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 * The zfs_mg_noalloc_threshold defines which metaslab groups should
69 * be eligible for allocation. The value is defined as a percentage of
70 * free space. Metaslab groups that have more free space than
71 * zfs_mg_noalloc_threshold are always eligible for allocations. Once
72 * a metaslab group's free space is less than or equal to the
73 * zfs_mg_noalloc_threshold the allocator will avoid allocating to that
74 * group unless all groups in the pool have reached zfs_mg_noalloc_threshold.
75 * Once all groups in the pool reach zfs_mg_noalloc_threshold then all
76 * groups are allowed to accept allocations. Gang blocks are always
77 * eligible to allocate on any metaslab group. The default value of 0 means
78 * no metaslab group will be excluded based on this criterion.
80 int zfs_mg_noalloc_threshold
= 0;
83 * Metaslab groups are considered eligible for allocations if their
84 * fragmenation metric (measured as a percentage) is less than or equal to
85 * zfs_mg_fragmentation_threshold. If a metaslab group exceeds this threshold
86 * then it will be skipped unless all metaslab groups within the metaslab
87 * class have also crossed this threshold.
89 int zfs_mg_fragmentation_threshold
= 85;
92 * Allow metaslabs to keep their active state as long as their fragmentation
93 * percentage is less than or equal to zfs_metaslab_fragmentation_threshold. An
94 * active metaslab that exceeds this threshold will no longer keep its active
95 * status allowing better metaslabs to be selected.
97 int zfs_metaslab_fragmentation_threshold
= 70;
100 * When set will load all metaslabs when pool is first opened.
102 int metaslab_debug_load
= 0;
105 * When set will prevent metaslabs from being unloaded.
107 int metaslab_debug_unload
= 0;
110 * Minimum size which forces the dynamic allocator to change
111 * it's allocation strategy. Once the space map cannot satisfy
112 * an allocation of this size then it switches to using more
113 * aggressive strategy (i.e search by size rather than offset).
115 uint64_t metaslab_df_alloc_threshold
= SPA_MAXBLOCKSIZE
;
118 * The minimum free space, in percent, which must be available
119 * in a space map to continue allocations in a first-fit fashion.
120 * Once the space_map's free space drops below this level we dynamically
121 * switch to using best-fit allocations.
123 int metaslab_df_free_pct
= 4;
126 * A metaslab is considered "free" if it contains a contiguous
127 * segment which is greater than metaslab_min_alloc_size.
129 uint64_t metaslab_min_alloc_size
= DMU_MAX_ACCESS
;
132 * Percentage of all cpus that can be used by the metaslab taskq.
134 int metaslab_load_pct
= 50;
137 * Determines how many txgs a metaslab may remain loaded without having any
138 * allocations from it. As long as a metaslab continues to be used we will
141 int metaslab_unload_delay
= TXG_SIZE
* 2;
144 * Max number of metaslabs per group to preload.
146 int metaslab_preload_limit
= SPA_DVAS_PER_BP
;
149 * Enable/disable preloading of metaslab.
151 int metaslab_preload_enabled
= B_TRUE
;
154 * Enable/disable fragmentation weighting on metaslabs.
156 int metaslab_fragmentation_factor_enabled
= B_TRUE
;
159 * Enable/disable lba weighting (i.e. outer tracks are given preference).
161 int metaslab_lba_weighting_enabled
= B_TRUE
;
164 * Enable/disable metaslab group biasing.
166 int metaslab_bias_enabled
= B_TRUE
;
168 static uint64_t metaslab_fragmentation(metaslab_t
*);
171 * ==========================================================================
173 * ==========================================================================
176 metaslab_class_create(spa_t
*spa
, metaslab_ops_t
*ops
)
178 metaslab_class_t
*mc
;
180 mc
= kmem_zalloc(sizeof (metaslab_class_t
), KM_PUSHPAGE
);
185 mutex_init(&mc
->mc_fastwrite_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
191 metaslab_class_destroy(metaslab_class_t
*mc
)
193 ASSERT(mc
->mc_rotor
== NULL
);
194 ASSERT(mc
->mc_alloc
== 0);
195 ASSERT(mc
->mc_deferred
== 0);
196 ASSERT(mc
->mc_space
== 0);
197 ASSERT(mc
->mc_dspace
== 0);
199 mutex_destroy(&mc
->mc_fastwrite_lock
);
200 kmem_free(mc
, sizeof (metaslab_class_t
));
204 metaslab_class_validate(metaslab_class_t
*mc
)
206 metaslab_group_t
*mg
;
210 * Must hold one of the spa_config locks.
212 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_READER
) ||
213 spa_config_held(mc
->mc_spa
, SCL_ALL
, RW_WRITER
));
215 if ((mg
= mc
->mc_rotor
) == NULL
)
220 ASSERT(vd
->vdev_mg
!= NULL
);
221 ASSERT3P(vd
->vdev_top
, ==, vd
);
222 ASSERT3P(mg
->mg_class
, ==, mc
);
223 ASSERT3P(vd
->vdev_ops
, !=, &vdev_hole_ops
);
224 } while ((mg
= mg
->mg_next
) != mc
->mc_rotor
);
230 metaslab_class_space_update(metaslab_class_t
*mc
, int64_t alloc_delta
,
231 int64_t defer_delta
, int64_t space_delta
, int64_t dspace_delta
)
233 atomic_add_64(&mc
->mc_alloc
, alloc_delta
);
234 atomic_add_64(&mc
->mc_deferred
, defer_delta
);
235 atomic_add_64(&mc
->mc_space
, space_delta
);
236 atomic_add_64(&mc
->mc_dspace
, dspace_delta
);
240 metaslab_class_get_alloc(metaslab_class_t
*mc
)
242 return (mc
->mc_alloc
);
246 metaslab_class_get_deferred(metaslab_class_t
*mc
)
248 return (mc
->mc_deferred
);
252 metaslab_class_get_space(metaslab_class_t
*mc
)
254 return (mc
->mc_space
);
258 metaslab_class_get_dspace(metaslab_class_t
*mc
)
260 return (spa_deflate(mc
->mc_spa
) ? mc
->mc_dspace
: mc
->mc_space
);
264 metaslab_class_histogram_verify(metaslab_class_t
*mc
)
266 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
270 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
273 mc_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
276 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
277 vdev_t
*tvd
= rvd
->vdev_child
[c
];
278 metaslab_group_t
*mg
= tvd
->vdev_mg
;
281 * Skip any holes, uninitialized top-levels, or
282 * vdevs that are not in this metalab class.
284 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
285 mg
->mg_class
!= mc
) {
289 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
290 mc_hist
[i
] += mg
->mg_histogram
[i
];
293 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
294 VERIFY3U(mc_hist
[i
], ==, mc
->mc_histogram
[i
]);
296 kmem_free(mc_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
300 * Calculate the metaslab class's fragmentation metric. The metric
301 * is weighted based on the space contribution of each metaslab group.
302 * The return value will be a number between 0 and 100 (inclusive), or
303 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
304 * zfs_frag_table for more information about the metric.
307 metaslab_class_fragmentation(metaslab_class_t
*mc
)
309 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
310 uint64_t fragmentation
= 0;
313 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
315 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
316 vdev_t
*tvd
= rvd
->vdev_child
[c
];
317 metaslab_group_t
*mg
= tvd
->vdev_mg
;
320 * Skip any holes, uninitialized top-levels, or
321 * vdevs that are not in this metalab class.
323 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
324 mg
->mg_class
!= mc
) {
329 * If a metaslab group does not contain a fragmentation
330 * metric then just bail out.
332 if (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
) {
333 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
334 return (ZFS_FRAG_INVALID
);
338 * Determine how much this metaslab_group is contributing
339 * to the overall pool fragmentation metric.
341 fragmentation
+= mg
->mg_fragmentation
*
342 metaslab_group_get_space(mg
);
344 fragmentation
/= metaslab_class_get_space(mc
);
346 ASSERT3U(fragmentation
, <=, 100);
347 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
348 return (fragmentation
);
352 * Calculate the amount of expandable space that is available in
353 * this metaslab class. If a device is expanded then its expandable
354 * space will be the amount of allocatable space that is currently not
355 * part of this metaslab class.
358 metaslab_class_expandable_space(metaslab_class_t
*mc
)
360 vdev_t
*rvd
= mc
->mc_spa
->spa_root_vdev
;
364 spa_config_enter(mc
->mc_spa
, SCL_VDEV
, FTAG
, RW_READER
);
365 for (c
= 0; c
< rvd
->vdev_children
; c
++) {
366 vdev_t
*tvd
= rvd
->vdev_child
[c
];
367 metaslab_group_t
*mg
= tvd
->vdev_mg
;
369 if (tvd
->vdev_ishole
|| tvd
->vdev_ms_shift
== 0 ||
370 mg
->mg_class
!= mc
) {
374 space
+= tvd
->vdev_max_asize
- tvd
->vdev_asize
;
376 spa_config_exit(mc
->mc_spa
, SCL_VDEV
, FTAG
);
381 * ==========================================================================
383 * ==========================================================================
386 metaslab_compare(const void *x1
, const void *x2
)
388 const metaslab_t
*m1
= x1
;
389 const metaslab_t
*m2
= x2
;
391 if (m1
->ms_weight
< m2
->ms_weight
)
393 if (m1
->ms_weight
> m2
->ms_weight
)
397 * If the weights are identical, use the offset to force uniqueness.
399 if (m1
->ms_start
< m2
->ms_start
)
401 if (m1
->ms_start
> m2
->ms_start
)
404 ASSERT3P(m1
, ==, m2
);
410 * Update the allocatable flag and the metaslab group's capacity.
411 * The allocatable flag is set to true if the capacity is below
412 * the zfs_mg_noalloc_threshold. If a metaslab group transitions
413 * from allocatable to non-allocatable or vice versa then the metaslab
414 * group's class is updated to reflect the transition.
417 metaslab_group_alloc_update(metaslab_group_t
*mg
)
419 vdev_t
*vd
= mg
->mg_vd
;
420 metaslab_class_t
*mc
= mg
->mg_class
;
421 vdev_stat_t
*vs
= &vd
->vdev_stat
;
422 boolean_t was_allocatable
;
424 ASSERT(vd
== vd
->vdev_top
);
426 mutex_enter(&mg
->mg_lock
);
427 was_allocatable
= mg
->mg_allocatable
;
429 mg
->mg_free_capacity
= ((vs
->vs_space
- vs
->vs_alloc
) * 100) /
433 * A metaslab group is considered allocatable if it has plenty
434 * of free space or is not heavily fragmented. We only take
435 * fragmentation into account if the metaslab group has a valid
436 * fragmentation metric (i.e. a value between 0 and 100).
438 mg
->mg_allocatable
= (mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
439 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
440 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
));
443 * The mc_alloc_groups maintains a count of the number of
444 * groups in this metaslab class that are still above the
445 * zfs_mg_noalloc_threshold. This is used by the allocating
446 * threads to determine if they should avoid allocations to
447 * a given group. The allocator will avoid allocations to a group
448 * if that group has reached or is below the zfs_mg_noalloc_threshold
449 * and there are still other groups that are above the threshold.
450 * When a group transitions from allocatable to non-allocatable or
451 * vice versa we update the metaslab class to reflect that change.
452 * When the mc_alloc_groups value drops to 0 that means that all
453 * groups have reached the zfs_mg_noalloc_threshold making all groups
454 * eligible for allocations. This effectively means that all devices
455 * are balanced again.
457 if (was_allocatable
&& !mg
->mg_allocatable
)
458 mc
->mc_alloc_groups
--;
459 else if (!was_allocatable
&& mg
->mg_allocatable
)
460 mc
->mc_alloc_groups
++;
462 mutex_exit(&mg
->mg_lock
);
466 metaslab_group_create(metaslab_class_t
*mc
, vdev_t
*vd
)
468 metaslab_group_t
*mg
;
470 mg
= kmem_zalloc(sizeof (metaslab_group_t
), KM_PUSHPAGE
);
471 mutex_init(&mg
->mg_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
472 avl_create(&mg
->mg_metaslab_tree
, metaslab_compare
,
473 sizeof (metaslab_t
), offsetof(struct metaslab
, ms_group_node
));
476 mg
->mg_activation_count
= 0;
478 mg
->mg_taskq
= taskq_create("metaslab_group_taskq", metaslab_load_pct
,
479 minclsyspri
, 10, INT_MAX
, TASKQ_THREADS_CPU_PCT
);
485 metaslab_group_destroy(metaslab_group_t
*mg
)
487 ASSERT(mg
->mg_prev
== NULL
);
488 ASSERT(mg
->mg_next
== NULL
);
490 * We may have gone below zero with the activation count
491 * either because we never activated in the first place or
492 * because we're done, and possibly removing the vdev.
494 ASSERT(mg
->mg_activation_count
<= 0);
496 taskq_destroy(mg
->mg_taskq
);
497 avl_destroy(&mg
->mg_metaslab_tree
);
498 mutex_destroy(&mg
->mg_lock
);
499 kmem_free(mg
, sizeof (metaslab_group_t
));
503 metaslab_group_activate(metaslab_group_t
*mg
)
505 metaslab_class_t
*mc
= mg
->mg_class
;
506 metaslab_group_t
*mgprev
, *mgnext
;
508 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
510 ASSERT(mc
->mc_rotor
!= mg
);
511 ASSERT(mg
->mg_prev
== NULL
);
512 ASSERT(mg
->mg_next
== NULL
);
513 ASSERT(mg
->mg_activation_count
<= 0);
515 if (++mg
->mg_activation_count
<= 0)
518 mg
->mg_aliquot
= metaslab_aliquot
* MAX(1, mg
->mg_vd
->vdev_children
);
519 metaslab_group_alloc_update(mg
);
521 if ((mgprev
= mc
->mc_rotor
) == NULL
) {
525 mgnext
= mgprev
->mg_next
;
526 mg
->mg_prev
= mgprev
;
527 mg
->mg_next
= mgnext
;
528 mgprev
->mg_next
= mg
;
529 mgnext
->mg_prev
= mg
;
535 metaslab_group_passivate(metaslab_group_t
*mg
)
537 metaslab_class_t
*mc
= mg
->mg_class
;
538 metaslab_group_t
*mgprev
, *mgnext
;
540 ASSERT(spa_config_held(mc
->mc_spa
, SCL_ALLOC
, RW_WRITER
));
542 if (--mg
->mg_activation_count
!= 0) {
543 ASSERT(mc
->mc_rotor
!= mg
);
544 ASSERT(mg
->mg_prev
== NULL
);
545 ASSERT(mg
->mg_next
== NULL
);
546 ASSERT(mg
->mg_activation_count
< 0);
550 taskq_wait(mg
->mg_taskq
);
551 metaslab_group_alloc_update(mg
);
553 mgprev
= mg
->mg_prev
;
554 mgnext
= mg
->mg_next
;
559 mc
->mc_rotor
= mgnext
;
560 mgprev
->mg_next
= mgnext
;
561 mgnext
->mg_prev
= mgprev
;
569 metaslab_group_get_space(metaslab_group_t
*mg
)
571 return ((1ULL << mg
->mg_vd
->vdev_ms_shift
) * mg
->mg_vd
->vdev_ms_count
);
575 metaslab_group_histogram_verify(metaslab_group_t
*mg
)
578 vdev_t
*vd
= mg
->mg_vd
;
579 uint64_t ashift
= vd
->vdev_ashift
;
582 if ((zfs_flags
& ZFS_DEBUG_HISTOGRAM_VERIFY
) == 0)
585 mg_hist
= kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
,
588 ASSERT3U(RANGE_TREE_HISTOGRAM_SIZE
, >=,
589 SPACE_MAP_HISTOGRAM_SIZE
+ ashift
);
591 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
592 metaslab_t
*msp
= vd
->vdev_ms
[m
];
594 if (msp
->ms_sm
== NULL
)
597 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++)
598 mg_hist
[i
+ ashift
] +=
599 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
602 for (i
= 0; i
< RANGE_TREE_HISTOGRAM_SIZE
; i
++)
603 VERIFY3U(mg_hist
[i
], ==, mg
->mg_histogram
[i
]);
605 kmem_free(mg_hist
, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE
);
609 metaslab_group_histogram_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
611 metaslab_class_t
*mc
= mg
->mg_class
;
612 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
615 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
616 if (msp
->ms_sm
== NULL
)
619 mutex_enter(&mg
->mg_lock
);
620 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
621 mg
->mg_histogram
[i
+ ashift
] +=
622 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
623 mc
->mc_histogram
[i
+ ashift
] +=
624 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
626 mutex_exit(&mg
->mg_lock
);
630 metaslab_group_histogram_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
632 metaslab_class_t
*mc
= mg
->mg_class
;
633 uint64_t ashift
= mg
->mg_vd
->vdev_ashift
;
636 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
637 if (msp
->ms_sm
== NULL
)
640 mutex_enter(&mg
->mg_lock
);
641 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
642 ASSERT3U(mg
->mg_histogram
[i
+ ashift
], >=,
643 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
644 ASSERT3U(mc
->mc_histogram
[i
+ ashift
], >=,
645 msp
->ms_sm
->sm_phys
->smp_histogram
[i
]);
647 mg
->mg_histogram
[i
+ ashift
] -=
648 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
649 mc
->mc_histogram
[i
+ ashift
] -=
650 msp
->ms_sm
->sm_phys
->smp_histogram
[i
];
652 mutex_exit(&mg
->mg_lock
);
656 metaslab_group_add(metaslab_group_t
*mg
, metaslab_t
*msp
)
658 ASSERT(msp
->ms_group
== NULL
);
659 mutex_enter(&mg
->mg_lock
);
662 avl_add(&mg
->mg_metaslab_tree
, msp
);
663 mutex_exit(&mg
->mg_lock
);
665 mutex_enter(&msp
->ms_lock
);
666 metaslab_group_histogram_add(mg
, msp
);
667 mutex_exit(&msp
->ms_lock
);
671 metaslab_group_remove(metaslab_group_t
*mg
, metaslab_t
*msp
)
673 mutex_enter(&msp
->ms_lock
);
674 metaslab_group_histogram_remove(mg
, msp
);
675 mutex_exit(&msp
->ms_lock
);
677 mutex_enter(&mg
->mg_lock
);
678 ASSERT(msp
->ms_group
== mg
);
679 avl_remove(&mg
->mg_metaslab_tree
, msp
);
680 msp
->ms_group
= NULL
;
681 mutex_exit(&mg
->mg_lock
);
685 metaslab_group_sort(metaslab_group_t
*mg
, metaslab_t
*msp
, uint64_t weight
)
688 * Although in principle the weight can be any value, in
689 * practice we do not use values in the range [1, 511].
691 ASSERT(weight
>= SPA_MINBLOCKSIZE
|| weight
== 0);
692 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
694 mutex_enter(&mg
->mg_lock
);
695 ASSERT(msp
->ms_group
== mg
);
696 avl_remove(&mg
->mg_metaslab_tree
, msp
);
697 msp
->ms_weight
= weight
;
698 avl_add(&mg
->mg_metaslab_tree
, msp
);
699 mutex_exit(&mg
->mg_lock
);
703 * Calculate the fragmentation for a given metaslab group. We can use
704 * a simple average here since all metaslabs within the group must have
705 * the same size. The return value will be a value between 0 and 100
706 * (inclusive), or ZFS_FRAG_INVALID if less than half of the metaslab in this
707 * group have a fragmentation metric.
710 metaslab_group_fragmentation(metaslab_group_t
*mg
)
712 vdev_t
*vd
= mg
->mg_vd
;
713 uint64_t fragmentation
= 0;
714 uint64_t valid_ms
= 0;
717 for (m
= 0; m
< vd
->vdev_ms_count
; m
++) {
718 metaslab_t
*msp
= vd
->vdev_ms
[m
];
720 if (msp
->ms_fragmentation
== ZFS_FRAG_INVALID
)
724 fragmentation
+= msp
->ms_fragmentation
;
727 if (valid_ms
<= vd
->vdev_ms_count
/ 2)
728 return (ZFS_FRAG_INVALID
);
730 fragmentation
/= valid_ms
;
731 ASSERT3U(fragmentation
, <=, 100);
732 return (fragmentation
);
736 * Determine if a given metaslab group should skip allocations. A metaslab
737 * group should avoid allocations if its free capacity is less than the
738 * zfs_mg_noalloc_threshold or its fragmentation metric is greater than
739 * zfs_mg_fragmentation_threshold and there is at least one metaslab group
740 * that can still handle allocations.
743 metaslab_group_allocatable(metaslab_group_t
*mg
)
745 vdev_t
*vd
= mg
->mg_vd
;
746 spa_t
*spa
= vd
->vdev_spa
;
747 metaslab_class_t
*mc
= mg
->mg_class
;
750 * We use two key metrics to determine if a metaslab group is
751 * considered allocatable -- free space and fragmentation. If
752 * the free space is greater than the free space threshold and
753 * the fragmentation is less than the fragmentation threshold then
754 * consider the group allocatable. There are two case when we will
755 * not consider these key metrics. The first is if the group is
756 * associated with a slog device and the second is if all groups
757 * in this metaslab class have already been consider ineligible
760 return ((mg
->mg_free_capacity
> zfs_mg_noalloc_threshold
&&
761 (mg
->mg_fragmentation
== ZFS_FRAG_INVALID
||
762 mg
->mg_fragmentation
<= zfs_mg_fragmentation_threshold
)) ||
763 mc
!= spa_normal_class(spa
) || mc
->mc_alloc_groups
== 0);
767 * ==========================================================================
768 * Range tree callbacks
769 * ==========================================================================
773 * Comparison function for the private size-ordered tree. Tree is sorted
774 * by size, larger sizes at the end of the tree.
777 metaslab_rangesize_compare(const void *x1
, const void *x2
)
779 const range_seg_t
*r1
= x1
;
780 const range_seg_t
*r2
= x2
;
781 uint64_t rs_size1
= r1
->rs_end
- r1
->rs_start
;
782 uint64_t rs_size2
= r2
->rs_end
- r2
->rs_start
;
784 if (rs_size1
< rs_size2
)
786 if (rs_size1
> rs_size2
)
789 if (r1
->rs_start
< r2
->rs_start
)
792 if (r1
->rs_start
> r2
->rs_start
)
799 * Create any block allocator specific components. The current allocators
800 * rely on using both a size-ordered range_tree_t and an array of uint64_t's.
803 metaslab_rt_create(range_tree_t
*rt
, void *arg
)
805 metaslab_t
*msp
= arg
;
807 ASSERT3P(rt
->rt_arg
, ==, msp
);
808 ASSERT(msp
->ms_tree
== NULL
);
810 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
811 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
815 * Destroy the block allocator specific components.
818 metaslab_rt_destroy(range_tree_t
*rt
, void *arg
)
820 metaslab_t
*msp
= arg
;
822 ASSERT3P(rt
->rt_arg
, ==, msp
);
823 ASSERT3P(msp
->ms_tree
, ==, rt
);
824 ASSERT0(avl_numnodes(&msp
->ms_size_tree
));
826 avl_destroy(&msp
->ms_size_tree
);
830 metaslab_rt_add(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
832 metaslab_t
*msp
= arg
;
834 ASSERT3P(rt
->rt_arg
, ==, msp
);
835 ASSERT3P(msp
->ms_tree
, ==, rt
);
836 VERIFY(!msp
->ms_condensing
);
837 avl_add(&msp
->ms_size_tree
, rs
);
841 metaslab_rt_remove(range_tree_t
*rt
, range_seg_t
*rs
, void *arg
)
843 metaslab_t
*msp
= arg
;
845 ASSERT3P(rt
->rt_arg
, ==, msp
);
846 ASSERT3P(msp
->ms_tree
, ==, rt
);
847 VERIFY(!msp
->ms_condensing
);
848 avl_remove(&msp
->ms_size_tree
, rs
);
852 metaslab_rt_vacate(range_tree_t
*rt
, void *arg
)
854 metaslab_t
*msp
= arg
;
856 ASSERT3P(rt
->rt_arg
, ==, msp
);
857 ASSERT3P(msp
->ms_tree
, ==, rt
);
860 * Normally one would walk the tree freeing nodes along the way.
861 * Since the nodes are shared with the range trees we can avoid
862 * walking all nodes and just reinitialize the avl tree. The nodes
863 * will be freed by the range tree, so we don't want to free them here.
865 avl_create(&msp
->ms_size_tree
, metaslab_rangesize_compare
,
866 sizeof (range_seg_t
), offsetof(range_seg_t
, rs_pp_node
));
869 static range_tree_ops_t metaslab_rt_ops
= {
878 * ==========================================================================
879 * Metaslab block operations
880 * ==========================================================================
884 * Return the maximum contiguous segment within the metaslab.
887 metaslab_block_maxsize(metaslab_t
*msp
)
889 avl_tree_t
*t
= &msp
->ms_size_tree
;
892 if (t
== NULL
|| (rs
= avl_last(t
)) == NULL
)
895 return (rs
->rs_end
- rs
->rs_start
);
899 metaslab_block_alloc(metaslab_t
*msp
, uint64_t size
)
902 range_tree_t
*rt
= msp
->ms_tree
;
904 VERIFY(!msp
->ms_condensing
);
906 start
= msp
->ms_ops
->msop_alloc(msp
, size
);
907 if (start
!= -1ULL) {
908 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
910 VERIFY0(P2PHASE(start
, 1ULL << vd
->vdev_ashift
));
911 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
912 VERIFY3U(range_tree_space(rt
) - size
, <=, msp
->ms_size
);
913 range_tree_remove(rt
, start
, size
);
919 * ==========================================================================
920 * Common allocator routines
921 * ==========================================================================
924 #if defined(WITH_FF_BLOCK_ALLOCATOR) || \
925 defined(WITH_DF_BLOCK_ALLOCATOR) || \
926 defined(WITH_CF_BLOCK_ALLOCATOR)
928 * This is a helper function that can be used by the allocator to find
929 * a suitable block to allocate. This will search the specified AVL
930 * tree looking for a block that matches the specified criteria.
933 metaslab_block_picker(avl_tree_t
*t
, uint64_t *cursor
, uint64_t size
,
936 range_seg_t
*rs
, rsearch
;
939 rsearch
.rs_start
= *cursor
;
940 rsearch
.rs_end
= *cursor
+ size
;
942 rs
= avl_find(t
, &rsearch
, &where
);
944 rs
= avl_nearest(t
, where
, AVL_AFTER
);
947 uint64_t offset
= P2ROUNDUP(rs
->rs_start
, align
);
949 if (offset
+ size
<= rs
->rs_end
) {
950 *cursor
= offset
+ size
;
953 rs
= AVL_NEXT(t
, rs
);
957 * If we know we've searched the whole map (*cursor == 0), give up.
958 * Otherwise, reset the cursor to the beginning and try again.
964 return (metaslab_block_picker(t
, cursor
, size
, align
));
966 #endif /* WITH_FF/DF/CF_BLOCK_ALLOCATOR */
968 #if defined(WITH_FF_BLOCK_ALLOCATOR)
970 * ==========================================================================
971 * The first-fit block allocator
972 * ==========================================================================
975 metaslab_ff_alloc(metaslab_t
*msp
, uint64_t size
)
978 * Find the largest power of 2 block size that evenly divides the
979 * requested size. This is used to try to allocate blocks with similar
980 * alignment from the same area of the metaslab (i.e. same cursor
981 * bucket) but it does not guarantee that other allocations sizes
982 * may exist in the same region.
984 uint64_t align
= size
& -size
;
985 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
986 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
988 return (metaslab_block_picker(t
, cursor
, size
, align
));
991 static metaslab_ops_t metaslab_ff_ops
= {
995 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ff_ops
;
996 #endif /* WITH_FF_BLOCK_ALLOCATOR */
998 #if defined(WITH_DF_BLOCK_ALLOCATOR)
1000 * ==========================================================================
1001 * Dynamic block allocator -
1002 * Uses the first fit allocation scheme until space get low and then
1003 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1004 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1005 * ==========================================================================
1008 metaslab_df_alloc(metaslab_t
*msp
, uint64_t size
)
1011 * Find the largest power of 2 block size that evenly divides the
1012 * requested size. This is used to try to allocate blocks with similar
1013 * alignment from the same area of the metaslab (i.e. same cursor
1014 * bucket) but it does not guarantee that other allocations sizes
1015 * may exist in the same region.
1017 uint64_t align
= size
& -size
;
1018 uint64_t *cursor
= &msp
->ms_lbas
[highbit64(align
) - 1];
1019 range_tree_t
*rt
= msp
->ms_tree
;
1020 avl_tree_t
*t
= &rt
->rt_root
;
1021 uint64_t max_size
= metaslab_block_maxsize(msp
);
1022 int free_pct
= range_tree_space(rt
) * 100 / msp
->ms_size
;
1024 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1025 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1027 if (max_size
< size
)
1031 * If we're running low on space switch to using the size
1032 * sorted AVL tree (best-fit).
1034 if (max_size
< metaslab_df_alloc_threshold
||
1035 free_pct
< metaslab_df_free_pct
) {
1036 t
= &msp
->ms_size_tree
;
1040 return (metaslab_block_picker(t
, cursor
, size
, 1ULL));
1043 static metaslab_ops_t metaslab_df_ops
= {
1047 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_df_ops
;
1048 #endif /* WITH_DF_BLOCK_ALLOCATOR */
1050 #if defined(WITH_CF_BLOCK_ALLOCATOR)
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
= {
1095 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_cf_ops
;
1096 #endif /* WITH_CF_BLOCK_ALLOCATOR */
1098 #if defined(WITH_NDF_BLOCK_ALLOCATOR)
1100 * ==========================================================================
1101 * New dynamic fit allocator -
1102 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1103 * contiguous blocks. If no region is found then just use the largest segment
1105 * ==========================================================================
1109 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1110 * to request from the allocator.
1112 uint64_t metaslab_ndf_clump_shift
= 4;
1115 metaslab_ndf_alloc(metaslab_t
*msp
, uint64_t size
)
1117 avl_tree_t
*t
= &msp
->ms_tree
->rt_root
;
1119 range_seg_t
*rs
, rsearch
;
1120 uint64_t hbit
= highbit64(size
);
1121 uint64_t *cursor
= &msp
->ms_lbas
[hbit
- 1];
1122 uint64_t max_size
= metaslab_block_maxsize(msp
);
1124 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1125 ASSERT3U(avl_numnodes(t
), ==, avl_numnodes(&msp
->ms_size_tree
));
1127 if (max_size
< size
)
1130 rsearch
.rs_start
= *cursor
;
1131 rsearch
.rs_end
= *cursor
+ size
;
1133 rs
= avl_find(t
, &rsearch
, &where
);
1134 if (rs
== NULL
|| (rs
->rs_end
- rs
->rs_start
) < size
) {
1135 t
= &msp
->ms_size_tree
;
1137 rsearch
.rs_start
= 0;
1138 rsearch
.rs_end
= MIN(max_size
,
1139 1ULL << (hbit
+ metaslab_ndf_clump_shift
));
1140 rs
= avl_find(t
, &rsearch
, &where
);
1142 rs
= avl_nearest(t
, where
, AVL_AFTER
);
1146 if ((rs
->rs_end
- rs
->rs_start
) >= size
) {
1147 *cursor
= rs
->rs_start
+ size
;
1148 return (rs
->rs_start
);
1153 static metaslab_ops_t metaslab_ndf_ops
= {
1157 metaslab_ops_t
*zfs_metaslab_ops
= &metaslab_ndf_ops
;
1158 #endif /* WITH_NDF_BLOCK_ALLOCATOR */
1162 * ==========================================================================
1164 * ==========================================================================
1168 * Wait for any in-progress metaslab loads to complete.
1171 metaslab_load_wait(metaslab_t
*msp
)
1173 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1175 while (msp
->ms_loading
) {
1176 ASSERT(!msp
->ms_loaded
);
1177 cv_wait(&msp
->ms_load_cv
, &msp
->ms_lock
);
1182 metaslab_load(metaslab_t
*msp
)
1187 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1188 ASSERT(!msp
->ms_loaded
);
1189 ASSERT(!msp
->ms_loading
);
1191 msp
->ms_loading
= B_TRUE
;
1194 * If the space map has not been allocated yet, then treat
1195 * all the space in the metaslab as free and add it to the
1198 if (msp
->ms_sm
!= NULL
)
1199 error
= space_map_load(msp
->ms_sm
, msp
->ms_tree
, SM_FREE
);
1201 range_tree_add(msp
->ms_tree
, msp
->ms_start
, msp
->ms_size
);
1203 msp
->ms_loaded
= (error
== 0);
1204 msp
->ms_loading
= B_FALSE
;
1206 if (msp
->ms_loaded
) {
1207 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1208 range_tree_walk(msp
->ms_defertree
[t
],
1209 range_tree_remove
, msp
->ms_tree
);
1212 cv_broadcast(&msp
->ms_load_cv
);
1217 metaslab_unload(metaslab_t
*msp
)
1219 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1220 range_tree_vacate(msp
->ms_tree
, NULL
, NULL
);
1221 msp
->ms_loaded
= B_FALSE
;
1222 msp
->ms_weight
&= ~METASLAB_ACTIVE_MASK
;
1226 metaslab_init(metaslab_group_t
*mg
, uint64_t id
, uint64_t object
, uint64_t txg
)
1228 vdev_t
*vd
= mg
->mg_vd
;
1229 objset_t
*mos
= vd
->vdev_spa
->spa_meta_objset
;
1232 msp
= kmem_zalloc(sizeof (metaslab_t
), KM_PUSHPAGE
);
1233 mutex_init(&msp
->ms_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1234 cv_init(&msp
->ms_load_cv
, NULL
, CV_DEFAULT
, NULL
);
1236 msp
->ms_start
= id
<< vd
->vdev_ms_shift
;
1237 msp
->ms_size
= 1ULL << vd
->vdev_ms_shift
;
1240 * We only open space map objects that already exist. All others
1241 * will be opened when we finally allocate an object for it.
1244 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, object
, msp
->ms_start
,
1245 msp
->ms_size
, vd
->vdev_ashift
, &msp
->ms_lock
));
1246 ASSERT(msp
->ms_sm
!= NULL
);
1250 * We create the main range tree here, but we don't create the
1251 * alloctree and freetree until metaslab_sync_done(). This serves
1252 * two purposes: it allows metaslab_sync_done() to detect the
1253 * addition of new space; and for debugging, it ensures that we'd
1254 * data fault on any attempt to use this metaslab before it's ready.
1256 msp
->ms_tree
= range_tree_create(&metaslab_rt_ops
, msp
, &msp
->ms_lock
);
1257 metaslab_group_add(mg
, msp
);
1259 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1260 msp
->ms_ops
= mg
->mg_class
->mc_ops
;
1263 * If we're opening an existing pool (txg == 0) or creating
1264 * a new one (txg == TXG_INITIAL), all space is available now.
1265 * If we're adding space to an existing pool, the new space
1266 * does not become available until after this txg has synced.
1268 if (txg
<= TXG_INITIAL
)
1269 metaslab_sync_done(msp
, 0);
1272 * If metaslab_debug_load is set and we're initializing a metaslab
1273 * that has an allocated space_map object then load the its space
1274 * map so that can verify frees.
1276 if (metaslab_debug_load
&& msp
->ms_sm
!= NULL
) {
1277 mutex_enter(&msp
->ms_lock
);
1278 VERIFY0(metaslab_load(msp
));
1279 mutex_exit(&msp
->ms_lock
);
1283 vdev_dirty(vd
, 0, NULL
, txg
);
1284 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
1291 metaslab_fini(metaslab_t
*msp
)
1295 metaslab_group_t
*mg
= msp
->ms_group
;
1297 metaslab_group_remove(mg
, msp
);
1299 mutex_enter(&msp
->ms_lock
);
1301 VERIFY(msp
->ms_group
== NULL
);
1302 vdev_space_update(mg
->mg_vd
, -space_map_allocated(msp
->ms_sm
),
1304 space_map_close(msp
->ms_sm
);
1306 metaslab_unload(msp
);
1307 range_tree_destroy(msp
->ms_tree
);
1309 for (t
= 0; t
< TXG_SIZE
; t
++) {
1310 range_tree_destroy(msp
->ms_alloctree
[t
]);
1311 range_tree_destroy(msp
->ms_freetree
[t
]);
1314 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1315 range_tree_destroy(msp
->ms_defertree
[t
]);
1318 ASSERT0(msp
->ms_deferspace
);
1320 mutex_exit(&msp
->ms_lock
);
1321 cv_destroy(&msp
->ms_load_cv
);
1322 mutex_destroy(&msp
->ms_lock
);
1324 kmem_free(msp
, sizeof (metaslab_t
));
1327 #define FRAGMENTATION_TABLE_SIZE 17
1330 * This table defines a segment size based fragmentation metric that will
1331 * allow each metaslab to derive its own fragmentation value. This is done
1332 * by calculating the space in each bucket of the spacemap histogram and
1333 * multiplying that by the fragmetation metric in this table. Doing
1334 * this for all buckets and dividing it by the total amount of free
1335 * space in this metaslab (i.e. the total free space in all buckets) gives
1336 * us the fragmentation metric. This means that a high fragmentation metric
1337 * equates to most of the free space being comprised of small segments.
1338 * Conversely, if the metric is low, then most of the free space is in
1339 * large segments. A 10% change in fragmentation equates to approximately
1340 * double the number of segments.
1342 * This table defines 0% fragmented space using 16MB segments. Testing has
1343 * shown that segments that are greater than or equal to 16MB do not suffer
1344 * from drastic performance problems. Using this value, we derive the rest
1345 * of the table. Since the fragmentation value is never stored on disk, it
1346 * is possible to change these calculations in the future.
1348 int zfs_frag_table
[FRAGMENTATION_TABLE_SIZE
] = {
1368 * Calclate the metaslab's fragmentation metric. A return value
1369 * of ZFS_FRAG_INVALID means that the metaslab has not been upgraded and does
1370 * not support this metric. Otherwise, the return value should be in the
1374 metaslab_fragmentation(metaslab_t
*msp
)
1376 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1377 uint64_t fragmentation
= 0;
1379 boolean_t feature_enabled
= spa_feature_is_enabled(spa
,
1380 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 vdev_t
*vd
= msp
->ms_group
->mg_vd
;
1400 if (spa_writeable(vd
->vdev_spa
)) {
1401 uint64_t txg
= spa_syncing_txg(spa
);
1403 msp
->ms_condense_wanted
= B_TRUE
;
1404 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1405 spa_dbgmsg(spa
, "txg %llu, requesting force condense: "
1406 "msp %p, vd %p", txg
, msp
, vd
);
1408 return (ZFS_FRAG_INVALID
);
1411 for (i
= 0; i
< SPACE_MAP_HISTOGRAM_SIZE
; i
++) {
1413 uint8_t shift
= msp
->ms_sm
->sm_shift
;
1414 int idx
= MIN(shift
- SPA_MINBLOCKSHIFT
+ i
,
1415 FRAGMENTATION_TABLE_SIZE
- 1);
1417 if (msp
->ms_sm
->sm_phys
->smp_histogram
[i
] == 0)
1420 space
= msp
->ms_sm
->sm_phys
->smp_histogram
[i
] << (i
+ shift
);
1423 ASSERT3U(idx
, <, FRAGMENTATION_TABLE_SIZE
);
1424 fragmentation
+= space
* zfs_frag_table
[idx
];
1428 fragmentation
/= total
;
1429 ASSERT3U(fragmentation
, <=, 100);
1430 return (fragmentation
);
1434 * Compute a weight -- a selection preference value -- for the given metaslab.
1435 * This is based on the amount of free space, the level of fragmentation,
1436 * the LBA range, and whether the metaslab is loaded.
1439 metaslab_weight(metaslab_t
*msp
)
1441 metaslab_group_t
*mg
= msp
->ms_group
;
1442 vdev_t
*vd
= mg
->mg_vd
;
1443 uint64_t weight
, space
;
1445 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1448 * This vdev is in the process of being removed so there is nothing
1449 * for us to do here.
1451 if (vd
->vdev_removing
) {
1452 ASSERT0(space_map_allocated(msp
->ms_sm
));
1453 ASSERT0(vd
->vdev_ms_shift
);
1458 * The baseline weight is the metaslab's free space.
1460 space
= msp
->ms_size
- space_map_allocated(msp
->ms_sm
);
1462 msp
->ms_fragmentation
= metaslab_fragmentation(msp
);
1463 if (metaslab_fragmentation_factor_enabled
&&
1464 msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
) {
1466 * Use the fragmentation information to inversely scale
1467 * down the baseline weight. We need to ensure that we
1468 * don't exclude this metaslab completely when it's 100%
1469 * fragmented. To avoid this we reduce the fragmented value
1472 space
= (space
* (100 - (msp
->ms_fragmentation
- 1))) / 100;
1475 * If space < SPA_MINBLOCKSIZE, then we will not allocate from
1476 * this metaslab again. The fragmentation metric may have
1477 * decreased the space to something smaller than
1478 * SPA_MINBLOCKSIZE, so reset the space to SPA_MINBLOCKSIZE
1479 * so that we can consume any remaining space.
1481 if (space
> 0 && space
< SPA_MINBLOCKSIZE
)
1482 space
= SPA_MINBLOCKSIZE
;
1487 * Modern disks have uniform bit density and constant angular velocity.
1488 * Therefore, the outer recording zones are faster (higher bandwidth)
1489 * than the inner zones by the ratio of outer to inner track diameter,
1490 * which is typically around 2:1. We account for this by assigning
1491 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
1492 * In effect, this means that we'll select the metaslab with the most
1493 * free bandwidth rather than simply the one with the most free space.
1495 if (metaslab_lba_weighting_enabled
) {
1496 weight
= 2 * weight
- (msp
->ms_id
* weight
) / vd
->vdev_ms_count
;
1497 ASSERT(weight
>= space
&& weight
<= 2 * space
);
1501 * If this metaslab is one we're actively using, adjust its
1502 * weight to make it preferable to any inactive metaslab so
1503 * we'll polish it off. If the fragmentation on this metaslab
1504 * has exceed our threshold, then don't mark it active.
1506 if (msp
->ms_loaded
&& msp
->ms_fragmentation
!= ZFS_FRAG_INVALID
&&
1507 msp
->ms_fragmentation
<= zfs_metaslab_fragmentation_threshold
) {
1508 weight
|= (msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1515 metaslab_activate(metaslab_t
*msp
, uint64_t activation_weight
)
1517 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1519 if ((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0) {
1520 metaslab_load_wait(msp
);
1521 if (!msp
->ms_loaded
) {
1522 int error
= metaslab_load(msp
);
1524 metaslab_group_sort(msp
->ms_group
, msp
, 0);
1529 metaslab_group_sort(msp
->ms_group
, msp
,
1530 msp
->ms_weight
| activation_weight
);
1532 ASSERT(msp
->ms_loaded
);
1533 ASSERT(msp
->ms_weight
& METASLAB_ACTIVE_MASK
);
1539 metaslab_passivate(metaslab_t
*msp
, uint64_t size
)
1542 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
1543 * this metaslab again. In that case, it had better be empty,
1544 * or we would be leaving space on the table.
1546 ASSERT(size
>= SPA_MINBLOCKSIZE
|| range_tree_space(msp
->ms_tree
) == 0);
1547 metaslab_group_sort(msp
->ms_group
, msp
, MIN(msp
->ms_weight
, size
));
1548 ASSERT((msp
->ms_weight
& METASLAB_ACTIVE_MASK
) == 0);
1552 metaslab_preload(void *arg
)
1554 metaslab_t
*msp
= arg
;
1555 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1557 ASSERT(!MUTEX_HELD(&msp
->ms_group
->mg_lock
));
1559 mutex_enter(&msp
->ms_lock
);
1560 metaslab_load_wait(msp
);
1561 if (!msp
->ms_loaded
)
1562 (void) metaslab_load(msp
);
1565 * Set the ms_access_txg value so that we don't unload it right away.
1567 msp
->ms_access_txg
= spa_syncing_txg(spa
) + metaslab_unload_delay
+ 1;
1568 mutex_exit(&msp
->ms_lock
);
1572 metaslab_group_preload(metaslab_group_t
*mg
)
1574 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
1576 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
1579 if (spa_shutting_down(spa
) || !metaslab_preload_enabled
) {
1580 taskq_wait(mg
->mg_taskq
);
1584 mutex_enter(&mg
->mg_lock
);
1586 * Load the next potential metaslabs
1589 while (msp
!= NULL
) {
1590 metaslab_t
*msp_next
= AVL_NEXT(t
, msp
);
1593 * We preload only the maximum number of metaslabs specified
1594 * by metaslab_preload_limit. If a metaslab is being forced
1595 * to condense then we preload it too. This will ensure
1596 * that force condensing happens in the next txg.
1598 if (++m
> metaslab_preload_limit
&& !msp
->ms_condense_wanted
) {
1604 * We must drop the metaslab group lock here to preserve
1605 * lock ordering with the ms_lock (when grabbing both
1606 * the mg_lock and the ms_lock, the ms_lock must be taken
1607 * first). As a result, it is possible that the ordering
1608 * of the metaslabs within the avl tree may change before
1609 * we reacquire the lock. The metaslab cannot be removed from
1610 * the tree while we're in syncing context so it is safe to
1611 * drop the mg_lock here. If the metaslabs are reordered
1612 * nothing will break -- we just may end up loading a
1613 * less than optimal one.
1615 mutex_exit(&mg
->mg_lock
);
1616 VERIFY(taskq_dispatch(mg
->mg_taskq
, metaslab_preload
,
1617 msp
, TQ_PUSHPAGE
) != 0);
1618 mutex_enter(&mg
->mg_lock
);
1621 mutex_exit(&mg
->mg_lock
);
1625 * Determine if the space map's on-disk footprint is past our tolerance
1626 * for inefficiency. We would like to use the following criteria to make
1629 * 1. The size of the space map object should not dramatically increase as a
1630 * result of writing out the free space range tree.
1632 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
1633 * times the size than the free space range tree representation
1634 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
1636 * Checking the first condition is tricky since we don't want to walk
1637 * the entire AVL tree calculating the estimated on-disk size. Instead we
1638 * use the size-ordered range tree in the metaslab and calculate the
1639 * size required to write out the largest segment in our free tree. If the
1640 * size required to represent that segment on disk is larger than the space
1641 * map object then we avoid condensing this map.
1643 * To determine the second criterion we use a best-case estimate and assume
1644 * each segment can be represented on-disk as a single 64-bit entry. We refer
1645 * to this best-case estimate as the space map's minimal form.
1648 metaslab_should_condense(metaslab_t
*msp
)
1650 space_map_t
*sm
= msp
->ms_sm
;
1652 uint64_t size
, entries
, segsz
;
1654 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1655 ASSERT(msp
->ms_loaded
);
1658 * Use the ms_size_tree range tree, which is ordered by size, to
1659 * obtain the largest segment in the free tree. We always condense
1660 * metaslabs that are empty and metaslabs for which a condense
1661 * request has been made.
1663 rs
= avl_last(&msp
->ms_size_tree
);
1664 if (rs
== NULL
|| msp
->ms_condense_wanted
)
1668 * Calculate the number of 64-bit entries this segment would
1669 * require when written to disk. If this single segment would be
1670 * larger on-disk than the entire current on-disk structure, then
1671 * clearly condensing will increase the on-disk structure size.
1673 size
= (rs
->rs_end
- rs
->rs_start
) >> sm
->sm_shift
;
1674 entries
= size
/ (MIN(size
, SM_RUN_MAX
));
1675 segsz
= entries
* sizeof (uint64_t);
1677 return (segsz
<= space_map_length(msp
->ms_sm
) &&
1678 space_map_length(msp
->ms_sm
) >= (zfs_condense_pct
*
1679 sizeof (uint64_t) * avl_numnodes(&msp
->ms_tree
->rt_root
)) / 100);
1683 * Condense the on-disk space map representation to its minimized form.
1684 * The minimized form consists of a small number of allocations followed by
1685 * the entries of the free range tree.
1688 metaslab_condense(metaslab_t
*msp
, uint64_t txg
, dmu_tx_t
*tx
)
1690 spa_t
*spa
= msp
->ms_group
->mg_vd
->vdev_spa
;
1691 range_tree_t
*freetree
= msp
->ms_freetree
[txg
& TXG_MASK
];
1692 range_tree_t
*condense_tree
;
1693 space_map_t
*sm
= msp
->ms_sm
;
1696 ASSERT(MUTEX_HELD(&msp
->ms_lock
));
1697 ASSERT3U(spa_sync_pass(spa
), ==, 1);
1698 ASSERT(msp
->ms_loaded
);
1701 spa_dbgmsg(spa
, "condensing: txg %llu, msp[%llu] %p, "
1702 "smp size %llu, segments %lu, forcing condense=%s", txg
,
1703 msp
->ms_id
, msp
, space_map_length(msp
->ms_sm
),
1704 avl_numnodes(&msp
->ms_tree
->rt_root
),
1705 msp
->ms_condense_wanted
? "TRUE" : "FALSE");
1707 msp
->ms_condense_wanted
= B_FALSE
;
1710 * Create an range tree that is 100% allocated. We remove segments
1711 * that have been freed in this txg, any deferred frees that exist,
1712 * and any allocation in the future. Removing segments should be
1713 * a relatively inexpensive operation since we expect these trees to
1714 * have a small number of nodes.
1716 condense_tree
= range_tree_create(NULL
, NULL
, &msp
->ms_lock
);
1717 range_tree_add(condense_tree
, msp
->ms_start
, msp
->ms_size
);
1720 * Remove what's been freed in this txg from the condense_tree.
1721 * Since we're in sync_pass 1, we know that all the frees from
1722 * this txg are in the freetree.
1724 range_tree_walk(freetree
, range_tree_remove
, condense_tree
);
1726 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1727 range_tree_walk(msp
->ms_defertree
[t
],
1728 range_tree_remove
, condense_tree
);
1731 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1732 range_tree_walk(msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
],
1733 range_tree_remove
, condense_tree
);
1737 * We're about to drop the metaslab's lock thus allowing
1738 * other consumers to change it's content. Set the
1739 * metaslab's ms_condensing flag to ensure that
1740 * allocations on this metaslab do not occur while we're
1741 * in the middle of committing it to disk. This is only critical
1742 * for the ms_tree as all other range trees use per txg
1743 * views of their content.
1745 msp
->ms_condensing
= B_TRUE
;
1747 mutex_exit(&msp
->ms_lock
);
1748 space_map_truncate(sm
, tx
);
1749 mutex_enter(&msp
->ms_lock
);
1752 * While we would ideally like to create a space_map representation
1753 * that consists only of allocation records, doing so can be
1754 * prohibitively expensive because the in-core free tree can be
1755 * large, and therefore computationally expensive to subtract
1756 * from the condense_tree. Instead we sync out two trees, a cheap
1757 * allocation only tree followed by the in-core free tree. While not
1758 * optimal, this is typically close to optimal, and much cheaper to
1761 space_map_write(sm
, condense_tree
, SM_ALLOC
, tx
);
1762 range_tree_vacate(condense_tree
, NULL
, NULL
);
1763 range_tree_destroy(condense_tree
);
1765 space_map_write(sm
, msp
->ms_tree
, SM_FREE
, tx
);
1766 msp
->ms_condensing
= B_FALSE
;
1770 * Write a metaslab to disk in the context of the specified transaction group.
1773 metaslab_sync(metaslab_t
*msp
, uint64_t txg
)
1775 metaslab_group_t
*mg
= msp
->ms_group
;
1776 vdev_t
*vd
= mg
->mg_vd
;
1777 spa_t
*spa
= vd
->vdev_spa
;
1778 objset_t
*mos
= spa_meta_objset(spa
);
1779 range_tree_t
*alloctree
= msp
->ms_alloctree
[txg
& TXG_MASK
];
1780 range_tree_t
**freetree
= &msp
->ms_freetree
[txg
& TXG_MASK
];
1781 range_tree_t
**freed_tree
=
1782 &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1784 uint64_t object
= space_map_object(msp
->ms_sm
);
1786 ASSERT(!vd
->vdev_ishole
);
1789 * This metaslab has just been added so there's no work to do now.
1791 if (*freetree
== NULL
) {
1792 ASSERT3P(alloctree
, ==, NULL
);
1796 ASSERT3P(alloctree
, !=, NULL
);
1797 ASSERT3P(*freetree
, !=, NULL
);
1798 ASSERT3P(*freed_tree
, !=, NULL
);
1801 * Normally, we don't want to process a metaslab if there
1802 * are no allocations or frees to perform. However, if the metaslab
1803 * is being forced to condense we need to let it through.
1805 if (range_tree_space(alloctree
) == 0 &&
1806 range_tree_space(*freetree
) == 0 &&
1807 !msp
->ms_condense_wanted
)
1811 * The only state that can actually be changing concurrently with
1812 * metaslab_sync() is the metaslab's ms_tree. No other thread can
1813 * be modifying this txg's alloctree, freetree, freed_tree, or
1814 * space_map_phys_t. Therefore, we only hold ms_lock to satify
1815 * space_map ASSERTs. We drop it whenever we call into the DMU,
1816 * because the DMU can call down to us (e.g. via zio_free()) at
1820 tx
= dmu_tx_create_assigned(spa_get_dsl(spa
), txg
);
1822 if (msp
->ms_sm
== NULL
) {
1823 uint64_t new_object
;
1825 new_object
= space_map_alloc(mos
, tx
);
1826 VERIFY3U(new_object
, !=, 0);
1828 VERIFY0(space_map_open(&msp
->ms_sm
, mos
, new_object
,
1829 msp
->ms_start
, msp
->ms_size
, vd
->vdev_ashift
,
1831 ASSERT(msp
->ms_sm
!= NULL
);
1834 mutex_enter(&msp
->ms_lock
);
1836 if (msp
->ms_loaded
&& spa_sync_pass(spa
) == 1 &&
1837 metaslab_should_condense(msp
)) {
1838 metaslab_condense(msp
, txg
, tx
);
1840 space_map_write(msp
->ms_sm
, alloctree
, SM_ALLOC
, tx
);
1841 space_map_write(msp
->ms_sm
, *freetree
, SM_FREE
, tx
);
1844 metaslab_group_histogram_verify(mg
);
1845 metaslab_class_histogram_verify(mg
->mg_class
);
1846 metaslab_group_histogram_remove(mg
, msp
);
1847 if (msp
->ms_loaded
) {
1849 * When the space map is loaded, we have an accruate
1850 * histogram in the range tree. This gives us an opportunity
1851 * to bring the space map's histogram up-to-date so we clear
1852 * it first before updating it.
1854 space_map_histogram_clear(msp
->ms_sm
);
1855 space_map_histogram_add(msp
->ms_sm
, msp
->ms_tree
, tx
);
1858 * Since the space map is not loaded we simply update the
1859 * exisiting histogram with what was freed in this txg. This
1860 * means that the on-disk histogram may not have an accurate
1861 * view of the free space but it's close enough to allow
1862 * us to make allocation decisions.
1864 space_map_histogram_add(msp
->ms_sm
, *freetree
, tx
);
1866 metaslab_group_histogram_add(mg
, msp
);
1867 metaslab_group_histogram_verify(mg
);
1868 metaslab_class_histogram_verify(mg
->mg_class
);
1871 * For sync pass 1, we avoid traversing this txg's free range tree
1872 * and instead will just swap the pointers for freetree and
1873 * freed_tree. We can safely do this since the freed_tree is
1874 * guaranteed to be empty on the initial pass.
1876 if (spa_sync_pass(spa
) == 1) {
1877 range_tree_swap(freetree
, freed_tree
);
1879 range_tree_vacate(*freetree
, range_tree_add
, *freed_tree
);
1881 range_tree_vacate(alloctree
, NULL
, NULL
);
1883 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1884 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1886 mutex_exit(&msp
->ms_lock
);
1888 if (object
!= space_map_object(msp
->ms_sm
)) {
1889 object
= space_map_object(msp
->ms_sm
);
1890 dmu_write(mos
, vd
->vdev_ms_array
, sizeof (uint64_t) *
1891 msp
->ms_id
, sizeof (uint64_t), &object
, tx
);
1897 * Called after a transaction group has completely synced to mark
1898 * all of the metaslab's free space as usable.
1901 metaslab_sync_done(metaslab_t
*msp
, uint64_t txg
)
1903 metaslab_group_t
*mg
= msp
->ms_group
;
1904 vdev_t
*vd
= mg
->mg_vd
;
1905 range_tree_t
**freed_tree
;
1906 range_tree_t
**defer_tree
;
1907 int64_t alloc_delta
, defer_delta
;
1910 ASSERT(!vd
->vdev_ishole
);
1912 mutex_enter(&msp
->ms_lock
);
1915 * If this metaslab is just becoming available, initialize its
1916 * alloctrees, freetrees, and defertree and add its capacity to
1919 if (msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
] == NULL
) {
1920 for (t
= 0; t
< TXG_SIZE
; t
++) {
1921 ASSERT(msp
->ms_alloctree
[t
] == NULL
);
1922 ASSERT(msp
->ms_freetree
[t
] == NULL
);
1924 msp
->ms_alloctree
[t
] = range_tree_create(NULL
, msp
,
1926 msp
->ms_freetree
[t
] = range_tree_create(NULL
, msp
,
1930 for (t
= 0; t
< TXG_DEFER_SIZE
; t
++) {
1931 ASSERT(msp
->ms_defertree
[t
] == NULL
);
1933 msp
->ms_defertree
[t
] = range_tree_create(NULL
, msp
,
1937 vdev_space_update(vd
, 0, 0, msp
->ms_size
);
1940 freed_tree
= &msp
->ms_freetree
[TXG_CLEAN(txg
) & TXG_MASK
];
1941 defer_tree
= &msp
->ms_defertree
[txg
% TXG_DEFER_SIZE
];
1943 alloc_delta
= space_map_alloc_delta(msp
->ms_sm
);
1944 defer_delta
= range_tree_space(*freed_tree
) -
1945 range_tree_space(*defer_tree
);
1947 vdev_space_update(vd
, alloc_delta
+ defer_delta
, defer_delta
, 0);
1949 ASSERT0(range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]));
1950 ASSERT0(range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]));
1953 * If there's a metaslab_load() in progress, wait for it to complete
1954 * so that we have a consistent view of the in-core space map.
1956 metaslab_load_wait(msp
);
1959 * Move the frees from the defer_tree back to the free
1960 * range tree (if it's loaded). Swap the freed_tree and the
1961 * defer_tree -- this is safe to do because we've just emptied out
1964 range_tree_vacate(*defer_tree
,
1965 msp
->ms_loaded
? range_tree_add
: NULL
, msp
->ms_tree
);
1966 range_tree_swap(freed_tree
, defer_tree
);
1968 space_map_update(msp
->ms_sm
);
1970 msp
->ms_deferspace
+= defer_delta
;
1971 ASSERT3S(msp
->ms_deferspace
, >=, 0);
1972 ASSERT3S(msp
->ms_deferspace
, <=, msp
->ms_size
);
1973 if (msp
->ms_deferspace
!= 0) {
1975 * Keep syncing this metaslab until all deferred frees
1976 * are back in circulation.
1978 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
+ 1);
1981 if (msp
->ms_loaded
&& msp
->ms_access_txg
< txg
) {
1982 for (t
= 1; t
< TXG_CONCURRENT_STATES
; t
++) {
1983 VERIFY0(range_tree_space(
1984 msp
->ms_alloctree
[(txg
+ t
) & TXG_MASK
]));
1987 if (!metaslab_debug_unload
)
1988 metaslab_unload(msp
);
1991 metaslab_group_sort(mg
, msp
, metaslab_weight(msp
));
1992 mutex_exit(&msp
->ms_lock
);
1996 metaslab_sync_reassess(metaslab_group_t
*mg
)
1998 metaslab_group_alloc_update(mg
);
1999 mg
->mg_fragmentation
= metaslab_group_fragmentation(mg
);
2002 * Preload the next potential metaslabs
2004 metaslab_group_preload(mg
);
2008 metaslab_distance(metaslab_t
*msp
, dva_t
*dva
)
2010 uint64_t ms_shift
= msp
->ms_group
->mg_vd
->vdev_ms_shift
;
2011 uint64_t offset
= DVA_GET_OFFSET(dva
) >> ms_shift
;
2012 uint64_t start
= msp
->ms_id
;
2014 if (msp
->ms_group
->mg_vd
->vdev_id
!= DVA_GET_VDEV(dva
))
2015 return (1ULL << 63);
2018 return ((start
- offset
) << ms_shift
);
2020 return ((offset
- start
) << ms_shift
);
2025 metaslab_group_alloc(metaslab_group_t
*mg
, uint64_t psize
, uint64_t asize
,
2026 uint64_t txg
, uint64_t min_distance
, dva_t
*dva
, int d
)
2028 spa_t
*spa
= mg
->mg_vd
->vdev_spa
;
2029 metaslab_t
*msp
= NULL
;
2030 uint64_t offset
= -1ULL;
2031 avl_tree_t
*t
= &mg
->mg_metaslab_tree
;
2032 uint64_t activation_weight
;
2033 uint64_t target_distance
;
2036 activation_weight
= METASLAB_WEIGHT_PRIMARY
;
2037 for (i
= 0; i
< d
; i
++) {
2038 if (DVA_GET_VDEV(&dva
[i
]) == mg
->mg_vd
->vdev_id
) {
2039 activation_weight
= METASLAB_WEIGHT_SECONDARY
;
2045 boolean_t was_active
;
2047 mutex_enter(&mg
->mg_lock
);
2048 for (msp
= avl_first(t
); msp
; msp
= AVL_NEXT(t
, msp
)) {
2049 if (msp
->ms_weight
< asize
) {
2050 spa_dbgmsg(spa
, "%s: failed to meet weight "
2051 "requirement: vdev %llu, txg %llu, mg %p, "
2052 "msp %p, psize %llu, asize %llu, "
2053 "weight %llu", spa_name(spa
),
2054 mg
->mg_vd
->vdev_id
, txg
,
2055 mg
, msp
, psize
, asize
, msp
->ms_weight
);
2056 mutex_exit(&mg
->mg_lock
);
2061 * If the selected metaslab is condensing, skip it.
2063 if (msp
->ms_condensing
)
2066 was_active
= msp
->ms_weight
& METASLAB_ACTIVE_MASK
;
2067 if (activation_weight
== METASLAB_WEIGHT_PRIMARY
)
2070 target_distance
= min_distance
+
2071 (space_map_allocated(msp
->ms_sm
) != 0 ? 0 :
2074 for (i
= 0; i
< d
; i
++)
2075 if (metaslab_distance(msp
, &dva
[i
]) <
2081 mutex_exit(&mg
->mg_lock
);
2085 mutex_enter(&msp
->ms_lock
);
2088 * Ensure that the metaslab we have selected is still
2089 * capable of handling our request. It's possible that
2090 * another thread may have changed the weight while we
2091 * were blocked on the metaslab lock.
2093 if (msp
->ms_weight
< asize
|| (was_active
&&
2094 !(msp
->ms_weight
& METASLAB_ACTIVE_MASK
) &&
2095 activation_weight
== METASLAB_WEIGHT_PRIMARY
)) {
2096 mutex_exit(&msp
->ms_lock
);
2100 if ((msp
->ms_weight
& METASLAB_WEIGHT_SECONDARY
) &&
2101 activation_weight
== METASLAB_WEIGHT_PRIMARY
) {
2102 metaslab_passivate(msp
,
2103 msp
->ms_weight
& ~METASLAB_ACTIVE_MASK
);
2104 mutex_exit(&msp
->ms_lock
);
2108 if (metaslab_activate(msp
, activation_weight
) != 0) {
2109 mutex_exit(&msp
->ms_lock
);
2114 * If this metaslab is currently condensing then pick again as
2115 * we can't manipulate this metaslab until it's committed
2118 if (msp
->ms_condensing
) {
2119 mutex_exit(&msp
->ms_lock
);
2123 if ((offset
= metaslab_block_alloc(msp
, asize
)) != -1ULL)
2126 metaslab_passivate(msp
, metaslab_block_maxsize(msp
));
2127 mutex_exit(&msp
->ms_lock
);
2130 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2131 vdev_dirty(mg
->mg_vd
, VDD_METASLAB
, msp
, txg
);
2133 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, asize
);
2134 msp
->ms_access_txg
= txg
+ metaslab_unload_delay
;
2136 mutex_exit(&msp
->ms_lock
);
2142 * Allocate a block for the specified i/o.
2145 metaslab_alloc_dva(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
,
2146 dva_t
*dva
, int d
, dva_t
*hintdva
, uint64_t txg
, int flags
)
2148 metaslab_group_t
*mg
, *fast_mg
, *rotor
;
2152 int zio_lock
= B_FALSE
;
2153 boolean_t allocatable
;
2154 uint64_t offset
= -1ULL;
2158 ASSERT(!DVA_IS_VALID(&dva
[d
]));
2161 * For testing, make some blocks above a certain size be gang blocks.
2163 if (psize
>= metaslab_gang_bang
&& (ddi_get_lbolt() & 3) == 0)
2164 return (SET_ERROR(ENOSPC
));
2166 if (flags
& METASLAB_FASTWRITE
)
2167 mutex_enter(&mc
->mc_fastwrite_lock
);
2170 * Start at the rotor and loop through all mgs until we find something.
2171 * Note that there's no locking on mc_rotor or mc_aliquot because
2172 * nothing actually breaks if we miss a few updates -- we just won't
2173 * allocate quite as evenly. It all balances out over time.
2175 * If we are doing ditto or log blocks, try to spread them across
2176 * consecutive vdevs. If we're forced to reuse a vdev before we've
2177 * allocated all of our ditto blocks, then try and spread them out on
2178 * that vdev as much as possible. If it turns out to not be possible,
2179 * gradually lower our standards until anything becomes acceptable.
2180 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
2181 * gives us hope of containing our fault domains to something we're
2182 * able to reason about. Otherwise, any two top-level vdev failures
2183 * will guarantee the loss of data. With consecutive allocation,
2184 * only two adjacent top-level vdev failures will result in data loss.
2186 * If we are doing gang blocks (hintdva is non-NULL), try to keep
2187 * ourselves on the same vdev as our gang block header. That
2188 * way, we can hope for locality in vdev_cache, plus it makes our
2189 * fault domains something tractable.
2192 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&hintdva
[d
]));
2195 * It's possible the vdev we're using as the hint no
2196 * longer exists (i.e. removed). Consult the rotor when
2202 if (flags
& METASLAB_HINTBP_AVOID
&&
2203 mg
->mg_next
!= NULL
)
2208 } else if (d
!= 0) {
2209 vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
- 1]));
2210 mg
= vd
->vdev_mg
->mg_next
;
2211 } else if (flags
& METASLAB_FASTWRITE
) {
2212 mg
= fast_mg
= mc
->mc_rotor
;
2215 if (fast_mg
->mg_vd
->vdev_pending_fastwrite
<
2216 mg
->mg_vd
->vdev_pending_fastwrite
)
2218 } while ((fast_mg
= fast_mg
->mg_next
) != mc
->mc_rotor
);
2225 * If the hint put us into the wrong metaslab class, or into a
2226 * metaslab group that has been passivated, just follow the rotor.
2228 if (mg
->mg_class
!= mc
|| mg
->mg_activation_count
<= 0)
2235 ASSERT(mg
->mg_activation_count
== 1);
2240 * Don't allocate from faulted devices.
2243 spa_config_enter(spa
, SCL_ZIO
, FTAG
, RW_READER
);
2244 allocatable
= vdev_allocatable(vd
);
2245 spa_config_exit(spa
, SCL_ZIO
, FTAG
);
2247 allocatable
= vdev_allocatable(vd
);
2251 * Determine if the selected metaslab group is eligible
2252 * for allocations. If we're ganging or have requested
2253 * an allocation for the smallest gang block size
2254 * then we don't want to avoid allocating to the this
2255 * metaslab group. If we're in this condition we should
2256 * try to allocate from any device possible so that we
2257 * don't inadvertently return ENOSPC and suspend the pool
2258 * even though space is still available.
2260 if (allocatable
&& CAN_FASTGANG(flags
) &&
2261 psize
> SPA_GANGBLOCKSIZE
)
2262 allocatable
= metaslab_group_allocatable(mg
);
2268 * Avoid writing single-copy data to a failing vdev
2269 * unless the user instructs us that it is okay.
2271 if ((vd
->vdev_stat
.vs_write_errors
> 0 ||
2272 vd
->vdev_state
< VDEV_STATE_HEALTHY
) &&
2273 d
== 0 && dshift
== 3 && vd
->vdev_children
== 0) {
2278 ASSERT(mg
->mg_class
== mc
);
2280 distance
= vd
->vdev_asize
>> dshift
;
2281 if (distance
<= (1ULL << vd
->vdev_ms_shift
))
2286 asize
= vdev_psize_to_asize(vd
, psize
);
2287 ASSERT(P2PHASE(asize
, 1ULL << vd
->vdev_ashift
) == 0);
2289 offset
= metaslab_group_alloc(mg
, psize
, asize
, txg
, distance
,
2291 if (offset
!= -1ULL) {
2293 * If we've just selected this metaslab group,
2294 * figure out whether the corresponding vdev is
2295 * over- or under-used relative to the pool,
2296 * and set an allocation bias to even it out.
2298 if (mc
->mc_aliquot
== 0 && metaslab_bias_enabled
) {
2299 vdev_stat_t
*vs
= &vd
->vdev_stat
;
2302 vu
= (vs
->vs_alloc
* 100) / (vs
->vs_space
+ 1);
2303 cu
= (mc
->mc_alloc
* 100) / (mc
->mc_space
+ 1);
2306 * Calculate how much more or less we should
2307 * try to allocate from this device during
2308 * this iteration around the rotor.
2309 * For example, if a device is 80% full
2310 * and the pool is 20% full then we should
2311 * reduce allocations by 60% on this device.
2313 * mg_bias = (20 - 80) * 512K / 100 = -307K
2315 * This reduces allocations by 307K for this
2318 mg
->mg_bias
= ((cu
- vu
) *
2319 (int64_t)mg
->mg_aliquot
) / 100;
2320 } else if (!metaslab_bias_enabled
) {
2324 if ((flags
& METASLAB_FASTWRITE
) ||
2325 atomic_add_64_nv(&mc
->mc_aliquot
, asize
) >=
2326 mg
->mg_aliquot
+ mg
->mg_bias
) {
2327 mc
->mc_rotor
= mg
->mg_next
;
2331 DVA_SET_VDEV(&dva
[d
], vd
->vdev_id
);
2332 DVA_SET_OFFSET(&dva
[d
], offset
);
2333 DVA_SET_GANG(&dva
[d
], !!(flags
& METASLAB_GANG_HEADER
));
2334 DVA_SET_ASIZE(&dva
[d
], asize
);
2336 if (flags
& METASLAB_FASTWRITE
) {
2337 atomic_add_64(&vd
->vdev_pending_fastwrite
,
2339 mutex_exit(&mc
->mc_fastwrite_lock
);
2345 mc
->mc_rotor
= mg
->mg_next
;
2347 } while ((mg
= mg
->mg_next
) != rotor
);
2351 ASSERT(dshift
< 64);
2355 if (!allocatable
&& !zio_lock
) {
2361 bzero(&dva
[d
], sizeof (dva_t
));
2363 if (flags
& METASLAB_FASTWRITE
)
2364 mutex_exit(&mc
->mc_fastwrite_lock
);
2366 return (SET_ERROR(ENOSPC
));
2370 * Free the block represented by DVA in the context of the specified
2371 * transaction group.
2374 metaslab_free_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
, boolean_t now
)
2376 uint64_t vdev
= DVA_GET_VDEV(dva
);
2377 uint64_t offset
= DVA_GET_OFFSET(dva
);
2378 uint64_t size
= DVA_GET_ASIZE(dva
);
2382 ASSERT(DVA_IS_VALID(dva
));
2384 if (txg
> spa_freeze_txg(spa
))
2387 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2388 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
) {
2389 cmn_err(CE_WARN
, "metaslab_free_dva(): bad DVA %llu:%llu",
2390 (u_longlong_t
)vdev
, (u_longlong_t
)offset
);
2395 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2397 if (DVA_GET_GANG(dva
))
2398 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2400 mutex_enter(&msp
->ms_lock
);
2403 range_tree_remove(msp
->ms_alloctree
[txg
& TXG_MASK
],
2406 VERIFY(!msp
->ms_condensing
);
2407 VERIFY3U(offset
, >=, msp
->ms_start
);
2408 VERIFY3U(offset
+ size
, <=, msp
->ms_start
+ msp
->ms_size
);
2409 VERIFY3U(range_tree_space(msp
->ms_tree
) + size
, <=,
2411 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2412 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2413 range_tree_add(msp
->ms_tree
, offset
, size
);
2415 if (range_tree_space(msp
->ms_freetree
[txg
& TXG_MASK
]) == 0)
2416 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2417 range_tree_add(msp
->ms_freetree
[txg
& TXG_MASK
],
2421 mutex_exit(&msp
->ms_lock
);
2425 * Intent log support: upon opening the pool after a crash, notify the SPA
2426 * of blocks that the intent log has allocated for immediate write, but
2427 * which are still considered free by the SPA because the last transaction
2428 * group didn't commit yet.
2431 metaslab_claim_dva(spa_t
*spa
, const dva_t
*dva
, uint64_t txg
)
2433 uint64_t vdev
= DVA_GET_VDEV(dva
);
2434 uint64_t offset
= DVA_GET_OFFSET(dva
);
2435 uint64_t size
= DVA_GET_ASIZE(dva
);
2440 ASSERT(DVA_IS_VALID(dva
));
2442 if ((vd
= vdev_lookup_top(spa
, vdev
)) == NULL
||
2443 (offset
>> vd
->vdev_ms_shift
) >= vd
->vdev_ms_count
)
2444 return (SET_ERROR(ENXIO
));
2446 msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2448 if (DVA_GET_GANG(dva
))
2449 size
= vdev_psize_to_asize(vd
, SPA_GANGBLOCKSIZE
);
2451 mutex_enter(&msp
->ms_lock
);
2453 if ((txg
!= 0 && spa_writeable(spa
)) || !msp
->ms_loaded
)
2454 error
= metaslab_activate(msp
, METASLAB_WEIGHT_SECONDARY
);
2456 if (error
== 0 && !range_tree_contains(msp
->ms_tree
, offset
, size
))
2457 error
= SET_ERROR(ENOENT
);
2459 if (error
|| txg
== 0) { /* txg == 0 indicates dry run */
2460 mutex_exit(&msp
->ms_lock
);
2464 VERIFY(!msp
->ms_condensing
);
2465 VERIFY0(P2PHASE(offset
, 1ULL << vd
->vdev_ashift
));
2466 VERIFY0(P2PHASE(size
, 1ULL << vd
->vdev_ashift
));
2467 VERIFY3U(range_tree_space(msp
->ms_tree
) - size
, <=, msp
->ms_size
);
2468 range_tree_remove(msp
->ms_tree
, offset
, size
);
2470 if (spa_writeable(spa
)) { /* don't dirty if we're zdb(1M) */
2471 if (range_tree_space(msp
->ms_alloctree
[txg
& TXG_MASK
]) == 0)
2472 vdev_dirty(vd
, VDD_METASLAB
, msp
, txg
);
2473 range_tree_add(msp
->ms_alloctree
[txg
& TXG_MASK
], offset
, size
);
2476 mutex_exit(&msp
->ms_lock
);
2482 metaslab_alloc(spa_t
*spa
, metaslab_class_t
*mc
, uint64_t psize
, blkptr_t
*bp
,
2483 int ndvas
, uint64_t txg
, blkptr_t
*hintbp
, int flags
)
2485 dva_t
*dva
= bp
->blk_dva
;
2486 dva_t
*hintdva
= hintbp
->blk_dva
;
2489 ASSERT(bp
->blk_birth
== 0);
2490 ASSERT(BP_PHYSICAL_BIRTH(bp
) == 0);
2492 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2494 if (mc
->mc_rotor
== NULL
) { /* no vdevs in this class */
2495 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2496 return (SET_ERROR(ENOSPC
));
2499 ASSERT(ndvas
> 0 && ndvas
<= spa_max_replication(spa
));
2500 ASSERT(BP_GET_NDVAS(bp
) == 0);
2501 ASSERT(hintbp
== NULL
|| ndvas
<= BP_GET_NDVAS(hintbp
));
2503 for (d
= 0; d
< ndvas
; d
++) {
2504 error
= metaslab_alloc_dva(spa
, mc
, psize
, dva
, d
, hintdva
,
2507 for (d
--; d
>= 0; d
--) {
2508 metaslab_free_dva(spa
, &dva
[d
], txg
, B_TRUE
);
2509 bzero(&dva
[d
], sizeof (dva_t
));
2511 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2516 ASSERT(BP_GET_NDVAS(bp
) == ndvas
);
2518 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2520 BP_SET_BIRTH(bp
, txg
, txg
);
2526 metaslab_free(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
, boolean_t now
)
2528 const dva_t
*dva
= bp
->blk_dva
;
2529 int d
, ndvas
= BP_GET_NDVAS(bp
);
2531 ASSERT(!BP_IS_HOLE(bp
));
2532 ASSERT(!now
|| bp
->blk_birth
>= spa_syncing_txg(spa
));
2534 spa_config_enter(spa
, SCL_FREE
, FTAG
, RW_READER
);
2536 for (d
= 0; d
< ndvas
; d
++)
2537 metaslab_free_dva(spa
, &dva
[d
], txg
, now
);
2539 spa_config_exit(spa
, SCL_FREE
, FTAG
);
2543 metaslab_claim(spa_t
*spa
, const blkptr_t
*bp
, uint64_t txg
)
2545 const dva_t
*dva
= bp
->blk_dva
;
2546 int ndvas
= BP_GET_NDVAS(bp
);
2549 ASSERT(!BP_IS_HOLE(bp
));
2553 * First do a dry run to make sure all DVAs are claimable,
2554 * so we don't have to unwind from partial failures below.
2556 if ((error
= metaslab_claim(spa
, bp
, 0)) != 0)
2560 spa_config_enter(spa
, SCL_ALLOC
, FTAG
, RW_READER
);
2562 for (d
= 0; d
< ndvas
; d
++)
2563 if ((error
= metaslab_claim_dva(spa
, &dva
[d
], txg
)) != 0)
2566 spa_config_exit(spa
, SCL_ALLOC
, FTAG
);
2568 ASSERT(error
== 0 || txg
== 0);
2574 metaslab_fastwrite_mark(spa_t
*spa
, const blkptr_t
*bp
)
2576 const dva_t
*dva
= bp
->blk_dva
;
2577 int ndvas
= BP_GET_NDVAS(bp
);
2578 uint64_t psize
= BP_GET_PSIZE(bp
);
2582 ASSERT(!BP_IS_HOLE(bp
));
2583 ASSERT(!BP_IS_EMBEDDED(bp
));
2586 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2588 for (d
= 0; d
< ndvas
; d
++) {
2589 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2591 atomic_add_64(&vd
->vdev_pending_fastwrite
, psize
);
2594 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2598 metaslab_fastwrite_unmark(spa_t
*spa
, const blkptr_t
*bp
)
2600 const dva_t
*dva
= bp
->blk_dva
;
2601 int ndvas
= BP_GET_NDVAS(bp
);
2602 uint64_t psize
= BP_GET_PSIZE(bp
);
2606 ASSERT(!BP_IS_HOLE(bp
));
2607 ASSERT(!BP_IS_EMBEDDED(bp
));
2610 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2612 for (d
= 0; d
< ndvas
; d
++) {
2613 if ((vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[d
]))) == NULL
)
2615 ASSERT3U(vd
->vdev_pending_fastwrite
, >=, psize
);
2616 atomic_sub_64(&vd
->vdev_pending_fastwrite
, psize
);
2619 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2623 metaslab_check_free(spa_t
*spa
, const blkptr_t
*bp
)
2627 if ((zfs_flags
& ZFS_DEBUG_ZIO_FREE
) == 0)
2630 spa_config_enter(spa
, SCL_VDEV
, FTAG
, RW_READER
);
2631 for (i
= 0; i
< BP_GET_NDVAS(bp
); i
++) {
2632 uint64_t vdev
= DVA_GET_VDEV(&bp
->blk_dva
[i
]);
2633 vdev_t
*vd
= vdev_lookup_top(spa
, vdev
);
2634 uint64_t offset
= DVA_GET_OFFSET(&bp
->blk_dva
[i
]);
2635 uint64_t size
= DVA_GET_ASIZE(&bp
->blk_dva
[i
]);
2636 metaslab_t
*msp
= vd
->vdev_ms
[offset
>> vd
->vdev_ms_shift
];
2639 range_tree_verify(msp
->ms_tree
, offset
, size
);
2641 for (j
= 0; j
< TXG_SIZE
; j
++)
2642 range_tree_verify(msp
->ms_freetree
[j
], offset
, size
);
2643 for (j
= 0; j
< TXG_DEFER_SIZE
; j
++)
2644 range_tree_verify(msp
->ms_defertree
[j
], offset
, size
);
2646 spa_config_exit(spa
, SCL_VDEV
, FTAG
);
2649 #if defined(_KERNEL) && defined(HAVE_SPL)
2650 module_param(metaslab_debug_load
, int, 0644);
2651 module_param(metaslab_debug_unload
, int, 0644);
2652 module_param(metaslab_preload_enabled
, int, 0644);
2653 module_param(zfs_mg_noalloc_threshold
, int, 0644);
2654 module_param(zfs_mg_fragmentation_threshold
, int, 0644);
2655 module_param(zfs_metaslab_fragmentation_threshold
, int, 0644);
2656 module_param(metaslab_fragmentation_factor_enabled
, int, 0644);
2657 module_param(metaslab_lba_weighting_enabled
, int, 0644);
2658 module_param(metaslab_bias_enabled
, int, 0644);
2660 MODULE_PARM_DESC(metaslab_debug_load
,
2661 "load all metaslabs when pool is first opened");
2662 MODULE_PARM_DESC(metaslab_debug_unload
,
2663 "prevent metaslabs from being unloaded");
2664 MODULE_PARM_DESC(metaslab_preload_enabled
,
2665 "preload potential metaslabs during reassessment");
2667 MODULE_PARM_DESC(zfs_mg_noalloc_threshold
,
2668 "percentage of free space for metaslab group to allow allocation");
2669 MODULE_PARM_DESC(zfs_mg_fragmentation_threshold
,
2670 "fragmentation for metaslab group to allow allocation");
2672 MODULE_PARM_DESC(zfs_metaslab_fragmentation_threshold
,
2673 "fragmentation for metaslab to allow allocation");
2674 MODULE_PARM_DESC(metaslab_fragmentation_factor_enabled
,
2675 "use the fragmentation metric to prefer less fragmented metaslabs");
2676 MODULE_PARM_DESC(metaslab_lba_weighting_enabled
,
2677 "prefer metaslabs with lower LBAs");
2678 MODULE_PARM_DESC(metaslab_bias_enabled
,
2679 "enable metaslab group biasing");
2680 #endif /* _KERNEL && HAVE_SPL */