1 // SPDX-License-Identifier: GPL-2.0
3 #include <linux/list_sort.h>
6 #include "block-group.h"
7 #include "space-info.h"
9 #include "free-space-cache.h"
10 #include "free-space-tree.h"
12 #include "transaction.h"
13 #include "ref-verify.h"
16 #include "delalloc-space.h"
22 * Return target flags in extended format or 0 if restripe for this chunk_type
25 * Should be called with balance_lock held
27 static u64
get_restripe_target(struct btrfs_fs_info
*fs_info
, u64 flags
)
29 struct btrfs_balance_control
*bctl
= fs_info
->balance_ctl
;
35 if (flags
& BTRFS_BLOCK_GROUP_DATA
&&
36 bctl
->data
.flags
& BTRFS_BALANCE_ARGS_CONVERT
) {
37 target
= BTRFS_BLOCK_GROUP_DATA
| bctl
->data
.target
;
38 } else if (flags
& BTRFS_BLOCK_GROUP_SYSTEM
&&
39 bctl
->sys
.flags
& BTRFS_BALANCE_ARGS_CONVERT
) {
40 target
= BTRFS_BLOCK_GROUP_SYSTEM
| bctl
->sys
.target
;
41 } else if (flags
& BTRFS_BLOCK_GROUP_METADATA
&&
42 bctl
->meta
.flags
& BTRFS_BALANCE_ARGS_CONVERT
) {
43 target
= BTRFS_BLOCK_GROUP_METADATA
| bctl
->meta
.target
;
50 * @flags: available profiles in extended format (see ctree.h)
52 * Return reduced profile in chunk format. If profile changing is in progress
53 * (either running or paused) picks the target profile (if it's already
54 * available), otherwise falls back to plain reducing.
56 static u64
btrfs_reduce_alloc_profile(struct btrfs_fs_info
*fs_info
, u64 flags
)
58 u64 num_devices
= fs_info
->fs_devices
->rw_devices
;
64 * See if restripe for this chunk_type is in progress, if so try to
65 * reduce to the target profile
67 spin_lock(&fs_info
->balance_lock
);
68 target
= get_restripe_target(fs_info
, flags
);
70 spin_unlock(&fs_info
->balance_lock
);
71 return extended_to_chunk(target
);
73 spin_unlock(&fs_info
->balance_lock
);
75 /* First, mask out the RAID levels which aren't possible */
76 for (raid_type
= 0; raid_type
< BTRFS_NR_RAID_TYPES
; raid_type
++) {
77 if (num_devices
>= btrfs_raid_array
[raid_type
].devs_min
)
78 allowed
|= btrfs_raid_array
[raid_type
].bg_flag
;
82 if (allowed
& BTRFS_BLOCK_GROUP_RAID6
)
83 allowed
= BTRFS_BLOCK_GROUP_RAID6
;
84 else if (allowed
& BTRFS_BLOCK_GROUP_RAID5
)
85 allowed
= BTRFS_BLOCK_GROUP_RAID5
;
86 else if (allowed
& BTRFS_BLOCK_GROUP_RAID10
)
87 allowed
= BTRFS_BLOCK_GROUP_RAID10
;
88 else if (allowed
& BTRFS_BLOCK_GROUP_RAID1
)
89 allowed
= BTRFS_BLOCK_GROUP_RAID1
;
90 else if (allowed
& BTRFS_BLOCK_GROUP_RAID0
)
91 allowed
= BTRFS_BLOCK_GROUP_RAID0
;
93 flags
&= ~BTRFS_BLOCK_GROUP_PROFILE_MASK
;
95 return extended_to_chunk(flags
| allowed
);
98 u64
btrfs_get_alloc_profile(struct btrfs_fs_info
*fs_info
, u64 orig_flags
)
105 seq
= read_seqbegin(&fs_info
->profiles_lock
);
107 if (flags
& BTRFS_BLOCK_GROUP_DATA
)
108 flags
|= fs_info
->avail_data_alloc_bits
;
109 else if (flags
& BTRFS_BLOCK_GROUP_SYSTEM
)
110 flags
|= fs_info
->avail_system_alloc_bits
;
111 else if (flags
& BTRFS_BLOCK_GROUP_METADATA
)
112 flags
|= fs_info
->avail_metadata_alloc_bits
;
113 } while (read_seqretry(&fs_info
->profiles_lock
, seq
));
115 return btrfs_reduce_alloc_profile(fs_info
, flags
);
118 void btrfs_get_block_group(struct btrfs_block_group
*cache
)
120 refcount_inc(&cache
->refs
);
123 void btrfs_put_block_group(struct btrfs_block_group
*cache
)
125 if (refcount_dec_and_test(&cache
->refs
)) {
126 WARN_ON(cache
->pinned
> 0);
128 * If there was a failure to cleanup a log tree, very likely due
129 * to an IO failure on a writeback attempt of one or more of its
130 * extent buffers, we could not do proper (and cheap) unaccounting
131 * of their reserved space, so don't warn on reserved > 0 in that
134 if (!(cache
->flags
& BTRFS_BLOCK_GROUP_METADATA
) ||
135 !BTRFS_FS_LOG_CLEANUP_ERROR(cache
->fs_info
))
136 WARN_ON(cache
->reserved
> 0);
139 * A block_group shouldn't be on the discard_list anymore.
140 * Remove the block_group from the discard_list to prevent us
141 * from causing a panic due to NULL pointer dereference.
143 if (WARN_ON(!list_empty(&cache
->discard_list
)))
144 btrfs_discard_cancel_work(&cache
->fs_info
->discard_ctl
,
148 * If not empty, someone is still holding mutex of
149 * full_stripe_lock, which can only be released by caller.
150 * And it will definitely cause use-after-free when caller
151 * tries to release full stripe lock.
153 * No better way to resolve, but only to warn.
155 WARN_ON(!RB_EMPTY_ROOT(&cache
->full_stripe_locks_root
.root
));
156 kfree(cache
->free_space_ctl
);
157 kfree(cache
->physical_map
);
163 * This adds the block group to the fs_info rb tree for the block group cache
165 static int btrfs_add_block_group_cache(struct btrfs_fs_info
*info
,
166 struct btrfs_block_group
*block_group
)
169 struct rb_node
*parent
= NULL
;
170 struct btrfs_block_group
*cache
;
171 bool leftmost
= true;
173 ASSERT(block_group
->length
!= 0);
175 write_lock(&info
->block_group_cache_lock
);
176 p
= &info
->block_group_cache_tree
.rb_root
.rb_node
;
180 cache
= rb_entry(parent
, struct btrfs_block_group
, cache_node
);
181 if (block_group
->start
< cache
->start
) {
183 } else if (block_group
->start
> cache
->start
) {
187 write_unlock(&info
->block_group_cache_lock
);
192 rb_link_node(&block_group
->cache_node
, parent
, p
);
193 rb_insert_color_cached(&block_group
->cache_node
,
194 &info
->block_group_cache_tree
, leftmost
);
196 write_unlock(&info
->block_group_cache_lock
);
202 * This will return the block group at or after bytenr if contains is 0, else
203 * it will return the block group that contains the bytenr
205 static struct btrfs_block_group
*block_group_cache_tree_search(
206 struct btrfs_fs_info
*info
, u64 bytenr
, int contains
)
208 struct btrfs_block_group
*cache
, *ret
= NULL
;
212 read_lock(&info
->block_group_cache_lock
);
213 n
= info
->block_group_cache_tree
.rb_root
.rb_node
;
216 cache
= rb_entry(n
, struct btrfs_block_group
, cache_node
);
217 end
= cache
->start
+ cache
->length
- 1;
218 start
= cache
->start
;
220 if (bytenr
< start
) {
221 if (!contains
&& (!ret
|| start
< ret
->start
))
224 } else if (bytenr
> start
) {
225 if (contains
&& bytenr
<= end
) {
236 btrfs_get_block_group(ret
);
237 read_unlock(&info
->block_group_cache_lock
);
243 * Return the block group that starts at or after bytenr
245 struct btrfs_block_group
*btrfs_lookup_first_block_group(
246 struct btrfs_fs_info
*info
, u64 bytenr
)
248 return block_group_cache_tree_search(info
, bytenr
, 0);
252 * Return the block group that contains the given bytenr
254 struct btrfs_block_group
*btrfs_lookup_block_group(
255 struct btrfs_fs_info
*info
, u64 bytenr
)
257 return block_group_cache_tree_search(info
, bytenr
, 1);
260 struct btrfs_block_group
*btrfs_next_block_group(
261 struct btrfs_block_group
*cache
)
263 struct btrfs_fs_info
*fs_info
= cache
->fs_info
;
264 struct rb_node
*node
;
266 read_lock(&fs_info
->block_group_cache_lock
);
268 /* If our block group was removed, we need a full search. */
269 if (RB_EMPTY_NODE(&cache
->cache_node
)) {
270 const u64 next_bytenr
= cache
->start
+ cache
->length
;
272 read_unlock(&fs_info
->block_group_cache_lock
);
273 btrfs_put_block_group(cache
);
274 return btrfs_lookup_first_block_group(fs_info
, next_bytenr
);
276 node
= rb_next(&cache
->cache_node
);
277 btrfs_put_block_group(cache
);
279 cache
= rb_entry(node
, struct btrfs_block_group
, cache_node
);
280 btrfs_get_block_group(cache
);
283 read_unlock(&fs_info
->block_group_cache_lock
);
288 * Check if we can do a NOCOW write for a given extent.
290 * @fs_info: The filesystem information object.
291 * @bytenr: Logical start address of the extent.
293 * Check if we can do a NOCOW write for the given extent, and increments the
294 * number of NOCOW writers in the block group that contains the extent, as long
295 * as the block group exists and it's currently not in read-only mode.
297 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
298 * is responsible for calling btrfs_dec_nocow_writers() later.
300 * Or NULL if we can not do a NOCOW write
302 struct btrfs_block_group
*btrfs_inc_nocow_writers(struct btrfs_fs_info
*fs_info
,
305 struct btrfs_block_group
*bg
;
306 bool can_nocow
= true;
308 bg
= btrfs_lookup_block_group(fs_info
, bytenr
);
312 spin_lock(&bg
->lock
);
316 atomic_inc(&bg
->nocow_writers
);
317 spin_unlock(&bg
->lock
);
320 btrfs_put_block_group(bg
);
324 /* No put on block group, done by btrfs_dec_nocow_writers(). */
329 * Decrement the number of NOCOW writers in a block group.
331 * @bg: The block group.
333 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
334 * and on the block group returned by that call. Typically this is called after
335 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
338 * After this call, the caller should not use the block group anymore. It it wants
339 * to use it, then it should get a reference on it before calling this function.
341 void btrfs_dec_nocow_writers(struct btrfs_block_group
*bg
)
343 if (atomic_dec_and_test(&bg
->nocow_writers
))
344 wake_up_var(&bg
->nocow_writers
);
346 /* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
347 btrfs_put_block_group(bg
);
350 void btrfs_wait_nocow_writers(struct btrfs_block_group
*bg
)
352 wait_var_event(&bg
->nocow_writers
, !atomic_read(&bg
->nocow_writers
));
355 void btrfs_dec_block_group_reservations(struct btrfs_fs_info
*fs_info
,
358 struct btrfs_block_group
*bg
;
360 bg
= btrfs_lookup_block_group(fs_info
, start
);
362 if (atomic_dec_and_test(&bg
->reservations
))
363 wake_up_var(&bg
->reservations
);
364 btrfs_put_block_group(bg
);
367 void btrfs_wait_block_group_reservations(struct btrfs_block_group
*bg
)
369 struct btrfs_space_info
*space_info
= bg
->space_info
;
373 if (!(bg
->flags
& BTRFS_BLOCK_GROUP_DATA
))
377 * Our block group is read only but before we set it to read only,
378 * some task might have had allocated an extent from it already, but it
379 * has not yet created a respective ordered extent (and added it to a
380 * root's list of ordered extents).
381 * Therefore wait for any task currently allocating extents, since the
382 * block group's reservations counter is incremented while a read lock
383 * on the groups' semaphore is held and decremented after releasing
384 * the read access on that semaphore and creating the ordered extent.
386 down_write(&space_info
->groups_sem
);
387 up_write(&space_info
->groups_sem
);
389 wait_var_event(&bg
->reservations
, !atomic_read(&bg
->reservations
));
392 struct btrfs_caching_control
*btrfs_get_caching_control(
393 struct btrfs_block_group
*cache
)
395 struct btrfs_caching_control
*ctl
;
397 spin_lock(&cache
->lock
);
398 if (!cache
->caching_ctl
) {
399 spin_unlock(&cache
->lock
);
403 ctl
= cache
->caching_ctl
;
404 refcount_inc(&ctl
->count
);
405 spin_unlock(&cache
->lock
);
409 void btrfs_put_caching_control(struct btrfs_caching_control
*ctl
)
411 if (refcount_dec_and_test(&ctl
->count
))
416 * When we wait for progress in the block group caching, its because our
417 * allocation attempt failed at least once. So, we must sleep and let some
418 * progress happen before we try again.
420 * This function will sleep at least once waiting for new free space to show
421 * up, and then it will check the block group free space numbers for our min
422 * num_bytes. Another option is to have it go ahead and look in the rbtree for
423 * a free extent of a given size, but this is a good start.
425 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
426 * any of the information in this block group.
428 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group
*cache
,
431 struct btrfs_caching_control
*caching_ctl
;
433 caching_ctl
= btrfs_get_caching_control(cache
);
437 wait_event(caching_ctl
->wait
, btrfs_block_group_done(cache
) ||
438 (cache
->free_space_ctl
->free_space
>= num_bytes
));
440 btrfs_put_caching_control(caching_ctl
);
443 static int btrfs_caching_ctl_wait_done(struct btrfs_block_group
*cache
,
444 struct btrfs_caching_control
*caching_ctl
)
446 wait_event(caching_ctl
->wait
, btrfs_block_group_done(cache
));
447 return cache
->cached
== BTRFS_CACHE_ERROR
? -EIO
: 0;
450 static int btrfs_wait_block_group_cache_done(struct btrfs_block_group
*cache
)
452 struct btrfs_caching_control
*caching_ctl
;
455 caching_ctl
= btrfs_get_caching_control(cache
);
457 return (cache
->cached
== BTRFS_CACHE_ERROR
) ? -EIO
: 0;
458 ret
= btrfs_caching_ctl_wait_done(cache
, caching_ctl
);
459 btrfs_put_caching_control(caching_ctl
);
463 #ifdef CONFIG_BTRFS_DEBUG
464 static void fragment_free_space(struct btrfs_block_group
*block_group
)
466 struct btrfs_fs_info
*fs_info
= block_group
->fs_info
;
467 u64 start
= block_group
->start
;
468 u64 len
= block_group
->length
;
469 u64 chunk
= block_group
->flags
& BTRFS_BLOCK_GROUP_METADATA
?
470 fs_info
->nodesize
: fs_info
->sectorsize
;
471 u64 step
= chunk
<< 1;
473 while (len
> chunk
) {
474 btrfs_remove_free_space(block_group
, start
, chunk
);
485 * This is only called by btrfs_cache_block_group, since we could have freed
486 * extents we need to check the pinned_extents for any extents that can't be
487 * used yet since their free space will be released as soon as the transaction
490 u64
add_new_free_space(struct btrfs_block_group
*block_group
, u64 start
, u64 end
)
492 struct btrfs_fs_info
*info
= block_group
->fs_info
;
493 u64 extent_start
, extent_end
, size
, total_added
= 0;
496 while (start
< end
) {
497 ret
= find_first_extent_bit(&info
->excluded_extents
, start
,
498 &extent_start
, &extent_end
,
499 EXTENT_DIRTY
| EXTENT_UPTODATE
,
504 if (extent_start
<= start
) {
505 start
= extent_end
+ 1;
506 } else if (extent_start
> start
&& extent_start
< end
) {
507 size
= extent_start
- start
;
509 ret
= btrfs_add_free_space_async_trimmed(block_group
,
511 BUG_ON(ret
); /* -ENOMEM or logic error */
512 start
= extent_end
+ 1;
521 ret
= btrfs_add_free_space_async_trimmed(block_group
, start
,
523 BUG_ON(ret
); /* -ENOMEM or logic error */
529 static int load_extent_tree_free(struct btrfs_caching_control
*caching_ctl
)
531 struct btrfs_block_group
*block_group
= caching_ctl
->block_group
;
532 struct btrfs_fs_info
*fs_info
= block_group
->fs_info
;
533 struct btrfs_root
*extent_root
;
534 struct btrfs_path
*path
;
535 struct extent_buffer
*leaf
;
536 struct btrfs_key key
;
543 path
= btrfs_alloc_path();
547 last
= max_t(u64
, block_group
->start
, BTRFS_SUPER_INFO_OFFSET
);
548 extent_root
= btrfs_extent_root(fs_info
, last
);
550 #ifdef CONFIG_BTRFS_DEBUG
552 * If we're fragmenting we don't want to make anybody think we can
553 * allocate from this block group until we've had a chance to fragment
556 if (btrfs_should_fragment_free_space(block_group
))
560 * We don't want to deadlock with somebody trying to allocate a new
561 * extent for the extent root while also trying to search the extent
562 * root to add free space. So we skip locking and search the commit
563 * root, since its read-only
565 path
->skip_locking
= 1;
566 path
->search_commit_root
= 1;
567 path
->reada
= READA_FORWARD
;
571 key
.type
= BTRFS_EXTENT_ITEM_KEY
;
574 ret
= btrfs_search_slot(NULL
, extent_root
, &key
, path
, 0, 0);
578 leaf
= path
->nodes
[0];
579 nritems
= btrfs_header_nritems(leaf
);
582 if (btrfs_fs_closing(fs_info
) > 1) {
587 if (path
->slots
[0] < nritems
) {
588 btrfs_item_key_to_cpu(leaf
, &key
, path
->slots
[0]);
590 ret
= btrfs_find_next_key(extent_root
, path
, &key
, 0, 0);
594 if (need_resched() ||
595 rwsem_is_contended(&fs_info
->commit_root_sem
)) {
596 btrfs_release_path(path
);
597 up_read(&fs_info
->commit_root_sem
);
598 mutex_unlock(&caching_ctl
->mutex
);
600 mutex_lock(&caching_ctl
->mutex
);
601 down_read(&fs_info
->commit_root_sem
);
605 ret
= btrfs_next_leaf(extent_root
, path
);
610 leaf
= path
->nodes
[0];
611 nritems
= btrfs_header_nritems(leaf
);
615 if (key
.objectid
< last
) {
618 key
.type
= BTRFS_EXTENT_ITEM_KEY
;
619 btrfs_release_path(path
);
623 if (key
.objectid
< block_group
->start
) {
628 if (key
.objectid
>= block_group
->start
+ block_group
->length
)
631 if (key
.type
== BTRFS_EXTENT_ITEM_KEY
||
632 key
.type
== BTRFS_METADATA_ITEM_KEY
) {
633 total_found
+= add_new_free_space(block_group
, last
,
635 if (key
.type
== BTRFS_METADATA_ITEM_KEY
)
636 last
= key
.objectid
+
639 last
= key
.objectid
+ key
.offset
;
641 if (total_found
> CACHING_CTL_WAKE_UP
) {
644 wake_up(&caching_ctl
->wait
);
651 total_found
+= add_new_free_space(block_group
, last
,
652 block_group
->start
+ block_group
->length
);
655 btrfs_free_path(path
);
659 static noinline
void caching_thread(struct btrfs_work
*work
)
661 struct btrfs_block_group
*block_group
;
662 struct btrfs_fs_info
*fs_info
;
663 struct btrfs_caching_control
*caching_ctl
;
666 caching_ctl
= container_of(work
, struct btrfs_caching_control
, work
);
667 block_group
= caching_ctl
->block_group
;
668 fs_info
= block_group
->fs_info
;
670 mutex_lock(&caching_ctl
->mutex
);
671 down_read(&fs_info
->commit_root_sem
);
673 if (btrfs_test_opt(fs_info
, SPACE_CACHE
)) {
674 ret
= load_free_space_cache(block_group
);
681 * We failed to load the space cache, set ourselves to
682 * CACHE_STARTED and carry on.
684 spin_lock(&block_group
->lock
);
685 block_group
->cached
= BTRFS_CACHE_STARTED
;
686 spin_unlock(&block_group
->lock
);
687 wake_up(&caching_ctl
->wait
);
691 * If we are in the transaction that populated the free space tree we
692 * can't actually cache from the free space tree as our commit root and
693 * real root are the same, so we could change the contents of the blocks
694 * while caching. Instead do the slow caching in this case, and after
695 * the transaction has committed we will be safe.
697 if (btrfs_fs_compat_ro(fs_info
, FREE_SPACE_TREE
) &&
698 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED
, &fs_info
->flags
)))
699 ret
= load_free_space_tree(caching_ctl
);
701 ret
= load_extent_tree_free(caching_ctl
);
703 spin_lock(&block_group
->lock
);
704 block_group
->caching_ctl
= NULL
;
705 block_group
->cached
= ret
? BTRFS_CACHE_ERROR
: BTRFS_CACHE_FINISHED
;
706 spin_unlock(&block_group
->lock
);
708 #ifdef CONFIG_BTRFS_DEBUG
709 if (btrfs_should_fragment_free_space(block_group
)) {
712 spin_lock(&block_group
->space_info
->lock
);
713 spin_lock(&block_group
->lock
);
714 bytes_used
= block_group
->length
- block_group
->used
;
715 block_group
->space_info
->bytes_used
+= bytes_used
>> 1;
716 spin_unlock(&block_group
->lock
);
717 spin_unlock(&block_group
->space_info
->lock
);
718 fragment_free_space(block_group
);
722 up_read(&fs_info
->commit_root_sem
);
723 btrfs_free_excluded_extents(block_group
);
724 mutex_unlock(&caching_ctl
->mutex
);
726 wake_up(&caching_ctl
->wait
);
728 btrfs_put_caching_control(caching_ctl
);
729 btrfs_put_block_group(block_group
);
732 int btrfs_cache_block_group(struct btrfs_block_group
*cache
, bool wait
)
734 struct btrfs_fs_info
*fs_info
= cache
->fs_info
;
735 struct btrfs_caching_control
*caching_ctl
= NULL
;
738 /* Allocator for zoned filesystems does not use the cache at all */
739 if (btrfs_is_zoned(fs_info
))
742 caching_ctl
= kzalloc(sizeof(*caching_ctl
), GFP_NOFS
);
746 INIT_LIST_HEAD(&caching_ctl
->list
);
747 mutex_init(&caching_ctl
->mutex
);
748 init_waitqueue_head(&caching_ctl
->wait
);
749 caching_ctl
->block_group
= cache
;
750 refcount_set(&caching_ctl
->count
, 2);
751 btrfs_init_work(&caching_ctl
->work
, caching_thread
, NULL
, NULL
);
753 spin_lock(&cache
->lock
);
754 if (cache
->cached
!= BTRFS_CACHE_NO
) {
757 caching_ctl
= cache
->caching_ctl
;
759 refcount_inc(&caching_ctl
->count
);
760 spin_unlock(&cache
->lock
);
763 WARN_ON(cache
->caching_ctl
);
764 cache
->caching_ctl
= caching_ctl
;
765 cache
->cached
= BTRFS_CACHE_STARTED
;
766 spin_unlock(&cache
->lock
);
768 write_lock(&fs_info
->block_group_cache_lock
);
769 refcount_inc(&caching_ctl
->count
);
770 list_add_tail(&caching_ctl
->list
, &fs_info
->caching_block_groups
);
771 write_unlock(&fs_info
->block_group_cache_lock
);
773 btrfs_get_block_group(cache
);
775 btrfs_queue_work(fs_info
->caching_workers
, &caching_ctl
->work
);
778 if (wait
&& caching_ctl
)
779 ret
= btrfs_caching_ctl_wait_done(cache
, caching_ctl
);
780 /* wait_event(caching_ctl->wait, space_cache_v1_done(cache)); */
782 btrfs_put_caching_control(caching_ctl
);
787 static void clear_avail_alloc_bits(struct btrfs_fs_info
*fs_info
, u64 flags
)
789 u64 extra_flags
= chunk_to_extended(flags
) &
790 BTRFS_EXTENDED_PROFILE_MASK
;
792 write_seqlock(&fs_info
->profiles_lock
);
793 if (flags
& BTRFS_BLOCK_GROUP_DATA
)
794 fs_info
->avail_data_alloc_bits
&= ~extra_flags
;
795 if (flags
& BTRFS_BLOCK_GROUP_METADATA
)
796 fs_info
->avail_metadata_alloc_bits
&= ~extra_flags
;
797 if (flags
& BTRFS_BLOCK_GROUP_SYSTEM
)
798 fs_info
->avail_system_alloc_bits
&= ~extra_flags
;
799 write_sequnlock(&fs_info
->profiles_lock
);
803 * Clear incompat bits for the following feature(s):
805 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
806 * in the whole filesystem
808 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
810 static void clear_incompat_bg_bits(struct btrfs_fs_info
*fs_info
, u64 flags
)
812 bool found_raid56
= false;
813 bool found_raid1c34
= false;
815 if ((flags
& BTRFS_BLOCK_GROUP_RAID56_MASK
) ||
816 (flags
& BTRFS_BLOCK_GROUP_RAID1C3
) ||
817 (flags
& BTRFS_BLOCK_GROUP_RAID1C4
)) {
818 struct list_head
*head
= &fs_info
->space_info
;
819 struct btrfs_space_info
*sinfo
;
821 list_for_each_entry_rcu(sinfo
, head
, list
) {
822 down_read(&sinfo
->groups_sem
);
823 if (!list_empty(&sinfo
->block_groups
[BTRFS_RAID_RAID5
]))
825 if (!list_empty(&sinfo
->block_groups
[BTRFS_RAID_RAID6
]))
827 if (!list_empty(&sinfo
->block_groups
[BTRFS_RAID_RAID1C3
]))
828 found_raid1c34
= true;
829 if (!list_empty(&sinfo
->block_groups
[BTRFS_RAID_RAID1C4
]))
830 found_raid1c34
= true;
831 up_read(&sinfo
->groups_sem
);
834 btrfs_clear_fs_incompat(fs_info
, RAID56
);
836 btrfs_clear_fs_incompat(fs_info
, RAID1C34
);
840 static int remove_block_group_item(struct btrfs_trans_handle
*trans
,
841 struct btrfs_path
*path
,
842 struct btrfs_block_group
*block_group
)
844 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
845 struct btrfs_root
*root
;
846 struct btrfs_key key
;
849 root
= btrfs_block_group_root(fs_info
);
850 key
.objectid
= block_group
->start
;
851 key
.type
= BTRFS_BLOCK_GROUP_ITEM_KEY
;
852 key
.offset
= block_group
->length
;
854 ret
= btrfs_search_slot(trans
, root
, &key
, path
, -1, 1);
860 ret
= btrfs_del_item(trans
, root
, path
);
864 int btrfs_remove_block_group(struct btrfs_trans_handle
*trans
,
865 u64 group_start
, struct extent_map
*em
)
867 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
868 struct btrfs_path
*path
;
869 struct btrfs_block_group
*block_group
;
870 struct btrfs_free_cluster
*cluster
;
872 struct kobject
*kobj
= NULL
;
876 struct btrfs_caching_control
*caching_ctl
= NULL
;
878 bool remove_rsv
= false;
880 block_group
= btrfs_lookup_block_group(fs_info
, group_start
);
881 BUG_ON(!block_group
);
882 BUG_ON(!block_group
->ro
);
884 trace_btrfs_remove_block_group(block_group
);
886 * Free the reserved super bytes from this block group before
889 btrfs_free_excluded_extents(block_group
);
890 btrfs_free_ref_tree_range(fs_info
, block_group
->start
,
891 block_group
->length
);
893 index
= btrfs_bg_flags_to_raid_index(block_group
->flags
);
894 factor
= btrfs_bg_type_to_factor(block_group
->flags
);
896 /* make sure this block group isn't part of an allocation cluster */
897 cluster
= &fs_info
->data_alloc_cluster
;
898 spin_lock(&cluster
->refill_lock
);
899 btrfs_return_cluster_to_free_space(block_group
, cluster
);
900 spin_unlock(&cluster
->refill_lock
);
903 * make sure this block group isn't part of a metadata
906 cluster
= &fs_info
->meta_alloc_cluster
;
907 spin_lock(&cluster
->refill_lock
);
908 btrfs_return_cluster_to_free_space(block_group
, cluster
);
909 spin_unlock(&cluster
->refill_lock
);
911 btrfs_clear_treelog_bg(block_group
);
912 btrfs_clear_data_reloc_bg(block_group
);
914 path
= btrfs_alloc_path();
921 * get the inode first so any iput calls done for the io_list
922 * aren't the final iput (no unlinks allowed now)
924 inode
= lookup_free_space_inode(block_group
, path
);
926 mutex_lock(&trans
->transaction
->cache_write_mutex
);
928 * Make sure our free space cache IO is done before removing the
931 spin_lock(&trans
->transaction
->dirty_bgs_lock
);
932 if (!list_empty(&block_group
->io_list
)) {
933 list_del_init(&block_group
->io_list
);
935 WARN_ON(!IS_ERR(inode
) && inode
!= block_group
->io_ctl
.inode
);
937 spin_unlock(&trans
->transaction
->dirty_bgs_lock
);
938 btrfs_wait_cache_io(trans
, block_group
, path
);
939 btrfs_put_block_group(block_group
);
940 spin_lock(&trans
->transaction
->dirty_bgs_lock
);
943 if (!list_empty(&block_group
->dirty_list
)) {
944 list_del_init(&block_group
->dirty_list
);
946 btrfs_put_block_group(block_group
);
948 spin_unlock(&trans
->transaction
->dirty_bgs_lock
);
949 mutex_unlock(&trans
->transaction
->cache_write_mutex
);
951 ret
= btrfs_remove_free_space_inode(trans
, inode
, block_group
);
955 write_lock(&fs_info
->block_group_cache_lock
);
956 rb_erase_cached(&block_group
->cache_node
,
957 &fs_info
->block_group_cache_tree
);
958 RB_CLEAR_NODE(&block_group
->cache_node
);
960 /* Once for the block groups rbtree */
961 btrfs_put_block_group(block_group
);
963 write_unlock(&fs_info
->block_group_cache_lock
);
965 down_write(&block_group
->space_info
->groups_sem
);
967 * we must use list_del_init so people can check to see if they
968 * are still on the list after taking the semaphore
970 list_del_init(&block_group
->list
);
971 if (list_empty(&block_group
->space_info
->block_groups
[index
])) {
972 kobj
= block_group
->space_info
->block_group_kobjs
[index
];
973 block_group
->space_info
->block_group_kobjs
[index
] = NULL
;
974 clear_avail_alloc_bits(fs_info
, block_group
->flags
);
976 up_write(&block_group
->space_info
->groups_sem
);
977 clear_incompat_bg_bits(fs_info
, block_group
->flags
);
983 if (block_group
->cached
== BTRFS_CACHE_STARTED
)
984 btrfs_wait_block_group_cache_done(block_group
);
986 write_lock(&fs_info
->block_group_cache_lock
);
987 caching_ctl
= btrfs_get_caching_control(block_group
);
989 struct btrfs_caching_control
*ctl
;
991 list_for_each_entry(ctl
, &fs_info
->caching_block_groups
, list
) {
992 if (ctl
->block_group
== block_group
) {
994 refcount_inc(&caching_ctl
->count
);
1000 list_del_init(&caching_ctl
->list
);
1001 write_unlock(&fs_info
->block_group_cache_lock
);
1004 /* Once for the caching bgs list and once for us. */
1005 btrfs_put_caching_control(caching_ctl
);
1006 btrfs_put_caching_control(caching_ctl
);
1009 spin_lock(&trans
->transaction
->dirty_bgs_lock
);
1010 WARN_ON(!list_empty(&block_group
->dirty_list
));
1011 WARN_ON(!list_empty(&block_group
->io_list
));
1012 spin_unlock(&trans
->transaction
->dirty_bgs_lock
);
1014 btrfs_remove_free_space_cache(block_group
);
1016 spin_lock(&block_group
->space_info
->lock
);
1017 list_del_init(&block_group
->ro_list
);
1019 if (btrfs_test_opt(fs_info
, ENOSPC_DEBUG
)) {
1020 WARN_ON(block_group
->space_info
->total_bytes
1021 < block_group
->length
);
1022 WARN_ON(block_group
->space_info
->bytes_readonly
1023 < block_group
->length
- block_group
->zone_unusable
);
1024 WARN_ON(block_group
->space_info
->bytes_zone_unusable
1025 < block_group
->zone_unusable
);
1026 WARN_ON(block_group
->space_info
->disk_total
1027 < block_group
->length
* factor
);
1028 WARN_ON(test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE
,
1029 &block_group
->runtime_flags
) &&
1030 block_group
->space_info
->active_total_bytes
1031 < block_group
->length
);
1033 block_group
->space_info
->total_bytes
-= block_group
->length
;
1034 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE
, &block_group
->runtime_flags
))
1035 block_group
->space_info
->active_total_bytes
-= block_group
->length
;
1036 block_group
->space_info
->bytes_readonly
-=
1037 (block_group
->length
- block_group
->zone_unusable
);
1038 block_group
->space_info
->bytes_zone_unusable
-=
1039 block_group
->zone_unusable
;
1040 block_group
->space_info
->disk_total
-= block_group
->length
* factor
;
1042 spin_unlock(&block_group
->space_info
->lock
);
1045 * Remove the free space for the block group from the free space tree
1046 * and the block group's item from the extent tree before marking the
1047 * block group as removed. This is to prevent races with tasks that
1048 * freeze and unfreeze a block group, this task and another task
1049 * allocating a new block group - the unfreeze task ends up removing
1050 * the block group's extent map before the task calling this function
1051 * deletes the block group item from the extent tree, allowing for
1052 * another task to attempt to create another block group with the same
1053 * item key (and failing with -EEXIST and a transaction abort).
1055 ret
= remove_block_group_free_space(trans
, block_group
);
1059 ret
= remove_block_group_item(trans
, path
, block_group
);
1063 spin_lock(&block_group
->lock
);
1064 set_bit(BLOCK_GROUP_FLAG_REMOVED
, &block_group
->runtime_flags
);
1067 * At this point trimming or scrub can't start on this block group,
1068 * because we removed the block group from the rbtree
1069 * fs_info->block_group_cache_tree so no one can't find it anymore and
1070 * even if someone already got this block group before we removed it
1071 * from the rbtree, they have already incremented block_group->frozen -
1072 * if they didn't, for the trimming case they won't find any free space
1073 * entries because we already removed them all when we called
1074 * btrfs_remove_free_space_cache().
1076 * And we must not remove the extent map from the fs_info->mapping_tree
1077 * to prevent the same logical address range and physical device space
1078 * ranges from being reused for a new block group. This is needed to
1079 * avoid races with trimming and scrub.
1081 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1082 * completely transactionless, so while it is trimming a range the
1083 * currently running transaction might finish and a new one start,
1084 * allowing for new block groups to be created that can reuse the same
1085 * physical device locations unless we take this special care.
1087 * There may also be an implicit trim operation if the file system
1088 * is mounted with -odiscard. The same protections must remain
1089 * in place until the extents have been discarded completely when
1090 * the transaction commit has completed.
1092 remove_em
= (atomic_read(&block_group
->frozen
) == 0);
1093 spin_unlock(&block_group
->lock
);
1096 struct extent_map_tree
*em_tree
;
1098 em_tree
= &fs_info
->mapping_tree
;
1099 write_lock(&em_tree
->lock
);
1100 remove_extent_mapping(em_tree
, em
);
1101 write_unlock(&em_tree
->lock
);
1102 /* once for the tree */
1103 free_extent_map(em
);
1107 /* Once for the lookup reference */
1108 btrfs_put_block_group(block_group
);
1110 btrfs_delayed_refs_rsv_release(fs_info
, 1);
1111 btrfs_free_path(path
);
1115 struct btrfs_trans_handle
*btrfs_start_trans_remove_block_group(
1116 struct btrfs_fs_info
*fs_info
, const u64 chunk_offset
)
1118 struct btrfs_root
*root
= btrfs_block_group_root(fs_info
);
1119 struct extent_map_tree
*em_tree
= &fs_info
->mapping_tree
;
1120 struct extent_map
*em
;
1121 struct map_lookup
*map
;
1122 unsigned int num_items
;
1124 read_lock(&em_tree
->lock
);
1125 em
= lookup_extent_mapping(em_tree
, chunk_offset
, 1);
1126 read_unlock(&em_tree
->lock
);
1127 ASSERT(em
&& em
->start
== chunk_offset
);
1130 * We need to reserve 3 + N units from the metadata space info in order
1131 * to remove a block group (done at btrfs_remove_chunk() and at
1132 * btrfs_remove_block_group()), which are used for:
1134 * 1 unit for adding the free space inode's orphan (located in the tree
1136 * 1 unit for deleting the block group item (located in the extent
1138 * 1 unit for deleting the free space item (located in tree of tree
1140 * N units for deleting N device extent items corresponding to each
1141 * stripe (located in the device tree).
1143 * In order to remove a block group we also need to reserve units in the
1144 * system space info in order to update the chunk tree (update one or
1145 * more device items and remove one chunk item), but this is done at
1146 * btrfs_remove_chunk() through a call to check_system_chunk().
1148 map
= em
->map_lookup
;
1149 num_items
= 3 + map
->num_stripes
;
1150 free_extent_map(em
);
1152 return btrfs_start_transaction_fallback_global_rsv(root
, num_items
);
1156 * Mark block group @cache read-only, so later write won't happen to block
1159 * If @force is not set, this function will only mark the block group readonly
1160 * if we have enough free space (1M) in other metadata/system block groups.
1161 * If @force is not set, this function will mark the block group readonly
1162 * without checking free space.
1164 * NOTE: This function doesn't care if other block groups can contain all the
1165 * data in this block group. That check should be done by relocation routine,
1166 * not this function.
1168 static int inc_block_group_ro(struct btrfs_block_group
*cache
, int force
)
1170 struct btrfs_space_info
*sinfo
= cache
->space_info
;
1174 spin_lock(&sinfo
->lock
);
1175 spin_lock(&cache
->lock
);
1177 if (cache
->swap_extents
) {
1188 num_bytes
= cache
->length
- cache
->reserved
- cache
->pinned
-
1189 cache
->bytes_super
- cache
->zone_unusable
- cache
->used
;
1192 * Data never overcommits, even in mixed mode, so do just the straight
1193 * check of left over space in how much we have allocated.
1197 } else if (sinfo
->flags
& BTRFS_BLOCK_GROUP_DATA
) {
1198 u64 sinfo_used
= btrfs_space_info_used(sinfo
, true);
1201 * Here we make sure if we mark this bg RO, we still have enough
1202 * free space as buffer.
1204 if (sinfo_used
+ num_bytes
<= sinfo
->total_bytes
)
1208 * We overcommit metadata, so we need to do the
1209 * btrfs_can_overcommit check here, and we need to pass in
1210 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1211 * leeway to allow us to mark this block group as read only.
1213 if (btrfs_can_overcommit(cache
->fs_info
, sinfo
, num_bytes
,
1214 BTRFS_RESERVE_NO_FLUSH
))
1219 sinfo
->bytes_readonly
+= num_bytes
;
1220 if (btrfs_is_zoned(cache
->fs_info
)) {
1221 /* Migrate zone_unusable bytes to readonly */
1222 sinfo
->bytes_readonly
+= cache
->zone_unusable
;
1223 sinfo
->bytes_zone_unusable
-= cache
->zone_unusable
;
1224 cache
->zone_unusable
= 0;
1227 list_add_tail(&cache
->ro_list
, &sinfo
->ro_bgs
);
1230 spin_unlock(&cache
->lock
);
1231 spin_unlock(&sinfo
->lock
);
1232 if (ret
== -ENOSPC
&& btrfs_test_opt(cache
->fs_info
, ENOSPC_DEBUG
)) {
1233 btrfs_info(cache
->fs_info
,
1234 "unable to make block group %llu ro", cache
->start
);
1235 btrfs_dump_space_info(cache
->fs_info
, cache
->space_info
, 0, 0);
1240 static bool clean_pinned_extents(struct btrfs_trans_handle
*trans
,
1241 struct btrfs_block_group
*bg
)
1243 struct btrfs_fs_info
*fs_info
= bg
->fs_info
;
1244 struct btrfs_transaction
*prev_trans
= NULL
;
1245 const u64 start
= bg
->start
;
1246 const u64 end
= start
+ bg
->length
- 1;
1249 spin_lock(&fs_info
->trans_lock
);
1250 if (trans
->transaction
->list
.prev
!= &fs_info
->trans_list
) {
1251 prev_trans
= list_last_entry(&trans
->transaction
->list
,
1252 struct btrfs_transaction
, list
);
1253 refcount_inc(&prev_trans
->use_count
);
1255 spin_unlock(&fs_info
->trans_lock
);
1258 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1259 * btrfs_finish_extent_commit(). If we are at transaction N, another
1260 * task might be running finish_extent_commit() for the previous
1261 * transaction N - 1, and have seen a range belonging to the block
1262 * group in pinned_extents before we were able to clear the whole block
1263 * group range from pinned_extents. This means that task can lookup for
1264 * the block group after we unpinned it from pinned_extents and removed
1265 * it, leading to a BUG_ON() at unpin_extent_range().
1267 mutex_lock(&fs_info
->unused_bg_unpin_mutex
);
1269 ret
= clear_extent_bits(&prev_trans
->pinned_extents
, start
, end
,
1275 ret
= clear_extent_bits(&trans
->transaction
->pinned_extents
, start
, end
,
1278 mutex_unlock(&fs_info
->unused_bg_unpin_mutex
);
1280 btrfs_put_transaction(prev_trans
);
1286 * Process the unused_bgs list and remove any that don't have any allocated
1287 * space inside of them.
1289 void btrfs_delete_unused_bgs(struct btrfs_fs_info
*fs_info
)
1291 struct btrfs_block_group
*block_group
;
1292 struct btrfs_space_info
*space_info
;
1293 struct btrfs_trans_handle
*trans
;
1294 const bool async_trim_enabled
= btrfs_test_opt(fs_info
, DISCARD_ASYNC
);
1297 if (!test_bit(BTRFS_FS_OPEN
, &fs_info
->flags
))
1300 if (btrfs_fs_closing(fs_info
))
1304 * Long running balances can keep us blocked here for eternity, so
1305 * simply skip deletion if we're unable to get the mutex.
1307 if (!mutex_trylock(&fs_info
->reclaim_bgs_lock
))
1310 spin_lock(&fs_info
->unused_bgs_lock
);
1311 while (!list_empty(&fs_info
->unused_bgs
)) {
1314 block_group
= list_first_entry(&fs_info
->unused_bgs
,
1315 struct btrfs_block_group
,
1317 list_del_init(&block_group
->bg_list
);
1319 space_info
= block_group
->space_info
;
1321 if (ret
|| btrfs_mixed_space_info(space_info
)) {
1322 btrfs_put_block_group(block_group
);
1325 spin_unlock(&fs_info
->unused_bgs_lock
);
1327 btrfs_discard_cancel_work(&fs_info
->discard_ctl
, block_group
);
1329 /* Don't want to race with allocators so take the groups_sem */
1330 down_write(&space_info
->groups_sem
);
1333 * Async discard moves the final block group discard to be prior
1334 * to the unused_bgs code path. Therefore, if it's not fully
1335 * trimmed, punt it back to the async discard lists.
1337 if (btrfs_test_opt(fs_info
, DISCARD_ASYNC
) &&
1338 !btrfs_is_free_space_trimmed(block_group
)) {
1339 trace_btrfs_skip_unused_block_group(block_group
);
1340 up_write(&space_info
->groups_sem
);
1341 /* Requeue if we failed because of async discard */
1342 btrfs_discard_queue_work(&fs_info
->discard_ctl
,
1347 spin_lock(&block_group
->lock
);
1348 if (block_group
->reserved
|| block_group
->pinned
||
1349 block_group
->used
|| block_group
->ro
||
1350 list_is_singular(&block_group
->list
)) {
1352 * We want to bail if we made new allocations or have
1353 * outstanding allocations in this block group. We do
1354 * the ro check in case balance is currently acting on
1357 trace_btrfs_skip_unused_block_group(block_group
);
1358 spin_unlock(&block_group
->lock
);
1359 up_write(&space_info
->groups_sem
);
1362 spin_unlock(&block_group
->lock
);
1364 /* We don't want to force the issue, only flip if it's ok. */
1365 ret
= inc_block_group_ro(block_group
, 0);
1366 up_write(&space_info
->groups_sem
);
1372 ret
= btrfs_zone_finish(block_group
);
1374 btrfs_dec_block_group_ro(block_group
);
1381 * Want to do this before we do anything else so we can recover
1382 * properly if we fail to join the transaction.
1384 trans
= btrfs_start_trans_remove_block_group(fs_info
,
1385 block_group
->start
);
1386 if (IS_ERR(trans
)) {
1387 btrfs_dec_block_group_ro(block_group
);
1388 ret
= PTR_ERR(trans
);
1393 * We could have pending pinned extents for this block group,
1394 * just delete them, we don't care about them anymore.
1396 if (!clean_pinned_extents(trans
, block_group
)) {
1397 btrfs_dec_block_group_ro(block_group
);
1402 * At this point, the block_group is read only and should fail
1403 * new allocations. However, btrfs_finish_extent_commit() can
1404 * cause this block_group to be placed back on the discard
1405 * lists because now the block_group isn't fully discarded.
1406 * Bail here and try again later after discarding everything.
1408 spin_lock(&fs_info
->discard_ctl
.lock
);
1409 if (!list_empty(&block_group
->discard_list
)) {
1410 spin_unlock(&fs_info
->discard_ctl
.lock
);
1411 btrfs_dec_block_group_ro(block_group
);
1412 btrfs_discard_queue_work(&fs_info
->discard_ctl
,
1416 spin_unlock(&fs_info
->discard_ctl
.lock
);
1418 /* Reset pinned so btrfs_put_block_group doesn't complain */
1419 spin_lock(&space_info
->lock
);
1420 spin_lock(&block_group
->lock
);
1422 btrfs_space_info_update_bytes_pinned(fs_info
, space_info
,
1423 -block_group
->pinned
);
1424 space_info
->bytes_readonly
+= block_group
->pinned
;
1425 block_group
->pinned
= 0;
1427 spin_unlock(&block_group
->lock
);
1428 spin_unlock(&space_info
->lock
);
1431 * The normal path here is an unused block group is passed here,
1432 * then trimming is handled in the transaction commit path.
1433 * Async discard interposes before this to do the trimming
1434 * before coming down the unused block group path as trimming
1435 * will no longer be done later in the transaction commit path.
1437 if (!async_trim_enabled
&& btrfs_test_opt(fs_info
, DISCARD_ASYNC
))
1441 * DISCARD can flip during remount. On zoned filesystems, we
1442 * need to reset sequential-required zones.
1444 trimming
= btrfs_test_opt(fs_info
, DISCARD_SYNC
) ||
1445 btrfs_is_zoned(fs_info
);
1447 /* Implicit trim during transaction commit. */
1449 btrfs_freeze_block_group(block_group
);
1452 * Btrfs_remove_chunk will abort the transaction if things go
1455 ret
= btrfs_remove_chunk(trans
, block_group
->start
);
1459 btrfs_unfreeze_block_group(block_group
);
1464 * If we're not mounted with -odiscard, we can just forget
1465 * about this block group. Otherwise we'll need to wait
1466 * until transaction commit to do the actual discard.
1469 spin_lock(&fs_info
->unused_bgs_lock
);
1471 * A concurrent scrub might have added us to the list
1472 * fs_info->unused_bgs, so use a list_move operation
1473 * to add the block group to the deleted_bgs list.
1475 list_move(&block_group
->bg_list
,
1476 &trans
->transaction
->deleted_bgs
);
1477 spin_unlock(&fs_info
->unused_bgs_lock
);
1478 btrfs_get_block_group(block_group
);
1481 btrfs_end_transaction(trans
);
1483 btrfs_put_block_group(block_group
);
1484 spin_lock(&fs_info
->unused_bgs_lock
);
1486 spin_unlock(&fs_info
->unused_bgs_lock
);
1487 mutex_unlock(&fs_info
->reclaim_bgs_lock
);
1491 btrfs_end_transaction(trans
);
1492 mutex_unlock(&fs_info
->reclaim_bgs_lock
);
1493 btrfs_put_block_group(block_group
);
1494 btrfs_discard_punt_unused_bgs_list(fs_info
);
1497 void btrfs_mark_bg_unused(struct btrfs_block_group
*bg
)
1499 struct btrfs_fs_info
*fs_info
= bg
->fs_info
;
1501 spin_lock(&fs_info
->unused_bgs_lock
);
1502 if (list_empty(&bg
->bg_list
)) {
1503 btrfs_get_block_group(bg
);
1504 trace_btrfs_add_unused_block_group(bg
);
1505 list_add_tail(&bg
->bg_list
, &fs_info
->unused_bgs
);
1507 spin_unlock(&fs_info
->unused_bgs_lock
);
1511 * We want block groups with a low number of used bytes to be in the beginning
1512 * of the list, so they will get reclaimed first.
1514 static int reclaim_bgs_cmp(void *unused
, const struct list_head
*a
,
1515 const struct list_head
*b
)
1517 const struct btrfs_block_group
*bg1
, *bg2
;
1519 bg1
= list_entry(a
, struct btrfs_block_group
, bg_list
);
1520 bg2
= list_entry(b
, struct btrfs_block_group
, bg_list
);
1522 return bg1
->used
> bg2
->used
;
1525 static inline bool btrfs_should_reclaim(struct btrfs_fs_info
*fs_info
)
1527 if (btrfs_is_zoned(fs_info
))
1528 return btrfs_zoned_should_reclaim(fs_info
);
1532 void btrfs_reclaim_bgs_work(struct work_struct
*work
)
1534 struct btrfs_fs_info
*fs_info
=
1535 container_of(work
, struct btrfs_fs_info
, reclaim_bgs_work
);
1536 struct btrfs_block_group
*bg
;
1537 struct btrfs_space_info
*space_info
;
1539 if (!test_bit(BTRFS_FS_OPEN
, &fs_info
->flags
))
1542 if (btrfs_fs_closing(fs_info
))
1545 if (!btrfs_should_reclaim(fs_info
))
1548 sb_start_write(fs_info
->sb
);
1550 if (!btrfs_exclop_start(fs_info
, BTRFS_EXCLOP_BALANCE
)) {
1551 sb_end_write(fs_info
->sb
);
1556 * Long running balances can keep us blocked here for eternity, so
1557 * simply skip reclaim if we're unable to get the mutex.
1559 if (!mutex_trylock(&fs_info
->reclaim_bgs_lock
)) {
1560 btrfs_exclop_finish(fs_info
);
1561 sb_end_write(fs_info
->sb
);
1565 spin_lock(&fs_info
->unused_bgs_lock
);
1567 * Sort happens under lock because we can't simply splice it and sort.
1568 * The block groups might still be in use and reachable via bg_list,
1569 * and their presence in the reclaim_bgs list must be preserved.
1571 list_sort(NULL
, &fs_info
->reclaim_bgs
, reclaim_bgs_cmp
);
1572 while (!list_empty(&fs_info
->reclaim_bgs
)) {
1576 bg
= list_first_entry(&fs_info
->reclaim_bgs
,
1577 struct btrfs_block_group
,
1579 list_del_init(&bg
->bg_list
);
1581 space_info
= bg
->space_info
;
1582 spin_unlock(&fs_info
->unused_bgs_lock
);
1584 /* Don't race with allocators so take the groups_sem */
1585 down_write(&space_info
->groups_sem
);
1587 spin_lock(&bg
->lock
);
1588 if (bg
->reserved
|| bg
->pinned
|| bg
->ro
) {
1590 * We want to bail if we made new allocations or have
1591 * outstanding allocations in this block group. We do
1592 * the ro check in case balance is currently acting on
1595 spin_unlock(&bg
->lock
);
1596 up_write(&space_info
->groups_sem
);
1599 spin_unlock(&bg
->lock
);
1601 /* Get out fast, in case we're unmounting the filesystem */
1602 if (btrfs_fs_closing(fs_info
)) {
1603 up_write(&space_info
->groups_sem
);
1608 * Cache the zone_unusable value before turning the block group
1609 * to read only. As soon as the blog group is read only it's
1610 * zone_unusable value gets moved to the block group's read-only
1611 * bytes and isn't available for calculations anymore.
1613 zone_unusable
= bg
->zone_unusable
;
1614 ret
= inc_block_group_ro(bg
, 0);
1615 up_write(&space_info
->groups_sem
);
1620 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1621 bg
->start
, div_u64(bg
->used
* 100, bg
->length
),
1622 div64_u64(zone_unusable
* 100, bg
->length
));
1623 trace_btrfs_reclaim_block_group(bg
);
1624 ret
= btrfs_relocate_chunk(fs_info
, bg
->start
);
1626 btrfs_dec_block_group_ro(bg
);
1627 btrfs_err(fs_info
, "error relocating chunk %llu",
1632 btrfs_put_block_group(bg
);
1633 spin_lock(&fs_info
->unused_bgs_lock
);
1635 spin_unlock(&fs_info
->unused_bgs_lock
);
1636 mutex_unlock(&fs_info
->reclaim_bgs_lock
);
1637 btrfs_exclop_finish(fs_info
);
1638 sb_end_write(fs_info
->sb
);
1641 void btrfs_reclaim_bgs(struct btrfs_fs_info
*fs_info
)
1643 spin_lock(&fs_info
->unused_bgs_lock
);
1644 if (!list_empty(&fs_info
->reclaim_bgs
))
1645 queue_work(system_unbound_wq
, &fs_info
->reclaim_bgs_work
);
1646 spin_unlock(&fs_info
->unused_bgs_lock
);
1649 void btrfs_mark_bg_to_reclaim(struct btrfs_block_group
*bg
)
1651 struct btrfs_fs_info
*fs_info
= bg
->fs_info
;
1653 spin_lock(&fs_info
->unused_bgs_lock
);
1654 if (list_empty(&bg
->bg_list
)) {
1655 btrfs_get_block_group(bg
);
1656 trace_btrfs_add_reclaim_block_group(bg
);
1657 list_add_tail(&bg
->bg_list
, &fs_info
->reclaim_bgs
);
1659 spin_unlock(&fs_info
->unused_bgs_lock
);
1662 static int read_bg_from_eb(struct btrfs_fs_info
*fs_info
, struct btrfs_key
*key
,
1663 struct btrfs_path
*path
)
1665 struct extent_map_tree
*em_tree
;
1666 struct extent_map
*em
;
1667 struct btrfs_block_group_item bg
;
1668 struct extent_buffer
*leaf
;
1673 slot
= path
->slots
[0];
1674 leaf
= path
->nodes
[0];
1676 em_tree
= &fs_info
->mapping_tree
;
1677 read_lock(&em_tree
->lock
);
1678 em
= lookup_extent_mapping(em_tree
, key
->objectid
, key
->offset
);
1679 read_unlock(&em_tree
->lock
);
1682 "logical %llu len %llu found bg but no related chunk",
1683 key
->objectid
, key
->offset
);
1687 if (em
->start
!= key
->objectid
|| em
->len
!= key
->offset
) {
1689 "block group %llu len %llu mismatch with chunk %llu len %llu",
1690 key
->objectid
, key
->offset
, em
->start
, em
->len
);
1695 read_extent_buffer(leaf
, &bg
, btrfs_item_ptr_offset(leaf
, slot
),
1697 flags
= btrfs_stack_block_group_flags(&bg
) &
1698 BTRFS_BLOCK_GROUP_TYPE_MASK
;
1700 if (flags
!= (em
->map_lookup
->type
& BTRFS_BLOCK_GROUP_TYPE_MASK
)) {
1702 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1703 key
->objectid
, key
->offset
, flags
,
1704 (BTRFS_BLOCK_GROUP_TYPE_MASK
& em
->map_lookup
->type
));
1709 free_extent_map(em
);
1713 static int find_first_block_group(struct btrfs_fs_info
*fs_info
,
1714 struct btrfs_path
*path
,
1715 struct btrfs_key
*key
)
1717 struct btrfs_root
*root
= btrfs_block_group_root(fs_info
);
1719 struct btrfs_key found_key
;
1721 btrfs_for_each_slot(root
, key
, &found_key
, path
, ret
) {
1722 if (found_key
.objectid
>= key
->objectid
&&
1723 found_key
.type
== BTRFS_BLOCK_GROUP_ITEM_KEY
) {
1724 return read_bg_from_eb(fs_info
, &found_key
, path
);
1730 static void set_avail_alloc_bits(struct btrfs_fs_info
*fs_info
, u64 flags
)
1732 u64 extra_flags
= chunk_to_extended(flags
) &
1733 BTRFS_EXTENDED_PROFILE_MASK
;
1735 write_seqlock(&fs_info
->profiles_lock
);
1736 if (flags
& BTRFS_BLOCK_GROUP_DATA
)
1737 fs_info
->avail_data_alloc_bits
|= extra_flags
;
1738 if (flags
& BTRFS_BLOCK_GROUP_METADATA
)
1739 fs_info
->avail_metadata_alloc_bits
|= extra_flags
;
1740 if (flags
& BTRFS_BLOCK_GROUP_SYSTEM
)
1741 fs_info
->avail_system_alloc_bits
|= extra_flags
;
1742 write_sequnlock(&fs_info
->profiles_lock
);
1746 * Map a physical disk address to a list of logical addresses
1748 * @fs_info: the filesystem
1749 * @chunk_start: logical address of block group
1750 * @bdev: physical device to resolve, can be NULL to indicate any device
1751 * @physical: physical address to map to logical addresses
1752 * @logical: return array of logical addresses which map to @physical
1753 * @naddrs: length of @logical
1754 * @stripe_len: size of IO stripe for the given block group
1756 * Maps a particular @physical disk address to a list of @logical addresses.
1757 * Used primarily to exclude those portions of a block group that contain super
1760 int btrfs_rmap_block(struct btrfs_fs_info
*fs_info
, u64 chunk_start
,
1761 struct block_device
*bdev
, u64 physical
, u64
**logical
,
1762 int *naddrs
, int *stripe_len
)
1764 struct extent_map
*em
;
1765 struct map_lookup
*map
;
1768 u64 data_stripe_length
;
1773 em
= btrfs_get_chunk_map(fs_info
, chunk_start
, 1);
1777 map
= em
->map_lookup
;
1778 data_stripe_length
= em
->orig_block_len
;
1779 io_stripe_size
= map
->stripe_len
;
1780 chunk_start
= em
->start
;
1782 /* For RAID5/6 adjust to a full IO stripe length */
1783 if (map
->type
& BTRFS_BLOCK_GROUP_RAID56_MASK
)
1784 io_stripe_size
= map
->stripe_len
* nr_data_stripes(map
);
1786 buf
= kcalloc(map
->num_stripes
, sizeof(u64
), GFP_NOFS
);
1792 for (i
= 0; i
< map
->num_stripes
; i
++) {
1793 bool already_inserted
= false;
1798 if (!in_range(physical
, map
->stripes
[i
].physical
,
1799 data_stripe_length
))
1802 if (bdev
&& map
->stripes
[i
].dev
->bdev
!= bdev
)
1805 stripe_nr
= physical
- map
->stripes
[i
].physical
;
1806 stripe_nr
= div64_u64_rem(stripe_nr
, map
->stripe_len
, &offset
);
1808 if (map
->type
& (BTRFS_BLOCK_GROUP_RAID0
|
1809 BTRFS_BLOCK_GROUP_RAID10
)) {
1810 stripe_nr
= stripe_nr
* map
->num_stripes
+ i
;
1811 stripe_nr
= div_u64(stripe_nr
, map
->sub_stripes
);
1814 * The remaining case would be for RAID56, multiply by
1815 * nr_data_stripes(). Alternatively, just use rmap_len below
1816 * instead of map->stripe_len
1819 bytenr
= chunk_start
+ stripe_nr
* io_stripe_size
+ offset
;
1821 /* Ensure we don't add duplicate addresses */
1822 for (j
= 0; j
< nr
; j
++) {
1823 if (buf
[j
] == bytenr
) {
1824 already_inserted
= true;
1829 if (!already_inserted
)
1835 *stripe_len
= io_stripe_size
;
1837 free_extent_map(em
);
1841 static int exclude_super_stripes(struct btrfs_block_group
*cache
)
1843 struct btrfs_fs_info
*fs_info
= cache
->fs_info
;
1844 const bool zoned
= btrfs_is_zoned(fs_info
);
1850 if (cache
->start
< BTRFS_SUPER_INFO_OFFSET
) {
1851 stripe_len
= BTRFS_SUPER_INFO_OFFSET
- cache
->start
;
1852 cache
->bytes_super
+= stripe_len
;
1853 ret
= btrfs_add_excluded_extent(fs_info
, cache
->start
,
1859 for (i
= 0; i
< BTRFS_SUPER_MIRROR_MAX
; i
++) {
1860 bytenr
= btrfs_sb_offset(i
);
1861 ret
= btrfs_rmap_block(fs_info
, cache
->start
, NULL
,
1862 bytenr
, &logical
, &nr
, &stripe_len
);
1866 /* Shouldn't have super stripes in sequential zones */
1869 "zoned: block group %llu must not contain super block",
1875 u64 len
= min_t(u64
, stripe_len
,
1876 cache
->start
+ cache
->length
- logical
[nr
]);
1878 cache
->bytes_super
+= len
;
1879 ret
= btrfs_add_excluded_extent(fs_info
, logical
[nr
],
1892 static struct btrfs_block_group
*btrfs_create_block_group_cache(
1893 struct btrfs_fs_info
*fs_info
, u64 start
)
1895 struct btrfs_block_group
*cache
;
1897 cache
= kzalloc(sizeof(*cache
), GFP_NOFS
);
1901 cache
->free_space_ctl
= kzalloc(sizeof(*cache
->free_space_ctl
),
1903 if (!cache
->free_space_ctl
) {
1908 cache
->start
= start
;
1910 cache
->fs_info
= fs_info
;
1911 cache
->full_stripe_len
= btrfs_full_stripe_len(fs_info
, start
);
1913 cache
->discard_index
= BTRFS_DISCARD_INDEX_UNUSED
;
1915 refcount_set(&cache
->refs
, 1);
1916 spin_lock_init(&cache
->lock
);
1917 init_rwsem(&cache
->data_rwsem
);
1918 INIT_LIST_HEAD(&cache
->list
);
1919 INIT_LIST_HEAD(&cache
->cluster_list
);
1920 INIT_LIST_HEAD(&cache
->bg_list
);
1921 INIT_LIST_HEAD(&cache
->ro_list
);
1922 INIT_LIST_HEAD(&cache
->discard_list
);
1923 INIT_LIST_HEAD(&cache
->dirty_list
);
1924 INIT_LIST_HEAD(&cache
->io_list
);
1925 INIT_LIST_HEAD(&cache
->active_bg_list
);
1926 btrfs_init_free_space_ctl(cache
, cache
->free_space_ctl
);
1927 atomic_set(&cache
->frozen
, 0);
1928 mutex_init(&cache
->free_space_lock
);
1929 btrfs_init_full_stripe_locks_tree(&cache
->full_stripe_locks_root
);
1935 * Iterate all chunks and verify that each of them has the corresponding block
1938 static int check_chunk_block_group_mappings(struct btrfs_fs_info
*fs_info
)
1940 struct extent_map_tree
*map_tree
= &fs_info
->mapping_tree
;
1941 struct extent_map
*em
;
1942 struct btrfs_block_group
*bg
;
1947 read_lock(&map_tree
->lock
);
1949 * lookup_extent_mapping will return the first extent map
1950 * intersecting the range, so setting @len to 1 is enough to
1951 * get the first chunk.
1953 em
= lookup_extent_mapping(map_tree
, start
, 1);
1954 read_unlock(&map_tree
->lock
);
1958 bg
= btrfs_lookup_block_group(fs_info
, em
->start
);
1961 "chunk start=%llu len=%llu doesn't have corresponding block group",
1962 em
->start
, em
->len
);
1964 free_extent_map(em
);
1967 if (bg
->start
!= em
->start
|| bg
->length
!= em
->len
||
1968 (bg
->flags
& BTRFS_BLOCK_GROUP_TYPE_MASK
) !=
1969 (em
->map_lookup
->type
& BTRFS_BLOCK_GROUP_TYPE_MASK
)) {
1971 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1973 em
->map_lookup
->type
& BTRFS_BLOCK_GROUP_TYPE_MASK
,
1974 bg
->start
, bg
->length
,
1975 bg
->flags
& BTRFS_BLOCK_GROUP_TYPE_MASK
);
1977 free_extent_map(em
);
1978 btrfs_put_block_group(bg
);
1981 start
= em
->start
+ em
->len
;
1982 free_extent_map(em
);
1983 btrfs_put_block_group(bg
);
1988 static int read_one_block_group(struct btrfs_fs_info
*info
,
1989 struct btrfs_block_group_item
*bgi
,
1990 const struct btrfs_key
*key
,
1993 struct btrfs_block_group
*cache
;
1994 const bool mixed
= btrfs_fs_incompat(info
, MIXED_GROUPS
);
1997 ASSERT(key
->type
== BTRFS_BLOCK_GROUP_ITEM_KEY
);
1999 cache
= btrfs_create_block_group_cache(info
, key
->objectid
);
2003 cache
->length
= key
->offset
;
2004 cache
->used
= btrfs_stack_block_group_used(bgi
);
2005 cache
->flags
= btrfs_stack_block_group_flags(bgi
);
2006 cache
->global_root_id
= btrfs_stack_block_group_chunk_objectid(bgi
);
2008 set_free_space_tree_thresholds(cache
);
2012 * When we mount with old space cache, we need to
2013 * set BTRFS_DC_CLEAR and set dirty flag.
2015 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2016 * truncate the old free space cache inode and
2018 * b) Setting 'dirty flag' makes sure that we flush
2019 * the new space cache info onto disk.
2021 if (btrfs_test_opt(info
, SPACE_CACHE
))
2022 cache
->disk_cache_state
= BTRFS_DC_CLEAR
;
2024 if (!mixed
&& ((cache
->flags
& BTRFS_BLOCK_GROUP_METADATA
) &&
2025 (cache
->flags
& BTRFS_BLOCK_GROUP_DATA
))) {
2027 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2033 ret
= btrfs_load_block_group_zone_info(cache
, false);
2035 btrfs_err(info
, "zoned: failed to load zone info of bg %llu",
2041 * We need to exclude the super stripes now so that the space info has
2042 * super bytes accounted for, otherwise we'll think we have more space
2043 * than we actually do.
2045 ret
= exclude_super_stripes(cache
);
2047 /* We may have excluded something, so call this just in case. */
2048 btrfs_free_excluded_extents(cache
);
2053 * For zoned filesystem, space after the allocation offset is the only
2054 * free space for a block group. So, we don't need any caching work.
2055 * btrfs_calc_zone_unusable() will set the amount of free space and
2056 * zone_unusable space.
2058 * For regular filesystem, check for two cases, either we are full, and
2059 * therefore don't need to bother with the caching work since we won't
2060 * find any space, or we are empty, and we can just add all the space
2061 * in and be done with it. This saves us _a_lot_ of time, particularly
2064 if (btrfs_is_zoned(info
)) {
2065 btrfs_calc_zone_unusable(cache
);
2066 /* Should not have any excluded extents. Just in case, though. */
2067 btrfs_free_excluded_extents(cache
);
2068 } else if (cache
->length
== cache
->used
) {
2069 cache
->cached
= BTRFS_CACHE_FINISHED
;
2070 btrfs_free_excluded_extents(cache
);
2071 } else if (cache
->used
== 0) {
2072 cache
->cached
= BTRFS_CACHE_FINISHED
;
2073 add_new_free_space(cache
, cache
->start
,
2074 cache
->start
+ cache
->length
);
2075 btrfs_free_excluded_extents(cache
);
2078 ret
= btrfs_add_block_group_cache(info
, cache
);
2080 btrfs_remove_free_space_cache(cache
);
2083 trace_btrfs_add_block_group(info
, cache
, 0);
2084 btrfs_add_bg_to_space_info(info
, cache
);
2086 set_avail_alloc_bits(info
, cache
->flags
);
2087 if (btrfs_chunk_writeable(info
, cache
->start
)) {
2088 if (cache
->used
== 0) {
2089 ASSERT(list_empty(&cache
->bg_list
));
2090 if (btrfs_test_opt(info
, DISCARD_ASYNC
))
2091 btrfs_discard_queue_work(&info
->discard_ctl
, cache
);
2093 btrfs_mark_bg_unused(cache
);
2096 inc_block_group_ro(cache
, 1);
2101 btrfs_put_block_group(cache
);
2105 static int fill_dummy_bgs(struct btrfs_fs_info
*fs_info
)
2107 struct extent_map_tree
*em_tree
= &fs_info
->mapping_tree
;
2108 struct rb_node
*node
;
2111 for (node
= rb_first_cached(&em_tree
->map
); node
; node
= rb_next(node
)) {
2112 struct extent_map
*em
;
2113 struct map_lookup
*map
;
2114 struct btrfs_block_group
*bg
;
2116 em
= rb_entry(node
, struct extent_map
, rb_node
);
2117 map
= em
->map_lookup
;
2118 bg
= btrfs_create_block_group_cache(fs_info
, em
->start
);
2124 /* Fill dummy cache as FULL */
2125 bg
->length
= em
->len
;
2126 bg
->flags
= map
->type
;
2127 bg
->cached
= BTRFS_CACHE_FINISHED
;
2129 bg
->flags
= map
->type
;
2130 ret
= btrfs_add_block_group_cache(fs_info
, bg
);
2132 * We may have some valid block group cache added already, in
2133 * that case we skip to the next one.
2135 if (ret
== -EEXIST
) {
2137 btrfs_put_block_group(bg
);
2142 btrfs_remove_free_space_cache(bg
);
2143 btrfs_put_block_group(bg
);
2147 btrfs_add_bg_to_space_info(fs_info
, bg
);
2149 set_avail_alloc_bits(fs_info
, bg
->flags
);
2152 btrfs_init_global_block_rsv(fs_info
);
2156 int btrfs_read_block_groups(struct btrfs_fs_info
*info
)
2158 struct btrfs_root
*root
= btrfs_block_group_root(info
);
2159 struct btrfs_path
*path
;
2161 struct btrfs_block_group
*cache
;
2162 struct btrfs_space_info
*space_info
;
2163 struct btrfs_key key
;
2168 return fill_dummy_bgs(info
);
2172 key
.type
= BTRFS_BLOCK_GROUP_ITEM_KEY
;
2173 path
= btrfs_alloc_path();
2177 cache_gen
= btrfs_super_cache_generation(info
->super_copy
);
2178 if (btrfs_test_opt(info
, SPACE_CACHE
) &&
2179 btrfs_super_generation(info
->super_copy
) != cache_gen
)
2181 if (btrfs_test_opt(info
, CLEAR_CACHE
))
2185 struct btrfs_block_group_item bgi
;
2186 struct extent_buffer
*leaf
;
2189 ret
= find_first_block_group(info
, path
, &key
);
2195 leaf
= path
->nodes
[0];
2196 slot
= path
->slots
[0];
2198 read_extent_buffer(leaf
, &bgi
, btrfs_item_ptr_offset(leaf
, slot
),
2201 btrfs_item_key_to_cpu(leaf
, &key
, slot
);
2202 btrfs_release_path(path
);
2203 ret
= read_one_block_group(info
, &bgi
, &key
, need_clear
);
2206 key
.objectid
+= key
.offset
;
2209 btrfs_release_path(path
);
2211 list_for_each_entry(space_info
, &info
->space_info
, list
) {
2214 for (i
= 0; i
< BTRFS_NR_RAID_TYPES
; i
++) {
2215 if (list_empty(&space_info
->block_groups
[i
]))
2217 cache
= list_first_entry(&space_info
->block_groups
[i
],
2218 struct btrfs_block_group
,
2220 btrfs_sysfs_add_block_group_type(cache
);
2223 if (!(btrfs_get_alloc_profile(info
, space_info
->flags
) &
2224 (BTRFS_BLOCK_GROUP_RAID10
|
2225 BTRFS_BLOCK_GROUP_RAID1_MASK
|
2226 BTRFS_BLOCK_GROUP_RAID56_MASK
|
2227 BTRFS_BLOCK_GROUP_DUP
)))
2230 * Avoid allocating from un-mirrored block group if there are
2231 * mirrored block groups.
2233 list_for_each_entry(cache
,
2234 &space_info
->block_groups
[BTRFS_RAID_RAID0
],
2236 inc_block_group_ro(cache
, 1);
2237 list_for_each_entry(cache
,
2238 &space_info
->block_groups
[BTRFS_RAID_SINGLE
],
2240 inc_block_group_ro(cache
, 1);
2243 btrfs_init_global_block_rsv(info
);
2244 ret
= check_chunk_block_group_mappings(info
);
2246 btrfs_free_path(path
);
2248 * We've hit some error while reading the extent tree, and have
2249 * rescue=ibadroots mount option.
2250 * Try to fill the tree using dummy block groups so that the user can
2251 * continue to mount and grab their data.
2253 if (ret
&& btrfs_test_opt(info
, IGNOREBADROOTS
))
2254 ret
= fill_dummy_bgs(info
);
2259 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2262 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2265 static int insert_block_group_item(struct btrfs_trans_handle
*trans
,
2266 struct btrfs_block_group
*block_group
)
2268 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2269 struct btrfs_block_group_item bgi
;
2270 struct btrfs_root
*root
= btrfs_block_group_root(fs_info
);
2271 struct btrfs_key key
;
2273 spin_lock(&block_group
->lock
);
2274 btrfs_set_stack_block_group_used(&bgi
, block_group
->used
);
2275 btrfs_set_stack_block_group_chunk_objectid(&bgi
,
2276 block_group
->global_root_id
);
2277 btrfs_set_stack_block_group_flags(&bgi
, block_group
->flags
);
2278 key
.objectid
= block_group
->start
;
2279 key
.type
= BTRFS_BLOCK_GROUP_ITEM_KEY
;
2280 key
.offset
= block_group
->length
;
2281 spin_unlock(&block_group
->lock
);
2283 return btrfs_insert_item(trans
, root
, &key
, &bgi
, sizeof(bgi
));
2286 static int insert_dev_extent(struct btrfs_trans_handle
*trans
,
2287 struct btrfs_device
*device
, u64 chunk_offset
,
2288 u64 start
, u64 num_bytes
)
2290 struct btrfs_fs_info
*fs_info
= device
->fs_info
;
2291 struct btrfs_root
*root
= fs_info
->dev_root
;
2292 struct btrfs_path
*path
;
2293 struct btrfs_dev_extent
*extent
;
2294 struct extent_buffer
*leaf
;
2295 struct btrfs_key key
;
2298 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA
, &device
->dev_state
));
2299 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT
, &device
->dev_state
));
2300 path
= btrfs_alloc_path();
2304 key
.objectid
= device
->devid
;
2305 key
.type
= BTRFS_DEV_EXTENT_KEY
;
2307 ret
= btrfs_insert_empty_item(trans
, root
, path
, &key
, sizeof(*extent
));
2311 leaf
= path
->nodes
[0];
2312 extent
= btrfs_item_ptr(leaf
, path
->slots
[0], struct btrfs_dev_extent
);
2313 btrfs_set_dev_extent_chunk_tree(leaf
, extent
, BTRFS_CHUNK_TREE_OBJECTID
);
2314 btrfs_set_dev_extent_chunk_objectid(leaf
, extent
,
2315 BTRFS_FIRST_CHUNK_TREE_OBJECTID
);
2316 btrfs_set_dev_extent_chunk_offset(leaf
, extent
, chunk_offset
);
2318 btrfs_set_dev_extent_length(leaf
, extent
, num_bytes
);
2319 btrfs_mark_buffer_dirty(leaf
);
2321 btrfs_free_path(path
);
2326 * This function belongs to phase 2.
2328 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2331 static int insert_dev_extents(struct btrfs_trans_handle
*trans
,
2332 u64 chunk_offset
, u64 chunk_size
)
2334 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2335 struct btrfs_device
*device
;
2336 struct extent_map
*em
;
2337 struct map_lookup
*map
;
2343 em
= btrfs_get_chunk_map(fs_info
, chunk_offset
, chunk_size
);
2347 map
= em
->map_lookup
;
2348 stripe_size
= em
->orig_block_len
;
2351 * Take the device list mutex to prevent races with the final phase of
2352 * a device replace operation that replaces the device object associated
2353 * with the map's stripes, because the device object's id can change
2354 * at any time during that final phase of the device replace operation
2355 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2356 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2357 * resulting in persisting a device extent item with such ID.
2359 mutex_lock(&fs_info
->fs_devices
->device_list_mutex
);
2360 for (i
= 0; i
< map
->num_stripes
; i
++) {
2361 device
= map
->stripes
[i
].dev
;
2362 dev_offset
= map
->stripes
[i
].physical
;
2364 ret
= insert_dev_extent(trans
, device
, chunk_offset
, dev_offset
,
2369 mutex_unlock(&fs_info
->fs_devices
->device_list_mutex
);
2371 free_extent_map(em
);
2376 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2379 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2382 void btrfs_create_pending_block_groups(struct btrfs_trans_handle
*trans
)
2384 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2385 struct btrfs_block_group
*block_group
;
2388 while (!list_empty(&trans
->new_bgs
)) {
2391 block_group
= list_first_entry(&trans
->new_bgs
,
2392 struct btrfs_block_group
,
2397 index
= btrfs_bg_flags_to_raid_index(block_group
->flags
);
2399 ret
= insert_block_group_item(trans
, block_group
);
2401 btrfs_abort_transaction(trans
, ret
);
2402 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED
,
2403 &block_group
->runtime_flags
)) {
2404 mutex_lock(&fs_info
->chunk_mutex
);
2405 ret
= btrfs_chunk_alloc_add_chunk_item(trans
, block_group
);
2406 mutex_unlock(&fs_info
->chunk_mutex
);
2408 btrfs_abort_transaction(trans
, ret
);
2410 ret
= insert_dev_extents(trans
, block_group
->start
,
2411 block_group
->length
);
2413 btrfs_abort_transaction(trans
, ret
);
2414 add_block_group_free_space(trans
, block_group
);
2417 * If we restriped during balance, we may have added a new raid
2418 * type, so now add the sysfs entries when it is safe to do so.
2419 * We don't have to worry about locking here as it's handled in
2420 * btrfs_sysfs_add_block_group_type.
2422 if (block_group
->space_info
->block_group_kobjs
[index
] == NULL
)
2423 btrfs_sysfs_add_block_group_type(block_group
);
2425 /* Already aborted the transaction if it failed. */
2427 btrfs_delayed_refs_rsv_release(fs_info
, 1);
2428 list_del_init(&block_group
->bg_list
);
2430 btrfs_trans_release_chunk_metadata(trans
);
2434 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2435 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2437 static u64
calculate_global_root_id(struct btrfs_fs_info
*fs_info
, u64 offset
)
2442 if (!btrfs_fs_incompat(fs_info
, EXTENT_TREE_V2
))
2443 return BTRFS_FIRST_CHUNK_TREE_OBJECTID
;
2445 /* If we have a smaller fs index based on 128MiB. */
2446 if (btrfs_super_total_bytes(fs_info
->super_copy
) <= (SZ_1G
* 10ULL))
2449 offset
= div64_u64(offset
, div
);
2450 div64_u64_rem(offset
, fs_info
->nr_global_roots
, &index
);
2454 struct btrfs_block_group
*btrfs_make_block_group(struct btrfs_trans_handle
*trans
,
2455 u64 bytes_used
, u64 type
,
2456 u64 chunk_offset
, u64 size
)
2458 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2459 struct btrfs_block_group
*cache
;
2462 btrfs_set_log_full_commit(trans
);
2464 cache
= btrfs_create_block_group_cache(fs_info
, chunk_offset
);
2466 return ERR_PTR(-ENOMEM
);
2468 cache
->length
= size
;
2469 set_free_space_tree_thresholds(cache
);
2470 cache
->used
= bytes_used
;
2471 cache
->flags
= type
;
2472 cache
->cached
= BTRFS_CACHE_FINISHED
;
2473 cache
->global_root_id
= calculate_global_root_id(fs_info
, cache
->start
);
2475 if (btrfs_fs_compat_ro(fs_info
, FREE_SPACE_TREE
))
2476 cache
->needs_free_space
= 1;
2478 ret
= btrfs_load_block_group_zone_info(cache
, true);
2480 btrfs_put_block_group(cache
);
2481 return ERR_PTR(ret
);
2484 ret
= exclude_super_stripes(cache
);
2486 /* We may have excluded something, so call this just in case */
2487 btrfs_free_excluded_extents(cache
);
2488 btrfs_put_block_group(cache
);
2489 return ERR_PTR(ret
);
2492 add_new_free_space(cache
, chunk_offset
, chunk_offset
+ size
);
2494 btrfs_free_excluded_extents(cache
);
2497 * Ensure the corresponding space_info object is created and
2498 * assigned to our block group. We want our bg to be added to the rbtree
2499 * with its ->space_info set.
2501 cache
->space_info
= btrfs_find_space_info(fs_info
, cache
->flags
);
2502 ASSERT(cache
->space_info
);
2504 ret
= btrfs_add_block_group_cache(fs_info
, cache
);
2506 btrfs_remove_free_space_cache(cache
);
2507 btrfs_put_block_group(cache
);
2508 return ERR_PTR(ret
);
2512 * Now that our block group has its ->space_info set and is inserted in
2513 * the rbtree, update the space info's counters.
2515 trace_btrfs_add_block_group(fs_info
, cache
, 1);
2516 btrfs_add_bg_to_space_info(fs_info
, cache
);
2517 btrfs_update_global_block_rsv(fs_info
);
2519 #ifdef CONFIG_BTRFS_DEBUG
2520 if (btrfs_should_fragment_free_space(cache
)) {
2521 u64 new_bytes_used
= size
- bytes_used
;
2523 cache
->space_info
->bytes_used
+= new_bytes_used
>> 1;
2524 fragment_free_space(cache
);
2528 list_add_tail(&cache
->bg_list
, &trans
->new_bgs
);
2529 trans
->delayed_ref_updates
++;
2530 btrfs_update_delayed_refs_rsv(trans
);
2532 set_avail_alloc_bits(fs_info
, type
);
2537 * Mark one block group RO, can be called several times for the same block
2540 * @cache: the destination block group
2541 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2542 * ensure we still have some free space after marking this
2545 int btrfs_inc_block_group_ro(struct btrfs_block_group
*cache
,
2546 bool do_chunk_alloc
)
2548 struct btrfs_fs_info
*fs_info
= cache
->fs_info
;
2549 struct btrfs_trans_handle
*trans
;
2550 struct btrfs_root
*root
= btrfs_block_group_root(fs_info
);
2553 bool dirty_bg_running
;
2556 * This can only happen when we are doing read-only scrub on read-only
2558 * In that case we should not start a new transaction on read-only fs.
2559 * Thus here we skip all chunk allocations.
2561 if (sb_rdonly(fs_info
->sb
)) {
2562 mutex_lock(&fs_info
->ro_block_group_mutex
);
2563 ret
= inc_block_group_ro(cache
, 0);
2564 mutex_unlock(&fs_info
->ro_block_group_mutex
);
2569 trans
= btrfs_join_transaction(root
);
2571 return PTR_ERR(trans
);
2573 dirty_bg_running
= false;
2576 * We're not allowed to set block groups readonly after the dirty
2577 * block group cache has started writing. If it already started,
2578 * back off and let this transaction commit.
2580 mutex_lock(&fs_info
->ro_block_group_mutex
);
2581 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN
, &trans
->transaction
->flags
)) {
2582 u64 transid
= trans
->transid
;
2584 mutex_unlock(&fs_info
->ro_block_group_mutex
);
2585 btrfs_end_transaction(trans
);
2587 ret
= btrfs_wait_for_commit(fs_info
, transid
);
2590 dirty_bg_running
= true;
2592 } while (dirty_bg_running
);
2594 if (do_chunk_alloc
) {
2596 * If we are changing raid levels, try to allocate a
2597 * corresponding block group with the new raid level.
2599 alloc_flags
= btrfs_get_alloc_profile(fs_info
, cache
->flags
);
2600 if (alloc_flags
!= cache
->flags
) {
2601 ret
= btrfs_chunk_alloc(trans
, alloc_flags
,
2604 * ENOSPC is allowed here, we may have enough space
2605 * already allocated at the new raid level to carry on
2614 ret
= inc_block_group_ro(cache
, 0);
2615 if (!do_chunk_alloc
|| ret
== -ETXTBSY
)
2619 alloc_flags
= btrfs_get_alloc_profile(fs_info
, cache
->space_info
->flags
);
2620 ret
= btrfs_chunk_alloc(trans
, alloc_flags
, CHUNK_ALLOC_FORCE
);
2624 * We have allocated a new chunk. We also need to activate that chunk to
2625 * grant metadata tickets for zoned filesystem.
2627 ret
= btrfs_zoned_activate_one_bg(fs_info
, cache
->space_info
, true);
2631 ret
= inc_block_group_ro(cache
, 0);
2632 if (ret
== -ETXTBSY
)
2635 if (cache
->flags
& BTRFS_BLOCK_GROUP_SYSTEM
) {
2636 alloc_flags
= btrfs_get_alloc_profile(fs_info
, cache
->flags
);
2637 mutex_lock(&fs_info
->chunk_mutex
);
2638 check_system_chunk(trans
, alloc_flags
);
2639 mutex_unlock(&fs_info
->chunk_mutex
);
2642 mutex_unlock(&fs_info
->ro_block_group_mutex
);
2644 btrfs_end_transaction(trans
);
2648 void btrfs_dec_block_group_ro(struct btrfs_block_group
*cache
)
2650 struct btrfs_space_info
*sinfo
= cache
->space_info
;
2655 spin_lock(&sinfo
->lock
);
2656 spin_lock(&cache
->lock
);
2658 if (btrfs_is_zoned(cache
->fs_info
)) {
2659 /* Migrate zone_unusable bytes back */
2660 cache
->zone_unusable
=
2661 (cache
->alloc_offset
- cache
->used
) +
2662 (cache
->length
- cache
->zone_capacity
);
2663 sinfo
->bytes_zone_unusable
+= cache
->zone_unusable
;
2664 sinfo
->bytes_readonly
-= cache
->zone_unusable
;
2666 num_bytes
= cache
->length
- cache
->reserved
-
2667 cache
->pinned
- cache
->bytes_super
-
2668 cache
->zone_unusable
- cache
->used
;
2669 sinfo
->bytes_readonly
-= num_bytes
;
2670 list_del_init(&cache
->ro_list
);
2672 spin_unlock(&cache
->lock
);
2673 spin_unlock(&sinfo
->lock
);
2676 static int update_block_group_item(struct btrfs_trans_handle
*trans
,
2677 struct btrfs_path
*path
,
2678 struct btrfs_block_group
*cache
)
2680 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2682 struct btrfs_root
*root
= btrfs_block_group_root(fs_info
);
2684 struct extent_buffer
*leaf
;
2685 struct btrfs_block_group_item bgi
;
2686 struct btrfs_key key
;
2688 key
.objectid
= cache
->start
;
2689 key
.type
= BTRFS_BLOCK_GROUP_ITEM_KEY
;
2690 key
.offset
= cache
->length
;
2692 ret
= btrfs_search_slot(trans
, root
, &key
, path
, 0, 1);
2699 leaf
= path
->nodes
[0];
2700 bi
= btrfs_item_ptr_offset(leaf
, path
->slots
[0]);
2701 btrfs_set_stack_block_group_used(&bgi
, cache
->used
);
2702 btrfs_set_stack_block_group_chunk_objectid(&bgi
,
2703 cache
->global_root_id
);
2704 btrfs_set_stack_block_group_flags(&bgi
, cache
->flags
);
2705 write_extent_buffer(leaf
, &bgi
, bi
, sizeof(bgi
));
2706 btrfs_mark_buffer_dirty(leaf
);
2708 btrfs_release_path(path
);
2713 static int cache_save_setup(struct btrfs_block_group
*block_group
,
2714 struct btrfs_trans_handle
*trans
,
2715 struct btrfs_path
*path
)
2717 struct btrfs_fs_info
*fs_info
= block_group
->fs_info
;
2718 struct btrfs_root
*root
= fs_info
->tree_root
;
2719 struct inode
*inode
= NULL
;
2720 struct extent_changeset
*data_reserved
= NULL
;
2722 int dcs
= BTRFS_DC_ERROR
;
2727 if (!btrfs_test_opt(fs_info
, SPACE_CACHE
))
2731 * If this block group is smaller than 100 megs don't bother caching the
2734 if (block_group
->length
< (100 * SZ_1M
)) {
2735 spin_lock(&block_group
->lock
);
2736 block_group
->disk_cache_state
= BTRFS_DC_WRITTEN
;
2737 spin_unlock(&block_group
->lock
);
2741 if (TRANS_ABORTED(trans
))
2744 inode
= lookup_free_space_inode(block_group
, path
);
2745 if (IS_ERR(inode
) && PTR_ERR(inode
) != -ENOENT
) {
2746 ret
= PTR_ERR(inode
);
2747 btrfs_release_path(path
);
2751 if (IS_ERR(inode
)) {
2755 if (block_group
->ro
)
2758 ret
= create_free_space_inode(trans
, block_group
, path
);
2765 * We want to set the generation to 0, that way if anything goes wrong
2766 * from here on out we know not to trust this cache when we load up next
2769 BTRFS_I(inode
)->generation
= 0;
2770 ret
= btrfs_update_inode(trans
, root
, BTRFS_I(inode
));
2773 * So theoretically we could recover from this, simply set the
2774 * super cache generation to 0 so we know to invalidate the
2775 * cache, but then we'd have to keep track of the block groups
2776 * that fail this way so we know we _have_ to reset this cache
2777 * before the next commit or risk reading stale cache. So to
2778 * limit our exposure to horrible edge cases lets just abort the
2779 * transaction, this only happens in really bad situations
2782 btrfs_abort_transaction(trans
, ret
);
2787 /* We've already setup this transaction, go ahead and exit */
2788 if (block_group
->cache_generation
== trans
->transid
&&
2789 i_size_read(inode
)) {
2790 dcs
= BTRFS_DC_SETUP
;
2794 if (i_size_read(inode
) > 0) {
2795 ret
= btrfs_check_trunc_cache_free_space(fs_info
,
2796 &fs_info
->global_block_rsv
);
2800 ret
= btrfs_truncate_free_space_cache(trans
, NULL
, inode
);
2805 spin_lock(&block_group
->lock
);
2806 if (block_group
->cached
!= BTRFS_CACHE_FINISHED
||
2807 !btrfs_test_opt(fs_info
, SPACE_CACHE
)) {
2809 * don't bother trying to write stuff out _if_
2810 * a) we're not cached,
2811 * b) we're with nospace_cache mount option,
2812 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2814 dcs
= BTRFS_DC_WRITTEN
;
2815 spin_unlock(&block_group
->lock
);
2818 spin_unlock(&block_group
->lock
);
2821 * We hit an ENOSPC when setting up the cache in this transaction, just
2822 * skip doing the setup, we've already cleared the cache so we're safe.
2824 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC
, &trans
->transaction
->flags
)) {
2830 * Try to preallocate enough space based on how big the block group is.
2831 * Keep in mind this has to include any pinned space which could end up
2832 * taking up quite a bit since it's not folded into the other space
2835 cache_size
= div_u64(block_group
->length
, SZ_256M
);
2840 cache_size
*= fs_info
->sectorsize
;
2842 ret
= btrfs_check_data_free_space(BTRFS_I(inode
), &data_reserved
, 0,
2847 ret
= btrfs_prealloc_file_range_trans(inode
, trans
, 0, 0, cache_size
,
2848 cache_size
, cache_size
,
2851 * Our cache requires contiguous chunks so that we don't modify a bunch
2852 * of metadata or split extents when writing the cache out, which means
2853 * we can enospc if we are heavily fragmented in addition to just normal
2854 * out of space conditions. So if we hit this just skip setting up any
2855 * other block groups for this transaction, maybe we'll unpin enough
2856 * space the next time around.
2859 dcs
= BTRFS_DC_SETUP
;
2860 else if (ret
== -ENOSPC
)
2861 set_bit(BTRFS_TRANS_CACHE_ENOSPC
, &trans
->transaction
->flags
);
2866 btrfs_release_path(path
);
2868 spin_lock(&block_group
->lock
);
2869 if (!ret
&& dcs
== BTRFS_DC_SETUP
)
2870 block_group
->cache_generation
= trans
->transid
;
2871 block_group
->disk_cache_state
= dcs
;
2872 spin_unlock(&block_group
->lock
);
2874 extent_changeset_free(data_reserved
);
2878 int btrfs_setup_space_cache(struct btrfs_trans_handle
*trans
)
2880 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2881 struct btrfs_block_group
*cache
, *tmp
;
2882 struct btrfs_transaction
*cur_trans
= trans
->transaction
;
2883 struct btrfs_path
*path
;
2885 if (list_empty(&cur_trans
->dirty_bgs
) ||
2886 !btrfs_test_opt(fs_info
, SPACE_CACHE
))
2889 path
= btrfs_alloc_path();
2893 /* Could add new block groups, use _safe just in case */
2894 list_for_each_entry_safe(cache
, tmp
, &cur_trans
->dirty_bgs
,
2896 if (cache
->disk_cache_state
== BTRFS_DC_CLEAR
)
2897 cache_save_setup(cache
, trans
, path
);
2900 btrfs_free_path(path
);
2905 * Transaction commit does final block group cache writeback during a critical
2906 * section where nothing is allowed to change the FS. This is required in
2907 * order for the cache to actually match the block group, but can introduce a
2908 * lot of latency into the commit.
2910 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2911 * There's a chance we'll have to redo some of it if the block group changes
2912 * again during the commit, but it greatly reduces the commit latency by
2913 * getting rid of the easy block groups while we're still allowing others to
2916 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle
*trans
)
2918 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
2919 struct btrfs_block_group
*cache
;
2920 struct btrfs_transaction
*cur_trans
= trans
->transaction
;
2923 struct btrfs_path
*path
= NULL
;
2925 struct list_head
*io
= &cur_trans
->io_bgs
;
2928 spin_lock(&cur_trans
->dirty_bgs_lock
);
2929 if (list_empty(&cur_trans
->dirty_bgs
)) {
2930 spin_unlock(&cur_trans
->dirty_bgs_lock
);
2933 list_splice_init(&cur_trans
->dirty_bgs
, &dirty
);
2934 spin_unlock(&cur_trans
->dirty_bgs_lock
);
2937 /* Make sure all the block groups on our dirty list actually exist */
2938 btrfs_create_pending_block_groups(trans
);
2941 path
= btrfs_alloc_path();
2949 * cache_write_mutex is here only to save us from balance or automatic
2950 * removal of empty block groups deleting this block group while we are
2951 * writing out the cache
2953 mutex_lock(&trans
->transaction
->cache_write_mutex
);
2954 while (!list_empty(&dirty
)) {
2955 bool drop_reserve
= true;
2957 cache
= list_first_entry(&dirty
, struct btrfs_block_group
,
2960 * This can happen if something re-dirties a block group that
2961 * is already under IO. Just wait for it to finish and then do
2964 if (!list_empty(&cache
->io_list
)) {
2965 list_del_init(&cache
->io_list
);
2966 btrfs_wait_cache_io(trans
, cache
, path
);
2967 btrfs_put_block_group(cache
);
2972 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2973 * it should update the cache_state. Don't delete until after
2976 * Since we're not running in the commit critical section
2977 * we need the dirty_bgs_lock to protect from update_block_group
2979 spin_lock(&cur_trans
->dirty_bgs_lock
);
2980 list_del_init(&cache
->dirty_list
);
2981 spin_unlock(&cur_trans
->dirty_bgs_lock
);
2985 cache_save_setup(cache
, trans
, path
);
2987 if (cache
->disk_cache_state
== BTRFS_DC_SETUP
) {
2988 cache
->io_ctl
.inode
= NULL
;
2989 ret
= btrfs_write_out_cache(trans
, cache
, path
);
2990 if (ret
== 0 && cache
->io_ctl
.inode
) {
2994 * The cache_write_mutex is protecting the
2995 * io_list, also refer to the definition of
2996 * btrfs_transaction::io_bgs for more details
2998 list_add_tail(&cache
->io_list
, io
);
3001 * If we failed to write the cache, the
3002 * generation will be bad and life goes on
3008 ret
= update_block_group_item(trans
, path
, cache
);
3010 * Our block group might still be attached to the list
3011 * of new block groups in the transaction handle of some
3012 * other task (struct btrfs_trans_handle->new_bgs). This
3013 * means its block group item isn't yet in the extent
3014 * tree. If this happens ignore the error, as we will
3015 * try again later in the critical section of the
3016 * transaction commit.
3018 if (ret
== -ENOENT
) {
3020 spin_lock(&cur_trans
->dirty_bgs_lock
);
3021 if (list_empty(&cache
->dirty_list
)) {
3022 list_add_tail(&cache
->dirty_list
,
3023 &cur_trans
->dirty_bgs
);
3024 btrfs_get_block_group(cache
);
3025 drop_reserve
= false;
3027 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3029 btrfs_abort_transaction(trans
, ret
);
3033 /* If it's not on the io list, we need to put the block group */
3035 btrfs_put_block_group(cache
);
3037 btrfs_delayed_refs_rsv_release(fs_info
, 1);
3039 * Avoid blocking other tasks for too long. It might even save
3040 * us from writing caches for block groups that are going to be
3043 mutex_unlock(&trans
->transaction
->cache_write_mutex
);
3046 mutex_lock(&trans
->transaction
->cache_write_mutex
);
3048 mutex_unlock(&trans
->transaction
->cache_write_mutex
);
3051 * Go through delayed refs for all the stuff we've just kicked off
3052 * and then loop back (just once)
3055 ret
= btrfs_run_delayed_refs(trans
, 0);
3056 if (!ret
&& loops
== 0) {
3058 spin_lock(&cur_trans
->dirty_bgs_lock
);
3059 list_splice_init(&cur_trans
->dirty_bgs
, &dirty
);
3061 * dirty_bgs_lock protects us from concurrent block group
3062 * deletes too (not just cache_write_mutex).
3064 if (!list_empty(&dirty
)) {
3065 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3068 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3072 spin_lock(&cur_trans
->dirty_bgs_lock
);
3073 list_splice_init(&dirty
, &cur_trans
->dirty_bgs
);
3074 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3075 btrfs_cleanup_dirty_bgs(cur_trans
, fs_info
);
3078 btrfs_free_path(path
);
3082 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle
*trans
)
3084 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
3085 struct btrfs_block_group
*cache
;
3086 struct btrfs_transaction
*cur_trans
= trans
->transaction
;
3089 struct btrfs_path
*path
;
3090 struct list_head
*io
= &cur_trans
->io_bgs
;
3092 path
= btrfs_alloc_path();
3097 * Even though we are in the critical section of the transaction commit,
3098 * we can still have concurrent tasks adding elements to this
3099 * transaction's list of dirty block groups. These tasks correspond to
3100 * endio free space workers started when writeback finishes for a
3101 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3102 * allocate new block groups as a result of COWing nodes of the root
3103 * tree when updating the free space inode. The writeback for the space
3104 * caches is triggered by an earlier call to
3105 * btrfs_start_dirty_block_groups() and iterations of the following
3107 * Also we want to do the cache_save_setup first and then run the
3108 * delayed refs to make sure we have the best chance at doing this all
3111 spin_lock(&cur_trans
->dirty_bgs_lock
);
3112 while (!list_empty(&cur_trans
->dirty_bgs
)) {
3113 cache
= list_first_entry(&cur_trans
->dirty_bgs
,
3114 struct btrfs_block_group
,
3118 * This can happen if cache_save_setup re-dirties a block group
3119 * that is already under IO. Just wait for it to finish and
3120 * then do it all again
3122 if (!list_empty(&cache
->io_list
)) {
3123 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3124 list_del_init(&cache
->io_list
);
3125 btrfs_wait_cache_io(trans
, cache
, path
);
3126 btrfs_put_block_group(cache
);
3127 spin_lock(&cur_trans
->dirty_bgs_lock
);
3131 * Don't remove from the dirty list until after we've waited on
3134 list_del_init(&cache
->dirty_list
);
3135 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3138 cache_save_setup(cache
, trans
, path
);
3141 ret
= btrfs_run_delayed_refs(trans
,
3142 (unsigned long) -1);
3144 if (!ret
&& cache
->disk_cache_state
== BTRFS_DC_SETUP
) {
3145 cache
->io_ctl
.inode
= NULL
;
3146 ret
= btrfs_write_out_cache(trans
, cache
, path
);
3147 if (ret
== 0 && cache
->io_ctl
.inode
) {
3149 list_add_tail(&cache
->io_list
, io
);
3152 * If we failed to write the cache, the
3153 * generation will be bad and life goes on
3159 ret
= update_block_group_item(trans
, path
, cache
);
3161 * One of the free space endio workers might have
3162 * created a new block group while updating a free space
3163 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3164 * and hasn't released its transaction handle yet, in
3165 * which case the new block group is still attached to
3166 * its transaction handle and its creation has not
3167 * finished yet (no block group item in the extent tree
3168 * yet, etc). If this is the case, wait for all free
3169 * space endio workers to finish and retry. This is a
3170 * very rare case so no need for a more efficient and
3173 if (ret
== -ENOENT
) {
3174 wait_event(cur_trans
->writer_wait
,
3175 atomic_read(&cur_trans
->num_writers
) == 1);
3176 ret
= update_block_group_item(trans
, path
, cache
);
3179 btrfs_abort_transaction(trans
, ret
);
3182 /* If its not on the io list, we need to put the block group */
3184 btrfs_put_block_group(cache
);
3185 btrfs_delayed_refs_rsv_release(fs_info
, 1);
3186 spin_lock(&cur_trans
->dirty_bgs_lock
);
3188 spin_unlock(&cur_trans
->dirty_bgs_lock
);
3191 * Refer to the definition of io_bgs member for details why it's safe
3192 * to use it without any locking
3194 while (!list_empty(io
)) {
3195 cache
= list_first_entry(io
, struct btrfs_block_group
,
3197 list_del_init(&cache
->io_list
);
3198 btrfs_wait_cache_io(trans
, cache
, path
);
3199 btrfs_put_block_group(cache
);
3202 btrfs_free_path(path
);
3206 static inline bool should_reclaim_block_group(struct btrfs_block_group
*bg
,
3209 const struct btrfs_space_info
*space_info
= bg
->space_info
;
3210 const int reclaim_thresh
= READ_ONCE(space_info
->bg_reclaim_threshold
);
3211 const u64 new_val
= bg
->used
;
3212 const u64 old_val
= new_val
+ bytes_freed
;
3215 if (reclaim_thresh
== 0)
3218 thresh
= div_factor_fine(bg
->length
, reclaim_thresh
);
3221 * If we were below the threshold before don't reclaim, we are likely a
3222 * brand new block group and we don't want to relocate new block groups.
3224 if (old_val
< thresh
)
3226 if (new_val
>= thresh
)
3231 int btrfs_update_block_group(struct btrfs_trans_handle
*trans
,
3232 u64 bytenr
, u64 num_bytes
, bool alloc
)
3234 struct btrfs_fs_info
*info
= trans
->fs_info
;
3235 struct btrfs_block_group
*cache
= NULL
;
3236 u64 total
= num_bytes
;
3242 /* Block accounting for super block */
3243 spin_lock(&info
->delalloc_root_lock
);
3244 old_val
= btrfs_super_bytes_used(info
->super_copy
);
3246 old_val
+= num_bytes
;
3248 old_val
-= num_bytes
;
3249 btrfs_set_super_bytes_used(info
->super_copy
, old_val
);
3250 spin_unlock(&info
->delalloc_root_lock
);
3255 cache
= btrfs_lookup_block_group(info
, bytenr
);
3260 factor
= btrfs_bg_type_to_factor(cache
->flags
);
3263 * If this block group has free space cache written out, we
3264 * need to make sure to load it if we are removing space. This
3265 * is because we need the unpinning stage to actually add the
3266 * space back to the block group, otherwise we will leak space.
3268 if (!alloc
&& !btrfs_block_group_done(cache
))
3269 btrfs_cache_block_group(cache
, true);
3271 byte_in_group
= bytenr
- cache
->start
;
3272 WARN_ON(byte_in_group
> cache
->length
);
3274 spin_lock(&cache
->space_info
->lock
);
3275 spin_lock(&cache
->lock
);
3277 if (btrfs_test_opt(info
, SPACE_CACHE
) &&
3278 cache
->disk_cache_state
< BTRFS_DC_CLEAR
)
3279 cache
->disk_cache_state
= BTRFS_DC_CLEAR
;
3281 old_val
= cache
->used
;
3282 num_bytes
= min(total
, cache
->length
- byte_in_group
);
3284 old_val
+= num_bytes
;
3285 cache
->used
= old_val
;
3286 cache
->reserved
-= num_bytes
;
3287 cache
->space_info
->bytes_reserved
-= num_bytes
;
3288 cache
->space_info
->bytes_used
+= num_bytes
;
3289 cache
->space_info
->disk_used
+= num_bytes
* factor
;
3290 spin_unlock(&cache
->lock
);
3291 spin_unlock(&cache
->space_info
->lock
);
3293 old_val
-= num_bytes
;
3294 cache
->used
= old_val
;
3295 cache
->pinned
+= num_bytes
;
3296 btrfs_space_info_update_bytes_pinned(info
,
3297 cache
->space_info
, num_bytes
);
3298 cache
->space_info
->bytes_used
-= num_bytes
;
3299 cache
->space_info
->disk_used
-= num_bytes
* factor
;
3301 reclaim
= should_reclaim_block_group(cache
, num_bytes
);
3302 spin_unlock(&cache
->lock
);
3303 spin_unlock(&cache
->space_info
->lock
);
3305 set_extent_dirty(&trans
->transaction
->pinned_extents
,
3306 bytenr
, bytenr
+ num_bytes
- 1,
3307 GFP_NOFS
| __GFP_NOFAIL
);
3310 spin_lock(&trans
->transaction
->dirty_bgs_lock
);
3311 if (list_empty(&cache
->dirty_list
)) {
3312 list_add_tail(&cache
->dirty_list
,
3313 &trans
->transaction
->dirty_bgs
);
3314 trans
->delayed_ref_updates
++;
3315 btrfs_get_block_group(cache
);
3317 spin_unlock(&trans
->transaction
->dirty_bgs_lock
);
3320 * No longer have used bytes in this block group, queue it for
3321 * deletion. We do this after adding the block group to the
3322 * dirty list to avoid races between cleaner kthread and space
3325 if (!alloc
&& old_val
== 0) {
3326 if (!btrfs_test_opt(info
, DISCARD_ASYNC
))
3327 btrfs_mark_bg_unused(cache
);
3328 } else if (!alloc
&& reclaim
) {
3329 btrfs_mark_bg_to_reclaim(cache
);
3332 btrfs_put_block_group(cache
);
3334 bytenr
+= num_bytes
;
3337 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3338 btrfs_update_delayed_refs_rsv(trans
);
3343 * btrfs_add_reserved_bytes - update the block_group and space info counters
3344 * @cache: The cache we are manipulating
3345 * @ram_bytes: The number of bytes of file content, and will be same to
3346 * @num_bytes except for the compress path.
3347 * @num_bytes: The number of bytes in question
3348 * @delalloc: The blocks are allocated for the delalloc write
3350 * This is called by the allocator when it reserves space. If this is a
3351 * reservation and the block group has become read only we cannot make the
3352 * reservation and return -EAGAIN, otherwise this function always succeeds.
3354 int btrfs_add_reserved_bytes(struct btrfs_block_group
*cache
,
3355 u64 ram_bytes
, u64 num_bytes
, int delalloc
)
3357 struct btrfs_space_info
*space_info
= cache
->space_info
;
3360 spin_lock(&space_info
->lock
);
3361 spin_lock(&cache
->lock
);
3365 cache
->reserved
+= num_bytes
;
3366 space_info
->bytes_reserved
+= num_bytes
;
3367 trace_btrfs_space_reservation(cache
->fs_info
, "space_info",
3368 space_info
->flags
, num_bytes
, 1);
3369 btrfs_space_info_update_bytes_may_use(cache
->fs_info
,
3370 space_info
, -ram_bytes
);
3372 cache
->delalloc_bytes
+= num_bytes
;
3375 * Compression can use less space than we reserved, so wake
3376 * tickets if that happens
3378 if (num_bytes
< ram_bytes
)
3379 btrfs_try_granting_tickets(cache
->fs_info
, space_info
);
3381 spin_unlock(&cache
->lock
);
3382 spin_unlock(&space_info
->lock
);
3387 * btrfs_free_reserved_bytes - update the block_group and space info counters
3388 * @cache: The cache we are manipulating
3389 * @num_bytes: The number of bytes in question
3390 * @delalloc: The blocks are allocated for the delalloc write
3392 * This is called by somebody who is freeing space that was never actually used
3393 * on disk. For example if you reserve some space for a new leaf in transaction
3394 * A and before transaction A commits you free that leaf, you call this with
3395 * reserve set to 0 in order to clear the reservation.
3397 void btrfs_free_reserved_bytes(struct btrfs_block_group
*cache
,
3398 u64 num_bytes
, int delalloc
)
3400 struct btrfs_space_info
*space_info
= cache
->space_info
;
3402 spin_lock(&space_info
->lock
);
3403 spin_lock(&cache
->lock
);
3405 space_info
->bytes_readonly
+= num_bytes
;
3406 cache
->reserved
-= num_bytes
;
3407 space_info
->bytes_reserved
-= num_bytes
;
3408 space_info
->max_extent_size
= 0;
3411 cache
->delalloc_bytes
-= num_bytes
;
3412 spin_unlock(&cache
->lock
);
3414 btrfs_try_granting_tickets(cache
->fs_info
, space_info
);
3415 spin_unlock(&space_info
->lock
);
3418 static void force_metadata_allocation(struct btrfs_fs_info
*info
)
3420 struct list_head
*head
= &info
->space_info
;
3421 struct btrfs_space_info
*found
;
3423 list_for_each_entry(found
, head
, list
) {
3424 if (found
->flags
& BTRFS_BLOCK_GROUP_METADATA
)
3425 found
->force_alloc
= CHUNK_ALLOC_FORCE
;
3429 static int should_alloc_chunk(struct btrfs_fs_info
*fs_info
,
3430 struct btrfs_space_info
*sinfo
, int force
)
3432 u64 bytes_used
= btrfs_space_info_used(sinfo
, false);
3435 if (force
== CHUNK_ALLOC_FORCE
)
3439 * in limited mode, we want to have some free space up to
3440 * about 1% of the FS size.
3442 if (force
== CHUNK_ALLOC_LIMITED
) {
3443 thresh
= btrfs_super_total_bytes(fs_info
->super_copy
);
3444 thresh
= max_t(u64
, SZ_64M
, div_factor_fine(thresh
, 1));
3446 if (sinfo
->total_bytes
- bytes_used
< thresh
)
3450 if (bytes_used
+ SZ_2M
< div_factor(sinfo
->total_bytes
, 8))
3455 int btrfs_force_chunk_alloc(struct btrfs_trans_handle
*trans
, u64 type
)
3457 u64 alloc_flags
= btrfs_get_alloc_profile(trans
->fs_info
, type
);
3459 return btrfs_chunk_alloc(trans
, alloc_flags
, CHUNK_ALLOC_FORCE
);
3462 static struct btrfs_block_group
*do_chunk_alloc(struct btrfs_trans_handle
*trans
, u64 flags
)
3464 struct btrfs_block_group
*bg
;
3468 * Check if we have enough space in the system space info because we
3469 * will need to update device items in the chunk btree and insert a new
3470 * chunk item in the chunk btree as well. This will allocate a new
3471 * system block group if needed.
3473 check_system_chunk(trans
, flags
);
3475 bg
= btrfs_create_chunk(trans
, flags
);
3481 ret
= btrfs_chunk_alloc_add_chunk_item(trans
, bg
);
3483 * Normally we are not expected to fail with -ENOSPC here, since we have
3484 * previously reserved space in the system space_info and allocated one
3485 * new system chunk if necessary. However there are three exceptions:
3487 * 1) We may have enough free space in the system space_info but all the
3488 * existing system block groups have a profile which can not be used
3489 * for extent allocation.
3491 * This happens when mounting in degraded mode. For example we have a
3492 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3493 * using the other device in degraded mode. If we then allocate a chunk,
3494 * we may have enough free space in the existing system space_info, but
3495 * none of the block groups can be used for extent allocation since they
3496 * have a RAID1 profile, and because we are in degraded mode with a
3497 * single device, we are forced to allocate a new system chunk with a
3498 * SINGLE profile. Making check_system_chunk() iterate over all system
3499 * block groups and check if they have a usable profile and enough space
3500 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3501 * try again after forcing allocation of a new system chunk. Like this
3502 * we avoid paying the cost of that search in normal circumstances, when
3503 * we were not mounted in degraded mode;
3505 * 2) We had enough free space info the system space_info, and one suitable
3506 * block group to allocate from when we called check_system_chunk()
3507 * above. However right after we called it, the only system block group
3508 * with enough free space got turned into RO mode by a running scrub,
3509 * and in this case we have to allocate a new one and retry. We only
3510 * need do this allocate and retry once, since we have a transaction
3511 * handle and scrub uses the commit root to search for block groups;
3513 * 3) We had one system block group with enough free space when we called
3514 * check_system_chunk(), but after that, right before we tried to
3515 * allocate the last extent buffer we needed, a discard operation came
3516 * in and it temporarily removed the last free space entry from the
3517 * block group (discard removes a free space entry, discards it, and
3518 * then adds back the entry to the block group cache).
3520 if (ret
== -ENOSPC
) {
3521 const u64 sys_flags
= btrfs_system_alloc_profile(trans
->fs_info
);
3522 struct btrfs_block_group
*sys_bg
;
3524 sys_bg
= btrfs_create_chunk(trans
, sys_flags
);
3525 if (IS_ERR(sys_bg
)) {
3526 ret
= PTR_ERR(sys_bg
);
3527 btrfs_abort_transaction(trans
, ret
);
3531 ret
= btrfs_chunk_alloc_add_chunk_item(trans
, sys_bg
);
3533 btrfs_abort_transaction(trans
, ret
);
3537 ret
= btrfs_chunk_alloc_add_chunk_item(trans
, bg
);
3539 btrfs_abort_transaction(trans
, ret
);
3543 btrfs_abort_transaction(trans
, ret
);
3547 btrfs_trans_release_chunk_metadata(trans
);
3550 return ERR_PTR(ret
);
3552 btrfs_get_block_group(bg
);
3557 * Chunk allocation is done in 2 phases:
3559 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3560 * the chunk, the chunk mapping, create its block group and add the items
3561 * that belong in the chunk btree to it - more specifically, we need to
3562 * update device items in the chunk btree and add a new chunk item to it.
3564 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3565 * group item to the extent btree and the device extent items to the devices
3568 * This is done to prevent deadlocks. For example when COWing a node from the
3569 * extent btree we are holding a write lock on the node's parent and if we
3570 * trigger chunk allocation and attempted to insert the new block group item
3571 * in the extent btree right way, we could deadlock because the path for the
3572 * insertion can include that parent node. At first glance it seems impossible
3573 * to trigger chunk allocation after starting a transaction since tasks should
3574 * reserve enough transaction units (metadata space), however while that is true
3575 * most of the time, chunk allocation may still be triggered for several reasons:
3577 * 1) When reserving metadata, we check if there is enough free space in the
3578 * metadata space_info and therefore don't trigger allocation of a new chunk.
3579 * However later when the task actually tries to COW an extent buffer from
3580 * the extent btree or from the device btree for example, it is forced to
3581 * allocate a new block group (chunk) because the only one that had enough
3582 * free space was just turned to RO mode by a running scrub for example (or
3583 * device replace, block group reclaim thread, etc), so we can not use it
3584 * for allocating an extent and end up being forced to allocate a new one;
3586 * 2) Because we only check that the metadata space_info has enough free bytes,
3587 * we end up not allocating a new metadata chunk in that case. However if
3588 * the filesystem was mounted in degraded mode, none of the existing block
3589 * groups might be suitable for extent allocation due to their incompatible
3590 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3591 * use a RAID1 profile, in degraded mode using a single device). In this case
3592 * when the task attempts to COW some extent buffer of the extent btree for
3593 * example, it will trigger allocation of a new metadata block group with a
3594 * suitable profile (SINGLE profile in the example of the degraded mount of
3595 * the RAID1 filesystem);
3597 * 3) The task has reserved enough transaction units / metadata space, but when
3598 * it attempts to COW an extent buffer from the extent or device btree for
3599 * example, it does not find any free extent in any metadata block group,
3600 * therefore forced to try to allocate a new metadata block group.
3601 * This is because some other task allocated all available extents in the
3602 * meanwhile - this typically happens with tasks that don't reserve space
3603 * properly, either intentionally or as a bug. One example where this is
3604 * done intentionally is fsync, as it does not reserve any transaction units
3605 * and ends up allocating a variable number of metadata extents for log
3606 * tree extent buffers;
3608 * 4) The task has reserved enough transaction units / metadata space, but right
3609 * before it tries to allocate the last extent buffer it needs, a discard
3610 * operation comes in and, temporarily, removes the last free space entry from
3611 * the only metadata block group that had free space (discard starts by
3612 * removing a free space entry from a block group, then does the discard
3613 * operation and, once it's done, it adds back the free space entry to the
3616 * We also need this 2 phases setup when adding a device to a filesystem with
3617 * a seed device - we must create new metadata and system chunks without adding
3618 * any of the block group items to the chunk, extent and device btrees. If we
3619 * did not do it this way, we would get ENOSPC when attempting to update those
3620 * btrees, since all the chunks from the seed device are read-only.
3622 * Phase 1 does the updates and insertions to the chunk btree because if we had
3623 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3624 * parallel, we risk having too many system chunks allocated by many tasks if
3625 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3626 * extreme case this leads to exhaustion of the system chunk array in the
3627 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3628 * and with RAID filesystems (so we have more device items in the chunk btree).
3629 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3630 * the system chunk array due to concurrent allocations") provides more details.
3632 * Allocation of system chunks does not happen through this function. A task that
3633 * needs to update the chunk btree (the only btree that uses system chunks), must
3634 * preallocate chunk space by calling either check_system_chunk() or
3635 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
3636 * metadata chunk or when removing a chunk, while the later is used before doing
3637 * a modification to the chunk btree - use cases for the later are adding,
3638 * removing and resizing a device as well as relocation of a system chunk.
3639 * See the comment below for more details.
3641 * The reservation of system space, done through check_system_chunk(), as well
3642 * as all the updates and insertions into the chunk btree must be done while
3643 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
3644 * an extent buffer from the chunks btree we never trigger allocation of a new
3645 * system chunk, which would result in a deadlock (trying to lock twice an
3646 * extent buffer of the chunk btree, first time before triggering the chunk
3647 * allocation and the second time during chunk allocation while attempting to
3648 * update the chunks btree). The system chunk array is also updated while holding
3649 * that mutex. The same logic applies to removing chunks - we must reserve system
3650 * space, update the chunk btree and the system chunk array in the superblock
3651 * while holding fs_info->chunk_mutex.
3653 * This function, btrfs_chunk_alloc(), belongs to phase 1.
3655 * If @force is CHUNK_ALLOC_FORCE:
3656 * - return 1 if it successfully allocates a chunk,
3657 * - return errors including -ENOSPC otherwise.
3658 * If @force is NOT CHUNK_ALLOC_FORCE:
3659 * - return 0 if it doesn't need to allocate a new chunk,
3660 * - return 1 if it successfully allocates a chunk,
3661 * - return errors including -ENOSPC otherwise.
3663 int btrfs_chunk_alloc(struct btrfs_trans_handle
*trans
, u64 flags
,
3664 enum btrfs_chunk_alloc_enum force
)
3666 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
3667 struct btrfs_space_info
*space_info
;
3668 struct btrfs_block_group
*ret_bg
;
3669 bool wait_for_alloc
= false;
3670 bool should_alloc
= false;
3671 bool from_extent_allocation
= false;
3674 if (force
== CHUNK_ALLOC_FORCE_FOR_EXTENT
) {
3675 from_extent_allocation
= true;
3676 force
= CHUNK_ALLOC_FORCE
;
3679 /* Don't re-enter if we're already allocating a chunk */
3680 if (trans
->allocating_chunk
)
3683 * Allocation of system chunks can not happen through this path, as we
3684 * could end up in a deadlock if we are allocating a data or metadata
3685 * chunk and there is another task modifying the chunk btree.
3687 * This is because while we are holding the chunk mutex, we will attempt
3688 * to add the new chunk item to the chunk btree or update an existing
3689 * device item in the chunk btree, while the other task that is modifying
3690 * the chunk btree is attempting to COW an extent buffer while holding a
3691 * lock on it and on its parent - if the COW operation triggers a system
3692 * chunk allocation, then we can deadlock because we are holding the
3693 * chunk mutex and we may need to access that extent buffer or its parent
3694 * in order to add the chunk item or update a device item.
3696 * Tasks that want to modify the chunk tree should reserve system space
3697 * before updating the chunk btree, by calling either
3698 * btrfs_reserve_chunk_metadata() or check_system_chunk().
3699 * It's possible that after a task reserves the space, it still ends up
3700 * here - this happens in the cases described above at do_chunk_alloc().
3701 * The task will have to either retry or fail.
3703 if (flags
& BTRFS_BLOCK_GROUP_SYSTEM
)
3706 space_info
= btrfs_find_space_info(fs_info
, flags
);
3710 spin_lock(&space_info
->lock
);
3711 if (force
< space_info
->force_alloc
)
3712 force
= space_info
->force_alloc
;
3713 should_alloc
= should_alloc_chunk(fs_info
, space_info
, force
);
3714 if (space_info
->full
) {
3715 /* No more free physical space */
3720 spin_unlock(&space_info
->lock
);
3722 } else if (!should_alloc
) {
3723 spin_unlock(&space_info
->lock
);
3725 } else if (space_info
->chunk_alloc
) {
3727 * Someone is already allocating, so we need to block
3728 * until this someone is finished and then loop to
3729 * recheck if we should continue with our allocation
3732 wait_for_alloc
= true;
3733 force
= CHUNK_ALLOC_NO_FORCE
;
3734 spin_unlock(&space_info
->lock
);
3735 mutex_lock(&fs_info
->chunk_mutex
);
3736 mutex_unlock(&fs_info
->chunk_mutex
);
3738 /* Proceed with allocation */
3739 space_info
->chunk_alloc
= 1;
3740 wait_for_alloc
= false;
3741 spin_unlock(&space_info
->lock
);
3745 } while (wait_for_alloc
);
3747 mutex_lock(&fs_info
->chunk_mutex
);
3748 trans
->allocating_chunk
= true;
3751 * If we have mixed data/metadata chunks we want to make sure we keep
3752 * allocating mixed chunks instead of individual chunks.
3754 if (btrfs_mixed_space_info(space_info
))
3755 flags
|= (BTRFS_BLOCK_GROUP_DATA
| BTRFS_BLOCK_GROUP_METADATA
);
3758 * if we're doing a data chunk, go ahead and make sure that
3759 * we keep a reasonable number of metadata chunks allocated in the
3762 if (flags
& BTRFS_BLOCK_GROUP_DATA
&& fs_info
->metadata_ratio
) {
3763 fs_info
->data_chunk_allocations
++;
3764 if (!(fs_info
->data_chunk_allocations
%
3765 fs_info
->metadata_ratio
))
3766 force_metadata_allocation(fs_info
);
3769 ret_bg
= do_chunk_alloc(trans
, flags
);
3770 trans
->allocating_chunk
= false;
3772 if (IS_ERR(ret_bg
)) {
3773 ret
= PTR_ERR(ret_bg
);
3774 } else if (from_extent_allocation
) {
3776 * New block group is likely to be used soon. Try to activate
3777 * it now. Failure is OK for now.
3779 btrfs_zone_activate(ret_bg
);
3783 btrfs_put_block_group(ret_bg
);
3785 spin_lock(&space_info
->lock
);
3788 space_info
->full
= 1;
3793 space_info
->max_extent_size
= 0;
3796 space_info
->force_alloc
= CHUNK_ALLOC_NO_FORCE
;
3798 space_info
->chunk_alloc
= 0;
3799 spin_unlock(&space_info
->lock
);
3800 mutex_unlock(&fs_info
->chunk_mutex
);
3805 static u64
get_profile_num_devs(struct btrfs_fs_info
*fs_info
, u64 type
)
3809 num_dev
= btrfs_raid_array
[btrfs_bg_flags_to_raid_index(type
)].devs_max
;
3811 num_dev
= fs_info
->fs_devices
->rw_devices
;
3816 static void reserve_chunk_space(struct btrfs_trans_handle
*trans
,
3820 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
3821 struct btrfs_space_info
*info
;
3826 * Needed because we can end up allocating a system chunk and for an
3827 * atomic and race free space reservation in the chunk block reserve.
3829 lockdep_assert_held(&fs_info
->chunk_mutex
);
3831 info
= btrfs_find_space_info(fs_info
, BTRFS_BLOCK_GROUP_SYSTEM
);
3832 spin_lock(&info
->lock
);
3833 left
= info
->total_bytes
- btrfs_space_info_used(info
, true);
3834 spin_unlock(&info
->lock
);
3836 if (left
< bytes
&& btrfs_test_opt(fs_info
, ENOSPC_DEBUG
)) {
3837 btrfs_info(fs_info
, "left=%llu, need=%llu, flags=%llu",
3839 btrfs_dump_space_info(fs_info
, info
, 0, 0);
3843 u64 flags
= btrfs_system_alloc_profile(fs_info
);
3844 struct btrfs_block_group
*bg
;
3847 * Ignore failure to create system chunk. We might end up not
3848 * needing it, as we might not need to COW all nodes/leafs from
3849 * the paths we visit in the chunk tree (they were already COWed
3850 * or created in the current transaction for example).
3852 bg
= btrfs_create_chunk(trans
, flags
);
3857 * We have a new chunk. We also need to activate it for
3860 ret
= btrfs_zoned_activate_one_bg(fs_info
, info
, true);
3865 * If we fail to add the chunk item here, we end up
3866 * trying again at phase 2 of chunk allocation, at
3867 * btrfs_create_pending_block_groups(). So ignore
3868 * any error here. An ENOSPC here could happen, due to
3869 * the cases described at do_chunk_alloc() - the system
3870 * block group we just created was just turned into RO
3871 * mode by a scrub for example, or a running discard
3872 * temporarily removed its free space entries, etc.
3874 btrfs_chunk_alloc_add_chunk_item(trans
, bg
);
3879 ret
= btrfs_block_rsv_add(fs_info
,
3880 &fs_info
->chunk_block_rsv
,
3881 bytes
, BTRFS_RESERVE_NO_FLUSH
);
3883 trans
->chunk_bytes_reserved
+= bytes
;
3888 * Reserve space in the system space for allocating or removing a chunk.
3889 * The caller must be holding fs_info->chunk_mutex.
3891 void check_system_chunk(struct btrfs_trans_handle
*trans
, u64 type
)
3893 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
3894 const u64 num_devs
= get_profile_num_devs(fs_info
, type
);
3897 /* num_devs device items to update and 1 chunk item to add or remove. */
3898 bytes
= btrfs_calc_metadata_size(fs_info
, num_devs
) +
3899 btrfs_calc_insert_metadata_size(fs_info
, 1);
3901 reserve_chunk_space(trans
, bytes
, type
);
3905 * Reserve space in the system space, if needed, for doing a modification to the
3908 * @trans: A transaction handle.
3909 * @is_item_insertion: Indicate if the modification is for inserting a new item
3910 * in the chunk btree or if it's for the deletion or update
3911 * of an existing item.
3913 * This is used in a context where we need to update the chunk btree outside
3914 * block group allocation and removal, to avoid a deadlock with a concurrent
3915 * task that is allocating a metadata or data block group and therefore needs to
3916 * update the chunk btree while holding the chunk mutex. After the update to the
3917 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
3920 void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle
*trans
,
3921 bool is_item_insertion
)
3923 struct btrfs_fs_info
*fs_info
= trans
->fs_info
;
3926 if (is_item_insertion
)
3927 bytes
= btrfs_calc_insert_metadata_size(fs_info
, 1);
3929 bytes
= btrfs_calc_metadata_size(fs_info
, 1);
3931 mutex_lock(&fs_info
->chunk_mutex
);
3932 reserve_chunk_space(trans
, bytes
, BTRFS_BLOCK_GROUP_SYSTEM
);
3933 mutex_unlock(&fs_info
->chunk_mutex
);
3936 void btrfs_put_block_group_cache(struct btrfs_fs_info
*info
)
3938 struct btrfs_block_group
*block_group
;
3940 block_group
= btrfs_lookup_first_block_group(info
, 0);
3941 while (block_group
) {
3942 btrfs_wait_block_group_cache_done(block_group
);
3943 spin_lock(&block_group
->lock
);
3944 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF
,
3945 &block_group
->runtime_flags
)) {
3946 struct inode
*inode
= block_group
->inode
;
3948 block_group
->inode
= NULL
;
3949 spin_unlock(&block_group
->lock
);
3951 ASSERT(block_group
->io_ctl
.inode
== NULL
);
3954 spin_unlock(&block_group
->lock
);
3956 block_group
= btrfs_next_block_group(block_group
);
3961 * Must be called only after stopping all workers, since we could have block
3962 * group caching kthreads running, and therefore they could race with us if we
3963 * freed the block groups before stopping them.
3965 int btrfs_free_block_groups(struct btrfs_fs_info
*info
)
3967 struct btrfs_block_group
*block_group
;
3968 struct btrfs_space_info
*space_info
;
3969 struct btrfs_caching_control
*caching_ctl
;
3972 write_lock(&info
->block_group_cache_lock
);
3973 while (!list_empty(&info
->caching_block_groups
)) {
3974 caching_ctl
= list_entry(info
->caching_block_groups
.next
,
3975 struct btrfs_caching_control
, list
);
3976 list_del(&caching_ctl
->list
);
3977 btrfs_put_caching_control(caching_ctl
);
3979 write_unlock(&info
->block_group_cache_lock
);
3981 spin_lock(&info
->unused_bgs_lock
);
3982 while (!list_empty(&info
->unused_bgs
)) {
3983 block_group
= list_first_entry(&info
->unused_bgs
,
3984 struct btrfs_block_group
,
3986 list_del_init(&block_group
->bg_list
);
3987 btrfs_put_block_group(block_group
);
3990 while (!list_empty(&info
->reclaim_bgs
)) {
3991 block_group
= list_first_entry(&info
->reclaim_bgs
,
3992 struct btrfs_block_group
,
3994 list_del_init(&block_group
->bg_list
);
3995 btrfs_put_block_group(block_group
);
3997 spin_unlock(&info
->unused_bgs_lock
);
3999 spin_lock(&info
->zone_active_bgs_lock
);
4000 while (!list_empty(&info
->zone_active_bgs
)) {
4001 block_group
= list_first_entry(&info
->zone_active_bgs
,
4002 struct btrfs_block_group
,
4004 list_del_init(&block_group
->active_bg_list
);
4005 btrfs_put_block_group(block_group
);
4007 spin_unlock(&info
->zone_active_bgs_lock
);
4009 write_lock(&info
->block_group_cache_lock
);
4010 while ((n
= rb_last(&info
->block_group_cache_tree
.rb_root
)) != NULL
) {
4011 block_group
= rb_entry(n
, struct btrfs_block_group
,
4013 rb_erase_cached(&block_group
->cache_node
,
4014 &info
->block_group_cache_tree
);
4015 RB_CLEAR_NODE(&block_group
->cache_node
);
4016 write_unlock(&info
->block_group_cache_lock
);
4018 down_write(&block_group
->space_info
->groups_sem
);
4019 list_del(&block_group
->list
);
4020 up_write(&block_group
->space_info
->groups_sem
);
4023 * We haven't cached this block group, which means we could
4024 * possibly have excluded extents on this block group.
4026 if (block_group
->cached
== BTRFS_CACHE_NO
||
4027 block_group
->cached
== BTRFS_CACHE_ERROR
)
4028 btrfs_free_excluded_extents(block_group
);
4030 btrfs_remove_free_space_cache(block_group
);
4031 ASSERT(block_group
->cached
!= BTRFS_CACHE_STARTED
);
4032 ASSERT(list_empty(&block_group
->dirty_list
));
4033 ASSERT(list_empty(&block_group
->io_list
));
4034 ASSERT(list_empty(&block_group
->bg_list
));
4035 ASSERT(refcount_read(&block_group
->refs
) == 1);
4036 ASSERT(block_group
->swap_extents
== 0);
4037 btrfs_put_block_group(block_group
);
4039 write_lock(&info
->block_group_cache_lock
);
4041 write_unlock(&info
->block_group_cache_lock
);
4043 btrfs_release_global_block_rsv(info
);
4045 while (!list_empty(&info
->space_info
)) {
4046 space_info
= list_entry(info
->space_info
.next
,
4047 struct btrfs_space_info
,
4051 * Do not hide this behind enospc_debug, this is actually
4052 * important and indicates a real bug if this happens.
4054 if (WARN_ON(space_info
->bytes_pinned
> 0 ||
4055 space_info
->bytes_may_use
> 0))
4056 btrfs_dump_space_info(info
, space_info
, 0, 0);
4059 * If there was a failure to cleanup a log tree, very likely due
4060 * to an IO failure on a writeback attempt of one or more of its
4061 * extent buffers, we could not do proper (and cheap) unaccounting
4062 * of their reserved space, so don't warn on bytes_reserved > 0 in
4065 if (!(space_info
->flags
& BTRFS_BLOCK_GROUP_METADATA
) ||
4066 !BTRFS_FS_LOG_CLEANUP_ERROR(info
)) {
4067 if (WARN_ON(space_info
->bytes_reserved
> 0))
4068 btrfs_dump_space_info(info
, space_info
, 0, 0);
4071 WARN_ON(space_info
->reclaim_size
> 0);
4072 list_del(&space_info
->list
);
4073 btrfs_sysfs_remove_space_info(space_info
);
4078 void btrfs_freeze_block_group(struct btrfs_block_group
*cache
)
4080 atomic_inc(&cache
->frozen
);
4083 void btrfs_unfreeze_block_group(struct btrfs_block_group
*block_group
)
4085 struct btrfs_fs_info
*fs_info
= block_group
->fs_info
;
4086 struct extent_map_tree
*em_tree
;
4087 struct extent_map
*em
;
4090 spin_lock(&block_group
->lock
);
4091 cleanup
= (atomic_dec_and_test(&block_group
->frozen
) &&
4092 test_bit(BLOCK_GROUP_FLAG_REMOVED
, &block_group
->runtime_flags
));
4093 spin_unlock(&block_group
->lock
);
4096 em_tree
= &fs_info
->mapping_tree
;
4097 write_lock(&em_tree
->lock
);
4098 em
= lookup_extent_mapping(em_tree
, block_group
->start
,
4100 BUG_ON(!em
); /* logic error, can't happen */
4101 remove_extent_mapping(em_tree
, em
);
4102 write_unlock(&em_tree
->lock
);
4104 /* once for us and once for the tree */
4105 free_extent_map(em
);
4106 free_extent_map(em
);
4109 * We may have left one free space entry and other possible
4110 * tasks trimming this block group have left 1 entry each one.
4113 __btrfs_remove_free_space_cache(block_group
->free_space_ctl
);
4117 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group
*bg
)
4121 spin_lock(&bg
->lock
);
4126 spin_unlock(&bg
->lock
);
4131 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group
*bg
, int amount
)
4133 spin_lock(&bg
->lock
);
4135 ASSERT(bg
->swap_extents
>= amount
);
4136 bg
->swap_extents
-= amount
;
4137 spin_unlock(&bg
->lock
);