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, 2014 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>
38 * Allow allocations to switch to gang blocks quickly. We do this to
39 * avoid having to load lots of space_maps in a given txg. There are,
40 * however, some cases where we want to avoid "fast" ganging and instead
41 * we want to do an exhaustive search of all metaslabs on this device.
42 * Currently we don't allow any gang, slog, or dump device related allocations
45 #define CAN_FASTGANG(flags) \
46 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
47 METASLAB_GANG_AVOID)))
49 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
50 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
51 #define METASLAB_ACTIVE_MASK \
52 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
54 uint64_t metaslab_aliquot
= 512ULL << 10;
55 uint64_t metaslab_gang_bang
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
58 * The in-core space map representation is more compact than its on-disk form.
59 * The zfs_condense_pct determines how much more compact the in-core
60 * space_map representation must be before we compact it on-disk.
61 * Values should be greater than or equal to 100.
63 int zfs_condense_pct
= 200;
66 * Condensing a metaslab is not guaranteed to actually reduce the amount of
67 * space used on disk. In particular, a space map uses data in increments of
68 * MAX(1 << ashift, SPACE_MAP_INITIAL_BLOCKSIZE), so a metaslab might use the
69 * same number of blocks after condensing. Since the goal of condensing is to
70 * reduce the number of IOPs required to read the space map, we only want to
71 * condense when we can be sure we will reduce the number of blocks used by the
72 * space map. Unfortunately, we cannot precisely compute whether or not this is
73 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
74 * we apply the following heuristic: do not condense a spacemap unless the
75 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
78 int zfs_metaslab_condense_block_threshold
= 4;
81 * The zfs_mg_noalloc_threshold defines which metaslab groups should
82 * be eligible for allocation. The value is defined as a percentage of
83 * free space. Metaslab groups that have more free space than
84 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
85 * a metaslab group's free space is less than or equal to the
86 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
87 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
88 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
89 * groups are allowed to accept allocations. Gang blocks are always
90 * eligible to allocate on any metaslab group. The default value of 0 means
91 * no metaslab group will be excluded based on this criterion.
93 int zfs_mg_noalloc_threshold
= 0;
96 * Metaslab groups are considered eligible for allocations if their
97 * fragmenation metric (measured as a percentage) is less than or equal to
98 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
99 * then it will be skipped unless all metaslab groups within the metaslab
100 * class have also crossed this threshold.
102 int zfs_mg_fragmentation_threshold
= 85;
105 * Allow metaslabs to keep their active state as long as their fragmentation
106 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
107 * active metaslab that exceeds this threshold will no longer keep its active
108 * status allowing better metaslabs to be selected.
110 int zfs_metaslab_fragmentation_threshold
= 70;
113 * When set will load all metaslabs when pool is first opened.
115 int metaslab_debug_load
= 0;
118 * When set will prevent metaslabs from being unloaded.
120 int metaslab_debug_unload
= 0;
123 * Minimum size which forces the dynamic allocator to change
124 * it's allocation strategy. Once the space map cannot satisfy
125 * an allocation of this size then it switches to using more
126 * aggressive strategy (i.e search by size rather than offset).
128 uint64_t metaslab_df_alloc_threshold
= SPA_MAXBLOCKSIZE
;
131 * The minimum free space, in percent, which must be available
132 * in a space map to continue allocations in a first-fit fashion.
133 * Once the space_map's free space drops below this level we dynamically
134 * switch to using best-fit allocations.
136 int metaslab_df_free_pct
= 4;
139 * A metaslab is considered "free" if it contains a contiguous
140 * segment which is greater than metaslab_min_alloc_size.
142 uint64_t metaslab_min_alloc_size
= DMU_MAX_ACCESS
;
145 * Percentage of all cpus that can be used by the metaslab taskq.
147 int metaslab_load_pct
= 50;
150 * Determines how many txgs a metaslab may remain loaded without having any
151 * allocations from it. As long as a metaslab continues to be used we will
154 int metaslab_unload_delay
= TXG_SIZE
* 2;
157 * Max number of metaslabs per group to preload.
159 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
162 * Enable/disable preloading of metaslab.
164 boolean_t metaslab_preload_enabled
= B_TRUE
;
167 * Enable/disable fragmentation weighting on metaslabs.
169 boolean_t metaslab_fragmentation_factor_enabled
= B_TRUE
;
172 * Enable/disable lba weighting (i.e. outer tracks are given preference).
174 boolean_t metaslab_lba_weighting_enabled
= B_TRUE
;
177 * Enable/disable metaslab group biasing.
179 boolean_t metaslab_bias_enabled
= B_TRUE
;
181 static uint64_t metaslab_fragmentation(metaslab_t
*);
184 * ==========================================================================
186 * ==========================================================================
189 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
191 metaslab_class_t
*mc
;
193 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
203 metaslab_class_destroy(metaslab_class_t
*mc
)
205 ASSERT(mc
->mc_rotor
== NULL
);
206 ASSERT(mc
->mc_alloc
== 0);
207 ASSERT(mc
->mc_deferred
== 0);
208 ASSERT(mc
->mc_space
== 0);
209 ASSERT(mc
->mc_dspace
== 0);
211 kmem_free(mc
, sizeof (metaslab_class_t
));
215 metaslab_class_validate(metaslab_class_t
*mc
)
217 metaslab_group_t
*mg
;
221 * Must hold one of the spa_config locks.
223 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
224 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
226 if ((mg
= mc
->mc_rotor
) == NULL
)
231 ASSERT(vd
->vdev_mg
!= NULL
);
232 ASSERT3P(vd
->vdev_top
, ==, vd
);
233 ASSERT3P(mg
->mg_class
, ==, mc
);
234 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
235 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
241 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
242 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
244 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
245 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
246 atomic_add_64(&mc
->mc_space
, space_delta
);
247 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
251 metaslab_class_get_alloc(metaslab_class_t
*mc
)
253 return (mc
->mc_alloc
);
257 metaslab_class_get_deferred(metaslab_class_t
*mc
)
259 return (mc
->mc_deferred
);
263 metaslab_class_get_space(metaslab_class_t
*mc
)
265 return (mc
->mc_space
);
269 metaslab_class_get_dspace(metaslab_class_t
*mc
)
271 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
275 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
277 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
281 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
284 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
287 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
288 vdev_t
*tvd
= rvd
->vdev_child
[c
];
289 metaslab_group_t
*mg
= tvd
->vdev_mg
;
292 * Skip any holes, uninitialized top-levels, or
293 * vdevs that are not in this metalab class.
295 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
296 mg
->mg_class
!= mc
) {
300 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
301 mc_hist
[i
] += mg
->mg_histogram
[i
];
304 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
305 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
307 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
311 * Calculate the metaslab class's fragmentation metric. The metric
312 * is weighted based on the space contribution of each metaslab group.
313 * The return value will be a number between 0 and 100 (inclusive), or
314 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
315 * zfs_frag_table for more information about the metric.
318 metaslab_class_fragmentation(metaslab_class_t
*mc
)
320 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
321 uint64_t fragmentation
= 0;
323 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
325 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
326 vdev_t
*tvd
= rvd
->vdev_child
[c
];
327 metaslab_group_t
*mg
= tvd
->vdev_mg
;
330 * Skip any holes, uninitialized top-levels, or
331 * vdevs that are not in this metalab class.
333 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
334 mg
->mg_class
!= mc
) {
339 * If a metaslab group does not contain a fragmentation
340 * metric then just bail out.
342 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
343 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
344 return (ZFS_FRAG_INVALID
);
348 * Determine how much this metaslab_group is contributing
349 * to the overall pool fragmentation metric.
351 fragmentation
+= mg
->mg_fragmentation
*
352 metaslab_group_get_space(mg
);
354 fragmentation
/= metaslab_class_get_space(mc
);
356 ASSERT3U(fragmentation
, <=, 100);
357 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
358 return (fragmentation
);
362 * Calculate the amount of expandable space that is available in
363 * this metaslab class. If a device is expanded then its expandable
364 * space will be the amount of allocatable space that is currently not
365 * part of this metaslab class.
368 metaslab_class_expandable_space(metaslab_class_t
*mc
)
370 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
373 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
374 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
375 vdev_t
*tvd
= rvd
->vdev_child
[c
];
376 metaslab_group_t
*mg
= tvd
->vdev_mg
;
378 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
379 mg
->mg_class
!= mc
) {
383 space
+= tvd
->vdev_max_asize
- tvd
->vdev_asize
;
385 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
390 * ==========================================================================
392 * ==========================================================================
395 metaslab_compare(const void *x1
, const void *x2
)
397 const metaslab_t
*m1
= x1
;
398 const metaslab_t
*m2
= x2
;
400 if (m1
->ms_weight
< m2
->ms_weight
)
402 if (m1
->ms_weight
> m2
->ms_weight
)
406 * If the weights are identical, use the offset to force uniqueness.
408 if (m1
->ms_start
< m2
->ms_start
)
410 if (m1
->ms_start
> m2
->ms_start
)
413 ASSERT3P(m1
, ==, m2
);
419 * Update the allocatable flag and the metaslab group's capacity.
420 * The allocatable flag is set to true if the capacity is below
421 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
422 * from allocatable to non-allocatable or vice versa then the metaslab
423 * group's class is updated to reflect the transition.
426 metaslab_group_alloc_update(metaslab_group_t
*mg
)
428 vdev_t
*vd
= mg
->mg_vd
;
429 metaslab_class_t
*mc
= mg
->mg_class
;
430 vdev_stat_t
*vs
= &vd
->vdev_stat
;
431 boolean_t was_allocatable
;
433 ASSERT(vd
== vd
->vdev_top
);
435 mutex_enter(&mg
->mg_lock
);
436 was_allocatable
= mg
->mg_allocatable
;
438 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
442 * A metaslab group is considered allocatable if it has plenty
443 * of free space or is not heavily fragmented. We only take
444 * fragmentation into account if the metaslab group has a valid
445 * fragmentation metric (i.e. a value between 0 and 100).
447 mg
->mg_allocatable
= (mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
448 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
449 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
452 * The mc_alloc_groups maintains a count of the number of
453 * groups in this metaslab class that are still above the
454 * zfs_mg_noalloc_threshold. This is used by the allocating
455 * threads to determine if they should avoid allocations to
456 * a given group. The allocator will avoid allocations to a group
457 * if that group has reached or is below the zfs_mg_noalloc_threshold
458 * and there are still other groups that are above the threshold.
459 * When a group transitions from allocatable to non-allocatable or
460 * vice versa we update the metaslab class to reflect that change.
461 * When the mc_alloc_groups value drops to 0 that means that all
462 * groups have reached the zfs_mg_noalloc_threshold making all groups
463 * eligible for allocations. This effectively means that all devices
464 * are balanced again.
466 if (was_allocatable
&& !mg
->mg_allocatable
)
467 mc
->mc_alloc_groups
--;
468 else if (!was_allocatable
&& mg
->mg_allocatable
)
469 mc
->mc_alloc_groups
++;
471 mutex_exit(&mg
->mg_lock
);
475 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
477 metaslab_group_t
*mg
;
479 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
480 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
481 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
482 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
485 mg
->mg_activation_count
= 0;
487 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
488 minclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
);
494 metaslab_group_destroy(metaslab_group_t
*mg
)
496 ASSERT(mg
->mg_prev
== NULL
);
497 ASSERT(mg
->mg_next
== NULL
);
499 * We may have gone below zero with the activation count
500 * either because we never activated in the first place or
501 * because we're done, and possibly removing the vdev.
503 ASSERT(mg
->mg_activation_count
<= 0);
505 taskq_destroy(mg
->mg_taskq
);
506 avl_destroy(&mg
->mg_metaslab_tree
);
507 mutex_destroy(&mg
->mg_lock
);
508 kmem_free(mg
, sizeof (metaslab_group_t
));
512 metaslab_group_activate(metaslab_group_t
*mg
)
514 metaslab_class_t
*mc
= mg
->mg_class
;
515 metaslab_group_t
*mgprev
, *mgnext
;
517 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
519 ASSERT(mc
->mc_rotor
!= mg
);
520 ASSERT(mg
->mg_prev
== NULL
);
521 ASSERT(mg
->mg_next
== NULL
);
522 ASSERT(mg
->mg_activation_count
<= 0);
524 if (++mg
->mg_activation_count
<= 0)
527 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
528 metaslab_group_alloc_update(mg
);
530 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
534 mgnext
= mgprev
->mg_next
;
535 mg
->mg_prev
= mgprev
;
536 mg
->mg_next
= mgnext
;
537 mgprev
->mg_next
= mg
;
538 mgnext
->mg_prev
= mg
;
544 metaslab_group_passivate(metaslab_group_t
*mg
)
546 metaslab_class_t
*mc
= mg
->mg_class
;
547 metaslab_group_t
*mgprev
, *mgnext
;
549 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
551 if (--mg
->mg_activation_count
!= 0) {
552 ASSERT(mc
->mc_rotor
!= mg
);
553 ASSERT(mg
->mg_prev
== NULL
);
554 ASSERT(mg
->mg_next
== NULL
);
555 ASSERT(mg
->mg_activation_count
< 0);
559 taskq_wait(mg
->mg_taskq
);
560 metaslab_group_alloc_update(mg
);
562 mgprev
= mg
->mg_prev
;
563 mgnext
= mg
->mg_next
;
568 mc
->mc_rotor
= mgnext
;
569 mgprev
->mg_next
= mgnext
;
570 mgnext
->mg_prev
= mgprev
;
578 metaslab_group_get_space(metaslab_group_t
*mg
)
580 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
584 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
587 vdev_t
*vd
= mg
->mg_vd
;
588 uint64_t ashift
= vd
->vdev_ashift
;
591 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
594 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
597 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
598 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
600 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
601 metaslab_t
*msp
= vd
->vdev_ms
[m
];
603 if (msp
->ms_sm
== NULL
)
606 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
607 mg_hist
[i
+ ashift
] +=
608 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
611 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
612 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
614 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
618 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
620 metaslab_class_t
*mc
= mg
->mg_class
;
621 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
623 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
624 if (msp
->ms_sm
== NULL
)
627 mutex_enter(&mg
->mg_lock
);
628 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
629 mg
->mg_histogram
[i
+ ashift
] +=
630 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
631 mc
->mc_histogram
[i
+ ashift
] +=
632 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
634 mutex_exit(&mg
->mg_lock
);
638 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
640 metaslab_class_t
*mc
= mg
->mg_class
;
641 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
643 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
644 if (msp
->ms_sm
== NULL
)
647 mutex_enter(&mg
->mg_lock
);
648 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
649 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
650 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
651 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
652 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
654 mg
->mg_histogram
[i
+ ashift
] -=
655 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
656 mc
->mc_histogram
[i
+ ashift
] -=
657 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
659 mutex_exit(&mg
->mg_lock
);
663 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
665 ASSERT(msp
->ms_group
== NULL
);
666 mutex_enter(&mg
->mg_lock
);
669 avl_add(&mg
->mg_metaslab_tree
, msp
);
670 mutex_exit(&mg
->mg_lock
);
672 mutex_enter(&msp
->ms_lock
);
673 metaslab_group_histogram_add(mg
, msp
);
674 mutex_exit(&msp
->ms_lock
);
678 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
680 mutex_enter(&msp
->ms_lock
);
681 metaslab_group_histogram_remove(mg
, msp
);
682 mutex_exit(&msp
->ms_lock
);
684 mutex_enter(&mg
->mg_lock
);
685 ASSERT(msp
->ms_group
== mg
);
686 avl_remove(&mg
->mg_metaslab_tree
, msp
);
687 msp
->ms_group
= NULL
;
688 mutex_exit(&mg
->mg_lock
);
692 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
695 * Although in principle the weight can be any value, in
696 * practice we do not use values in the range [1, 511].
698 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
699 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
701 mutex_enter(&mg
->mg_lock
);
702 ASSERT(msp
->ms_group
== mg
);
703 avl_remove(&mg
->mg_metaslab_tree
, msp
);
704 msp
->ms_weight
= weight
;
705 avl_add(&mg
->mg_metaslab_tree
, msp
);
706 mutex_exit(&mg
->mg_lock
);
710 * Calculate the fragmentation for a given metaslab group. We can use
711 * a simple average here since all metaslabs within the group must have
712 * the same size. The return value will be a value between 0 and 100
713 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
714 * group have a fragmentation metric.
717 metaslab_group_fragmentation(metaslab_group_t
*mg
)
719 vdev_t
*vd
= mg
->mg_vd
;
720 uint64_t fragmentation
= 0;
721 uint64_t valid_ms
= 0;
723 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
724 metaslab_t
*msp
= vd
->vdev_ms
[m
];
726 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
730 fragmentation
+= msp
->ms_fragmentation
;
733 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
734 return (ZFS_FRAG_INVALID
);
736 fragmentation
/= valid_ms
;
737 ASSERT3U(fragmentation
, <=, 100);
738 return (fragmentation
);
742 * Determine if a given metaslab group should skip allocations. A metaslab
743 * group should avoid allocations if its free capacity is less than the
744 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
745 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
746 * that can still handle allocations.
749 metaslab_group_allocatable(metaslab_group_t
*mg
)
751 vdev_t
*vd
= mg
->mg_vd
;
752 spa_t
*spa
= vd
->vdev_spa
;
753 metaslab_class_t
*mc
= mg
->mg_class
;
756 * We use two key metrics to determine if a metaslab group is
757 * considered allocatable -- free space and fragmentation. If
758 * the free space is greater than the free space threshold and
759 * the fragmentation is less than the fragmentation threshold then
760 * consider the group allocatable. There are two case when we will
761 * not consider these key metrics. The first is if the group is
762 * associated with a slog device and the second is if all groups
763 * in this metaslab class have already been consider ineligible
766 return ((mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
767 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
768 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
)) ||
769 mc
!= spa_normal_class(spa
) || mc
->mc_alloc_groups
== 0);
773 * ==========================================================================
774 * Range tree callbacks
775 * ==========================================================================
779 * Comparison function for the private size-ordered tree. Tree is sorted
780 * by size, larger sizes at the end of the tree.
783 metaslab_rangesize_compare(const void *x1
, const void *x2
)
785 const range_seg_t
*r1
= x1
;
786 const range_seg_t
*r2
= x2
;
787 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
788 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
790 if (rs_size1
< rs_size2
)
792 if (rs_size1
> rs_size2
)
795 if (r1
->rs_start
< r2
->rs_start
)
798 if (r1
->rs_start
> r2
->rs_start
)
805 * Create any block allocator specific components. The current allocators
806 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
809 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
811 metaslab_t
*msp
= arg
;
813 ASSERT3P(rt
->rt_arg
, ==, msp
);
814 ASSERT(msp
->ms_tree
== NULL
);
816 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
817 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
821 * Destroy the block allocator specific components.
824 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
826 metaslab_t
*msp
= arg
;
828 ASSERT3P(rt
->rt_arg
, ==, msp
);
829 ASSERT3P(msp
->ms_tree
, ==, rt
);
830 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
832 avl_destroy(&msp
->ms_size_tree
);
836 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
838 metaslab_t
*msp
= arg
;
840 ASSERT3P(rt
->rt_arg
, ==, msp
);
841 ASSERT3P(msp
->ms_tree
, ==, rt
);
842 VERIFY(!msp
->ms_condensing
);
843 avl_add(&msp
->ms_size_tree
, rs
);
847 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
849 metaslab_t
*msp
= arg
;
851 ASSERT3P(rt
->rt_arg
, ==, msp
);
852 ASSERT3P(msp
->ms_tree
, ==, rt
);
853 VERIFY(!msp
->ms_condensing
);
854 avl_remove(&msp
->ms_size_tree
, rs
);
858 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
860 metaslab_t
*msp
= arg
;
862 ASSERT3P(rt
->rt_arg
, ==, msp
);
863 ASSERT3P(msp
->ms_tree
, ==, rt
);
866 * Normally one would walk the tree freeing nodes along the way.
867 * Since the nodes are shared with the range trees we can avoid
868 * walking all nodes and just reinitialize the avl tree. The nodes
869 * will be freed by the range tree, so we don't want to free them here.
871 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
872 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
875 static range_tree_ops_t metaslab_rt_ops
= {
884 * ==========================================================================
885 * Metaslab block operations
886 * ==========================================================================
890 * Return the maximum contiguous segment within the metaslab.
893 metaslab_block_maxsize(metaslab_t
*msp
)
895 avl_tree_t
*t
= &msp
->ms_size_tree
;
898 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
901 return (rs
->rs_end
- rs
->rs_start
);
905 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
)
908 range_tree_t
*rt
= msp
->ms_tree
;
910 VERIFY(!msp
->ms_condensing
);
912 start
= msp
->ms_ops
->msop_alloc(msp
, size
);
913 if (start
!= -1ULL) {
914 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
916 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
917 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
918 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
919 range_tree_remove(rt
, start
, size
);
925 * ==========================================================================
926 * Common allocator routines
927 * ==========================================================================
931 * This is a helper function that can be used by the allocator to find
932 * a suitable block to allocate. This will search the specified AVL
933 * tree looking for a block that matches the specified criteria.
936 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
939 range_seg_t
*rs
, rsearch
;
942 rsearch
.rs_start
= *cursor
;
943 rsearch
.rs_end
= *cursor
+ size
;
945 rs
= avl_find(t
, &rsearch
, &where
);
947 rs
= avl_nearest(t
, where
, AVL_AFTER
);
950 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
952 if (offset
+ size
<= rs
->rs_end
) {
953 *cursor
= offset
+ size
;
956 rs
= AVL_NEXT(t
, rs
);
960 * If we know we've searched the whole map (*cursor == 0), give up.
961 * Otherwise, reset the cursor to the beginning and try again.
967 return (metaslab_block_picker(t
, cursor
, size
, align
));
971 * ==========================================================================
972 * The first-fit block allocator
973 * ==========================================================================
976 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
979 * Find the largest power of 2 block size that evenly divides the
980 * requested size. This is used to try to allocate blocks with similar
981 * alignment from the same area of the metaslab (i.e. same cursor
982 * bucket) but it does not guarantee that other allocations sizes
983 * may exist in the same region.
985 uint64_t align
= size
& -size
;
986 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
987 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
989 return (metaslab_block_picker(t
, cursor
, size
, align
));
992 static metaslab_ops_t metaslab_ff_ops
= {
997 * ==========================================================================
998 * Dynamic block allocator -
999 * Uses the first fit allocation scheme until space get low and then
1000 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1001 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1002 * ==========================================================================
1005 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1008 * Find the largest power of 2 block size that evenly divides the
1009 * requested size. This is used to try to allocate blocks with similar
1010 * alignment from the same area of the metaslab (i.e. same cursor
1011 * bucket) but it does not guarantee that other allocations sizes
1012 * may exist in the same region.
1014 uint64_t align
= size
& -size
;
1015 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1016 range_tree_t
*rt
= msp
->ms_tree
;
1017 avl_tree_t
*t
= &rt
->rt_root
;
1018 uint64_t max_size
= metaslab_block_maxsize(msp
);
1019 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1021 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1022 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1024 if (max_size
< size
)
1028 * If we're running low on space switch to using the size
1029 * sorted AVL tree (best-fit).
1031 if (max_size
< metaslab_df_alloc_threshold
||
1032 free_pct
< metaslab_df_free_pct
) {
1033 t
= &msp
->ms_size_tree
;
1037 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1040 static metaslab_ops_t metaslab_df_ops
= {
1045 * ==========================================================================
1046 * Cursor fit block allocator -
1047 * Select the largest region in the metaslab, set the cursor to the beginning
1048 * of the range and the cursor_end to the end of the range. As allocations
1049 * are made advance the cursor. Continue allocating from the cursor until
1050 * the range is exhausted and then find a new range.
1051 * ==========================================================================
1054 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1056 range_tree_t
*rt
= msp
->ms_tree
;
1057 avl_tree_t
*t
= &msp
->ms_size_tree
;
1058 uint64_t *cursor
= &msp
->ms_lbas
[0];
1059 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1060 uint64_t offset
= 0;
1062 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1063 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1065 ASSERT3U(*cursor_end
, >=, *cursor
);
1067 if ((*cursor
+ size
) > *cursor_end
) {
1070 rs
= avl_last(&msp
->ms_size_tree
);
1071 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1074 *cursor
= rs
->rs_start
;
1075 *cursor_end
= rs
->rs_end
;
1084 static metaslab_ops_t metaslab_cf_ops
= {
1089 * ==========================================================================
1090 * New dynamic fit allocator -
1091 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1092 * contiguous blocks. If no region is found then just use the largest segment
1094 * ==========================================================================
1098 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1099 * to request from the allocator.
1101 uint64_t metaslab_ndf_clump_shift
= 4;
1104 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1106 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1108 range_seg_t
*rs
, rsearch
;
1109 uint64_t hbit
= highbit64(size
);
1110 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1111 uint64_t max_size
= metaslab_block_maxsize(msp
);
1113 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1114 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1116 if (max_size
< size
)
1119 rsearch
.rs_start
= *cursor
;
1120 rsearch
.rs_end
= *cursor
+ size
;
1122 rs
= avl_find(t
, &rsearch
, &where
);
1123 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1124 t
= &msp
->ms_size_tree
;
1126 rsearch
.rs_start
= 0;
1127 rsearch
.rs_end
= MIN(max_size
,
1128 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1129 rs
= avl_find(t
, &rsearch
, &where
);
1131 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1135 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1136 *cursor
= rs
->rs_start
+ size
;
1137 return (rs
->rs_start
);
1142 static metaslab_ops_t metaslab_ndf_ops
= {
1146 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1149 * ==========================================================================
1151 * ==========================================================================
1155 * Wait for any in-progress metaslab loads to complete.
1158 metaslab_load_wait(metaslab_t
*msp
)
1160 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1162 while (msp
->ms_loading
) {
1163 ASSERT(!msp
->ms_loaded
);
1164 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1169 metaslab_load(metaslab_t
*msp
)
1173 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1174 ASSERT(!msp
->ms_loaded
);
1175 ASSERT(!msp
->ms_loading
);
1177 msp
->ms_loading
= B_TRUE
;
1180 * If the space map has not been allocated yet, then treat
1181 * all the space in the metaslab as free and add it to the
1184 if (msp
->ms_sm
!= NULL
)
1185 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1187 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1189 msp
->ms_loaded
= (error
== 0);
1190 msp
->ms_loading
= B_FALSE
;
1192 if (msp
->ms_loaded
) {
1193 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1194 range_tree_walk(msp
->ms_defertree
[t
],
1195 range_tree_remove
, msp
->ms_tree
);
1198 cv_broadcast(&msp
->ms_load_cv
);
1203 metaslab_unload(metaslab_t
*msp
)
1205 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1206 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1207 msp
->ms_loaded
= B_FALSE
;
1208 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1212 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
)
1214 vdev_t
*vd
= mg
->mg_vd
;
1215 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1218 msp
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1219 mutex_init(&msp
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1220 cv_init(&msp
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1222 msp
->ms_start
= id
<< vd
->vdev_ms_shift
;
1223 msp
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1226 * We only open space map objects that already exist. All others
1227 * will be opened when we finally allocate an object for it.
1230 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, object
, msp
->ms_start
,
1231 msp
->ms_size
, vd
->vdev_ashift
, &msp
->ms_lock
));
1232 ASSERT(msp
->ms_sm
!= NULL
);
1236 * We create the main range tree here, but we don't create the
1237 * alloctree and freetree until metaslab_sync_done(). This serves
1238 * two purposes: it allows metaslab_sync_done() to detect the
1239 * addition of new space; and for debugging, it ensures that we'd
1240 * data fault on any attempt to use this metaslab before it's ready.
1242 msp
->ms_tree
= range_tree_create(&metaslab_rt_ops
, msp
, &msp
->ms_lock
);
1243 metaslab_group_add(mg
, msp
);
1245 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1246 msp
->ms_ops
= mg
->mg_class
->mc_ops
;
1249 * If we're opening an existing pool (txg == 0) or creating
1250 * a new one (txg == TXG_INITIAL), all space is available now.
1251 * If we're adding space to an existing pool, the new space
1252 * does not become available until after this txg has synced.
1254 if (txg
<= TXG_INITIAL
)
1255 metaslab_sync_done(msp
, 0);
1258 * If metaslab_debug_load is set and we're initializing a metaslab
1259 * that has an allocated space_map object then load the its space
1260 * map so that can verify frees.
1262 if (metaslab_debug_load
&& msp
->ms_sm
!= NULL
) {
1263 mutex_enter(&msp
->ms_lock
);
1264 VERIFY0(metaslab_load(msp
));
1265 mutex_exit(&msp
->ms_lock
);
1269 vdev_dirty(vd
, 0, NULL
, txg
);
1270 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
1277 metaslab_fini(metaslab_t
*msp
)
1279 metaslab_group_t
*mg
= msp
->ms_group
;
1281 metaslab_group_remove(mg
, msp
);
1283 mutex_enter(&msp
->ms_lock
);
1285 VERIFY(msp
->ms_group
== NULL
);
1286 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1288 space_map_close(msp
->ms_sm
);
1290 metaslab_unload(msp
);
1291 range_tree_destroy(msp
->ms_tree
);
1293 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1294 range_tree_destroy(msp
->ms_alloctree
[t
]);
1295 range_tree_destroy(msp
->ms_freetree
[t
]);
1298 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1299 range_tree_destroy(msp
->ms_defertree
[t
]);
1302 ASSERT0(msp
->ms_deferspace
);
1304 mutex_exit(&msp
->ms_lock
);
1305 cv_destroy(&msp
->ms_load_cv
);
1306 mutex_destroy(&msp
->ms_lock
);
1308 kmem_free(msp
, sizeof (metaslab_t
));
1311 #define FRAGMENTATION_TABLE_SIZE 17
1314 * This table defines a segment size based fragmentation metric that will
1315 * allow each metaslab to derive its own fragmentation value. This is done
1316 * by calculating the space in each bucket of the spacemap histogram and
1317 * multiplying that by the fragmetation metric in this table. Doing
1318 * this for all buckets and dividing it by the total amount of free
1319 * space in this metaslab (i.e. the total free space in all buckets) gives
1320 * us the fragmentation metric. This means that a high fragmentation metric
1321 * equates to most of the free space being comprised of small segments.
1322 * Conversely, if the metric is low, then most of the free space is in
1323 * large segments. A 10% change in fragmentation equates to approximately
1324 * double the number of segments.
1326 * This table defines 0% fragmented space using 16MB segments. Testing has
1327 * shown that segments that are greater than or equal to 16MB do not suffer
1328 * from drastic performance problems. Using this value, we derive the rest
1329 * of the table. Since the fragmentation value is never stored on disk, it
1330 * is possible to change these calculations in the future.
1332 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1352 * Calclate the metaslab's fragmentation metric. A return value
1353 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1354 * not support this metric. Otherwise, the return value should be in the
1358 metaslab_fragmentation(metaslab_t
*msp
)
1360 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1361 uint64_t fragmentation
= 0;
1363 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1364 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1366 if (!feature_enabled
)
1367 return (ZFS_FRAG_INVALID
);
1370 * A null space map means that the entire metaslab is free
1371 * and thus is not fragmented.
1373 if (msp
->ms_sm
== NULL
)
1377 * If this metaslab's space_map has not been upgraded, flag it
1378 * so that we upgrade next time we encounter it.
1380 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1381 uint64_t txg
= spa_syncing_txg(spa
);
1382 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1384 msp
->ms_condense_wanted
= B_TRUE
;
1385 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1386 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1387 "msp %p, vd %p", txg
, msp
, vd
);
1388 return (ZFS_FRAG_INVALID
);
1391 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1393 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1394 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1395 FRAGMENTATION_TABLE_SIZE
- 1);
1397 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1400 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1403 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1404 fragmentation
+= space
* zfs_frag_table
[idx
];
1408 fragmentation
/= total
;
1409 ASSERT3U(fragmentation
, <=, 100);
1410 return (fragmentation
);
1414 * Compute a weight -- a selection preference value -- for the given metaslab.
1415 * This is based on the amount of free space, the level of fragmentation,
1416 * the LBA range, and whether the metaslab is loaded.
1419 metaslab_weight(metaslab_t
*msp
)
1421 metaslab_group_t
*mg
= msp
->ms_group
;
1422 vdev_t
*vd
= mg
->mg_vd
;
1423 uint64_t weight
, space
;
1425 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1428 * This vdev is in the process of being removed so there is nothing
1429 * for us to do here.
1431 if (vd
->vdev_removing
) {
1432 ASSERT0(space_map_allocated(msp
->ms_sm
));
1433 ASSERT0(vd
->vdev_ms_shift
);
1438 * The baseline weight is the metaslab's free space.
1440 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1442 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1443 if (metaslab_fragmentation_factor_enabled
&&
1444 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1446 * Use the fragmentation information to inversely scale
1447 * down the baseline weight. We need to ensure that we
1448 * don't exclude this metaslab completely when it's 100%
1449 * fragmented. To avoid this we reduce the fragmented value
1452 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1455 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1456 * this metaslab again. The fragmentation metric may have
1457 * decreased the space to something smaller than
1458 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1459 * so that we can consume any remaining space.
1461 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1462 space
= SPA_MINBLOCKSIZE
;
1467 * Modern disks have uniform bit density and constant angular velocity.
1468 * Therefore, the outer recording zones are faster (higher bandwidth)
1469 * than the inner zones by the ratio of outer to inner track diameter,
1470 * which is typically around 2:1. We account for this by assigning
1471 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1472 * In effect, this means that we'll select the metaslab with the most
1473 * free bandwidth rather than simply the one with the most free space.
1475 if (metaslab_lba_weighting_enabled
) {
1476 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1477 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1481 * If this metaslab is one we're actively using, adjust its
1482 * weight to make it preferable to any inactive metaslab so
1483 * we'll polish it off. If the fragmentation on this metaslab
1484 * has exceed our threshold, then don't mark it active.
1486 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1487 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1488 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1495 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1497 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1499 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1500 metaslab_load_wait(msp
);
1501 if (!msp
->ms_loaded
) {
1502 int error
= metaslab_load(msp
);
1504 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1509 metaslab_group_sort(msp
->ms_group
, msp
,
1510 msp
->ms_weight
| activation_weight
);
1512 ASSERT(msp
->ms_loaded
);
1513 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1519 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
1522 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1523 * this metaslab again. In that case, it had better be empty,
1524 * or we would be leaving space on the table.
1526 ASSERT(size
>= SPA_MINBLOCKSIZE
|| range_tree_space(msp
->ms_tree
) == 0);
1527 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
1528 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1532 metaslab_preload(void *arg
)
1534 metaslab_t
*msp
= arg
;
1535 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1537 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1539 mutex_enter(&msp
->ms_lock
);
1540 metaslab_load_wait(msp
);
1541 if (!msp
->ms_loaded
)
1542 (void) metaslab_load(msp
);
1545 * Set the ms_access_txg value so that we don't unload it right away.
1547 msp
->ms_access_txg
= spa_syncing_txg(spa
) + metaslab_unload_delay
+ 1;
1548 mutex_exit(&msp
->ms_lock
);
1552 metaslab_group_preload(metaslab_group_t
*mg
)
1554 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1556 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1559 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
1560 taskq_wait(mg
->mg_taskq
);
1564 mutex_enter(&mg
->mg_lock
);
1566 * Load the next potential metaslabs
1569 while (msp
!= NULL
) {
1570 metaslab_t
*msp_next
= AVL_NEXT(t
, msp
);
1573 * We preload only the maximum number of metaslabs specified
1574 * by metaslab_preload_limit. If a metaslab is being forced
1575 * to condense then we preload it too. This will ensure
1576 * that force condensing happens in the next txg.
1578 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
1584 * We must drop the metaslab group lock here to preserve
1585 * lock ordering with the ms_lock (when grabbing both
1586 * the mg_lock and the ms_lock, the ms_lock must be taken
1587 * first). As a result, it is possible that the ordering
1588 * of the metaslabs within the avl tree may change before
1589 * we reacquire the lock. The metaslab cannot be removed from
1590 * the tree while we're in syncing context so it is safe to
1591 * drop the mg_lock here. If the metaslabs are reordered
1592 * nothing will break -- we just may end up loading a
1593 * less than optimal one.
1595 mutex_exit(&mg
->mg_lock
);
1596 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
1597 msp
, TQ_SLEEP
) != NULL
);
1598 mutex_enter(&mg
->mg_lock
);
1601 mutex_exit(&mg
->mg_lock
);
1605 * Determine if the space map's on-disk footprint is past our tolerance
1606 * for inefficiency. We would like to use the following criteria to make
1609 * 1. The size of the space map object should not dramatically increase as a
1610 * result of writing out the free space range tree.
1612 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1613 * times the size than the free space range tree representation
1614 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1616 * 3. The on-disk size of the space map should actually decrease.
1618 * Checking the first condition is tricky since we don't want to walk
1619 * the entire AVL tree calculating the estimated on-disk size. Instead we
1620 * use the size-ordered range tree in the metaslab and calculate the
1621 * size required to write out the largest segment in our free tree. If the
1622 * size required to represent that segment on disk is larger than the space
1623 * map object then we avoid condensing this map.
1625 * To determine the second criterion we use a best-case estimate and assume
1626 * each segment can be represented on-disk as a single 64-bit entry. We refer
1627 * to this best-case estimate as the space map's minimal form.
1629 * Unfortunately, we cannot compute the on-disk size of the space map in this
1630 * context because we cannot accurately compute the effects of compression, etc.
1631 * Instead, we apply the heuristic described in the block comment for
1632 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1633 * is greater than a threshold number of blocks.
1636 metaslab_should_condense(metaslab_t
*msp
)
1638 space_map_t
*sm
= msp
->ms_sm
;
1640 uint64_t size
, entries
, segsz
, object_size
, optimal_size
, record_size
;
1641 dmu_object_info_t doi
;
1642 uint64_t vdev_blocksize
= 1 << msp
->ms_group
->mg_vd
->vdev_ashift
;
1644 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1645 ASSERT(msp
->ms_loaded
);
1648 * Use the ms_size_tree range tree, which is ordered by size, to
1649 * obtain the largest segment in the free tree. We always condense
1650 * metaslabs that are empty and metaslabs for which a condense
1651 * request has been made.
1653 rs
= avl_last(&msp
->ms_size_tree
);
1654 if (rs
== NULL
|| msp
->ms_condense_wanted
)
1658 * Calculate the number of 64-bit entries this segment would
1659 * require when written to disk. If this single segment would be
1660 * larger on-disk than the entire current on-disk structure, then
1661 * clearly condensing will increase the on-disk structure size.
1663 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
1664 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
1665 segsz
= entries
* sizeof (uint64_t);
1667 optimal_size
= sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
);
1668 object_size
= space_map_length(msp
->ms_sm
);
1670 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
1671 record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
1673 return (segsz
<= object_size
&&
1674 object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
1675 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
1679 * Condense the on-disk space map representation to its minimized form.
1680 * The minimized form consists of a small number of allocations followed by
1681 * the entries of the free range tree.
1684 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
1686 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1687 range_tree_t
*freetree
= msp
->ms_freetree
[txg
& TXG_MASK
];
1688 range_tree_t
*condense_tree
;
1689 space_map_t
*sm
= msp
->ms_sm
;
1691 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1692 ASSERT3U(spa_sync_pass(spa
), ==, 1);
1693 ASSERT(msp
->ms_loaded
);
1696 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, "
1697 "smp size %llu, segments %lu, forcing condense=%s", txg
,
1698 msp
->ms_id
, msp
, space_map_length(msp
->ms_sm
),
1699 avl_numnodes(&msp
->ms_tree
->rt_root
),
1700 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
1702 msp
->ms_condense_wanted
= B_FALSE
;
1705 * Create an range tree that is 100% allocated. We remove segments
1706 * that have been freed in this txg, any deferred frees that exist,
1707 * and any allocation in the future. Removing segments should be
1708 * a relatively inexpensive operation since we expect these trees to
1709 * have a small number of nodes.
1711 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
1712 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
1715 * Remove what's been freed in this txg from the condense_tree.
1716 * Since we're in sync_pass 1, we know that all the frees from
1717 * this txg are in the freetree.
1719 range_tree_walk(freetree
, range_tree_remove
, condense_tree
);
1721 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1722 range_tree_walk(msp
->ms_defertree
[t
],
1723 range_tree_remove
, condense_tree
);
1726 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1727 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
1728 range_tree_remove
, condense_tree
);
1732 * We're about to drop the metaslab's lock thus allowing
1733 * other consumers to change it's content. Set the
1734 * metaslab's ms_condensing flag to ensure that
1735 * allocations on this metaslab do not occur while we're
1736 * in the middle of committing it to disk. This is only critical
1737 * for the ms_tree as all other range trees use per txg
1738 * views of their content.
1740 msp
->ms_condensing
= B_TRUE
;
1742 mutex_exit(&msp
->ms_lock
);
1743 space_map_truncate(sm
, tx
);
1744 mutex_enter(&msp
->ms_lock
);
1747 * While we would ideally like to create a space_map representation
1748 * that consists only of allocation records, doing so can be
1749 * prohibitively expensive because the in-core free tree can be
1750 * large, and therefore computationally expensive to subtract
1751 * from the condense_tree. Instead we sync out two trees, a cheap
1752 * allocation only tree followed by the in-core free tree. While not
1753 * optimal, this is typically close to optimal, and much cheaper to
1756 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
1757 range_tree_vacate(condense_tree
, NULL
, NULL
);
1758 range_tree_destroy(condense_tree
);
1760 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
1761 msp
->ms_condensing
= B_FALSE
;
1765 * Write a metaslab to disk in the context of the specified transaction group.
1768 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
1770 metaslab_group_t
*mg
= msp
->ms_group
;
1771 vdev_t
*vd
= mg
->mg_vd
;
1772 spa_t
*spa
= vd
->vdev_spa
;
1773 objset_t
*mos
= spa_meta_objset(spa
);
1774 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
1775 range_tree_t
**freetree
= &msp
->ms_freetree
[txg
& TXG_MASK
];
1776 range_tree_t
**freed_tree
=
1777 &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1779 uint64_t object
= space_map_object(msp
->ms_sm
);
1781 ASSERT(!vd
->vdev_ishole
);
1784 * This metaslab has just been added so there's no work to do now.
1786 if (*freetree
== NULL
) {
1787 ASSERT3P(alloctree
, ==, NULL
);
1791 ASSERT3P(alloctree
, !=, NULL
);
1792 ASSERT3P(*freetree
, !=, NULL
);
1793 ASSERT3P(*freed_tree
, !=, NULL
);
1796 * Normally, we don't want to process a metaslab if there
1797 * are no allocations or frees to perform. However, if the metaslab
1798 * is being forced to condense we need to let it through.
1800 if (range_tree_space(alloctree
) == 0 &&
1801 range_tree_space(*freetree
) == 0 &&
1802 !msp
->ms_condense_wanted
)
1806 * The only state that can actually be changing concurrently with
1807 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1808 * be modifying this txg's alloctree, freetree, freed_tree, or
1809 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1810 * space_map ASSERTs. We drop it whenever we call into the DMU,
1811 * because the DMU can call down to us (e.g. via zio_free()) at
1815 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
1817 if (msp
->ms_sm
== NULL
) {
1818 uint64_t new_object
;
1820 new_object
= space_map_alloc(mos
, tx
);
1821 VERIFY3U(new_object
, !=, 0);
1823 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
1824 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
1826 ASSERT(msp
->ms_sm
!= NULL
);
1829 mutex_enter(&msp
->ms_lock
);
1831 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
1832 metaslab_should_condense(msp
)) {
1833 metaslab_condense(msp
, txg
, tx
);
1835 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
1836 space_map_write(msp
->ms_sm
, *freetree
, SM_FREE
, tx
);
1839 metaslab_group_histogram_verify(mg
);
1840 metaslab_class_histogram_verify(mg
->mg_class
);
1841 metaslab_group_histogram_remove(mg
, msp
);
1842 if (msp
->ms_loaded
) {
1844 * When the space map is loaded, we have an accruate
1845 * histogram in the range tree. This gives us an opportunity
1846 * to bring the space map's histogram up-to-date so we clear
1847 * it first before updating it.
1849 space_map_histogram_clear(msp
->ms_sm
);
1850 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
1853 * Since the space map is not loaded we simply update the
1854 * exisiting histogram with what was freed in this txg. This
1855 * means that the on-disk histogram may not have an accurate
1856 * view of the free space but it's close enough to allow
1857 * us to make allocation decisions.
1859 space_map_histogram_add(msp
->ms_sm
, *freetree
, tx
);
1861 metaslab_group_histogram_add(mg
, msp
);
1862 metaslab_group_histogram_verify(mg
);
1863 metaslab_class_histogram_verify(mg
->mg_class
);
1866 * For sync pass 1, we avoid traversing this txg's free range tree
1867 * and instead will just swap the pointers for freetree and
1868 * freed_tree. We can safely do this since the freed_tree is
1869 * guaranteed to be empty on the initial pass.
1871 if (spa_sync_pass(spa
) == 1) {
1872 range_tree_swap(freetree
, freed_tree
);
1874 range_tree_vacate(*freetree
, range_tree_add
, *freed_tree
);
1876 range_tree_vacate(alloctree
, NULL
, NULL
);
1878 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1879 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1881 mutex_exit(&msp
->ms_lock
);
1883 if (object
!= space_map_object(msp
->ms_sm
)) {
1884 object
= space_map_object(msp
->ms_sm
);
1885 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
1886 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
1892 * Called after a transaction group has completely synced to mark
1893 * all of the metaslab's free space as usable.
1896 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1898 metaslab_group_t
*mg
= msp
->ms_group
;
1899 vdev_t
*vd
= mg
->mg_vd
;
1900 range_tree_t
**freed_tree
;
1901 range_tree_t
**defer_tree
;
1902 int64_t alloc_delta
, defer_delta
;
1904 ASSERT(!vd
->vdev_ishole
);
1906 mutex_enter(&msp
->ms_lock
);
1909 * If this metaslab is just becoming available, initialize its
1910 * alloctrees, freetrees, and defertree and add its capacity to
1913 if (msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
] == NULL
) {
1914 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1915 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
1916 ASSERT(msp
->ms_freetree
[t
] == NULL
);
1918 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
1920 msp
->ms_freetree
[t
] = range_tree_create(NULL
, msp
,
1924 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1925 ASSERT(msp
->ms_defertree
[t
] == NULL
);
1927 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
1931 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
1934 freed_tree
= &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1935 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
1937 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
1938 defer_delta
= range_tree_space(*freed_tree
) -
1939 range_tree_space(*defer_tree
);
1941 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1943 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1944 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1947 * If there's a metaslab_load() in progress, wait for it to complete
1948 * so that we have a consistent view of the in-core space map.
1950 metaslab_load_wait(msp
);
1953 * Move the frees from the defer_tree back to the free
1954 * range tree (if it's loaded). Swap the freed_tree and the
1955 * defer_tree -- this is safe to do because we've just emptied out
1958 range_tree_vacate(*defer_tree
,
1959 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
1960 range_tree_swap(freed_tree
, defer_tree
);
1962 space_map_update(msp
->ms_sm
);
1964 msp
->ms_deferspace
+= defer_delta
;
1965 ASSERT3S(msp
->ms_deferspace
, >=, 0);
1966 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
1967 if (msp
->ms_deferspace
!= 0) {
1969 * Keep syncing this metaslab until all deferred frees
1970 * are back in circulation.
1972 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1975 if (msp
->ms_loaded
&& msp
->ms_access_txg
< txg
) {
1976 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1977 VERIFY0(range_tree_space(
1978 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
1981 if (!metaslab_debug_unload
)
1982 metaslab_unload(msp
);
1985 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
1986 mutex_exit(&msp
->ms_lock
);
1990 metaslab_sync_reassess(metaslab_group_t
*mg
)
1992 metaslab_group_alloc_update(mg
);
1993 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
1996 * Preload the next potential metaslabs
1998 metaslab_group_preload(mg
);
2002 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2004 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2005 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2006 uint64_t start
= msp
->ms_id
;
2008 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2009 return (1ULL << 63);
2012 return ((start
- offset
) << ms_shift
);
2014 return ((offset
- start
) << ms_shift
);
2019 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
2020 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2022 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2023 metaslab_t
*msp
= NULL
;
2024 uint64_t offset
= -1ULL;
2025 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2026 uint64_t activation_weight
;
2027 uint64_t target_distance
;
2030 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2031 for (i
= 0; i
< d
; i
++) {
2032 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2033 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2039 boolean_t was_active
;
2041 mutex_enter(&mg
->mg_lock
);
2042 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
2043 if (msp
->ms_weight
< asize
) {
2044 spa_dbgmsg(spa
, "%s: failed to meet weight "
2045 "requirement: vdev %llu, txg %llu, mg %p, "
2046 "msp %p, psize %llu, asize %llu, "
2047 "weight %llu", spa_name(spa
),
2048 mg
->mg_vd
->vdev_id
, txg
,
2049 mg
, msp
, psize
, asize
, msp
->ms_weight
);
2050 mutex_exit(&mg
->mg_lock
);
2055 * If the selected metaslab is condensing, skip it.
2057 if (msp
->ms_condensing
)
2060 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2061 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2064 target_distance
= min_distance
+
2065 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2068 for (i
= 0; i
< d
; i
++)
2069 if (metaslab_distance(msp
, &dva
[i
]) <
2075 mutex_exit(&mg
->mg_lock
);
2079 mutex_enter(&msp
->ms_lock
);
2082 * Ensure that the metaslab we have selected is still
2083 * capable of handling our request. It's possible that
2084 * another thread may have changed the weight while we
2085 * were blocked on the metaslab lock.
2087 if (msp
->ms_weight
< asize
|| (was_active
&&
2088 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
2089 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
2090 mutex_exit(&msp
->ms_lock
);
2094 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2095 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2096 metaslab_passivate(msp
,
2097 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2098 mutex_exit(&msp
->ms_lock
);
2102 if (metaslab_activate(msp
, activation_weight
) != 0) {
2103 mutex_exit(&msp
->ms_lock
);
2108 * If this metaslab is currently condensing then pick again as
2109 * we can't manipulate this metaslab until it's committed
2112 if (msp
->ms_condensing
) {
2113 mutex_exit(&msp
->ms_lock
);
2117 if ((offset
= metaslab_block_alloc(msp
, asize
)) != -1ULL)
2120 metaslab_passivate(msp
, metaslab_block_maxsize(msp
));
2121 mutex_exit(&msp
->ms_lock
);
2124 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2125 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2127 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, asize
);
2128 msp
->ms_access_txg
= txg
+ metaslab_unload_delay
;
2130 mutex_exit(&msp
->ms_lock
);
2136 * Allocate a block for the specified i/o.
2139 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2140 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
2142 metaslab_group_t
*mg
, *rotor
;
2146 int zio_lock
= B_FALSE
;
2147 boolean_t allocatable
;
2148 uint64_t offset
= -1ULL;
2152 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2155 * For testing, make some blocks above a certain size be gang blocks.
2157 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
2158 return (SET_ERROR(ENOSPC
));
2161 * Start at the rotor and loop through all mgs until we find something.
2162 * Note that there's no locking on mc_rotor or mc_aliquot because
2163 * nothing actually breaks if we miss a few updates -- we just won't
2164 * allocate quite as evenly. It all balances out over time.
2166 * If we are doing ditto or log blocks, try to spread them across
2167 * consecutive vdevs. If we're forced to reuse a vdev before we've
2168 * allocated all of our ditto blocks, then try and spread them out on
2169 * that vdev as much as possible. If it turns out to not be possible,
2170 * gradually lower our standards until anything becomes acceptable.
2171 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2172 * gives us hope of containing our fault domains to something we're
2173 * able to reason about. Otherwise, any two top-level vdev failures
2174 * will guarantee the loss of data. With consecutive allocation,
2175 * only two adjacent top-level vdev failures will result in data loss.
2177 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2178 * ourselves on the same vdev as our gang block header. That
2179 * way, we can hope for locality in vdev_cache, plus it makes our
2180 * fault domains something tractable.
2183 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
2186 * It's possible the vdev we're using as the hint no
2187 * longer exists (i.e. removed). Consult the rotor when
2193 if (flags
& METASLAB_HINTBP_AVOID
&&
2194 mg
->mg_next
!= NULL
)
2199 } else if (d
!= 0) {
2200 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
2201 mg
= vd
->vdev_mg
->mg_next
;
2207 * If the hint put us into the wrong metaslab class, or into a
2208 * metaslab group that has been passivated, just follow the rotor.
2210 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
2217 ASSERT(mg
->mg_activation_count
== 1);
2222 * Don't allocate from faulted devices.
2225 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
2226 allocatable
= vdev_allocatable(vd
);
2227 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
2229 allocatable
= vdev_allocatable(vd
);
2233 * Determine if the selected metaslab group is eligible
2234 * for allocations. If we're ganging or have requested
2235 * an allocation for the smallest gang block size
2236 * then we don't want to avoid allocating to the this
2237 * metaslab group. If we're in this condition we should
2238 * try to allocate from any device possible so that we
2239 * don't inadvertently return ENOSPC and suspend the pool
2240 * even though space is still available.
2242 if (allocatable
&& CAN_FASTGANG(flags
) &&
2243 psize
> SPA_GANGBLOCKSIZE
)
2244 allocatable
= metaslab_group_allocatable(mg
);
2250 * Avoid writing single-copy data to a failing vdev
2251 * unless the user instructs us that it is okay.
2253 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
2254 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
2255 d
== 0 && dshift
== 3 && vd
->vdev_children
== 0) {
2260 ASSERT(mg
->mg_class
== mc
);
2262 distance
= vd
->vdev_asize
>> dshift
;
2263 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
2268 asize
= vdev_psize_to_asize(vd
, psize
);
2269 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
2271 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
2273 if (offset
!= -1ULL) {
2275 * If we've just selected this metaslab group,
2276 * figure out whether the corresponding vdev is
2277 * over- or under-used relative to the pool,
2278 * and set an allocation bias to even it out.
2280 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
2281 vdev_stat_t
*vs
= &vd
->vdev_stat
;
2284 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
2285 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
2288 * Calculate how much more or less we should
2289 * try to allocate from this device during
2290 * this iteration around the rotor.
2291 * For example, if a device is 80% full
2292 * and the pool is 20% full then we should
2293 * reduce allocations by 60% on this device.
2295 * mg_bias = (20 - 80) * 512K / 100 = -307K
2297 * This reduces allocations by 307K for this
2300 mg
->mg_bias
= ((cu
- vu
) *
2301 (int64_t)mg
->mg_aliquot
) / 100;
2302 } else if (!metaslab_bias_enabled
) {
2306 if (atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
2307 mg
->mg_aliquot
+ mg
->mg_bias
) {
2308 mc
->mc_rotor
= mg
->mg_next
;
2312 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
2313 DVA_SET_OFFSET(&dva
[d
], offset
);
2314 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
2315 DVA_SET_ASIZE(&dva
[d
], asize
);
2320 mc
->mc_rotor
= mg
->mg_next
;
2322 } while ((mg
= mg
->mg_next
) != rotor
);
2326 ASSERT(dshift
< 64);
2330 if (!allocatable
&& !zio_lock
) {
2336 bzero(&dva
[d
], sizeof (dva_t
));
2338 return (SET_ERROR(ENOSPC
));
2342 * Free the block represented by DVA in the context of the specified
2343 * transaction group.
2346 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
2348 uint64_t vdev
= DVA_GET_VDEV(dva
);
2349 uint64_t offset
= DVA_GET_OFFSET(dva
);
2350 uint64_t size
= DVA_GET_ASIZE(dva
);
2354 ASSERT(DVA_IS_VALID(dva
));
2356 if (txg
> spa_freeze_txg(spa
))
2359 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2360 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
2361 cmn_err(CE_WARN
, "metaslab_free_dva(): bad DVA %llu:%llu",
2362 (u_longlong_t
)vdev
, (u_longlong_t
)offset
);
2367 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2369 if (DVA_GET_GANG(dva
))
2370 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2372 mutex_enter(&msp
->ms_lock
);
2375 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
2378 VERIFY(!msp
->ms_condensing
);
2379 VERIFY3U(offset
, >=, msp
->ms_start
);
2380 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
2381 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
2383 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2384 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2385 range_tree_add(msp
->ms_tree
, offset
, size
);
2387 if (range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]) == 0)
2388 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2389 range_tree_add(msp
->ms_freetree
[txg
& TXG_MASK
],
2393 mutex_exit(&msp
->ms_lock
);
2397 * Intent log support: upon opening the pool after a crash, notify the SPA
2398 * of blocks that the intent log has allocated for immediate write, but
2399 * which are still considered free by the SPA because the last transaction
2400 * group didn't commit yet.
2403 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
2405 uint64_t vdev
= DVA_GET_VDEV(dva
);
2406 uint64_t offset
= DVA_GET_OFFSET(dva
);
2407 uint64_t size
= DVA_GET_ASIZE(dva
);
2412 ASSERT(DVA_IS_VALID(dva
));
2414 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2415 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
2416 return (SET_ERROR(ENXIO
));
2418 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2420 if (DVA_GET_GANG(dva
))
2421 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2423 mutex_enter(&msp
->ms_lock
);
2425 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
2426 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
2428 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
2429 error
= SET_ERROR(ENOENT
);
2431 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
2432 mutex_exit(&msp
->ms_lock
);
2436 VERIFY(!msp
->ms_condensing
);
2437 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2438 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2439 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
2440 range_tree_remove(msp
->ms_tree
, offset
, size
);
2442 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
2443 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2444 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2445 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
2448 mutex_exit(&msp
->ms_lock
);
2454 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
2455 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
2457 dva_t
*dva
= bp
->blk_dva
;
2458 dva_t
*hintdva
= hintbp
->blk_dva
;
2461 ASSERT(bp
->blk_birth
== 0);
2462 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
2464 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2466 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
2467 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2468 return (SET_ERROR(ENOSPC
));
2471 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
2472 ASSERT(BP_GET_NDVAS(bp
) == 0);
2473 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
2475 for (int d
= 0; d
< ndvas
; d
++) {
2476 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
2479 for (d
--; d
>= 0; d
--) {
2480 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
2481 bzero(&dva
[d
], sizeof (dva_t
));
2483 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2488 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
2490 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2492 BP_SET_BIRTH(bp
, txg
, txg
);
2498 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
2500 const dva_t
*dva
= bp
->blk_dva
;
2501 int ndvas
= BP_GET_NDVAS(bp
);
2503 ASSERT(!BP_IS_HOLE(bp
));
2504 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
2506 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
2508 for (int d
= 0; d
< ndvas
; d
++)
2509 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
2511 spa_config_exit(spa
, SCL_FREE
, FTAG
);
2515 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
2517 const dva_t
*dva
= bp
->blk_dva
;
2518 int ndvas
= BP_GET_NDVAS(bp
);
2521 ASSERT(!BP_IS_HOLE(bp
));
2525 * First do a dry run to make sure all DVAs are claimable,
2526 * so we don't have to unwind from partial failures below.
2528 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
2532 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2534 for (int d
= 0; d
< ndvas
; d
++)
2535 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
2538 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2540 ASSERT(error
== 0 || txg
== 0);
2546 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
2548 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
2551 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2552 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2553 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
2554 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
2555 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
2556 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2557 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2560 range_tree_verify(msp
->ms_tree
, offset
, size
);
2562 for (int j
= 0; j
< TXG_SIZE
; j
++)
2563 range_tree_verify(msp
->ms_freetree
[j
], offset
, size
);
2564 for (int j
= 0; j
< TXG_DEFER_SIZE
; j
++)
2565 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
2567 spa_config_exit(spa
, SCL_VDEV
, FTAG
);