ms_lock (when grabbing both
1632 * the mg_lock and the ms_lock, the ms_lock must be taken
1633 * first). As a result, it is possible that the ordering
1634 * of the metaslabs within the avl tree may change before
1635 * we reacquire the lock. The metaslab cannot be removed from
1636 * the tree while we're in syncing context so it is safe to
1637 * drop the mg_lock here. If the metaslabs are reordered
1638 * nothing will break -- we just may end up loading a
1639 * less than optimal one.
1641 mutex_exit(&mg
->mg_lock
);
1642 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
1643 msp
, TQ_SLEEP
) != 0);
1644 mutex_enter(&mg
->mg_lock
);
1647 mutex_exit(&mg
->mg_lock
);
1651 * Determine if the space map's on-disk footprint is past our tolerance
1652 * for inefficiency. We would like to use the following criteria to make
1655 * 1. The size of the space map object should not dramatically increase as a
1656 * result of writing out the free space range tree.
1658 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1659 * times the size than the free space range tree representation
1660 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1662 * 3. The on-disk size of the space map should actually decrease.
1664 * Checking the first condition is tricky since we don't want to walk
1665 * the entire AVL tree calculating the estimated on-disk size. Instead we
1666 * use the size-ordered range tree in the metaslab and calculate the
1667 * size required to write out the largest segment in our free tree. If the
1668 * size required to represent that segment on disk is larger than the space
1669 * map object then we avoid condensing this map.
1671 * To determine the second criterion we use a best-case estimate and assume
1672 * each segment can be represented on-disk as a single 64-bit entry. We refer
1673 * to this best-case estimate as the space map's minimal form.
1675 * Unfortunately, we cannot compute the on-disk size of the space map in this
1676 * context because we cannot accurately compute the effects of compression, etc.
1677 * Instead, we apply the heuristic described in the block comment for
1678 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1679 * is greater than a threshold number of blocks.
1682 metaslab_should_condense(metaslab_t
*msp
)
1684 space_map_t
*sm
= msp
->ms_sm
;
1686 uint64_t size
, entries
, segsz
, object_size
, optimal_size
, record_size
;
1687 dmu_object_info_t doi
;
1688 uint64_t vdev_blocksize
= 1 << msp
->ms_group
->mg_vd
->vdev_ashift
;
1690 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1691 ASSERT(msp
->ms_loaded
);
1694 * Use the ms_size_tree range tree, which is ordered by size, to
1695 * obtain the largest segment in the free tree. We always condense
1696 * metaslabs that are empty and metaslabs for which a condense
1697 * request has been made.
1699 rs
= avl_last(&msp
->ms_size_tree
);
1700 if (rs
== NULL
|| msp
->ms_condense_wanted
)
1704 * Calculate the number of 64-bit entries this segment would
1705 * require when written to disk. If this single segment would be
1706 * larger on-disk than the entire current on-disk structure, then
1707 * clearly condensing will increase the on-disk structure size.
1709 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
1710 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
1711 segsz
= entries
* sizeof (uint64_t);
1713 optimal_size
= sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
);
1714 object_size
= space_map_length(msp
->ms_sm
);
1716 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
1717 record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
1719 return (segsz
<= object_size
&&
1720 object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
1721 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
1725 * Condense the on-disk space map representation to its minimized form.
1726 * The minimized form consists of a small number of allocations followed by
1727 * the entries of the free range tree.
1730 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
1732 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1733 range_tree_t
*freetree
= msp
->ms_freetree
[txg
& TXG_MASK
];
1734 range_tree_t
*condense_tree
;
1735 space_map_t
*sm
= msp
->ms_sm
;
1738 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1739 ASSERT3U(spa_sync_pass(spa
), ==, 1);
1740 ASSERT(msp
->ms_loaded
);
1743 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
1744 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
1745 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
1746 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
1747 space_map_length(msp
->ms_sm
), avl_numnodes(&msp
->ms_tree
->rt_root
),
1748 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
1750 msp
->ms_condense_wanted
= B_FALSE
;
1753 * Create an range tree that is 100% allocated. We remove segments
1754 * that have been freed in this txg, any deferred frees that exist,
1755 * and any allocation in the future. Removing segments should be
1756 * a relatively inexpensive operation since we expect these trees to
1757 * have a small number of nodes.
1759 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
1760 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
1763 * Remove what's been freed in this txg from the condense_tree.
1764 * Since we're in sync_pass 1, we know that all the frees from
1765 * this txg are in the freetree.
1767 range_tree_walk(freetree
, range_tree_remove
, condense_tree
);
1769 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1770 range_tree_walk(msp
->ms_defertree
[t
],
1771 range_tree_remove
, condense_tree
);
1774 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1775 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
1776 range_tree_remove
, condense_tree
);
1780 * We're about to drop the metaslab's lock thus allowing
1781 * other consumers to change it's content. Set the
1782 * metaslab's ms_condensing flag to ensure that
1783 * allocations on this metaslab do not occur while we're
1784 * in the middle of committing it to disk. This is only critical
1785 * for the ms_tree as all other range trees use per txg
1786 * views of their content.
1788 msp
->ms_condensing
= B_TRUE
;
1790 mutex_exit(&msp
->ms_lock
);
1791 space_map_truncate(sm
, tx
);
1792 mutex_enter(&msp
->ms_lock
);
1795 * While we would ideally like to create a space_map representation
1796 * that consists only of allocation records, doing so can be
1797 * prohibitively expensive because the in-core free tree can be
1798 * large, and therefore computationally expensive to subtract
1799 * from the condense_tree. Instead we sync out two trees, a cheap
1800 * allocation only tree followed by the in-core free tree. While not
1801 * optimal, this is typically close to optimal, and much cheaper to
1804 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
1805 range_tree_vacate(condense_tree
, NULL
, NULL
);
1806 range_tree_destroy(condense_tree
);
1808 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
1809 msp
->ms_condensing
= B_FALSE
;
1813 * Write a metaslab to disk in the context of the specified transaction group.
1816 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
1818 metaslab_group_t
*mg
= msp
->ms_group
;
1819 vdev_t
*vd
= mg
->mg_vd
;
1820 spa_t
*spa
= vd
->vdev_spa
;
1821 objset_t
*mos
= spa_meta_objset(spa
);
1822 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
1823 range_tree_t
**freetree
= &msp
->ms_freetree
[txg
& TXG_MASK
];
1824 range_tree_t
**freed_tree
=
1825 &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1827 uint64_t object
= space_map_object(msp
->ms_sm
);
1829 ASSERT(!vd
->vdev_ishole
);
1832 * This metaslab has just been added so there's no work to do now.
1834 if (*freetree
== NULL
) {
1835 ASSERT3P(alloctree
, ==, NULL
);
1839 ASSERT3P(alloctree
, !=, NULL
);
1840 ASSERT3P(*freetree
, !=, NULL
);
1841 ASSERT3P(*freed_tree
, !=, NULL
);
1844 * Normally, we don't want to process a metaslab if there
1845 * are no allocations or frees to perform. However, if the metaslab
1846 * is being forced to condense we need to let it through.
1848 if (range_tree_space(alloctree
) == 0 &&
1849 range_tree_space(*freetree
) == 0 &&
1850 !msp
->ms_condense_wanted
)
1854 * The only state that can actually be changing concurrently with
1855 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1856 * be modifying this txg's alloctree, freetree, freed_tree, or
1857 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1858 * space_map ASSERTs. We drop it whenever we call into the DMU,
1859 * because the DMU can call down to us (e.g. via zio_free()) at
1863 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
1865 if (msp
->ms_sm
== NULL
) {
1866 uint64_t new_object
;
1868 new_object
= space_map_alloc(mos
, tx
);
1869 VERIFY3U(new_object
, !=, 0);
1871 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
1872 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
1874 ASSERT(msp
->ms_sm
!= NULL
);
1877 mutex_enter(&msp
->ms_lock
);
1880 * Note: metaslab_condense() clears the space_map's histogram.
1881 * Therefore we muse verify and remove this histogram before
1884 metaslab_group_histogram_verify(mg
);
1885 metaslab_class_histogram_verify(mg
->mg_class
);
1886 metaslab_group_histogram_remove(mg
, msp
);
1888 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
1889 metaslab_should_condense(msp
)) {
1890 metaslab_condense(msp
, txg
, tx
);
1892 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
1893 space_map_write(msp
->ms_sm
, *freetree
, SM_FREE
, tx
);
1896 if (msp
->ms_loaded
) {
1898 * When the space map is loaded, we have an accruate
1899 * histogram in the range tree. This gives us an opportunity
1900 * to bring the space map's histogram up-to-date so we clear
1901 * it first before updating it.
1903 space_map_histogram_clear(msp
->ms_sm
);
1904 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
1907 * Since the space map is not loaded we simply update the
1908 * exisiting histogram with what was freed in this txg. This
1909 * means that the on-disk histogram may not have an accurate
1910 * view of the free space but it's close enough to allow
1911 * us to make allocation decisions.
1913 space_map_histogram_add(msp
->ms_sm
, *freetree
, tx
);
1915 metaslab_group_histogram_add(mg
, msp
);
1916 metaslab_group_histogram_verify(mg
);
1917 metaslab_class_histogram_verify(mg
->mg_class
);
1920 * For sync pass 1, we avoid traversing this txg's free range tree
1921 * and instead will just swap the pointers for freetree and
1922 * freed_tree. We can safely do this since the freed_tree is
1923 * guaranteed to be empty on the initial pass.
1925 if (spa_sync_pass(spa
) == 1) {
1926 range_tree_swap(freetree
, freed_tree
);
1928 range_tree_vacate(*freetree
, range_tree_add
, *freed_tree
);
1930 range_tree_vacate(alloctree
, NULL
, NULL
);
1932 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1933 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1935 mutex_exit(&msp
->ms_lock
);
1937 if (object
!= space_map_object(msp
->ms_sm
)) {
1938 object
= space_map_object(msp
->ms_sm
);
1939 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
1940 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
1946 * Called after a transaction group has completely synced to mark
1947 * all of the metaslab's free space as usable.
1950 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1952 metaslab_group_t
*mg
= msp
->ms_group
;
1953 vdev_t
*vd
= mg
->mg_vd
;
1954 range_tree_t
**freed_tree
;
1955 range_tree_t
**defer_tree
;
1956 int64_t alloc_delta
, defer_delta
;
1959 ASSERT(!vd
->vdev_ishole
);
1961 mutex_enter(&msp
->ms_lock
);
1964 * If this metaslab is just becoming available, initialize its
1965 * alloctrees, freetrees, and defertree and add its capacity to
1968 if (msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
] == NULL
) {
1969 for (t
= 0; t
< TXG_SIZE
; t
++) {
1970 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
1971 ASSERT(msp
->ms_freetree
[t
] == NULL
);
1973 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
1975 msp
->ms_freetree
[t
] = range_tree_create(NULL
, msp
,
1979 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1980 ASSERT(msp
->ms_defertree
[t
] == NULL
);
1982 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
1986 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
1989 freed_tree
= &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1990 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
1992 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
1993 defer_delta
= range_tree_space(*freed_tree
) -
1994 range_tree_space(*defer_tree
);
1996 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1998 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1999 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
2002 * If there's a metaslab_load() in progress, wait for it to complete
2003 * so that we have a consistent view of the in-core space map.
2005 metaslab_load_wait(msp
);
2008 * Move the frees from the defer_tree back to the free
2009 * range tree (if it's loaded). Swap the freed_tree and the
2010 * defer_tree -- this is safe to do because we've just emptied out
2013 range_tree_vacate(*defer_tree
,
2014 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
2015 range_tree_swap(freed_tree
, defer_tree
);
2017 space_map_update(msp
->ms_sm
);
2019 msp
->ms_deferspace
+= defer_delta
;
2020 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2021 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2022 if (msp
->ms_deferspace
!= 0) {
2024 * Keep syncing this metaslab until all deferred frees
2025 * are back in circulation.
2027 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2030 if (msp
->ms_loaded
&& msp
->ms_access_txg
< txg
) {
2031 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2032 VERIFY0(range_tree_space(
2033 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
2036 if (!metaslab_debug_unload
)
2037 metaslab_unload(msp
);
2040 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
2041 mutex_exit(&msp
->ms_lock
);
2045 metaslab_sync_reassess(metaslab_group_t
*mg
)
2047 metaslab_group_alloc_update(mg
);
2048 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2051 * Preload the next potential metaslabs
2053 metaslab_group_preload(mg
);
2057 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2059 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2060 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2061 uint64_t start
= msp
->ms_id
;
2063 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2064 return (1ULL << 63);
2067 return ((start
- offset
) << ms_shift
);
2069 return ((offset
- start
) << ms_shift
);
2074 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
2075 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2077 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2078 metaslab_t
*msp
= NULL
;
2079 uint64_t offset
= -1ULL;
2080 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2081 uint64_t activation_weight
;
2082 uint64_t target_distance
;
2085 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2086 for (i
= 0; i
< d
; i
++) {
2087 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2088 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2094 boolean_t was_active
;
2096 mutex_enter(&mg
->mg_lock
);
2097 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
2098 if (msp
->ms_weight
< asize
) {
2099 spa_dbgmsg(spa
, "%s: failed to meet weight "
2100 "requirement: vdev %llu, txg %llu, mg %p, "
2101 "msp %p, psize %llu, asize %llu, "
2102 "weight %llu", spa_name(spa
),
2103 mg
->mg_vd
->vdev_id
, txg
,
2104 mg
, msp
, psize
, asize
, msp
->ms_weight
);
2105 mutex_exit(&mg
->mg_lock
);
2110 * If the selected metaslab is condensing, skip it.
2112 if (msp
->ms_condensing
)
2115 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2116 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2119 target_distance
= min_distance
+
2120 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2123 for (i
= 0; i
< d
; i
++)
2124 if (metaslab_distance(msp
, &dva
[i
]) <
2130 mutex_exit(&mg
->mg_lock
);
2134 mutex_enter(&msp
->ms_lock
);
2137 * Ensure that the metaslab we have selected is still
2138 * capable of handling our request. It's possible that
2139 * another thread may have changed the weight while we
2140 * were blocked on the metaslab lock.
2142 if (msp
->ms_weight
< asize
|| (was_active
&&
2143 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
2144 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
2145 mutex_exit(&msp
->ms_lock
);
2149 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2150 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2151 metaslab_passivate(msp
,
2152 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2153 mutex_exit(&msp
->ms_lock
);
2157 if (metaslab_activate(msp
, activation_weight
) != 0) {
2158 mutex_exit(&msp
->ms_lock
);
2163 * If this metaslab is currently condensing then pick again as
2164 * we can't manipulate this metaslab until it's committed
2167 if (msp
->ms_condensing
) {
2168 mutex_exit(&msp
->ms_lock
);
2172 if ((offset
= metaslab_block_alloc(msp
, asize
)) != -1ULL)
2175 metaslab_passivate(msp
, metaslab_block_maxsize(msp
));
2176 mutex_exit(&msp
->ms_lock
);
2179 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2180 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2182 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, asize
);
2183 msp
->ms_access_txg
= txg
+ metaslab_unload_delay
;
2185 mutex_exit(&msp
->ms_lock
);
2191 * Allocate a block for the specified i/o.
2194 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2195 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
2197 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
2201 int zio_lock
= B_FALSE
;
2202 boolean_t allocatable
;
2203 uint64_t offset
= -1ULL;
2207 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2210 * For testing, make some blocks above a certain size be gang blocks.
2212 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
2213 return (SET_ERROR(ENOSPC
));
2216 * Start at the rotor and loop through all mgs until we find something.
2217 * Note that there's no locking on mc_rotor or mc_aliquot because
2218 * nothing actually breaks if we miss a few updates -- we just won't
2219 * allocate quite as evenly. It all balances out over time.
2221 * If we are doing ditto or log blocks, try to spread them across
2222 * consecutive vdevs. If we're forced to reuse a vdev before we've
2223 * allocated all of our ditto blocks, then try and spread them out on
2224 * that vdev as much as possible. If it turns out to not be possible,
2225 * gradually lower our standards until anything becomes acceptable.
2226 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2227 * gives us hope of containing our fault domains to something we're
2228 * able to reason about. Otherwise, any two top-level vdev failures
2229 * will guarantee the loss of data. With consecutive allocation,
2230 * only two adjacent top-level vdev failures will result in data loss.
2232 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2233 * ourselves on the same vdev as our gang block header. That
2234 * way, we can hope for locality in vdev_cache, plus it makes our
2235 * fault domains something tractable.
2238 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
2241 * It's possible the vdev we're using as the hint no
2242 * longer exists (i.e. removed). Consult the rotor when
2248 if (flags
& METASLAB_HINTBP_AVOID
&&
2249 mg
->mg_next
!= NULL
)
2254 } else if (d
!= 0) {
2255 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
2256 mg
= vd
->vdev_mg
->mg_next
;
2257 } else if (flags
& METASLAB_FASTWRITE
) {
2258 mg
= fast_mg
= mc
->mc_rotor
;
2261 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
2262 mg
->mg_vd
->vdev_pending_fastwrite
)
2264 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
2271 * If the hint put us into the wrong metaslab class, or into a
2272 * metaslab group that has been passivated, just follow the rotor.
2274 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
2281 ASSERT(mg
->mg_activation_count
== 1);
2286 * Don't allocate from faulted devices.
2289 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
2290 allocatable
= vdev_allocatable(vd
);
2291 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
2293 allocatable
= vdev_allocatable(vd
);
2297 * Determine if the selected metaslab group is eligible
2298 * for allocations. If we're ganging or have requested
2299 * an allocation for the smallest gang block size
2300 * then we don't want to avoid allocating to the this
2301 * metaslab group. If we're in this condition we should
2302 * try to allocate from any device possible so that we
2303 * don't inadvertently return ENOSPC and suspend the pool
2304 * even though space is still available.
2306 if (allocatable
&& CAN_FASTGANG(flags
) &&
2307 psize
> SPA_GANGBLOCKSIZE
)
2308 allocatable
= metaslab_group_allocatable(mg
);
2314 * Avoid writing single-copy data to a failing vdev
2315 * unless the user instructs us that it is okay.
2317 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
2318 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
2319 d
== 0 && dshift
== 3 && vd
->vdev_children
== 0) {
2324 ASSERT(mg
->mg_class
== mc
);
2326 distance
= vd
->vdev_asize
>> dshift
;
2327 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
2332 asize
= vdev_psize_to_asize(vd
, psize
);
2333 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
2335 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
2337 if (offset
!= -1ULL) {
2339 * If we've just selected this metaslab group,
2340 * figure out whether the corresponding vdev is
2341 * over- or under-used relative to the pool,
2342 * and set an allocation bias to even it out.
2344 * Bias is also used to compensate for unequally
2345 * sized vdevs so that space is allocated fairly.
2347 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
2348 vdev_stat_t
*vs
= &vd
->vdev_stat
;
2349 int64_t vs_free
= vs
->vs_space
- vs
->vs_alloc
;
2350 int64_t mc_free
= mc
->mc_space
- mc
->mc_alloc
;
2354 * Calculate how much more or less we should
2355 * try to allocate from this device during
2356 * this iteration around the rotor.
2358 * This basically introduces a zero-centered
2359 * bias towards the devices with the most
2360 * free space, while compensating for vdev
2364 * vdev V1 = 16M/128M
2365 * vdev V2 = 16M/128M
2366 * ratio(V1) = 100% ratio(V2) = 100%
2368 * vdev V1 = 16M/128M
2369 * vdev V2 = 64M/128M
2370 * ratio(V1) = 127% ratio(V2) = 72%
2372 * vdev V1 = 16M/128M
2373 * vdev V2 = 64M/512M
2374 * ratio(V1) = 40% ratio(V2) = 160%
2376 ratio
= (vs_free
* mc
->mc_alloc_groups
* 100) /
2378 mg
->mg_bias
= ((ratio
- 100) *
2379 (int64_t)mg
->mg_aliquot
) / 100;
2380 } else if (!metaslab_bias_enabled
) {
2384 if ((flags
& METASLAB_FASTWRITE
) ||
2385 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
2386 mg
->mg_aliquot
+ mg
->mg_bias
) {
2387 mc
->mc_rotor
= mg
->mg_next
;
2391 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
2392 DVA_SET_OFFSET(&dva
[d
], offset
);
2393 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
2394 DVA_SET_ASIZE(&dva
[d
], asize
);
2396 if (flags
& METASLAB_FASTWRITE
) {
2397 atomic_add_64(&vd
->vdev_pending_fastwrite
,
2404 mc
->mc_rotor
= mg
->mg_next
;
2406 } while ((mg
= mg
->mg_next
) != rotor
);
2410 ASSERT(dshift
< 64);
2414 if (!allocatable
&& !zio_lock
) {
2420 bzero(&dva
[d
], sizeof (dva_t
));
2422 return (SET_ERROR(ENOSPC
));
2426 * Free the block represented by DVA in the context of the specified
2427 * transaction group.
2430 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
2432 uint64_t vdev
= DVA_GET_VDEV(dva
);
2433 uint64_t offset
= DVA_GET_OFFSET(dva
);
2434 uint64_t size
= DVA_GET_ASIZE(dva
);
2438 if (txg
> spa_freeze_txg(spa
))
2441 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
2442 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
2443 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
2444 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
2445 (u_longlong_t
)size
);
2449 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2451 if (DVA_GET_GANG(dva
))
2452 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2454 mutex_enter(&msp
->ms_lock
);
2457 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
2460 VERIFY(!msp
->ms_condensing
);
2461 VERIFY3U(offset
, >=, msp
->ms_start
);
2462 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
2463 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
2465 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2466 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2467 range_tree_add(msp
->ms_tree
, offset
, size
);
2469 if (range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]) == 0)
2470 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2471 range_tree_add(msp
->ms_freetree
[txg
& TXG_MASK
],
2475 mutex_exit(&msp
->ms_lock
);
2479 * Intent log support: upon opening the pool after a crash, notify the SPA
2480 * of blocks that the intent log has allocated for immediate write, but
2481 * which are still considered free by the SPA because the last transaction
2482 * group didn't commit yet.
2485 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
2487 uint64_t vdev
= DVA_GET_VDEV(dva
);
2488 uint64_t offset
= DVA_GET_OFFSET(dva
);
2489 uint64_t size
= DVA_GET_ASIZE(dva
);
2494 ASSERT(DVA_IS_VALID(dva
));
2496 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2497 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
2498 return (SET_ERROR(ENXIO
));
2500 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2502 if (DVA_GET_GANG(dva
))
2503 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2505 mutex_enter(&msp
->ms_lock
);
2507 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
2508 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
2510 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
2511 error
= SET_ERROR(ENOENT
);
2513 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
2514 mutex_exit(&msp
->ms_lock
);
2518 VERIFY(!msp
->ms_condensing
);
2519 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2520 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2521 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
2522 range_tree_remove(msp
->ms_tree
, offset
, size
);
2524 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
2525 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2526 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2527 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
2530 mutex_exit(&msp
->ms_lock
);
2536 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
2537 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
2539 dva_t
*dva
= bp
->blk_dva
;
2540 dva_t
*hintdva
= hintbp
->blk_dva
;
2543 ASSERT(bp
->blk_birth
== 0);
2544 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
2546 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2548 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
2549 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2550 return (SET_ERROR(ENOSPC
));
2553 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
2554 ASSERT(BP_GET_NDVAS(bp
) == 0);
2555 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
2557 for (d
= 0; d
< ndvas
; d
++) {
2558 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
2561 for (d
--; d
>= 0; d
--) {
2562 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
2563 bzero(&dva
[d
], sizeof (dva_t
));
2565 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2570 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
2572 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2574 BP_SET_BIRTH(bp
, txg
, txg
);
2580 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
2582 const dva_t
*dva
= bp
->blk_dva
;
2583 int d
, ndvas
= BP_GET_NDVAS(bp
);
2585 ASSERT(!BP_IS_HOLE(bp
));
2586 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
2588 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
2590 for (d
= 0; d
< ndvas
; d
++)
2591 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
2593 spa_config_exit(spa
, SCL_FREE
, FTAG
);
2597 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
2599 const dva_t
*dva
= bp
->blk_dva
;
2600 int ndvas
= BP_GET_NDVAS(bp
);
2603 ASSERT(!BP_IS_HOLE(bp
));
2607 * First do a dry run to make sure all DVAs are claimable,
2608 * so we don't have to unwind from partial failures below.
2610 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
2614 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2616 for (d
= 0; d
< ndvas
; d
++)
2617 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
2620 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2622 ASSERT(error
== 0 || txg
== 0);
2628 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
2630 const dva_t
*dva
= bp
->blk_dva
;
2631 int ndvas
= BP_GET_NDVAS(bp
);
2632 uint64_t psize
= BP_GET_PSIZE(bp
);
2636 ASSERT(!BP_IS_HOLE(bp
));
2637 ASSERT(!BP_IS_EMBEDDED(bp
));
2640 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2642 for (d
= 0; d
< ndvas
; d
++) {
2643 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2645 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
2648 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2652 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
2654 const dva_t
*dva
= bp
->blk_dva
;
2655 int ndvas
= BP_GET_NDVAS(bp
);
2656 uint64_t psize
= BP_GET_PSIZE(bp
);
2660 ASSERT(!BP_IS_HOLE(bp
));
2661 ASSERT(!BP_IS_EMBEDDED(bp
));
2664 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2666 for (d
= 0; d
< ndvas
; d
++) {
2667 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2669 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
2670 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
2673 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2677 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
2681 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
2684 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2685 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2686 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
2687 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
2688 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
2689 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2690 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2693 range_tree_verify(msp
->ms_tree
, offset
, size
);
2695 for (j
= 0; j
< TXG_SIZE
; j
++)
2696 range_tree_verify(msp
->ms_freetree
[j
], offset
, size
);
2697 for (j
= 0; j
< TXG_DEFER_SIZE
; j
++)
2698 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
2700 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2703 #if defined(_KERNEL) && defined(HAVE_SPL)
2704 module_param(metaslab_aliquot
, ulong
, 0644);
2705 module_param(metaslab_debug_load
, int, 0644);
2706 module_param(metaslab_debug_unload
, int, 0644);
2707 module_param(metaslab_preload_enabled
, int, 0644);
2708 module_param(zfs_mg_noalloc_threshold
, int, 0644);
2709 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
2710 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
2711 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
2712 module_param(metaslab_lba_weighting_enabled
, int, 0644);
2713 module_param(metaslab_bias_enabled
, int, 0644);
2715 MODULE_PARM_DESC(metaslab_aliquot
,
2716 "allocation granularity (a.k.a. stripe size)");
2717 MODULE_PARM_DESC(metaslab_debug_load
,
2718 "load all metaslabs when pool is first opened");
2719 MODULE_PARM_DESC(metaslab_debug_unload
,
2720 "prevent metaslabs from being unloaded");
2721 MODULE_PARM_DESC(metaslab_preload_enabled
,
2722 "preload potential metaslabs during reassessment");
2724 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
2725 "percentage of free space for metaslab group to allow allocation");
2726 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
2727 "fragmentation for metaslab group to allow allocation");
2729 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
2730 "fragmentation for metaslab to allow allocation");
2731 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
2732 "use the fragmentation metric to prefer less fragmented metaslabs");
2733 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
2734 "prefer metaslabs with lower LBAs");
2735 MODULE_PARM_DESC(metaslab_bias_enabled
,
2736 "enable metaslab group biasing");
2737 #endif /* _KERNEL && HAVE_SPL */