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.
25 * Copyright (c) 2014 Integros [integros.com]
28 #include <sys/zfs_context.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
39 * Allow allocations to switch to gang blocks quickly. We do this to
40 * avoid having to load lots of space_maps in a given txg. There are,
41 * however, some cases where we want to avoid "fast" ganging and instead
42 * we want to do an exhaustive search of all metaslabs on this device.
43 * Currently we don't allow any gang, slog, or dump device related allocations
46 #define CAN_FASTGANG(flags) \
47 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
48 METASLAB_GANG_AVOID)))
50 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
51 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
52 #define METASLAB_ACTIVE_MASK \
53 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
55 uint64_t metaslab_aliquot
= 512ULL << 10;
56 uint64_t metaslab_gang_bang
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
59 * The in-core space map representation is more compact than its on-disk form.
60 * The zfs_condense_pct determines how much more compact the in-core
61 * space_map representation must be before we compact it on-disk.
62 * Values should be greater than or equal to 100.
64 int zfs_condense_pct
= 200;
67 * Condensing a metaslab is not guaranteed to actually reduce the amount of
68 * space used on disk. In particular, a space map uses data in increments of
69 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
70 * same number of blocks after condensing. Since the goal of condensing is to
71 * reduce the number of IOPs required to read the space map, we only want to
72 * condense when we can be sure we will reduce the number of blocks used by the
73 * space map. Unfortunately, we cannot precisely compute whether or not this is
74 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
75 * we apply the following heuristic: do not condense a spacemap unless the
76 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
79 int zfs_metaslab_condense_block_threshold
= 4;
82 * The zfs_mg_noalloc_threshold defines which metaslab groups should
83 * be eligible for allocation. The value is defined as a percentage of
84 * free space. Metaslab groups that have more free space than
85 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
86 * a metaslab group's free space is less than or equal to the
87 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
88 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
89 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
90 * groups are allowed to accept allocations. Gang blocks are always
91 * eligible to allocate on any metaslab group. The default value of 0 means
92 * no metaslab group will be excluded based on this criterion.
94 int zfs_mg_noalloc_threshold
= 0;
97 * Metaslab groups are considered eligible for allocations if their
98 * fragmenation metric (measured as a percentage) is less than or equal to
99 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
100 * then it will be skipped unless all metaslab groups within the metaslab
101 * class have also crossed this threshold.
103 int zfs_mg_fragmentation_threshold
= 85;
106 * Allow metaslabs to keep their active state as long as their fragmentation
107 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
108 * active metaslab that exceeds this threshold will no longer keep its active
109 * status allowing better metaslabs to be selected.
111 int zfs_metaslab_fragmentation_threshold
= 70;
114 * When set will load all metaslabs when pool is first opened.
116 int metaslab_debug_load
= 0;
119 * When set will prevent metaslabs from being unloaded.
121 int metaslab_debug_unload
= 0;
124 * Minimum size which forces the dynamic allocator to change
125 * it's allocation strategy. Once the space map cannot satisfy
126 * an allocation of this size then it switches to using more
127 * aggressive strategy (i.e search by size rather than offset).
129 uint64_t metaslab_df_alloc_threshold
= SPA_OLD_MAXBLOCKSIZE
;
132 * The minimum free space, in percent, which must be available
133 * in a space map to continue allocations in a first-fit fashion.
134 * Once the space_map's free space drops below this level we dynamically
135 * switch to using best-fit allocations.
137 int metaslab_df_free_pct
= 4;
140 * A metaslab is considered "free" if it contains a contiguous
141 * segment which is greater than metaslab_min_alloc_size.
143 uint64_t metaslab_min_alloc_size
= DMU_MAX_ACCESS
;
146 * Percentage of all cpus that can be used by the metaslab taskq.
148 int metaslab_load_pct
= 50;
151 * Determines how many txgs a metaslab may remain loaded without having any
152 * allocations from it. As long as a metaslab continues to be used we will
155 int metaslab_unload_delay
= TXG_SIZE
* 2;
158 * Max number of metaslabs per group to preload.
160 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
163 * Enable/disable preloading of metaslab.
165 boolean_t metaslab_preload_enabled
= B_TRUE
;
168 * Enable/disable fragmentation weighting on metaslabs.
170 boolean_t metaslab_fragmentation_factor_enabled
= B_TRUE
;
173 * Enable/disable lba weighting (i.e. outer tracks are given preference).
175 boolean_t metaslab_lba_weighting_enabled
= B_TRUE
;
178 * Enable/disable metaslab group biasing.
180 boolean_t metaslab_bias_enabled
= B_TRUE
;
182 static uint64_t metaslab_fragmentation(metaslab_t
*);
185 * ==========================================================================
187 * ==========================================================================
190 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
192 metaslab_class_t
*mc
;
194 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
204 metaslab_class_destroy(metaslab_class_t
*mc
)
206 ASSERT(mc
->mc_rotor
== NULL
);
207 ASSERT(mc
->mc_alloc
== 0);
208 ASSERT(mc
->mc_deferred
== 0);
209 ASSERT(mc
->mc_space
== 0);
210 ASSERT(mc
->mc_dspace
== 0);
212 kmem_free(mc
, sizeof (metaslab_class_t
));
216 metaslab_class_validate(metaslab_class_t
*mc
)
218 metaslab_group_t
*mg
;
222 * Must hold one of the spa_config locks.
224 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
225 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
227 if ((mg
= mc
->mc_rotor
) == NULL
)
232 ASSERT(vd
->vdev_mg
!= NULL
);
233 ASSERT3P(vd
->vdev_top
, ==, vd
);
234 ASSERT3P(mg
->mg_class
, ==, mc
);
235 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
236 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
242 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
243 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
245 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
246 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
247 atomic_add_64(&mc
->mc_space
, space_delta
);
248 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
252 metaslab_class_get_alloc(metaslab_class_t
*mc
)
254 return (mc
->mc_alloc
);
258 metaslab_class_get_deferred(metaslab_class_t
*mc
)
260 return (mc
->mc_deferred
);
264 metaslab_class_get_space(metaslab_class_t
*mc
)
266 return (mc
->mc_space
);
270 metaslab_class_get_dspace(metaslab_class_t
*mc
)
272 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
276 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
278 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
282 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
285 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
288 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
289 vdev_t
*tvd
= rvd
->vdev_child
[c
];
290 metaslab_group_t
*mg
= tvd
->vdev_mg
;
293 * Skip any holes, uninitialized top-levels, or
294 * vdevs that are not in this metalab class.
296 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
297 mg
->mg_class
!= mc
) {
301 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
302 mc_hist
[i
] += mg
->mg_histogram
[i
];
305 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
306 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
308 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
312 * Calculate the metaslab class's fragmentation metric. The metric
313 * is weighted based on the space contribution of each metaslab group.
314 * The return value will be a number between 0 and 100 (inclusive), or
315 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
316 * zfs_frag_table for more information about the metric.
319 metaslab_class_fragmentation(metaslab_class_t
*mc
)
321 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
322 uint64_t fragmentation
= 0;
324 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
326 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
327 vdev_t
*tvd
= rvd
->vdev_child
[c
];
328 metaslab_group_t
*mg
= tvd
->vdev_mg
;
331 * Skip any holes, uninitialized top-levels, or
332 * vdevs that are not in this metalab class.
334 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
335 mg
->mg_class
!= mc
) {
340 * If a metaslab group does not contain a fragmentation
341 * metric then just bail out.
343 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
344 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
345 return (ZFS_FRAG_INVALID
);
349 * Determine how much this metaslab_group is contributing
350 * to the overall pool fragmentation metric.
352 fragmentation
+= mg
->mg_fragmentation
*
353 metaslab_group_get_space(mg
);
355 fragmentation
/= metaslab_class_get_space(mc
);
357 ASSERT3U(fragmentation
, <=, 100);
358 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
359 return (fragmentation
);
363 * Calculate the amount of expandable space that is available in
364 * this metaslab class. If a device is expanded then its expandable
365 * space will be the amount of allocatable space that is currently not
366 * part of this metaslab class.
369 metaslab_class_expandable_space(metaslab_class_t
*mc
)
371 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
374 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
375 for (int c
= 0; c
< rvd
->vdev_children
; c
++) {
376 vdev_t
*tvd
= rvd
->vdev_child
[c
];
377 metaslab_group_t
*mg
= tvd
->vdev_mg
;
379 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
380 mg
->mg_class
!= mc
) {
385 * Calculate if we have enough space to add additional
386 * metaslabs. We report the expandable space in terms
387 * of the metaslab size since that's the unit of expansion.
389 space
+= P2ALIGN(tvd
->vdev_max_asize
- tvd
->vdev_asize
,
390 1ULL << tvd
->vdev_ms_shift
);
392 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
397 * ==========================================================================
399 * ==========================================================================
402 metaslab_compare(const void *x1
, const void *x2
)
404 const metaslab_t
*m1
= x1
;
405 const metaslab_t
*m2
= x2
;
407 if (m1
->ms_weight
< m2
->ms_weight
)
409 if (m1
->ms_weight
> m2
->ms_weight
)
413 * If the weights are identical, use the offset to force uniqueness.
415 if (m1
->ms_start
< m2
->ms_start
)
417 if (m1
->ms_start
> m2
->ms_start
)
420 ASSERT3P(m1
, ==, m2
);
426 * Update the allocatable flag and the metaslab group's capacity.
427 * The allocatable flag is set to true if the capacity is below
428 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
429 * from allocatable to non-allocatable or vice versa then the metaslab
430 * group's class is updated to reflect the transition.
433 metaslab_group_alloc_update(metaslab_group_t
*mg
)
435 vdev_t
*vd
= mg
->mg_vd
;
436 metaslab_class_t
*mc
= mg
->mg_class
;
437 vdev_stat_t
*vs
= &vd
->vdev_stat
;
438 boolean_t was_allocatable
;
440 ASSERT(vd
== vd
->vdev_top
);
442 mutex_enter(&mg
->mg_lock
);
443 was_allocatable
= mg
->mg_allocatable
;
445 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
449 * A metaslab group is considered allocatable if it has plenty
450 * of free space or is not heavily fragmented. We only take
451 * fragmentation into account if the metaslab group has a valid
452 * fragmentation metric (i.e. a value between 0 and 100).
454 mg
->mg_allocatable
= (mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
455 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
456 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
459 * The mc_alloc_groups maintains a count of the number of
460 * groups in this metaslab class that are still above the
461 * zfs_mg_noalloc_threshold. This is used by the allocating
462 * threads to determine if they should avoid allocations to
463 * a given group. The allocator will avoid allocations to a group
464 * if that group has reached or is below the zfs_mg_noalloc_threshold
465 * and there are still other groups that are above the threshold.
466 * When a group transitions from allocatable to non-allocatable or
467 * vice versa we update the metaslab class to reflect that change.
468 * When the mc_alloc_groups value drops to 0 that means that all
469 * groups have reached the zfs_mg_noalloc_threshold making all groups
470 * eligible for allocations. This effectively means that all devices
471 * are balanced again.
473 if (was_allocatable
&& !mg
->mg_allocatable
)
474 mc
->mc_alloc_groups
--;
475 else if (!was_allocatable
&& mg
->mg_allocatable
)
476 mc
->mc_alloc_groups
++;
478 mutex_exit(&mg
->mg_lock
);
482 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
484 metaslab_group_t
*mg
;
486 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_SLEEP
);
487 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
488 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
489 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
492 mg
->mg_activation_count
= 0;
494 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
495 minclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
);
501 metaslab_group_destroy(metaslab_group_t
*mg
)
503 ASSERT(mg
->mg_prev
== NULL
);
504 ASSERT(mg
->mg_next
== NULL
);
506 * We may have gone below zero with the activation count
507 * either because we never activated in the first place or
508 * because we're done, and possibly removing the vdev.
510 ASSERT(mg
->mg_activation_count
<= 0);
512 taskq_destroy(mg
->mg_taskq
);
513 avl_destroy(&mg
->mg_metaslab_tree
);
514 mutex_destroy(&mg
->mg_lock
);
515 kmem_free(mg
, sizeof (metaslab_group_t
));
519 metaslab_group_activate(metaslab_group_t
*mg
)
521 metaslab_class_t
*mc
= mg
->mg_class
;
522 metaslab_group_t
*mgprev
, *mgnext
;
524 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
526 ASSERT(mc
->mc_rotor
!= mg
);
527 ASSERT(mg
->mg_prev
== NULL
);
528 ASSERT(mg
->mg_next
== NULL
);
529 ASSERT(mg
->mg_activation_count
<= 0);
531 if (++mg
->mg_activation_count
<= 0)
534 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
535 metaslab_group_alloc_update(mg
);
537 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
541 mgnext
= mgprev
->mg_next
;
542 mg
->mg_prev
= mgprev
;
543 mg
->mg_next
= mgnext
;
544 mgprev
->mg_next
= mg
;
545 mgnext
->mg_prev
= mg
;
551 metaslab_group_passivate(metaslab_group_t
*mg
)
553 metaslab_class_t
*mc
= mg
->mg_class
;
554 metaslab_group_t
*mgprev
, *mgnext
;
556 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
558 if (--mg
->mg_activation_count
!= 0) {
559 ASSERT(mc
->mc_rotor
!= mg
);
560 ASSERT(mg
->mg_prev
== NULL
);
561 ASSERT(mg
->mg_next
== NULL
);
562 ASSERT(mg
->mg_activation_count
< 0);
566 taskq_wait(mg
->mg_taskq
);
567 metaslab_group_alloc_update(mg
);
569 mgprev
= mg
->mg_prev
;
570 mgnext
= mg
->mg_next
;
575 mc
->mc_rotor
= mgnext
;
576 mgprev
->mg_next
= mgnext
;
577 mgnext
->mg_prev
= mgprev
;
585 metaslab_group_get_space(metaslab_group_t
*mg
)
587 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
591 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
594 vdev_t
*vd
= mg
->mg_vd
;
595 uint64_t ashift
= vd
->vdev_ashift
;
598 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
601 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
604 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
605 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
607 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
608 metaslab_t
*msp
= vd
->vdev_ms
[m
];
610 if (msp
->ms_sm
== NULL
)
613 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
614 mg_hist
[i
+ ashift
] +=
615 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
618 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
619 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
621 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
625 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
627 metaslab_class_t
*mc
= mg
->mg_class
;
628 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
630 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
631 if (msp
->ms_sm
== NULL
)
634 mutex_enter(&mg
->mg_lock
);
635 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
636 mg
->mg_histogram
[i
+ ashift
] +=
637 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
638 mc
->mc_histogram
[i
+ ashift
] +=
639 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
641 mutex_exit(&mg
->mg_lock
);
645 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
647 metaslab_class_t
*mc
= mg
->mg_class
;
648 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 (int 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;
730 for (int m
= 0; m
< vd
->vdev_ms_count
; m
++) {
731 metaslab_t
*msp
= vd
->vdev_ms
[m
];
733 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
737 fragmentation
+= msp
->ms_fragmentation
;
740 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
741 return (ZFS_FRAG_INVALID
);
743 fragmentation
/= valid_ms
;
744 ASSERT3U(fragmentation
, <=, 100);
745 return (fragmentation
);
749 * Determine if a given metaslab group should skip allocations. A metaslab
750 * group should avoid allocations if its free capacity is less than the
751 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
752 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
753 * that can still handle allocations.
756 metaslab_group_allocatable(metaslab_group_t
*mg
)
758 vdev_t
*vd
= mg
->mg_vd
;
759 spa_t
*spa
= vd
->vdev_spa
;
760 metaslab_class_t
*mc
= mg
->mg_class
;
763 * We use two key metrics to determine if a metaslab group is
764 * considered allocatable -- free space and fragmentation. If
765 * the free space is greater than the free space threshold and
766 * the fragmentation is less than the fragmentation threshold then
767 * consider the group allocatable. There are two case when we will
768 * not consider these key metrics. The first is if the group is
769 * associated with a slog device and the second is if all groups
770 * in this metaslab class have already been consider ineligible
773 return ((mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
774 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
775 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
)) ||
776 mc
!= spa_normal_class(spa
) || mc
->mc_alloc_groups
== 0);
780 * ==========================================================================
781 * Range tree callbacks
782 * ==========================================================================
786 * Comparison function for the private size-ordered tree. Tree is sorted
787 * by size, larger sizes at the end of the tree.
790 metaslab_rangesize_compare(const void *x1
, const void *x2
)
792 const range_seg_t
*r1
= x1
;
793 const range_seg_t
*r2
= x2
;
794 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
795 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
797 if (rs_size1
< rs_size2
)
799 if (rs_size1
> rs_size2
)
802 if (r1
->rs_start
< r2
->rs_start
)
805 if (r1
->rs_start
> r2
->rs_start
)
812 * Create any block allocator specific components. The current allocators
813 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
816 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
818 metaslab_t
*msp
= arg
;
820 ASSERT3P(rt
->rt_arg
, ==, msp
);
821 ASSERT(msp
->ms_tree
== NULL
);
823 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
824 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
828 * Destroy the block allocator specific components.
831 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
833 metaslab_t
*msp
= arg
;
835 ASSERT3P(rt
->rt_arg
, ==, msp
);
836 ASSERT3P(msp
->ms_tree
, ==, rt
);
837 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
839 avl_destroy(&msp
->ms_size_tree
);
843 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
845 metaslab_t
*msp
= arg
;
847 ASSERT3P(rt
->rt_arg
, ==, msp
);
848 ASSERT3P(msp
->ms_tree
, ==, rt
);
849 VERIFY(!msp
->ms_condensing
);
850 avl_add(&msp
->ms_size_tree
, rs
);
854 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
856 metaslab_t
*msp
= arg
;
858 ASSERT3P(rt
->rt_arg
, ==, msp
);
859 ASSERT3P(msp
->ms_tree
, ==, rt
);
860 VERIFY(!msp
->ms_condensing
);
861 avl_remove(&msp
->ms_size_tree
, rs
);
865 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
867 metaslab_t
*msp
= arg
;
869 ASSERT3P(rt
->rt_arg
, ==, msp
);
870 ASSERT3P(msp
->ms_tree
, ==, rt
);
873 * Normally one would walk the tree freeing nodes along the way.
874 * Since the nodes are shared with the range trees we can avoid
875 * walking all nodes and just reinitialize the avl tree. The nodes
876 * will be freed by the range tree, so we don't want to free them here.
878 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
879 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
882 static range_tree_ops_t metaslab_rt_ops
= {
891 * ==========================================================================
892 * Metaslab block operations
893 * ==========================================================================
897 * Return the maximum contiguous segment within the metaslab.
900 metaslab_block_maxsize(metaslab_t
*msp
)
902 avl_tree_t
*t
= &msp
->ms_size_tree
;
905 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
908 return (rs
->rs_end
- rs
->rs_start
);
912 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
)
915 range_tree_t
*rt
= msp
->ms_tree
;
917 VERIFY(!msp
->ms_condensing
);
919 start
= msp
->ms_ops
->msop_alloc(msp
, size
);
920 if (start
!= -1ULL) {
921 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
923 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
924 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
925 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
926 range_tree_remove(rt
, start
, size
);
932 * ==========================================================================
933 * Common allocator routines
934 * ==========================================================================
938 * This is a helper function that can be used by the allocator to find
939 * a suitable block to allocate. This will search the specified AVL
940 * tree looking for a block that matches the specified criteria.
943 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
946 range_seg_t
*rs
, rsearch
;
949 rsearch
.rs_start
= *cursor
;
950 rsearch
.rs_end
= *cursor
+ size
;
952 rs
= avl_find(t
, &rsearch
, &where
);
954 rs
= avl_nearest(t
, where
, AVL_AFTER
);
957 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
959 if (offset
+ size
<= rs
->rs_end
) {
960 *cursor
= offset
+ size
;
963 rs
= AVL_NEXT(t
, rs
);
967 * If we know we've searched the whole map (*cursor == 0), give up.
968 * Otherwise, reset the cursor to the beginning and try again.
974 return (metaslab_block_picker(t
, cursor
, size
, align
));
978 * ==========================================================================
979 * The first-fit block allocator
980 * ==========================================================================
983 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
986 * Find the largest power of 2 block size that evenly divides the
987 * requested size. This is used to try to allocate blocks with similar
988 * alignment from the same area of the metaslab (i.e. same cursor
989 * bucket) but it does not guarantee that other allocations sizes
990 * may exist in the same region.
992 uint64_t align
= size
& -size
;
993 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
994 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
996 return (metaslab_block_picker(t
, cursor
, size
, align
));
999 static metaslab_ops_t metaslab_ff_ops
= {
1004 * ==========================================================================
1005 * Dynamic block allocator -
1006 * Uses the first fit allocation scheme until space get low and then
1007 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1008 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1009 * ==========================================================================
1012 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1015 * Find the largest power of 2 block size that evenly divides the
1016 * requested size. This is used to try to allocate blocks with similar
1017 * alignment from the same area of the metaslab (i.e. same cursor
1018 * bucket) but it does not guarantee that other allocations sizes
1019 * may exist in the same region.
1021 uint64_t align
= size
& -size
;
1022 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1023 range_tree_t
*rt
= msp
->ms_tree
;
1024 avl_tree_t
*t
= &rt
->rt_root
;
1025 uint64_t max_size
= metaslab_block_maxsize(msp
);
1026 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1028 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1029 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1031 if (max_size
< size
)
1035 * If we're running low on space switch to using the size
1036 * sorted AVL tree (best-fit).
1038 if (max_size
< metaslab_df_alloc_threshold
||
1039 free_pct
< metaslab_df_free_pct
) {
1040 t
= &msp
->ms_size_tree
;
1044 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1047 static metaslab_ops_t metaslab_df_ops
= {
1052 * ==========================================================================
1053 * Cursor fit block allocator -
1054 * Select the largest region in the metaslab, set the cursor to the beginning
1055 * of the range and the cursor_end to the end of the range. As allocations
1056 * are made advance the cursor. Continue allocating from the cursor until
1057 * the range is exhausted and then find a new range.
1058 * ==========================================================================
1061 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1063 range_tree_t
*rt
= msp
->ms_tree
;
1064 avl_tree_t
*t
= &msp
->ms_size_tree
;
1065 uint64_t *cursor
= &msp
->ms_lbas
[0];
1066 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1067 uint64_t offset
= 0;
1069 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1070 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1072 ASSERT3U(*cursor_end
, >=, *cursor
);
1074 if ((*cursor
+ size
) > *cursor_end
) {
1077 rs
= avl_last(&msp
->ms_size_tree
);
1078 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1081 *cursor
= rs
->rs_start
;
1082 *cursor_end
= rs
->rs_end
;
1091 static metaslab_ops_t metaslab_cf_ops
= {
1096 * ==========================================================================
1097 * New dynamic fit allocator -
1098 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1099 * contiguous blocks. If no region is found then just use the largest segment
1101 * ==========================================================================
1105 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1106 * to request from the allocator.
1108 uint64_t metaslab_ndf_clump_shift
= 4;
1111 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1113 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1115 range_seg_t
*rs
, rsearch
;
1116 uint64_t hbit
= highbit64(size
);
1117 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1118 uint64_t max_size
= metaslab_block_maxsize(msp
);
1120 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1121 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1123 if (max_size
< size
)
1126 rsearch
.rs_start
= *cursor
;
1127 rsearch
.rs_end
= *cursor
+ size
;
1129 rs
= avl_find(t
, &rsearch
, &where
);
1130 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1131 t
= &msp
->ms_size_tree
;
1133 rsearch
.rs_start
= 0;
1134 rsearch
.rs_end
= MIN(max_size
,
1135 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1136 rs
= avl_find(t
, &rsearch
, &where
);
1138 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1142 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1143 *cursor
= rs
->rs_start
+ size
;
1144 return (rs
->rs_start
);
1149 static metaslab_ops_t metaslab_ndf_ops
= {
1153 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1156 * ==========================================================================
1158 * ==========================================================================
1162 * Wait for any in-progress metaslab loads to complete.
1165 metaslab_load_wait(metaslab_t
*msp
)
1167 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1169 while (msp
->ms_loading
) {
1170 ASSERT(!msp
->ms_loaded
);
1171 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1176 metaslab_load(metaslab_t
*msp
)
1180 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1181 ASSERT(!msp
->ms_loaded
);
1182 ASSERT(!msp
->ms_loading
);
1184 msp
->ms_loading
= B_TRUE
;
1187 * If the space map has not been allocated yet, then treat
1188 * all the space in the metaslab as free and add it to the
1191 if (msp
->ms_sm
!= NULL
)
1192 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1194 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1196 msp
->ms_loaded
= (error
== 0);
1197 msp
->ms_loading
= B_FALSE
;
1199 if (msp
->ms_loaded
) {
1200 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1201 range_tree_walk(msp
->ms_defertree
[t
],
1202 range_tree_remove
, msp
->ms_tree
);
1205 cv_broadcast(&msp
->ms_load_cv
);
1210 metaslab_unload(metaslab_t
*msp
)
1212 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1213 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1214 msp
->ms_loaded
= B_FALSE
;
1215 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1219 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1222 vdev_t
*vd
= mg
->mg_vd
;
1223 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1227 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1228 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1229 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1231 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1232 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1235 * We only open space map objects that already exist. All others
1236 * will be opened when we finally allocate an object for it.
1239 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1240 ms
->ms_size
, vd
->vdev_ashift
, &ms
->ms_lock
);
1243 kmem_free(ms
, sizeof (metaslab_t
));
1247 ASSERT(ms
->ms_sm
!= NULL
);
1251 * We create the main range tree here, but we don't create the
1252 * alloctree and freetree until metaslab_sync_done(). This serves
1253 * two purposes: it allows metaslab_sync_done() to detect the
1254 * addition of new space; and for debugging, it ensures that we'd
1255 * data fault on any attempt to use this metaslab before it's ready.
1257 ms
->ms_tree
= range_tree_create(&metaslab_rt_ops
, ms
, &ms
->ms_lock
);
1258 metaslab_group_add(mg
, ms
);
1260 ms
->ms_fragmentation
= metaslab_fragmentation(ms
);
1261 ms
->ms_ops
= mg
->mg_class
->mc_ops
;
1264 * If we're opening an existing pool (txg == 0) or creating
1265 * a new one (txg == TXG_INITIAL), all space is available now.
1266 * If we're adding space to an existing pool, the new space
1267 * does not become available until after this txg has synced.
1269 if (txg
<= TXG_INITIAL
)
1270 metaslab_sync_done(ms
, 0);
1273 * If metaslab_debug_load is set and we're initializing a metaslab
1274 * that has an allocated space_map object then load the its space
1275 * map so that can verify frees.
1277 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1278 mutex_enter(&ms
->ms_lock
);
1279 VERIFY0(metaslab_load(ms
));
1280 mutex_exit(&ms
->ms_lock
);
1284 vdev_dirty(vd
, 0, NULL
, txg
);
1285 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1294 metaslab_fini(metaslab_t
*msp
)
1296 metaslab_group_t
*mg
= msp
->ms_group
;
1298 metaslab_group_remove(mg
, msp
);
1300 mutex_enter(&msp
->ms_lock
);
1302 VERIFY(msp
->ms_group
== NULL
);
1303 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1305 space_map_close(msp
->ms_sm
);
1307 metaslab_unload(msp
);
1308 range_tree_destroy(msp
->ms_tree
);
1310 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1311 range_tree_destroy(msp
->ms_alloctree
[t
]);
1312 range_tree_destroy(msp
->ms_freetree
[t
]);
1315 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1316 range_tree_destroy(msp
->ms_defertree
[t
]);
1319 ASSERT0(msp
->ms_deferspace
);
1321 mutex_exit(&msp
->ms_lock
);
1322 cv_destroy(&msp
->ms_load_cv
);
1323 mutex_destroy(&msp
->ms_lock
);
1325 kmem_free(msp
, sizeof (metaslab_t
));
1328 #define FRAGMENTATION_TABLE_SIZE 17
1331 * This table defines a segment size based fragmentation metric that will
1332 * allow each metaslab to derive its own fragmentation value. This is done
1333 * by calculating the space in each bucket of the spacemap histogram and
1334 * multiplying that by the fragmetation metric in this table. Doing
1335 * this for all buckets and dividing it by the total amount of free
1336 * space in this metaslab (i.e. the total free space in all buckets) gives
1337 * us the fragmentation metric. This means that a high fragmentation metric
1338 * equates to most of the free space being comprised of small segments.
1339 * Conversely, if the metric is low, then most of the free space is in
1340 * large segments. A 10% change in fragmentation equates to approximately
1341 * double the number of segments.
1343 * This table defines 0% fragmented space using 16MB segments. Testing has
1344 * shown that segments that are greater than or equal to 16MB do not suffer
1345 * from drastic performance problems. Using this value, we derive the rest
1346 * of the table. Since the fragmentation value is never stored on disk, it
1347 * is possible to change these calculations in the future.
1349 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1369 * Calclate the metaslab's fragmentation metric. A return value
1370 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1371 * not support this metric. Otherwise, the return value should be in the
1375 metaslab_fragmentation(metaslab_t
*msp
)
1377 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1378 uint64_t fragmentation
= 0;
1380 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1381 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1383 if (!feature_enabled
)
1384 return (ZFS_FRAG_INVALID
);
1387 * A null space map means that the entire metaslab is free
1388 * and thus is not fragmented.
1390 if (msp
->ms_sm
== NULL
)
1394 * If this metaslab's space_map has not been upgraded, flag it
1395 * so that we upgrade next time we encounter it.
1397 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1398 uint64_t txg
= spa_syncing_txg(spa
);
1399 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1401 if (spa_writeable(spa
)) {
1402 msp
->ms_condense_wanted
= B_TRUE
;
1403 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1404 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1405 "msp %p, vd %p", txg
, msp
, vd
);
1407 return (ZFS_FRAG_INVALID
);
1410 for (int i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1412 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1413 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1414 FRAGMENTATION_TABLE_SIZE
- 1);
1416 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1419 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1422 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1423 fragmentation
+= space
* zfs_frag_table
[idx
];
1427 fragmentation
/= total
;
1428 ASSERT3U(fragmentation
, <=, 100);
1429 return (fragmentation
);
1433 * Compute a weight -- a selection preference value -- for the given metaslab.
1434 * This is based on the amount of free space, the level of fragmentation,
1435 * the LBA range, and whether the metaslab is loaded.
1438 metaslab_weight(metaslab_t
*msp
)
1440 metaslab_group_t
*mg
= msp
->ms_group
;
1441 vdev_t
*vd
= mg
->mg_vd
;
1442 uint64_t weight
, space
;
1444 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1447 * This vdev is in the process of being removed so there is nothing
1448 * for us to do here.
1450 if (vd
->vdev_removing
) {
1451 ASSERT0(space_map_allocated(msp
->ms_sm
));
1452 ASSERT0(vd
->vdev_ms_shift
);
1457 * The baseline weight is the metaslab's free space.
1459 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1461 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1462 if (metaslab_fragmentation_factor_enabled
&&
1463 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1465 * Use the fragmentation information to inversely scale
1466 * down the baseline weight. We need to ensure that we
1467 * don't exclude this metaslab completely when it's 100%
1468 * fragmented. To avoid this we reduce the fragmented value
1471 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1474 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1475 * this metaslab again. The fragmentation metric may have
1476 * decreased the space to something smaller than
1477 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1478 * so that we can consume any remaining space.
1480 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1481 space
= SPA_MINBLOCKSIZE
;
1486 * Modern disks have uniform bit density and constant angular velocity.
1487 * Therefore, the outer recording zones are faster (higher bandwidth)
1488 * than the inner zones by the ratio of outer to inner track diameter,
1489 * which is typically around 2:1. We account for this by assigning
1490 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1491 * In effect, this means that we'll select the metaslab with the most
1492 * free bandwidth rather than simply the one with the most free space.
1494 if (metaslab_lba_weighting_enabled
) {
1495 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1496 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1500 * If this metaslab is one we're actively using, adjust its
1501 * weight to make it preferable to any inactive metaslab so
1502 * we'll polish it off. If the fragmentation on this metaslab
1503 * has exceed our threshold, then don't mark it active.
1505 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1506 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1507 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1514 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1516 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1518 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1519 metaslab_load_wait(msp
);
1520 if (!msp
->ms_loaded
) {
1521 int error
= metaslab_load(msp
);
1523 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1528 metaslab_group_sort(msp
->ms_group
, msp
,
1529 msp
->ms_weight
| activation_weight
);
1531 ASSERT(msp
->ms_loaded
);
1532 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1538 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
1541 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1542 * this metaslab again. In that case, it had better be empty,
1543 * or we would be leaving space on the table.
1545 ASSERT(size
>= SPA_MINBLOCKSIZE
|| range_tree_space(msp
->ms_tree
) == 0);
1546 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
1547 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1551 metaslab_preload(void *arg
)
1553 metaslab_t
*msp
= arg
;
1554 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1556 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1558 mutex_enter(&msp
->ms_lock
);
1559 metaslab_load_wait(msp
);
1560 if (!msp
->ms_loaded
)
1561 (void) metaslab_load(msp
);
1564 * Set the ms_access_txg value so that we don't unload it right away.
1566 msp
->ms_access_txg
= spa_syncing_txg(spa
) + metaslab_unload_delay
+ 1;
1567 mutex_exit(&msp
->ms_lock
);
1571 metaslab_group_preload(metaslab_group_t
*mg
)
1573 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1575 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1578 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
1579 taskq_wait(mg
->mg_taskq
);
1583 mutex_enter(&mg
->mg_lock
);
1585 * Load the next potential metaslabs
1588 while (msp
!= NULL
) {
1589 metaslab_t
*msp_next
= AVL_NEXT(t
, msp
);
1592 * We preload only the maximum number of metaslabs specified
1593 * by metaslab_preload_limit. If a metaslab is being forced
1594 * to condense then we preload it too. This will ensure
1595 * that force condensing happens in the next txg.
1597 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
1603 * We must drop the metaslab group lock here to preserve
1604 * lock ordering with the ms_lock (when grabbing both
1605 * the mg_lock and the ms_lock, the ms_lock must be taken
1606 * first). As a result, it is possible that the ordering
1607 * of the metaslabs within the avl tree may change before
1608 * we reacquire the lock. The metaslab cannot be removed from
1609 * the tree while we're in syncing context so it is safe to
1610 * drop the mg_lock here. If the metaslabs are reordered
1611 * nothing will break -- we just may end up loading a
1612 * less than optimal one.
1614 mutex_exit(&mg
->mg_lock
);
1615 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
1616 msp
, TQ_SLEEP
) != NULL
);
1617 mutex_enter(&mg
->mg_lock
);
1620 mutex_exit(&mg
->mg_lock
);
1624 * Determine if the space map's on-disk footprint is past our tolerance
1625 * for inefficiency. We would like to use the following criteria to make
1628 * 1. The size of the space map object should not dramatically increase as a
1629 * result of writing out the free space range tree.
1631 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1632 * times the size than the free space range tree representation
1633 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1635 * 3. The on-disk size of the space map should actually decrease.
1637 * Checking the first condition is tricky since we don't want to walk
1638 * the entire AVL tree calculating the estimated on-disk size. Instead we
1639 * use the size-ordered range tree in the metaslab and calculate the
1640 * size required to write out the largest segment in our free tree. If the
1641 * size required to represent that segment on disk is larger than the space
1642 * map object then we avoid condensing this map.
1644 * To determine the second criterion we use a best-case estimate and assume
1645 * each segment can be represented on-disk as a single 64-bit entry. We refer
1646 * to this best-case estimate as the space map's minimal form.
1648 * Unfortunately, we cannot compute the on-disk size of the space map in this
1649 * context because we cannot accurately compute the effects of compression, etc.
1650 * Instead, we apply the heuristic described in the block comment for
1651 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1652 * is greater than a threshold number of blocks.
1655 metaslab_should_condense(metaslab_t
*msp
)
1657 space_map_t
*sm
= msp
->ms_sm
;
1659 uint64_t size
, entries
, segsz
, object_size
, optimal_size
, record_size
;
1660 dmu_object_info_t doi
;
1661 uint64_t vdev_blocksize
= 1 << msp
->ms_group
->mg_vd
->vdev_ashift
;
1663 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1664 ASSERT(msp
->ms_loaded
);
1667 * Use the ms_size_tree range tree, which is ordered by size, to
1668 * obtain the largest segment in the free tree. We always condense
1669 * metaslabs that are empty and metaslabs for which a condense
1670 * request has been made.
1672 rs
= avl_last(&msp
->ms_size_tree
);
1673 if (rs
== NULL
|| msp
->ms_condense_wanted
)
1677 * Calculate the number of 64-bit entries this segment would
1678 * require when written to disk. If this single segment would be
1679 * larger on-disk than the entire current on-disk structure, then
1680 * clearly condensing will increase the on-disk structure size.
1682 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
1683 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
1684 segsz
= entries
* sizeof (uint64_t);
1686 optimal_size
= sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
);
1687 object_size
= space_map_length(msp
->ms_sm
);
1689 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
1690 record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
1692 return (segsz
<= object_size
&&
1693 object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
1694 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
1698 * Condense the on-disk space map representation to its minimized form.
1699 * The minimized form consists of a small number of allocations followed by
1700 * the entries of the free range tree.
1703 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
1705 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1706 range_tree_t
*freetree
= msp
->ms_freetree
[txg
& TXG_MASK
];
1707 range_tree_t
*condense_tree
;
1708 space_map_t
*sm
= msp
->ms_sm
;
1710 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1711 ASSERT3U(spa_sync_pass(spa
), ==, 1);
1712 ASSERT(msp
->ms_loaded
);
1715 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
1716 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg
,
1717 msp
->ms_id
, msp
, msp
->ms_group
->mg_vd
->vdev_id
,
1718 msp
->ms_group
->mg_vd
->vdev_spa
->spa_name
,
1719 space_map_length(msp
->ms_sm
), avl_numnodes(&msp
->ms_tree
->rt_root
),
1720 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
1722 msp
->ms_condense_wanted
= B_FALSE
;
1725 * Create an range tree that is 100% allocated. We remove segments
1726 * that have been freed in this txg, any deferred frees that exist,
1727 * and any allocation in the future. Removing segments should be
1728 * a relatively inexpensive operation since we expect these trees to
1729 * have a small number of nodes.
1731 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
1732 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
1735 * Remove what's been freed in this txg from the condense_tree.
1736 * Since we're in sync_pass 1, we know that all the frees from
1737 * this txg are in the freetree.
1739 range_tree_walk(freetree
, range_tree_remove
, condense_tree
);
1741 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1742 range_tree_walk(msp
->ms_defertree
[t
],
1743 range_tree_remove
, condense_tree
);
1746 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1747 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
1748 range_tree_remove
, condense_tree
);
1752 * We're about to drop the metaslab's lock thus allowing
1753 * other consumers to change it's content. Set the
1754 * metaslab's ms_condensing flag to ensure that
1755 * allocations on this metaslab do not occur while we're
1756 * in the middle of committing it to disk. This is only critical
1757 * for the ms_tree as all other range trees use per txg
1758 * views of their content.
1760 msp
->ms_condensing
= B_TRUE
;
1762 mutex_exit(&msp
->ms_lock
);
1763 space_map_truncate(sm
, tx
);
1764 mutex_enter(&msp
->ms_lock
);
1767 * While we would ideally like to create a space_map representation
1768 * that consists only of allocation records, doing so can be
1769 * prohibitively expensive because the in-core free tree can be
1770 * large, and therefore computationally expensive to subtract
1771 * from the condense_tree. Instead we sync out two trees, a cheap
1772 * allocation only tree followed by the in-core free tree. While not
1773 * optimal, this is typically close to optimal, and much cheaper to
1776 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
1777 range_tree_vacate(condense_tree
, NULL
, NULL
);
1778 range_tree_destroy(condense_tree
);
1780 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
1781 msp
->ms_condensing
= B_FALSE
;
1785 * Write a metaslab to disk in the context of the specified transaction group.
1788 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
1790 metaslab_group_t
*mg
= msp
->ms_group
;
1791 vdev_t
*vd
= mg
->mg_vd
;
1792 spa_t
*spa
= vd
->vdev_spa
;
1793 objset_t
*mos
= spa_meta_objset(spa
);
1794 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
1795 range_tree_t
**freetree
= &msp
->ms_freetree
[txg
& TXG_MASK
];
1796 range_tree_t
**freed_tree
=
1797 &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1799 uint64_t object
= space_map_object(msp
->ms_sm
);
1801 ASSERT(!vd
->vdev_ishole
);
1804 * This metaslab has just been added so there's no work to do now.
1806 if (*freetree
== NULL
) {
1807 ASSERT3P(alloctree
, ==, NULL
);
1811 ASSERT3P(alloctree
, !=, NULL
);
1812 ASSERT3P(*freetree
, !=, NULL
);
1813 ASSERT3P(*freed_tree
, !=, NULL
);
1816 * Normally, we don't want to process a metaslab if there
1817 * are no allocations or frees to perform. However, if the metaslab
1818 * is being forced to condense we need to let it through.
1820 if (range_tree_space(alloctree
) == 0 &&
1821 range_tree_space(*freetree
) == 0 &&
1822 !msp
->ms_condense_wanted
)
1826 * The only state that can actually be changing concurrently with
1827 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1828 * be modifying this txg's alloctree, freetree, freed_tree, or
1829 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1830 * space_map ASSERTs. We drop it whenever we call into the DMU,
1831 * because the DMU can call down to us (e.g. via zio_free()) at
1835 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
1837 if (msp
->ms_sm
== NULL
) {
1838 uint64_t new_object
;
1840 new_object
= space_map_alloc(mos
, tx
);
1841 VERIFY3U(new_object
, !=, 0);
1843 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
1844 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
1846 ASSERT(msp
->ms_sm
!= NULL
);
1849 mutex_enter(&msp
->ms_lock
);
1852 * Note: metaslab_condense() clears the space_map's histogram.
1853 * Therefore we must verify and remove this histogram before
1856 metaslab_group_histogram_verify(mg
);
1857 metaslab_class_histogram_verify(mg
->mg_class
);
1858 metaslab_group_histogram_remove(mg
, msp
);
1860 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
1861 metaslab_should_condense(msp
)) {
1862 metaslab_condense(msp
, txg
, tx
);
1864 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
1865 space_map_write(msp
->ms_sm
, *freetree
, SM_FREE
, tx
);
1868 if (msp
->ms_loaded
) {
1870 * When the space map is loaded, we have an accruate
1871 * histogram in the range tree. This gives us an opportunity
1872 * to bring the space map's histogram up-to-date so we clear
1873 * it first before updating it.
1875 space_map_histogram_clear(msp
->ms_sm
);
1876 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
1879 * Since the space map is not loaded we simply update the
1880 * exisiting histogram with what was freed in this txg. This
1881 * means that the on-disk histogram may not have an accurate
1882 * view of the free space but it's close enough to allow
1883 * us to make allocation decisions.
1885 space_map_histogram_add(msp
->ms_sm
, *freetree
, tx
);
1887 metaslab_group_histogram_add(mg
, msp
);
1888 metaslab_group_histogram_verify(mg
);
1889 metaslab_class_histogram_verify(mg
->mg_class
);
1892 * For sync pass 1, we avoid traversing this txg's free range tree
1893 * and instead will just swap the pointers for freetree and
1894 * freed_tree. We can safely do this since the freed_tree is
1895 * guaranteed to be empty on the initial pass.
1897 if (spa_sync_pass(spa
) == 1) {
1898 range_tree_swap(freetree
, freed_tree
);
1900 range_tree_vacate(*freetree
, range_tree_add
, *freed_tree
);
1902 range_tree_vacate(alloctree
, NULL
, NULL
);
1904 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1905 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1907 mutex_exit(&msp
->ms_lock
);
1909 if (object
!= space_map_object(msp
->ms_sm
)) {
1910 object
= space_map_object(msp
->ms_sm
);
1911 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
1912 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
1918 * Called after a transaction group has completely synced to mark
1919 * all of the metaslab's free space as usable.
1922 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1924 metaslab_group_t
*mg
= msp
->ms_group
;
1925 vdev_t
*vd
= mg
->mg_vd
;
1926 range_tree_t
**freed_tree
;
1927 range_tree_t
**defer_tree
;
1928 int64_t alloc_delta
, defer_delta
;
1930 ASSERT(!vd
->vdev_ishole
);
1932 mutex_enter(&msp
->ms_lock
);
1935 * If this metaslab is just becoming available, initialize its
1936 * alloctrees, freetrees, and defertree and add its capacity to
1939 if (msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
] == NULL
) {
1940 for (int t
= 0; t
< TXG_SIZE
; t
++) {
1941 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
1942 ASSERT(msp
->ms_freetree
[t
] == NULL
);
1944 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
1946 msp
->ms_freetree
[t
] = range_tree_create(NULL
, msp
,
1950 for (int t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1951 ASSERT(msp
->ms_defertree
[t
] == NULL
);
1953 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
1957 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
1960 freed_tree
= &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1961 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
1963 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
1964 defer_delta
= range_tree_space(*freed_tree
) -
1965 range_tree_space(*defer_tree
);
1967 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1969 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1970 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1973 * If there's a metaslab_load() in progress, wait for it to complete
1974 * so that we have a consistent view of the in-core space map.
1976 metaslab_load_wait(msp
);
1979 * Move the frees from the defer_tree back to the free
1980 * range tree (if it's loaded). Swap the freed_tree and the
1981 * defer_tree -- this is safe to do because we've just emptied out
1984 range_tree_vacate(*defer_tree
,
1985 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
1986 range_tree_swap(freed_tree
, defer_tree
);
1988 space_map_update(msp
->ms_sm
);
1990 msp
->ms_deferspace
+= defer_delta
;
1991 ASSERT3S(msp
->ms_deferspace
, >=, 0);
1992 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
1993 if (msp
->ms_deferspace
!= 0) {
1995 * Keep syncing this metaslab until all deferred frees
1996 * are back in circulation.
1998 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2001 if (msp
->ms_loaded
&& msp
->ms_access_txg
< txg
) {
2002 for (int t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2003 VERIFY0(range_tree_space(
2004 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
2007 if (!metaslab_debug_unload
)
2008 metaslab_unload(msp
);
2011 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
2012 mutex_exit(&msp
->ms_lock
);
2016 metaslab_sync_reassess(metaslab_group_t
*mg
)
2018 metaslab_group_alloc_update(mg
);
2019 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2022 * Preload the next potential metaslabs
2024 metaslab_group_preload(mg
);
2028 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2030 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2031 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2032 uint64_t start
= msp
->ms_id
;
2034 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2035 return (1ULL << 63);
2038 return ((start
- offset
) << ms_shift
);
2040 return ((offset
- start
) << ms_shift
);
2045 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
2046 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2048 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2049 metaslab_t
*msp
= NULL
;
2050 uint64_t offset
= -1ULL;
2051 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2052 uint64_t activation_weight
;
2053 uint64_t target_distance
;
2056 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2057 for (i
= 0; i
< d
; i
++) {
2058 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2059 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2065 boolean_t was_active
;
2067 mutex_enter(&mg
->mg_lock
);
2068 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
2069 if (msp
->ms_weight
< asize
) {
2070 spa_dbgmsg(spa
, "%s: failed to meet weight "
2071 "requirement: vdev %llu, txg %llu, mg %p, "
2072 "msp %p, psize %llu, asize %llu, "
2073 "weight %llu", spa_name(spa
),
2074 mg
->mg_vd
->vdev_id
, txg
,
2075 mg
, msp
, psize
, asize
, msp
->ms_weight
);
2076 mutex_exit(&mg
->mg_lock
);
2081 * If the selected metaslab is condensing, skip it.
2083 if (msp
->ms_condensing
)
2086 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2087 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2090 target_distance
= min_distance
+
2091 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2094 for (i
= 0; i
< d
; i
++)
2095 if (metaslab_distance(msp
, &dva
[i
]) <
2101 mutex_exit(&mg
->mg_lock
);
2105 mutex_enter(&msp
->ms_lock
);
2108 * Ensure that the metaslab we have selected is still
2109 * capable of handling our request. It's possible that
2110 * another thread may have changed the weight while we
2111 * were blocked on the metaslab lock.
2113 if (msp
->ms_weight
< asize
|| (was_active
&&
2114 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
2115 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
2116 mutex_exit(&msp
->ms_lock
);
2120 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2121 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2122 metaslab_passivate(msp
,
2123 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2124 mutex_exit(&msp
->ms_lock
);
2128 if (metaslab_activate(msp
, activation_weight
) != 0) {
2129 mutex_exit(&msp
->ms_lock
);
2134 * If this metaslab is currently condensing then pick again as
2135 * we can't manipulate this metaslab until it's committed
2138 if (msp
->ms_condensing
) {
2139 mutex_exit(&msp
->ms_lock
);
2143 if ((offset
= metaslab_block_alloc(msp
, asize
)) != -1ULL)
2146 metaslab_passivate(msp
, metaslab_block_maxsize(msp
));
2147 mutex_exit(&msp
->ms_lock
);
2150 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2151 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2153 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, asize
);
2154 msp
->ms_access_txg
= txg
+ metaslab_unload_delay
;
2156 mutex_exit(&msp
->ms_lock
);
2162 * Allocate a block for the specified i/o.
2165 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2166 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
2168 metaslab_group_t
*mg
, *rotor
;
2172 int zio_lock
= B_FALSE
;
2173 boolean_t allocatable
;
2174 uint64_t offset
= -1ULL;
2178 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2181 * For testing, make some blocks above a certain size be gang blocks.
2183 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
2184 return (SET_ERROR(ENOSPC
));
2187 * Start at the rotor and loop through all mgs until we find something.
2188 * Note that there's no locking on mc_rotor or mc_aliquot because
2189 * nothing actually breaks if we miss a few updates -- we just won't
2190 * allocate quite as evenly. It all balances out over time.
2192 * If we are doing ditto or log blocks, try to spread them across
2193 * consecutive vdevs. If we're forced to reuse a vdev before we've
2194 * allocated all of our ditto blocks, then try and spread them out on
2195 * that vdev as much as possible. If it turns out to not be possible,
2196 * gradually lower our standards until anything becomes acceptable.
2197 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2198 * gives us hope of containing our fault domains to something we're
2199 * able to reason about. Otherwise, any two top-level vdev failures
2200 * will guarantee the loss of data. With consecutive allocation,
2201 * only two adjacent top-level vdev failures will result in data loss.
2203 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2204 * ourselves on the same vdev as our gang block header. That
2205 * way, we can hope for locality in vdev_cache, plus it makes our
2206 * fault domains something tractable.
2209 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
2212 * It's possible the vdev we're using as the hint no
2213 * longer exists (i.e. removed). Consult the rotor when
2219 if (flags
& METASLAB_HINTBP_AVOID
&&
2220 mg
->mg_next
!= NULL
)
2225 } else if (d
!= 0) {
2226 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
2227 mg
= vd
->vdev_mg
->mg_next
;
2233 * If the hint put us into the wrong metaslab class, or into a
2234 * metaslab group that has been passivated, just follow the rotor.
2236 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
2243 ASSERT(mg
->mg_activation_count
== 1);
2248 * Don't allocate from faulted devices.
2251 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
2252 allocatable
= vdev_allocatable(vd
);
2253 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
2255 allocatable
= vdev_allocatable(vd
);
2259 * Determine if the selected metaslab group is eligible
2260 * for allocations. If we're ganging or have requested
2261 * an allocation for the smallest gang block size
2262 * then we don't want to avoid allocating to the this
2263 * metaslab group. If we're in this condition we should
2264 * try to allocate from any device possible so that we
2265 * don't inadvertently return ENOSPC and suspend the pool
2266 * even though space is still available.
2268 if (allocatable
&& CAN_FASTGANG(flags
) &&
2269 psize
> SPA_GANGBLOCKSIZE
)
2270 allocatable
= metaslab_group_allocatable(mg
);
2276 * Avoid writing single-copy data to a failing vdev
2277 * unless the user instructs us that it is okay.
2279 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
2280 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
2281 d
== 0 && dshift
== 3 && vd
->vdev_children
== 0) {
2286 ASSERT(mg
->mg_class
== mc
);
2288 distance
= vd
->vdev_asize
>> dshift
;
2289 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
2294 asize
= vdev_psize_to_asize(vd
, psize
);
2295 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
2297 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
2299 if (offset
!= -1ULL) {
2301 * If we've just selected this metaslab group,
2302 * figure out whether the corresponding vdev is
2303 * over- or under-used relative to the pool,
2304 * and set an allocation bias to even it out.
2306 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
2307 vdev_stat_t
*vs
= &vd
->vdev_stat
;
2310 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
2311 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
2314 * Calculate how much more or less we should
2315 * try to allocate from this device during
2316 * this iteration around the rotor.
2317 * For example, if a device is 80% full
2318 * and the pool is 20% full then we should
2319 * reduce allocations by 60% on this device.
2321 * mg_bias = (20 - 80) * 512K / 100 = -307K
2323 * This reduces allocations by 307K for this
2326 mg
->mg_bias
= ((cu
- vu
) *
2327 (int64_t)mg
->mg_aliquot
) / 100;
2328 } else if (!metaslab_bias_enabled
) {
2332 if (atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
2333 mg
->mg_aliquot
+ mg
->mg_bias
) {
2334 mc
->mc_rotor
= mg
->mg_next
;
2338 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
2339 DVA_SET_OFFSET(&dva
[d
], offset
);
2340 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
2341 DVA_SET_ASIZE(&dva
[d
], asize
);
2346 mc
->mc_rotor
= mg
->mg_next
;
2348 } while ((mg
= mg
->mg_next
) != rotor
);
2352 ASSERT(dshift
< 64);
2356 if (!allocatable
&& !zio_lock
) {
2362 bzero(&dva
[d
], sizeof (dva_t
));
2364 return (SET_ERROR(ENOSPC
));
2368 * Free the block represented by DVA in the context of the specified
2369 * transaction group.
2372 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
2374 uint64_t vdev
= DVA_GET_VDEV(dva
);
2375 uint64_t offset
= DVA_GET_OFFSET(dva
);
2376 uint64_t size
= DVA_GET_ASIZE(dva
);
2380 ASSERT(DVA_IS_VALID(dva
));
2382 if (txg
> spa_freeze_txg(spa
))
2385 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2386 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
2387 cmn_err(CE_WARN
, "metaslab_free_dva(): bad DVA %llu:%llu",
2388 (u_longlong_t
)vdev
, (u_longlong_t
)offset
);
2393 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2395 if (DVA_GET_GANG(dva
))
2396 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2398 mutex_enter(&msp
->ms_lock
);
2401 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
2404 VERIFY(!msp
->ms_condensing
);
2405 VERIFY3U(offset
, >=, msp
->ms_start
);
2406 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
2407 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
2409 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2410 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2411 range_tree_add(msp
->ms_tree
, offset
, size
);
2413 if (range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]) == 0)
2414 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2415 range_tree_add(msp
->ms_freetree
[txg
& TXG_MASK
],
2419 mutex_exit(&msp
->ms_lock
);
2423 * Intent log support: upon opening the pool after a crash, notify the SPA
2424 * of blocks that the intent log has allocated for immediate write, but
2425 * which are still considered free by the SPA because the last transaction
2426 * group didn't commit yet.
2429 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
2431 uint64_t vdev
= DVA_GET_VDEV(dva
);
2432 uint64_t offset
= DVA_GET_OFFSET(dva
);
2433 uint64_t size
= DVA_GET_ASIZE(dva
);
2438 ASSERT(DVA_IS_VALID(dva
));
2440 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2441 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
2442 return (SET_ERROR(ENXIO
));
2444 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2446 if (DVA_GET_GANG(dva
))
2447 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2449 mutex_enter(&msp
->ms_lock
);
2451 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
2452 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
2454 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
2455 error
= SET_ERROR(ENOENT
);
2457 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
2458 mutex_exit(&msp
->ms_lock
);
2462 VERIFY(!msp
->ms_condensing
);
2463 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2464 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2465 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
2466 range_tree_remove(msp
->ms_tree
, offset
, size
);
2468 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
2469 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2470 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2471 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
2474 mutex_exit(&msp
->ms_lock
);
2480 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
2481 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
2483 dva_t
*dva
= bp
->blk_dva
;
2484 dva_t
*hintdva
= hintbp
->blk_dva
;
2487 ASSERT(bp
->blk_birth
== 0);
2488 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
2490 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2492 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
2493 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2494 return (SET_ERROR(ENOSPC
));
2497 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
2498 ASSERT(BP_GET_NDVAS(bp
) == 0);
2499 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
2501 for (int d
= 0; d
< ndvas
; d
++) {
2502 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
2505 for (d
--; d
>= 0; d
--) {
2506 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
2507 bzero(&dva
[d
], sizeof (dva_t
));
2509 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2514 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
2516 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2518 BP_SET_BIRTH(bp
, txg
, txg
);
2524 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
2526 const dva_t
*dva
= bp
->blk_dva
;
2527 int ndvas
= BP_GET_NDVAS(bp
);
2529 ASSERT(!BP_IS_HOLE(bp
));
2530 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
2532 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
2534 for (int d
= 0; d
< ndvas
; d
++)
2535 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
2537 spa_config_exit(spa
, SCL_FREE
, FTAG
);
2541 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
2543 const dva_t
*dva
= bp
->blk_dva
;
2544 int ndvas
= BP_GET_NDVAS(bp
);
2547 ASSERT(!BP_IS_HOLE(bp
));
2551 * First do a dry run to make sure all DVAs are claimable,
2552 * so we don't have to unwind from partial failures below.
2554 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
2558 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2560 for (int d
= 0; d
< ndvas
; d
++)
2561 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
2564 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2566 ASSERT(error
== 0 || txg
== 0);
2572 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
2574 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
2577 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2578 for (int i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2579 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
2580 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
2581 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
2582 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2583 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2586 range_tree_verify(msp
->ms_tree
, offset
, size
);
2588 for (int j
= 0; j
< TXG_SIZE
; j
++)
2589 range_tree_verify(msp
->ms_freetree
[j
], offset
, size
);
2590 for (int j
= 0; j
< TXG_DEFER_SIZE
; j
++)
2591 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
2593 spa_config_exit(spa
, SCL_VDEV
, FTAG
);