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>
37 #define WITH_DF_BLOCK_ALLOCATOR
40 * Allow allocations to switch to gang blocks quickly. We do this to
41 * avoid having to load lots of space_maps in a given txg. There are,
42 * however, some cases where we want to avoid "fast" ganging and instead
43 * we want to do an exhaustive search of all metaslabs on this device.
44 * Currently we don't allow any gang, slog, or dump device related allocations
47 #define CAN_FASTGANG(flags) \
48 (!((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER | \
49 METASLAB_GANG_AVOID)))
51 #define METASLAB_WEIGHT_PRIMARY (1ULL << 63)
52 #define METASLAB_WEIGHT_SECONDARY (1ULL << 62)
53 #define METASLAB_ACTIVE_MASK \
54 (METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
56 uint64_t metaslab_aliquot
= 512ULL << 10;
57 uint64_t metaslab_gang_bang
= SPA_MAXBLOCKSIZE
+ 1; /* force gang blocks */
60 * The in-core space map representation is more compact than its on-disk form.
61 * The zfs_condense_pct determines how much more compact the in-core
62 * space_map representation must be before we compact it on-disk.
63 * Values should be greater than or equal to 100.
65 int zfs_condense_pct
= 200;
68 * Condensing a metaslab is not guaranteed to actually reduce the amount of
69 * space used on disk. In particular, a space map uses data in increments of
70 * MAX(1 << ashift, space_map_blksz), so a metaslab might use the
71 * same number of blocks after condensing. Since the goal of condensing is to
72 * reduce the number of IOPs required to read the space map, we only want to
73 * condense when we can be sure we will reduce the number of blocks used by the
74 * space map. Unfortunately, we cannot precisely compute whether or not this is
75 * the case in metaslab_should_condense since we are holding ms_lock. Instead,
76 * we apply the following heuristic: do not condense a spacemap unless the
77 * uncondensed size consumes greater than zfs_metaslab_condense_block_threshold
80 int zfs_metaslab_condense_block_threshold
= 4;
83 * The zfs_mg_noalloc_threshold defines which metaslab groups should
84 * be eligible for allocation. The value is defined as a percentage of
85 * free space. Metaslab groups that have more free space than
86 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
87 * a metaslab group's free space is less than or equal to the
88 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
89 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
90 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
91 * groups are allowed to accept allocations. Gang blocks are always
92 * eligible to allocate on any metaslab group. The default value of 0 means
93 * no metaslab group will be excluded based on this criterion.
95 int zfs_mg_noalloc_threshold
= 0;
98 * Metaslab groups are considered eligible for allocations if their
99 * fragmenation metric (measured as a percentage) is less than or equal to
100 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
101 * then it will be skipped unless all metaslab groups within the metaslab
102 * class have also crossed this threshold.
104 int zfs_mg_fragmentation_threshold
= 85;
107 * Allow metaslabs to keep their active state as long as their fragmentation
108 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
109 * active metaslab that exceeds this threshold will no longer keep its active
110 * status allowing better metaslabs to be selected.
112 int zfs_metaslab_fragmentation_threshold
= 70;
115 * When set will load all metaslabs when pool is first opened.
117 int metaslab_debug_load
= 0;
120 * When set will prevent metaslabs from being unloaded.
122 int metaslab_debug_unload
= 0;
125 * Minimum size which forces the dynamic allocator to change
126 * it's allocation strategy. Once the space map cannot satisfy
127 * an allocation of this size then it switches to using more
128 * aggressive strategy (i.e search by size rather than offset).
130 uint64_t metaslab_df_alloc_threshold
= SPA_MAXBLOCKSIZE
;
133 * The minimum free space, in percent, which must be available
134 * in a space map to continue allocations in a first-fit fashion.
135 * Once the space_map's free space drops below this level we dynamically
136 * switch to using best-fit allocations.
138 int metaslab_df_free_pct
= 4;
141 * Percentage of all cpus that can be used by the metaslab taskq.
143 int metaslab_load_pct
= 50;
146 * Determines how many txgs a metaslab may remain loaded without having any
147 * allocations from it. As long as a metaslab continues to be used we will
150 int metaslab_unload_delay
= TXG_SIZE
* 2;
153 * Max number of metaslabs per group to preload.
155 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
158 * Enable/disable preloading of metaslab.
160 int metaslab_preload_enabled
= B_TRUE
;
163 * Enable/disable fragmentation weighting on metaslabs.
165 int metaslab_fragmentation_factor_enabled
= B_TRUE
;
168 * Enable/disable lba weighting (i.e. outer tracks are given preference).
170 int metaslab_lba_weighting_enabled
= B_TRUE
;
173 * Enable/disable metaslab group biasing.
175 int metaslab_bias_enabled
= B_TRUE
;
177 static uint64_t metaslab_fragmentation(metaslab_t
*);
180 * ==========================================================================
182 * ==========================================================================
185 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
187 metaslab_class_t
*mc
;
189 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_SLEEP
);
194 mutex_init(&mc
->mc_fastwrite_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
200 metaslab_class_destroy(metaslab_class_t
*mc
)
202 ASSERT(mc
->mc_rotor
== NULL
);
203 ASSERT(mc
->mc_alloc
== 0);
204 ASSERT(mc
->mc_deferred
== 0);
205 ASSERT(mc
->mc_space
== 0);
206 ASSERT(mc
->mc_dspace
== 0);
208 mutex_destroy(&mc
->mc_fastwrite_lock
);
209 kmem_free(mc
, sizeof (metaslab_class_t
));
213 metaslab_class_validate(metaslab_class_t
*mc
)
215 metaslab_group_t
*mg
;
219 * Must hold one of the spa_config locks.
221 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
222 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
224 if ((mg
= mc
->mc_rotor
) == NULL
)
229 ASSERT(vd
->vdev_mg
!= NULL
);
230 ASSERT3P(vd
->vdev_top
, ==, vd
);
231 ASSERT3P(mg
->mg_class
, ==, mc
);
232 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
233 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
239 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
240 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
242 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
243 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
244 atomic_add_64(&mc
->mc_space
, space_delta
);
245 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
249 metaslab_class_get_alloc(metaslab_class_t
*mc
)
251 return (mc
->mc_alloc
);
255 metaslab_class_get_deferred(metaslab_class_t
*mc
)
257 return (mc
->mc_deferred
);
261 metaslab_class_get_space(metaslab_class_t
*mc
)
263 return (mc
->mc_space
);
267 metaslab_class_get_dspace(metaslab_class_t
*mc
)
269 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
273 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
275 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
279 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
282 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
285 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
286 vdev_t
*tvd
= rvd
->vdev_child
[c
];
287 metaslab_group_t
*mg
= tvd
->vdev_mg
;
290 * Skip any holes, uninitialized top-levels, or
291 * vdevs that are not in this metalab class.
293 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
294 mg
->mg_class
!= mc
) {
298 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
299 mc_hist
[i
] += mg
->mg_histogram
[i
];
302 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
303 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
305 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
309 * Calculate the metaslab class's fragmentation metric. The metric
310 * is weighted based on the space contribution of each metaslab group.
311 * The return value will be a number between 0 and 100 (inclusive), or
312 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
313 * zfs_frag_table for more information about the metric.
316 metaslab_class_fragmentation(metaslab_class_t
*mc
)
318 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
319 uint64_t fragmentation
= 0;
322 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
324 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
325 vdev_t
*tvd
= rvd
->vdev_child
[c
];
326 metaslab_group_t
*mg
= tvd
->vdev_mg
;
329 * Skip any holes, uninitialized top-levels, or
330 * vdevs that are not in this metalab class.
332 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
333 mg
->mg_class
!= mc
) {
338 * If a metaslab group does not contain a fragmentation
339 * metric then just bail out.
341 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
342 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
343 return (ZFS_FRAG_INVALID
);
347 * Determine how much this metaslab_group is contributing
348 * to the overall pool fragmentation metric.
350 fragmentation
+= mg
->mg_fragmentation
*
351 metaslab_group_get_space(mg
);
353 fragmentation
/= metaslab_class_get_space(mc
);
355 ASSERT3U(fragmentation
, <=, 100);
356 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
357 return (fragmentation
);
361 * Calculate the amount of expandable space that is available in
362 * this metaslab class. If a device is expanded then its expandable
363 * space will be the amount of allocatable space that is currently not
364 * part of this metaslab class.
367 metaslab_class_expandable_space(metaslab_class_t
*mc
)
369 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
373 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
374 for (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_outstanding(mg
->mg_taskq
, 0);
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 (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
;
624 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
625 if (msp
->ms_sm
== NULL
)
628 mutex_enter(&mg
->mg_lock
);
629 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
630 mg
->mg_histogram
[i
+ ashift
] +=
631 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
632 mc
->mc_histogram
[i
+ ashift
] +=
633 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
635 mutex_exit(&mg
->mg_lock
);
639 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
641 metaslab_class_t
*mc
= mg
->mg_class
;
642 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
645 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
646 if (msp
->ms_sm
== NULL
)
649 mutex_enter(&mg
->mg_lock
);
650 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
651 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
652 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
653 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
654 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
656 mg
->mg_histogram
[i
+ ashift
] -=
657 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
658 mc
->mc_histogram
[i
+ ashift
] -=
659 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
661 mutex_exit(&mg
->mg_lock
);
665 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
667 ASSERT(msp
->ms_group
== NULL
);
668 mutex_enter(&mg
->mg_lock
);
671 avl_add(&mg
->mg_metaslab_tree
, msp
);
672 mutex_exit(&mg
->mg_lock
);
674 mutex_enter(&msp
->ms_lock
);
675 metaslab_group_histogram_add(mg
, msp
);
676 mutex_exit(&msp
->ms_lock
);
680 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
682 mutex_enter(&msp
->ms_lock
);
683 metaslab_group_histogram_remove(mg
, msp
);
684 mutex_exit(&msp
->ms_lock
);
686 mutex_enter(&mg
->mg_lock
);
687 ASSERT(msp
->ms_group
== mg
);
688 avl_remove(&mg
->mg_metaslab_tree
, msp
);
689 msp
->ms_group
= NULL
;
690 mutex_exit(&mg
->mg_lock
);
694 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
697 * Although in principle the weight can be any value, in
698 * practice we do not use values in the range [1, 511].
700 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
701 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
703 mutex_enter(&mg
->mg_lock
);
704 ASSERT(msp
->ms_group
== mg
);
705 avl_remove(&mg
->mg_metaslab_tree
, msp
);
706 msp
->ms_weight
= weight
;
707 avl_add(&mg
->mg_metaslab_tree
, msp
);
708 mutex_exit(&mg
->mg_lock
);
712 * Calculate the fragmentation for a given metaslab group. We can use
713 * a simple average here since all metaslabs within the group must have
714 * the same size. The return value will be a value between 0 and 100
715 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
716 * group have a fragmentation metric.
719 metaslab_group_fragmentation(metaslab_group_t
*mg
)
721 vdev_t
*vd
= mg
->mg_vd
;
722 uint64_t fragmentation
= 0;
723 uint64_t valid_ms
= 0;
726 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
727 metaslab_t
*msp
= vd
->vdev_ms
[m
];
729 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
733 fragmentation
+= msp
->ms_fragmentation
;
736 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
737 return (ZFS_FRAG_INVALID
);
739 fragmentation
/= valid_ms
;
740 ASSERT3U(fragmentation
, <=, 100);
741 return (fragmentation
);
745 * Determine if a given metaslab group should skip allocations. A metaslab
746 * group should avoid allocations if its free capacity is less than the
747 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
748 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
749 * that can still handle allocations.
752 metaslab_group_allocatable(metaslab_group_t
*mg
)
754 vdev_t
*vd
= mg
->mg_vd
;
755 spa_t
*spa
= vd
->vdev_spa
;
756 metaslab_class_t
*mc
= mg
->mg_class
;
759 * We use two key metrics to determine if a metaslab group is
760 * considered allocatable -- free space and fragmentation. If
761 * the free space is greater than the free space threshold and
762 * the fragmentation is less than the fragmentation threshold then
763 * consider the group allocatable. There are two case when we will
764 * not consider these key metrics. The first is if the group is
765 * associated with a slog device and the second is if all groups
766 * in this metaslab class have already been consider ineligible
769 return ((mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
770 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
771 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
)) ||
772 mc
!= spa_normal_class(spa
) || mc
->mc_alloc_groups
== 0);
776 * ==========================================================================
777 * Range tree callbacks
778 * ==========================================================================
782 * Comparison function for the private size-ordered tree. Tree is sorted
783 * by size, larger sizes at the end of the tree.
786 metaslab_rangesize_compare(const void *x1
, const void *x2
)
788 const range_seg_t
*r1
= x1
;
789 const range_seg_t
*r2
= x2
;
790 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
791 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
793 if (rs_size1
< rs_size2
)
795 if (rs_size1
> rs_size2
)
798 if (r1
->rs_start
< r2
->rs_start
)
801 if (r1
->rs_start
> r2
->rs_start
)
808 * Create any block allocator specific components. The current allocators
809 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
812 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
814 metaslab_t
*msp
= arg
;
816 ASSERT3P(rt
->rt_arg
, ==, msp
);
817 ASSERT(msp
->ms_tree
== NULL
);
819 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
820 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
824 * Destroy the block allocator specific components.
827 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
829 metaslab_t
*msp
= arg
;
831 ASSERT3P(rt
->rt_arg
, ==, msp
);
832 ASSERT3P(msp
->ms_tree
, ==, rt
);
833 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
835 avl_destroy(&msp
->ms_size_tree
);
839 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
841 metaslab_t
*msp
= arg
;
843 ASSERT3P(rt
->rt_arg
, ==, msp
);
844 ASSERT3P(msp
->ms_tree
, ==, rt
);
845 VERIFY(!msp
->ms_condensing
);
846 avl_add(&msp
->ms_size_tree
, rs
);
850 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
852 metaslab_t
*msp
= arg
;
854 ASSERT3P(rt
->rt_arg
, ==, msp
);
855 ASSERT3P(msp
->ms_tree
, ==, rt
);
856 VERIFY(!msp
->ms_condensing
);
857 avl_remove(&msp
->ms_size_tree
, rs
);
861 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
863 metaslab_t
*msp
= arg
;
865 ASSERT3P(rt
->rt_arg
, ==, msp
);
866 ASSERT3P(msp
->ms_tree
, ==, rt
);
869 * Normally one would walk the tree freeing nodes along the way.
870 * Since the nodes are shared with the range trees we can avoid
871 * walking all nodes and just reinitialize the avl tree. The nodes
872 * will be freed by the range tree, so we don't want to free them here.
874 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
875 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
878 static range_tree_ops_t metaslab_rt_ops
= {
887 * ==========================================================================
888 * Metaslab block operations
889 * ==========================================================================
893 * Return the maximum contiguous segment within the metaslab.
896 metaslab_block_maxsize(metaslab_t
*msp
)
898 avl_tree_t
*t
= &msp
->ms_size_tree
;
901 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
904 return (rs
->rs_end
- rs
->rs_start
);
908 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
)
911 range_tree_t
*rt
= msp
->ms_tree
;
913 VERIFY(!msp
->ms_condensing
);
915 start
= msp
->ms_ops
->msop_alloc(msp
, size
);
916 if (start
!= -1ULL) {
917 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
919 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
920 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
921 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
922 range_tree_remove(rt
, start
, size
);
928 * ==========================================================================
929 * Common allocator routines
930 * ==========================================================================
933 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
934 defined(WITH_DF_BLOCK_ALLOCATOR) || \
935 defined(WITH_CF_BLOCK_ALLOCATOR)
937 * This is a helper function that can be used by the allocator to find
938 * a suitable block to allocate. This will search the specified AVL
939 * tree looking for a block that matches the specified criteria.
942 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
945 range_seg_t
*rs
, rsearch
;
948 rsearch
.rs_start
= *cursor
;
949 rsearch
.rs_end
= *cursor
+ size
;
951 rs
= avl_find(t
, &rsearch
, &where
);
953 rs
= avl_nearest(t
, where
, AVL_AFTER
);
956 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
958 if (offset
+ size
<= rs
->rs_end
) {
959 *cursor
= offset
+ size
;
962 rs
= AVL_NEXT(t
, rs
);
966 * If we know we've searched the whole map (*cursor == 0), give up.
967 * Otherwise, reset the cursor to the beginning and try again.
973 return (metaslab_block_picker(t
, cursor
, size
, align
));
975 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
977 #if defined(WITH_FF_BLOCK_ALLOCATOR)
979 * ==========================================================================
980 * The first-fit block allocator
981 * ==========================================================================
984 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
987 * Find the largest power of 2 block size that evenly divides the
988 * requested size. This is used to try to allocate blocks with similar
989 * alignment from the same area of the metaslab (i.e. same cursor
990 * bucket) but it does not guarantee that other allocations sizes
991 * may exist in the same region.
993 uint64_t align
= size
& -size
;
994 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
995 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
997 return (metaslab_block_picker(t
, cursor
, size
, align
));
1000 static metaslab_ops_t metaslab_ff_ops
= {
1004 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
1005 #endif /* WITH_FF_BLOCK_ALLOCATOR */
1007 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1009 * ==========================================================================
1010 * Dynamic block allocator -
1011 * Uses the first fit allocation scheme until space get low and then
1012 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1013 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1014 * ==========================================================================
1017 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1020 * Find the largest power of 2 block size that evenly divides the
1021 * requested size. This is used to try to allocate blocks with similar
1022 * alignment from the same area of the metaslab (i.e. same cursor
1023 * bucket) but it does not guarantee that other allocations sizes
1024 * may exist in the same region.
1026 uint64_t align
= size
& -size
;
1027 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1028 range_tree_t
*rt
= msp
->ms_tree
;
1029 avl_tree_t
*t
= &rt
->rt_root
;
1030 uint64_t max_size
= metaslab_block_maxsize(msp
);
1031 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1033 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1034 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1036 if (max_size
< size
)
1040 * If we're running low on space switch to using the size
1041 * sorted AVL tree (best-fit).
1043 if (max_size
< metaslab_df_alloc_threshold
||
1044 free_pct
< metaslab_df_free_pct
) {
1045 t
= &msp
->ms_size_tree
;
1049 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1052 static metaslab_ops_t metaslab_df_ops
= {
1056 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1057 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1059 #if defined(WITH_CF_BLOCK_ALLOCATOR)
1061 * ==========================================================================
1062 * Cursor fit block allocator -
1063 * Select the largest region in the metaslab, set the cursor to the beginning
1064 * of the range and the cursor_end to the end of the range. As allocations
1065 * are made advance the cursor. Continue allocating from the cursor until
1066 * the range is exhausted and then find a new range.
1067 * ==========================================================================
1070 metaslab_cf_alloc(metaslab_t
*msp
, uint64_t size
)
1072 range_tree_t
*rt
= msp
->ms_tree
;
1073 avl_tree_t
*t
= &msp
->ms_size_tree
;
1074 uint64_t *cursor
= &msp
->ms_lbas
[0];
1075 uint64_t *cursor_end
= &msp
->ms_lbas
[1];
1076 uint64_t offset
= 0;
1078 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1079 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&rt
->rt_root
));
1081 ASSERT3U(*cursor_end
, >=, *cursor
);
1083 if ((*cursor
+ size
) > *cursor_end
) {
1086 rs
= avl_last(&msp
->ms_size_tree
);
1087 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
)
1090 *cursor
= rs
->rs_start
;
1091 *cursor_end
= rs
->rs_end
;
1100 static metaslab_ops_t metaslab_cf_ops
= {
1104 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1105 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1107 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1109 * ==========================================================================
1110 * New dynamic fit allocator -
1111 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1112 * contiguous blocks. If no region is found then just use the largest segment
1114 * ==========================================================================
1118 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1119 * to request from the allocator.
1121 uint64_t metaslab_ndf_clump_shift
= 4;
1124 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1126 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1128 range_seg_t
*rs
, rsearch
;
1129 uint64_t hbit
= highbit64(size
);
1130 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1131 uint64_t max_size
= metaslab_block_maxsize(msp
);
1133 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1134 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1136 if (max_size
< size
)
1139 rsearch
.rs_start
= *cursor
;
1140 rsearch
.rs_end
= *cursor
+ size
;
1142 rs
= avl_find(t
, &rsearch
, &where
);
1143 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1144 t
= &msp
->ms_size_tree
;
1146 rsearch
.rs_start
= 0;
1147 rsearch
.rs_end
= MIN(max_size
,
1148 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1149 rs
= avl_find(t
, &rsearch
, &where
);
1151 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1155 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1156 *cursor
= rs
->rs_start
+ size
;
1157 return (rs
->rs_start
);
1162 static metaslab_ops_t metaslab_ndf_ops
= {
1166 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1167 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1171 * ==========================================================================
1173 * ==========================================================================
1177 * Wait for any in-progress metaslab loads to complete.
1180 metaslab_load_wait(metaslab_t
*msp
)
1182 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1184 while (msp
->ms_loading
) {
1185 ASSERT(!msp
->ms_loaded
);
1186 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1191 metaslab_load(metaslab_t
*msp
)
1196 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1197 ASSERT(!msp
->ms_loaded
);
1198 ASSERT(!msp
->ms_loading
);
1200 msp
->ms_loading
= B_TRUE
;
1203 * If the space map has not been allocated yet, then treat
1204 * all the space in the metaslab as free and add it to the
1207 if (msp
->ms_sm
!= NULL
)
1208 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1210 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1212 msp
->ms_loaded
= (error
== 0);
1213 msp
->ms_loading
= B_FALSE
;
1215 if (msp
->ms_loaded
) {
1216 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1217 range_tree_walk(msp
->ms_defertree
[t
],
1218 range_tree_remove
, msp
->ms_tree
);
1221 cv_broadcast(&msp
->ms_load_cv
);
1226 metaslab_unload(metaslab_t
*msp
)
1228 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1229 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1230 msp
->ms_loaded
= B_FALSE
;
1231 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1235 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
,
1238 vdev_t
*vd
= mg
->mg_vd
;
1239 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1243 ms
= kmem_zalloc(sizeof (metaslab_t
), KM_SLEEP
);
1244 mutex_init(&ms
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1245 cv_init(&ms
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1247 ms
->ms_start
= id
<< vd
->vdev_ms_shift
;
1248 ms
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1251 * We only open space map objects that already exist. All others
1252 * will be opened when we finally allocate an object for it.
1255 error
= space_map_open(&ms
->ms_sm
, mos
, object
, ms
->ms_start
,
1256 ms
->ms_size
, vd
->vdev_ashift
, &ms
->ms_lock
);
1259 kmem_free(ms
, sizeof (metaslab_t
));
1263 ASSERT(ms
->ms_sm
!= NULL
);
1267 * We create the main range tree here, but we don't create the
1268 * alloctree and freetree until metaslab_sync_done(). This serves
1269 * two purposes: it allows metaslab_sync_done() to detect the
1270 * addition of new space; and for debugging, it ensures that we'd
1271 * data fault on any attempt to use this metaslab before it's ready.
1273 ms
->ms_tree
= range_tree_create(&metaslab_rt_ops
, ms
, &ms
->ms_lock
);
1274 metaslab_group_add(mg
, ms
);
1276 ms
->ms_fragmentation
= metaslab_fragmentation(ms
);
1277 ms
->ms_ops
= mg
->mg_class
->mc_ops
;
1280 * If we're opening an existing pool (txg == 0) or creating
1281 * a new one (txg == TXG_INITIAL), all space is available now.
1282 * If we're adding space to an existing pool, the new space
1283 * does not become available until after this txg has synced.
1285 if (txg
<= TXG_INITIAL
)
1286 metaslab_sync_done(ms
, 0);
1289 * If metaslab_debug_load is set and we're initializing a metaslab
1290 * that has an allocated space_map object then load the its space
1291 * map so that can verify frees.
1293 if (metaslab_debug_load
&& ms
->ms_sm
!= NULL
) {
1294 mutex_enter(&ms
->ms_lock
);
1295 VERIFY0(metaslab_load(ms
));
1296 mutex_exit(&ms
->ms_lock
);
1300 vdev_dirty(vd
, 0, NULL
, txg
);
1301 vdev_dirty(vd
, VDD_METASLAB
, ms
, txg
);
1310 metaslab_fini(metaslab_t
*msp
)
1314 metaslab_group_t
*mg
= msp
->ms_group
;
1316 metaslab_group_remove(mg
, msp
);
1318 mutex_enter(&msp
->ms_lock
);
1320 VERIFY(msp
->ms_group
== NULL
);
1321 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1323 space_map_close(msp
->ms_sm
);
1325 metaslab_unload(msp
);
1326 range_tree_destroy(msp
->ms_tree
);
1328 for (t
= 0; t
< TXG_SIZE
; t
++) {
1329 range_tree_destroy(msp
->ms_alloctree
[t
]);
1330 range_tree_destroy(msp
->ms_freetree
[t
]);
1333 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1334 range_tree_destroy(msp
->ms_defertree
[t
]);
1337 ASSERT0(msp
->ms_deferspace
);
1339 mutex_exit(&msp
->ms_lock
);
1340 cv_destroy(&msp
->ms_load_cv
);
1341 mutex_destroy(&msp
->ms_lock
);
1343 kmem_free(msp
, sizeof (metaslab_t
));
1346 #define FRAGMENTATION_TABLE_SIZE 17
1349 * This table defines a segment size based fragmentation metric that will
1350 * allow each metaslab to derive its own fragmentation value. This is done
1351 * by calculating the space in each bucket of the spacemap histogram and
1352 * multiplying that by the fragmetation metric in this table. Doing
1353 * this for all buckets and dividing it by the total amount of free
1354 * space in this metaslab (i.e. the total free space in all buckets) gives
1355 * us the fragmentation metric. This means that a high fragmentation metric
1356 * equates to most of the free space being comprised of small segments.
1357 * Conversely, if the metric is low, then most of the free space is in
1358 * large segments. A 10% change in fragmentation equates to approximately
1359 * double the number of segments.
1361 * This table defines 0% fragmented space using 16MB segments. Testing has
1362 * shown that segments that are greater than or equal to 16MB do not suffer
1363 * from drastic performance problems. Using this value, we derive the rest
1364 * of the table. Since the fragmentation value is never stored on disk, it
1365 * is possible to change these calculations in the future.
1367 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1387 * Calclate the metaslab's fragmentation metric. A return value
1388 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1389 * not support this metric. Otherwise, the return value should be in the
1393 metaslab_fragmentation(metaslab_t
*msp
)
1395 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1396 uint64_t fragmentation
= 0;
1398 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1399 SPA_FEATURE_SPACEMAP_HISTOGRAM
);
1402 if (!feature_enabled
)
1403 return (ZFS_FRAG_INVALID
);
1406 * A null space map means that the entire metaslab is free
1407 * and thus is not fragmented.
1409 if (msp
->ms_sm
== NULL
)
1413 * If this metaslab's space_map has not been upgraded, flag it
1414 * so that we upgrade next time we encounter it.
1416 if (msp
->ms_sm
->sm_dbuf
->db_size
!= sizeof (space_map_phys_t
)) {
1417 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1419 if (spa_writeable(vd
->vdev_spa
)) {
1420 uint64_t txg
= spa_syncing_txg(spa
);
1422 msp
->ms_condense_wanted
= B_TRUE
;
1423 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1424 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1425 "msp %p, vd %p", txg
, msp
, vd
);
1427 return (ZFS_FRAG_INVALID
);
1430 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1432 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1433 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1434 FRAGMENTATION_TABLE_SIZE
- 1);
1436 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1439 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1442 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1443 fragmentation
+= space
* zfs_frag_table
[idx
];
1447 fragmentation
/= total
;
1448 ASSERT3U(fragmentation
, <=, 100);
1449 return (fragmentation
);
1453 * Compute a weight -- a selection preference value -- for the given metaslab.
1454 * This is based on the amount of free space, the level of fragmentation,
1455 * the LBA range, and whether the metaslab is loaded.
1458 metaslab_weight(metaslab_t
*msp
)
1460 metaslab_group_t
*mg
= msp
->ms_group
;
1461 vdev_t
*vd
= mg
->mg_vd
;
1462 uint64_t weight
, space
;
1464 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1467 * This vdev is in the process of being removed so there is nothing
1468 * for us to do here.
1470 if (vd
->vdev_removing
) {
1471 ASSERT0(space_map_allocated(msp
->ms_sm
));
1472 ASSERT0(vd
->vdev_ms_shift
);
1477 * The baseline weight is the metaslab's free space.
1479 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1481 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1482 if (metaslab_fragmentation_factor_enabled
&&
1483 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1485 * Use the fragmentation information to inversely scale
1486 * down the baseline weight. We need to ensure that we
1487 * don't exclude this metaslab completely when it's 100%
1488 * fragmented. To avoid this we reduce the fragmented value
1491 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1494 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1495 * this metaslab again. The fragmentation metric may have
1496 * decreased the space to something smaller than
1497 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1498 * so that we can consume any remaining space.
1500 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1501 space
= SPA_MINBLOCKSIZE
;
1506 * Modern disks have uniform bit density and constant angular velocity.
1507 * Therefore, the outer recording zones are faster (higher bandwidth)
1508 * than the inner zones by the ratio of outer to inner track diameter,
1509 * which is typically around 2:1. We account for this by assigning
1510 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1511 * In effect, this means that we'll select the metaslab with the most
1512 * free bandwidth rather than simply the one with the most free space.
1514 if (metaslab_lba_weighting_enabled
) {
1515 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1516 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1520 * If this metaslab is one we're actively using, adjust its
1521 * weight to make it preferable to any inactive metaslab so
1522 * we'll polish it off. If the fragmentation on this metaslab
1523 * has exceed our threshold, then don't mark it active.
1525 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1526 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1527 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1534 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1536 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1538 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1539 metaslab_load_wait(msp
);
1540 if (!msp
->ms_loaded
) {
1541 int error
= metaslab_load(msp
);
1543 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1548 metaslab_group_sort(msp
->ms_group
, msp
,
1549 msp
->ms_weight
| activation_weight
);
1551 ASSERT(msp
->ms_loaded
);
1552 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1558 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
1561 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1562 * this metaslab again. In that case, it had better be empty,
1563 * or we would be leaving space on the table.
1565 ASSERT(size
>= SPA_MINBLOCKSIZE
|| range_tree_space(msp
->ms_tree
) == 0);
1566 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
1567 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1571 metaslab_preload(void *arg
)
1573 metaslab_t
*msp
= arg
;
1574 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1576 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1578 mutex_enter(&msp
->ms_lock
);
1579 metaslab_load_wait(msp
);
1580 if (!msp
->ms_loaded
)
1581 (void) metaslab_load(msp
);
1584 * Set the ms_access_txg value so that we don't unload it right away.
1586 msp
->ms_access_txg
= spa_syncing_txg(spa
) + metaslab_unload_delay
+ 1;
1587 mutex_exit(&msp
->ms_lock
);
1591 metaslab_group_preload(metaslab_group_t
*mg
)
1593 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1595 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1598 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
1599 taskq_wait_outstanding(mg
->mg_taskq
, 0);
1603 mutex_enter(&mg
->mg_lock
);
1605 * Load the next potential metaslabs
1608 while (msp
!= NULL
) {
1609 metaslab_t
*msp_next
= AVL_NEXT(t
, msp
);
1612 * We preload only the maximum number of metaslabs specified
1613 * by metaslab_preload_limit. If a metaslab is being forced
1614 * to condense then we preload it too. This will ensure
1615 * that force condensing happens in the next txg.
1617 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
1623 * We must drop the metaslab group lock here to preserve
1624 * lock ordering with the ms_lock (when grabbing both
1625 * the mg_lock and the ms_lock, the ms_lock must be taken
1626 * first). As a result, it is possible that the ordering
1627 * of the metaslabs within the avl tree may change before
1628 * we reacquire the lock. The metaslab cannot be removed from
1629 * the tree while we're in syncing context so it is safe to
1630 * drop the mg_lock here. If the metaslabs are reordered
1631 * nothing will break -- we just may end up loading a
1632 * less than optimal one.
1634 mutex_exit(&mg
->mg_lock
);
1635 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
1636 msp
, TQ_SLEEP
) != 0);
1637 mutex_enter(&mg
->mg_lock
);
1640 mutex_exit(&mg
->mg_lock
);
1644 * Determine if the space map's on-disk footprint is past our tolerance
1645 * for inefficiency. We would like to use the following criteria to make
1648 * 1. The size of the space map object should not dramatically increase as a
1649 * result of writing out the free space range tree.
1651 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1652 * times the size than the free space range tree representation
1653 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1655 * 3. The on-disk size of the space map should actually decrease.
1657 * Checking the first condition is tricky since we don't want to walk
1658 * the entire AVL tree calculating the estimated on-disk size. Instead we
1659 * use the size-ordered range tree in the metaslab and calculate the
1660 * size required to write out the largest segment in our free tree. If the
1661 * size required to represent that segment on disk is larger than the space
1662 * map object then we avoid condensing this map.
1664 * To determine the second criterion we use a best-case estimate and assume
1665 * each segment can be represented on-disk as a single 64-bit entry. We refer
1666 * to this best-case estimate as the space map's minimal form.
1668 * Unfortunately, we cannot compute the on-disk size of the space map in this
1669 * context because we cannot accurately compute the effects of compression, etc.
1670 * Instead, we apply the heuristic described in the block comment for
1671 * zfs_metaslab_condense_block_threshold - we only condense if the space used
1672 * is greater than a threshold number of blocks.
1675 metaslab_should_condense(metaslab_t
*msp
)
1677 space_map_t
*sm
= msp
->ms_sm
;
1679 uint64_t size
, entries
, segsz
, object_size
, optimal_size
, record_size
;
1680 dmu_object_info_t doi
;
1681 uint64_t vdev_blocksize
= 1 << msp
->ms_group
->mg_vd
->vdev_ashift
;
1683 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1684 ASSERT(msp
->ms_loaded
);
1687 * Use the ms_size_tree range tree, which is ordered by size, to
1688 * obtain the largest segment in the free tree. We always condense
1689 * metaslabs that are empty and metaslabs for which a condense
1690 * request has been made.
1692 rs
= avl_last(&msp
->ms_size_tree
);
1693 if (rs
== NULL
|| msp
->ms_condense_wanted
)
1697 * Calculate the number of 64-bit entries this segment would
1698 * require when written to disk. If this single segment would be
1699 * larger on-disk than the entire current on-disk structure, then
1700 * clearly condensing will increase the on-disk structure size.
1702 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
1703 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
1704 segsz
= entries
* sizeof (uint64_t);
1706 optimal_size
= sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
);
1707 object_size
= space_map_length(msp
->ms_sm
);
1709 dmu_object_info_from_db(sm
->sm_dbuf
, &doi
);
1710 record_size
= MAX(doi
.doi_data_block_size
, vdev_blocksize
);
1712 return (segsz
<= object_size
&&
1713 object_size
>= (optimal_size
* zfs_condense_pct
/ 100) &&
1714 object_size
> zfs_metaslab_condense_block_threshold
* record_size
);
1718 * Condense the on-disk space map representation to its minimized form.
1719 * The minimized form consists of a small number of allocations followed by
1720 * the entries of the free range tree.
1723 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
1725 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1726 range_tree_t
*freetree
= msp
->ms_freetree
[txg
& TXG_MASK
];
1727 range_tree_t
*condense_tree
;
1728 space_map_t
*sm
= msp
->ms_sm
;
1731 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1732 ASSERT3U(spa_sync_pass(spa
), ==, 1);
1733 ASSERT(msp
->ms_loaded
);
1736 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, "
1737 "smp size %llu, segments %lu, forcing condense=%s", txg
,
1738 msp
->ms_id
, msp
, space_map_length(msp
->ms_sm
),
1739 avl_numnodes(&msp
->ms_tree
->rt_root
),
1740 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
1742 msp
->ms_condense_wanted
= B_FALSE
;
1745 * Create an range tree that is 100% allocated. We remove segments
1746 * that have been freed in this txg, any deferred frees that exist,
1747 * and any allocation in the future. Removing segments should be
1748 * a relatively inexpensive operation since we expect these trees to
1749 * have a small number of nodes.
1751 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
1752 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
1755 * Remove what's been freed in this txg from the condense_tree.
1756 * Since we're in sync_pass 1, we know that all the frees from
1757 * this txg are in the freetree.
1759 range_tree_walk(freetree
, range_tree_remove
, condense_tree
);
1761 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1762 range_tree_walk(msp
->ms_defertree
[t
],
1763 range_tree_remove
, condense_tree
);
1766 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1767 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
1768 range_tree_remove
, condense_tree
);
1772 * We're about to drop the metaslab's lock thus allowing
1773 * other consumers to change it's content. Set the
1774 * metaslab's ms_condensing flag to ensure that
1775 * allocations on this metaslab do not occur while we're
1776 * in the middle of committing it to disk. This is only critical
1777 * for the ms_tree as all other range trees use per txg
1778 * views of their content.
1780 msp
->ms_condensing
= B_TRUE
;
1782 mutex_exit(&msp
->ms_lock
);
1783 space_map_truncate(sm
, tx
);
1784 mutex_enter(&msp
->ms_lock
);
1787 * While we would ideally like to create a space_map representation
1788 * that consists only of allocation records, doing so can be
1789 * prohibitively expensive because the in-core free tree can be
1790 * large, and therefore computationally expensive to subtract
1791 * from the condense_tree. Instead we sync out two trees, a cheap
1792 * allocation only tree followed by the in-core free tree. While not
1793 * optimal, this is typically close to optimal, and much cheaper to
1796 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
1797 range_tree_vacate(condense_tree
, NULL
, NULL
);
1798 range_tree_destroy(condense_tree
);
1800 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
1801 msp
->ms_condensing
= B_FALSE
;
1805 * Write a metaslab to disk in the context of the specified transaction group.
1808 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
1810 metaslab_group_t
*mg
= msp
->ms_group
;
1811 vdev_t
*vd
= mg
->mg_vd
;
1812 spa_t
*spa
= vd
->vdev_spa
;
1813 objset_t
*mos
= spa_meta_objset(spa
);
1814 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
1815 range_tree_t
**freetree
= &msp
->ms_freetree
[txg
& TXG_MASK
];
1816 range_tree_t
**freed_tree
=
1817 &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1819 uint64_t object
= space_map_object(msp
->ms_sm
);
1821 ASSERT(!vd
->vdev_ishole
);
1824 * This metaslab has just been added so there's no work to do now.
1826 if (*freetree
== NULL
) {
1827 ASSERT3P(alloctree
, ==, NULL
);
1831 ASSERT3P(alloctree
, !=, NULL
);
1832 ASSERT3P(*freetree
, !=, NULL
);
1833 ASSERT3P(*freed_tree
, !=, NULL
);
1836 * Normally, we don't want to process a metaslab if there
1837 * are no allocations or frees to perform. However, if the metaslab
1838 * is being forced to condense we need to let it through.
1840 if (range_tree_space(alloctree
) == 0 &&
1841 range_tree_space(*freetree
) == 0 &&
1842 !msp
->ms_condense_wanted
)
1846 * The only state that can actually be changing concurrently with
1847 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1848 * be modifying this txg's alloctree, freetree, freed_tree, or
1849 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1850 * space_map ASSERTs. We drop it whenever we call into the DMU,
1851 * because the DMU can call down to us (e.g. via zio_free()) at
1855 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
1857 if (msp
->ms_sm
== NULL
) {
1858 uint64_t new_object
;
1860 new_object
= space_map_alloc(mos
, tx
);
1861 VERIFY3U(new_object
, !=, 0);
1863 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
1864 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
1866 ASSERT(msp
->ms_sm
!= NULL
);
1869 mutex_enter(&msp
->ms_lock
);
1872 * Note: metaslab_condense() clears the space_map's histogram.
1873 * Therefore we muse verify and remove this histogram before
1876 metaslab_group_histogram_verify(mg
);
1877 metaslab_class_histogram_verify(mg
->mg_class
);
1878 metaslab_group_histogram_remove(mg
, msp
);
1880 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
1881 metaslab_should_condense(msp
)) {
1882 metaslab_condense(msp
, txg
, tx
);
1884 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
1885 space_map_write(msp
->ms_sm
, *freetree
, SM_FREE
, tx
);
1888 if (msp
->ms_loaded
) {
1890 * When the space map is loaded, we have an accruate
1891 * histogram in the range tree. This gives us an opportunity
1892 * to bring the space map's histogram up-to-date so we clear
1893 * it first before updating it.
1895 space_map_histogram_clear(msp
->ms_sm
);
1896 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
1899 * Since the space map is not loaded we simply update the
1900 * exisiting histogram with what was freed in this txg. This
1901 * means that the on-disk histogram may not have an accurate
1902 * view of the free space but it's close enough to allow
1903 * us to make allocation decisions.
1905 space_map_histogram_add(msp
->ms_sm
, *freetree
, tx
);
1907 metaslab_group_histogram_add(mg
, msp
);
1908 metaslab_group_histogram_verify(mg
);
1909 metaslab_class_histogram_verify(mg
->mg_class
);
1912 * For sync pass 1, we avoid traversing this txg's free range tree
1913 * and instead will just swap the pointers for freetree and
1914 * freed_tree. We can safely do this since the freed_tree is
1915 * guaranteed to be empty on the initial pass.
1917 if (spa_sync_pass(spa
) == 1) {
1918 range_tree_swap(freetree
, freed_tree
);
1920 range_tree_vacate(*freetree
, range_tree_add
, *freed_tree
);
1922 range_tree_vacate(alloctree
, NULL
, NULL
);
1924 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1925 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1927 mutex_exit(&msp
->ms_lock
);
1929 if (object
!= space_map_object(msp
->ms_sm
)) {
1930 object
= space_map_object(msp
->ms_sm
);
1931 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
1932 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
1938 * Called after a transaction group has completely synced to mark
1939 * all of the metaslab's free space as usable.
1942 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1944 metaslab_group_t
*mg
= msp
->ms_group
;
1945 vdev_t
*vd
= mg
->mg_vd
;
1946 range_tree_t
**freed_tree
;
1947 range_tree_t
**defer_tree
;
1948 int64_t alloc_delta
, defer_delta
;
1951 ASSERT(!vd
->vdev_ishole
);
1953 mutex_enter(&msp
->ms_lock
);
1956 * If this metaslab is just becoming available, initialize its
1957 * alloctrees, freetrees, and defertree and add its capacity to
1960 if (msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
] == NULL
) {
1961 for (t
= 0; t
< TXG_SIZE
; t
++) {
1962 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
1963 ASSERT(msp
->ms_freetree
[t
] == NULL
);
1965 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
1967 msp
->ms_freetree
[t
] = range_tree_create(NULL
, msp
,
1971 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1972 ASSERT(msp
->ms_defertree
[t
] == NULL
);
1974 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
1978 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
1981 freed_tree
= &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1982 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
1984 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
1985 defer_delta
= range_tree_space(*freed_tree
) -
1986 range_tree_space(*defer_tree
);
1988 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1990 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1991 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1994 * If there's a metaslab_load() in progress, wait for it to complete
1995 * so that we have a consistent view of the in-core space map.
1997 metaslab_load_wait(msp
);
2000 * Move the frees from the defer_tree back to the free
2001 * range tree (if it's loaded). Swap the freed_tree and the
2002 * defer_tree -- this is safe to do because we've just emptied out
2005 range_tree_vacate(*defer_tree
,
2006 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
2007 range_tree_swap(freed_tree
, defer_tree
);
2009 space_map_update(msp
->ms_sm
);
2011 msp
->ms_deferspace
+= defer_delta
;
2012 ASSERT3S(msp
->ms_deferspace
, >=, 0);
2013 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
2014 if (msp
->ms_deferspace
!= 0) {
2016 * Keep syncing this metaslab until all deferred frees
2017 * are back in circulation.
2019 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
2022 if (msp
->ms_loaded
&& msp
->ms_access_txg
< txg
) {
2023 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
2024 VERIFY0(range_tree_space(
2025 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
2028 if (!metaslab_debug_unload
)
2029 metaslab_unload(msp
);
2032 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
2033 mutex_exit(&msp
->ms_lock
);
2037 metaslab_sync_reassess(metaslab_group_t
*mg
)
2039 metaslab_group_alloc_update(mg
);
2040 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2043 * Preload the next potential metaslabs
2045 metaslab_group_preload(mg
);
2049 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2051 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2052 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2053 uint64_t start
= msp
->ms_id
;
2055 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2056 return (1ULL << 63);
2059 return ((start
- offset
) << ms_shift
);
2061 return ((offset
- start
) << ms_shift
);
2066 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
2067 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2069 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2070 metaslab_t
*msp
= NULL
;
2071 uint64_t offset
= -1ULL;
2072 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2073 uint64_t activation_weight
;
2074 uint64_t target_distance
;
2077 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2078 for (i
= 0; i
< d
; i
++) {
2079 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2080 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2086 boolean_t was_active
;
2088 mutex_enter(&mg
->mg_lock
);
2089 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
2090 if (msp
->ms_weight
< asize
) {
2091 spa_dbgmsg(spa
, "%s: failed to meet weight "
2092 "requirement: vdev %llu, txg %llu, mg %p, "
2093 "msp %p, psize %llu, asize %llu, "
2094 "weight %llu", spa_name(spa
),
2095 mg
->mg_vd
->vdev_id
, txg
,
2096 mg
, msp
, psize
, asize
, msp
->ms_weight
);
2097 mutex_exit(&mg
->mg_lock
);
2102 * If the selected metaslab is condensing, skip it.
2104 if (msp
->ms_condensing
)
2107 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2108 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2111 target_distance
= min_distance
+
2112 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2115 for (i
= 0; i
< d
; i
++)
2116 if (metaslab_distance(msp
, &dva
[i
]) <
2122 mutex_exit(&mg
->mg_lock
);
2126 mutex_enter(&msp
->ms_lock
);
2129 * Ensure that the metaslab we have selected is still
2130 * capable of handling our request. It's possible that
2131 * another thread may have changed the weight while we
2132 * were blocked on the metaslab lock.
2134 if (msp
->ms_weight
< asize
|| (was_active
&&
2135 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
2136 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
2137 mutex_exit(&msp
->ms_lock
);
2141 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2142 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2143 metaslab_passivate(msp
,
2144 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2145 mutex_exit(&msp
->ms_lock
);
2149 if (metaslab_activate(msp
, activation_weight
) != 0) {
2150 mutex_exit(&msp
->ms_lock
);
2155 * If this metaslab is currently condensing then pick again as
2156 * we can't manipulate this metaslab until it's committed
2159 if (msp
->ms_condensing
) {
2160 mutex_exit(&msp
->ms_lock
);
2164 if ((offset
= metaslab_block_alloc(msp
, asize
)) != -1ULL)
2167 metaslab_passivate(msp
, metaslab_block_maxsize(msp
));
2168 mutex_exit(&msp
->ms_lock
);
2171 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2172 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2174 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, asize
);
2175 msp
->ms_access_txg
= txg
+ metaslab_unload_delay
;
2177 mutex_exit(&msp
->ms_lock
);
2183 * Allocate a block for the specified i/o.
2186 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2187 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
2189 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
2193 int zio_lock
= B_FALSE
;
2194 boolean_t allocatable
;
2195 uint64_t offset
= -1ULL;
2199 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2202 * For testing, make some blocks above a certain size be gang blocks.
2204 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
2205 return (SET_ERROR(ENOSPC
));
2207 if (flags
& METASLAB_FASTWRITE
)
2208 mutex_enter(&mc
->mc_fastwrite_lock
);
2211 * Start at the rotor and loop through all mgs until we find something.
2212 * Note that there's no locking on mc_rotor or mc_aliquot because
2213 * nothing actually breaks if we miss a few updates -- we just won't
2214 * allocate quite as evenly. It all balances out over time.
2216 * If we are doing ditto or log blocks, try to spread them across
2217 * consecutive vdevs. If we're forced to reuse a vdev before we've
2218 * allocated all of our ditto blocks, then try and spread them out on
2219 * that vdev as much as possible. If it turns out to not be possible,
2220 * gradually lower our standards until anything becomes acceptable.
2221 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2222 * gives us hope of containing our fault domains to something we're
2223 * able to reason about. Otherwise, any two top-level vdev failures
2224 * will guarantee the loss of data. With consecutive allocation,
2225 * only two adjacent top-level vdev failures will result in data loss.
2227 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2228 * ourselves on the same vdev as our gang block header. That
2229 * way, we can hope for locality in vdev_cache, plus it makes our
2230 * fault domains something tractable.
2233 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
2236 * It's possible the vdev we're using as the hint no
2237 * longer exists (i.e. removed). Consult the rotor when
2243 if (flags
& METASLAB_HINTBP_AVOID
&&
2244 mg
->mg_next
!= NULL
)
2249 } else if (d
!= 0) {
2250 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
2251 mg
= vd
->vdev_mg
->mg_next
;
2252 } else if (flags
& METASLAB_FASTWRITE
) {
2253 mg
= fast_mg
= mc
->mc_rotor
;
2256 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
2257 mg
->mg_vd
->vdev_pending_fastwrite
)
2259 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
2266 * If the hint put us into the wrong metaslab class, or into a
2267 * metaslab group that has been passivated, just follow the rotor.
2269 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
2276 ASSERT(mg
->mg_activation_count
== 1);
2281 * Don't allocate from faulted devices.
2284 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
2285 allocatable
= vdev_allocatable(vd
);
2286 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
2288 allocatable
= vdev_allocatable(vd
);
2292 * Determine if the selected metaslab group is eligible
2293 * for allocations. If we're ganging or have requested
2294 * an allocation for the smallest gang block size
2295 * then we don't want to avoid allocating to the this
2296 * metaslab group. If we're in this condition we should
2297 * try to allocate from any device possible so that we
2298 * don't inadvertently return ENOSPC and suspend the pool
2299 * even though space is still available.
2301 if (allocatable
&& CAN_FASTGANG(flags
) &&
2302 psize
> SPA_GANGBLOCKSIZE
)
2303 allocatable
= metaslab_group_allocatable(mg
);
2309 * Avoid writing single-copy data to a failing vdev
2310 * unless the user instructs us that it is okay.
2312 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
2313 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
2314 d
== 0 && dshift
== 3 && vd
->vdev_children
== 0) {
2319 ASSERT(mg
->mg_class
== mc
);
2321 distance
= vd
->vdev_asize
>> dshift
;
2322 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
2327 asize
= vdev_psize_to_asize(vd
, psize
);
2328 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
2330 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
2332 if (offset
!= -1ULL) {
2334 * If we've just selected this metaslab group,
2335 * figure out whether the corresponding vdev is
2336 * over- or under-used relative to the pool,
2337 * and set an allocation bias to even it out.
2339 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
2340 vdev_stat_t
*vs
= &vd
->vdev_stat
;
2343 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
2344 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
2347 * Calculate how much more or less we should
2348 * try to allocate from this device during
2349 * this iteration around the rotor.
2350 * For example, if a device is 80% full
2351 * and the pool is 20% full then we should
2352 * reduce allocations by 60% on this device.
2354 * mg_bias = (20 - 80) * 512K / 100 = -307K
2356 * This reduces allocations by 307K for this
2359 mg
->mg_bias
= ((cu
- vu
) *
2360 (int64_t)mg
->mg_aliquot
) / 100;
2361 } else if (!metaslab_bias_enabled
) {
2365 if ((flags
& METASLAB_FASTWRITE
) ||
2366 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
2367 mg
->mg_aliquot
+ mg
->mg_bias
) {
2368 mc
->mc_rotor
= mg
->mg_next
;
2372 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
2373 DVA_SET_OFFSET(&dva
[d
], offset
);
2374 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
2375 DVA_SET_ASIZE(&dva
[d
], asize
);
2377 if (flags
& METASLAB_FASTWRITE
) {
2378 atomic_add_64(&vd
->vdev_pending_fastwrite
,
2380 mutex_exit(&mc
->mc_fastwrite_lock
);
2386 mc
->mc_rotor
= mg
->mg_next
;
2388 } while ((mg
= mg
->mg_next
) != rotor
);
2392 ASSERT(dshift
< 64);
2396 if (!allocatable
&& !zio_lock
) {
2402 bzero(&dva
[d
], sizeof (dva_t
));
2404 if (flags
& METASLAB_FASTWRITE
)
2405 mutex_exit(&mc
->mc_fastwrite_lock
);
2407 return (SET_ERROR(ENOSPC
));
2411 * Free the block represented by DVA in the context of the specified
2412 * transaction group.
2415 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
2417 uint64_t vdev
= DVA_GET_VDEV(dva
);
2418 uint64_t offset
= DVA_GET_OFFSET(dva
);
2419 uint64_t size
= DVA_GET_ASIZE(dva
);
2423 if (txg
> spa_freeze_txg(spa
))
2426 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
|| !DVA_IS_VALID(dva
) ||
2427 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
2428 zfs_panic_recover("metaslab_free_dva(): bad DVA %llu:%llu:%llu",
2429 (u_longlong_t
)vdev
, (u_longlong_t
)offset
,
2430 (u_longlong_t
)size
);
2434 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2436 if (DVA_GET_GANG(dva
))
2437 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2439 mutex_enter(&msp
->ms_lock
);
2442 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
2445 VERIFY(!msp
->ms_condensing
);
2446 VERIFY3U(offset
, >=, msp
->ms_start
);
2447 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
2448 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
2450 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2451 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2452 range_tree_add(msp
->ms_tree
, offset
, size
);
2454 if (range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]) == 0)
2455 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2456 range_tree_add(msp
->ms_freetree
[txg
& TXG_MASK
],
2460 mutex_exit(&msp
->ms_lock
);
2464 * Intent log support: upon opening the pool after a crash, notify the SPA
2465 * of blocks that the intent log has allocated for immediate write, but
2466 * which are still considered free by the SPA because the last transaction
2467 * group didn't commit yet.
2470 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
2472 uint64_t vdev
= DVA_GET_VDEV(dva
);
2473 uint64_t offset
= DVA_GET_OFFSET(dva
);
2474 uint64_t size
= DVA_GET_ASIZE(dva
);
2479 ASSERT(DVA_IS_VALID(dva
));
2481 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2482 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
2483 return (SET_ERROR(ENXIO
));
2485 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2487 if (DVA_GET_GANG(dva
))
2488 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2490 mutex_enter(&msp
->ms_lock
);
2492 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
2493 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
2495 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
2496 error
= SET_ERROR(ENOENT
);
2498 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
2499 mutex_exit(&msp
->ms_lock
);
2503 VERIFY(!msp
->ms_condensing
);
2504 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2505 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2506 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
2507 range_tree_remove(msp
->ms_tree
, offset
, size
);
2509 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
2510 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2511 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2512 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
2515 mutex_exit(&msp
->ms_lock
);
2521 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
2522 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
2524 dva_t
*dva
= bp
->blk_dva
;
2525 dva_t
*hintdva
= hintbp
->blk_dva
;
2528 ASSERT(bp
->blk_birth
== 0);
2529 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
2531 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2533 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
2534 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2535 return (SET_ERROR(ENOSPC
));
2538 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
2539 ASSERT(BP_GET_NDVAS(bp
) == 0);
2540 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
2542 for (d
= 0; d
< ndvas
; d
++) {
2543 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
2546 for (d
--; d
>= 0; d
--) {
2547 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
2548 bzero(&dva
[d
], sizeof (dva_t
));
2550 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2555 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
2557 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2559 BP_SET_BIRTH(bp
, txg
, txg
);
2565 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
2567 const dva_t
*dva
= bp
->blk_dva
;
2568 int d
, ndvas
= BP_GET_NDVAS(bp
);
2570 ASSERT(!BP_IS_HOLE(bp
));
2571 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
2573 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
2575 for (d
= 0; d
< ndvas
; d
++)
2576 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
2578 spa_config_exit(spa
, SCL_FREE
, FTAG
);
2582 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
2584 const dva_t
*dva
= bp
->blk_dva
;
2585 int ndvas
= BP_GET_NDVAS(bp
);
2588 ASSERT(!BP_IS_HOLE(bp
));
2592 * First do a dry run to make sure all DVAs are claimable,
2593 * so we don't have to unwind from partial failures below.
2595 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
2599 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2601 for (d
= 0; d
< ndvas
; d
++)
2602 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
2605 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2607 ASSERT(error
== 0 || txg
== 0);
2613 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
2615 const dva_t
*dva
= bp
->blk_dva
;
2616 int ndvas
= BP_GET_NDVAS(bp
);
2617 uint64_t psize
= BP_GET_PSIZE(bp
);
2621 ASSERT(!BP_IS_HOLE(bp
));
2622 ASSERT(!BP_IS_EMBEDDED(bp
));
2625 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2627 for (d
= 0; d
< ndvas
; d
++) {
2628 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2630 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
2633 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2637 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
2639 const dva_t
*dva
= bp
->blk_dva
;
2640 int ndvas
= BP_GET_NDVAS(bp
);
2641 uint64_t psize
= BP_GET_PSIZE(bp
);
2645 ASSERT(!BP_IS_HOLE(bp
));
2646 ASSERT(!BP_IS_EMBEDDED(bp
));
2649 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2651 for (d
= 0; d
< ndvas
; d
++) {
2652 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2654 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
2655 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
2658 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2662 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
2666 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
2669 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2670 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2671 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
2672 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
2673 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
2674 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2675 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2678 range_tree_verify(msp
->ms_tree
, offset
, size
);
2680 for (j
= 0; j
< TXG_SIZE
; j
++)
2681 range_tree_verify(msp
->ms_freetree
[j
], offset
, size
);
2682 for (j
= 0; j
< TXG_DEFER_SIZE
; j
++)
2683 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
2685 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2688 #if defined(_KERNEL) && defined(HAVE_SPL)
2689 module_param(metaslab_debug_load
, int, 0644);
2690 module_param(metaslab_debug_unload
, int, 0644);
2691 module_param(metaslab_preload_enabled
, int, 0644);
2692 module_param(zfs_mg_noalloc_threshold
, int, 0644);
2693 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
2694 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
2695 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
2696 module_param(metaslab_lba_weighting_enabled
, int, 0644);
2697 module_param(metaslab_bias_enabled
, int, 0644);
2699 MODULE_PARM_DESC(metaslab_debug_load
,
2700 "load all metaslabs when pool is first opened");
2701 MODULE_PARM_DESC(metaslab_debug_unload
,
2702 "prevent metaslabs from being unloaded");
2703 MODULE_PARM_DESC(metaslab_preload_enabled
,
2704 "preload potential metaslabs during reassessment");
2706 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
2707 "percentage of free space for metaslab group to allow allocation");
2708 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
2709 "fragmentation for metaslab group to allow allocation");
2711 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
2712 "fragmentation for metaslab to allow allocation");
2713 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
2714 "use the fragmentation metric to prefer less fragmented metaslabs");
2715 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
2716 "prefer metaslabs with lower LBAs");
2717 MODULE_PARM_DESC(metaslab_bias_enabled
,
2718 "enable metaslab group biasing");
2719 #endif /* _KERNEL && HAVE_SPL */