1 // SPDX-License-Identifier: GPL-2.0
5 #include "block-group.h"
6 #include "space-info.h"
8 #include "free-space-cache.h"
9 #include "free-space-tree.h"
11 #include "transaction.h"
12 #include "ref-verify.h"
15 #include "delalloc-space.h"
21 * Return target flags in extended format or 0 if restripe for this chunk_type
24 * Should be called with balance_lock held
26 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
28 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
34 if (flags & BTRFS_BLOCK_GROUP_DATA &&
35 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
36 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
37 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
38 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
39 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
40 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
41 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
42 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
49 * @flags: available profiles in extended format (see ctree.h)
51 * Return reduced profile in chunk format. If profile changing is in progress
52 * (either running or paused) picks the target profile (if it's already
53 * available), otherwise falls back to plain reducing.
55 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
57 u64 num_devices = fs_info->fs_devices->rw_devices;
63 * See if restripe for this chunk_type is in progress, if so try to
64 * reduce to the target profile
66 spin_lock(&fs_info->balance_lock);
67 target = get_restripe_target(fs_info, flags);
69 spin_unlock(&fs_info->balance_lock);
70 return extended_to_chunk(target);
72 spin_unlock(&fs_info->balance_lock);
74 /* First, mask out the RAID levels which aren't possible */
75 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
76 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
77 allowed |= btrfs_raid_array[raid_type].bg_flag;
81 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
82 allowed = BTRFS_BLOCK_GROUP_RAID6;
83 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
84 allowed = BTRFS_BLOCK_GROUP_RAID5;
85 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
86 allowed = BTRFS_BLOCK_GROUP_RAID10;
87 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
88 allowed = BTRFS_BLOCK_GROUP_RAID1;
89 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
90 allowed = BTRFS_BLOCK_GROUP_RAID0;
92 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
94 return extended_to_chunk(flags | allowed);
97 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
104 seq = read_seqbegin(&fs_info->profiles_lock);
106 if (flags & BTRFS_BLOCK_GROUP_DATA)
107 flags |= fs_info->avail_data_alloc_bits;
108 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
109 flags |= fs_info->avail_system_alloc_bits;
110 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
111 flags |= fs_info->avail_metadata_alloc_bits;
112 } while (read_seqretry(&fs_info->profiles_lock, seq));
114 return btrfs_reduce_alloc_profile(fs_info, flags);
117 void btrfs_get_block_group(struct btrfs_block_group *cache)
119 refcount_inc(&cache->refs);
122 void btrfs_put_block_group(struct btrfs_block_group *cache)
124 if (refcount_dec_and_test(&cache->refs)) {
125 WARN_ON(cache->pinned > 0);
126 WARN_ON(cache->reserved > 0);
129 * A block_group shouldn't be on the discard_list anymore.
130 * Remove the block_group from the discard_list to prevent us
131 * from causing a panic due to NULL pointer dereference.
133 if (WARN_ON(!list_empty(&cache->discard_list)))
134 btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
138 * If not empty, someone is still holding mutex of
139 * full_stripe_lock, which can only be released by caller.
140 * And it will definitely cause use-after-free when caller
141 * tries to release full stripe lock.
143 * No better way to resolve, but only to warn.
145 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
146 kfree(cache->free_space_ctl);
152 * This adds the block group to the fs_info rb tree for the block group cache
154 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
155 struct btrfs_block_group *block_group)
158 struct rb_node *parent = NULL;
159 struct btrfs_block_group *cache;
161 ASSERT(block_group->length != 0);
163 spin_lock(&info->block_group_cache_lock);
164 p = &info->block_group_cache_tree.rb_node;
168 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
169 if (block_group->start < cache->start) {
171 } else if (block_group->start > cache->start) {
174 spin_unlock(&info->block_group_cache_lock);
179 rb_link_node(&block_group->cache_node, parent, p);
180 rb_insert_color(&block_group->cache_node,
181 &info->block_group_cache_tree);
183 if (info->first_logical_byte > block_group->start)
184 info->first_logical_byte = block_group->start;
186 spin_unlock(&info->block_group_cache_lock);
192 * This will return the block group at or after bytenr if contains is 0, else
193 * it will return the block group that contains the bytenr
195 static struct btrfs_block_group *block_group_cache_tree_search(
196 struct btrfs_fs_info *info, u64 bytenr, int contains)
198 struct btrfs_block_group *cache, *ret = NULL;
202 spin_lock(&info->block_group_cache_lock);
203 n = info->block_group_cache_tree.rb_node;
206 cache = rb_entry(n, struct btrfs_block_group, cache_node);
207 end = cache->start + cache->length - 1;
208 start = cache->start;
210 if (bytenr < start) {
211 if (!contains && (!ret || start < ret->start))
214 } else if (bytenr > start) {
215 if (contains && bytenr <= end) {
226 btrfs_get_block_group(ret);
227 if (bytenr == 0 && info->first_logical_byte > ret->start)
228 info->first_logical_byte = ret->start;
230 spin_unlock(&info->block_group_cache_lock);
236 * Return the block group that starts at or after bytenr
238 struct btrfs_block_group *btrfs_lookup_first_block_group(
239 struct btrfs_fs_info *info, u64 bytenr)
241 return block_group_cache_tree_search(info, bytenr, 0);
245 * Return the block group that contains the given bytenr
247 struct btrfs_block_group *btrfs_lookup_block_group(
248 struct btrfs_fs_info *info, u64 bytenr)
250 return block_group_cache_tree_search(info, bytenr, 1);
253 struct btrfs_block_group *btrfs_next_block_group(
254 struct btrfs_block_group *cache)
256 struct btrfs_fs_info *fs_info = cache->fs_info;
257 struct rb_node *node;
259 spin_lock(&fs_info->block_group_cache_lock);
261 /* If our block group was removed, we need a full search. */
262 if (RB_EMPTY_NODE(&cache->cache_node)) {
263 const u64 next_bytenr = cache->start + cache->length;
265 spin_unlock(&fs_info->block_group_cache_lock);
266 btrfs_put_block_group(cache);
267 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
269 node = rb_next(&cache->cache_node);
270 btrfs_put_block_group(cache);
272 cache = rb_entry(node, struct btrfs_block_group, cache_node);
273 btrfs_get_block_group(cache);
276 spin_unlock(&fs_info->block_group_cache_lock);
280 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
282 struct btrfs_block_group *bg;
285 bg = btrfs_lookup_block_group(fs_info, bytenr);
289 spin_lock(&bg->lock);
293 atomic_inc(&bg->nocow_writers);
294 spin_unlock(&bg->lock);
296 /* No put on block group, done by btrfs_dec_nocow_writers */
298 btrfs_put_block_group(bg);
303 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
305 struct btrfs_block_group *bg;
307 bg = btrfs_lookup_block_group(fs_info, bytenr);
309 if (atomic_dec_and_test(&bg->nocow_writers))
310 wake_up_var(&bg->nocow_writers);
312 * Once for our lookup and once for the lookup done by a previous call
313 * to btrfs_inc_nocow_writers()
315 btrfs_put_block_group(bg);
316 btrfs_put_block_group(bg);
319 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
321 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
324 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
327 struct btrfs_block_group *bg;
329 bg = btrfs_lookup_block_group(fs_info, start);
331 if (atomic_dec_and_test(&bg->reservations))
332 wake_up_var(&bg->reservations);
333 btrfs_put_block_group(bg);
336 void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
338 struct btrfs_space_info *space_info = bg->space_info;
342 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
346 * Our block group is read only but before we set it to read only,
347 * some task might have had allocated an extent from it already, but it
348 * has not yet created a respective ordered extent (and added it to a
349 * root's list of ordered extents).
350 * Therefore wait for any task currently allocating extents, since the
351 * block group's reservations counter is incremented while a read lock
352 * on the groups' semaphore is held and decremented after releasing
353 * the read access on that semaphore and creating the ordered extent.
355 down_write(&space_info->groups_sem);
356 up_write(&space_info->groups_sem);
358 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
361 struct btrfs_caching_control *btrfs_get_caching_control(
362 struct btrfs_block_group *cache)
364 struct btrfs_caching_control *ctl;
366 spin_lock(&cache->lock);
367 if (!cache->caching_ctl) {
368 spin_unlock(&cache->lock);
372 ctl = cache->caching_ctl;
373 refcount_inc(&ctl->count);
374 spin_unlock(&cache->lock);
378 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
380 if (refcount_dec_and_test(&ctl->count))
385 * When we wait for progress in the block group caching, its because our
386 * allocation attempt failed at least once. So, we must sleep and let some
387 * progress happen before we try again.
389 * This function will sleep at least once waiting for new free space to show
390 * up, and then it will check the block group free space numbers for our min
391 * num_bytes. Another option is to have it go ahead and look in the rbtree for
392 * a free extent of a given size, but this is a good start.
394 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
395 * any of the information in this block group.
397 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
400 struct btrfs_caching_control *caching_ctl;
402 caching_ctl = btrfs_get_caching_control(cache);
406 wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
407 (cache->free_space_ctl->free_space >= num_bytes));
409 btrfs_put_caching_control(caching_ctl);
412 int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
414 struct btrfs_caching_control *caching_ctl;
417 caching_ctl = btrfs_get_caching_control(cache);
419 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
421 wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
422 if (cache->cached == BTRFS_CACHE_ERROR)
424 btrfs_put_caching_control(caching_ctl);
428 static bool space_cache_v1_done(struct btrfs_block_group *cache)
432 spin_lock(&cache->lock);
433 ret = cache->cached != BTRFS_CACHE_FAST;
434 spin_unlock(&cache->lock);
439 void btrfs_wait_space_cache_v1_finished(struct btrfs_block_group *cache,
440 struct btrfs_caching_control *caching_ctl)
442 wait_event(caching_ctl->wait, space_cache_v1_done(cache));
445 #ifdef CONFIG_BTRFS_DEBUG
446 static void fragment_free_space(struct btrfs_block_group *block_group)
448 struct btrfs_fs_info *fs_info = block_group->fs_info;
449 u64 start = block_group->start;
450 u64 len = block_group->length;
451 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
452 fs_info->nodesize : fs_info->sectorsize;
453 u64 step = chunk << 1;
455 while (len > chunk) {
456 btrfs_remove_free_space(block_group, start, chunk);
467 * This is only called by btrfs_cache_block_group, since we could have freed
468 * extents we need to check the pinned_extents for any extents that can't be
469 * used yet since their free space will be released as soon as the transaction
472 u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
474 struct btrfs_fs_info *info = block_group->fs_info;
475 u64 extent_start, extent_end, size, total_added = 0;
478 while (start < end) {
479 ret = find_first_extent_bit(&info->excluded_extents, start,
480 &extent_start, &extent_end,
481 EXTENT_DIRTY | EXTENT_UPTODATE,
486 if (extent_start <= start) {
487 start = extent_end + 1;
488 } else if (extent_start > start && extent_start < end) {
489 size = extent_start - start;
491 ret = btrfs_add_free_space_async_trimmed(block_group,
493 BUG_ON(ret); /* -ENOMEM or logic error */
494 start = extent_end + 1;
503 ret = btrfs_add_free_space_async_trimmed(block_group, start,
505 BUG_ON(ret); /* -ENOMEM or logic error */
511 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
513 struct btrfs_block_group *block_group = caching_ctl->block_group;
514 struct btrfs_fs_info *fs_info = block_group->fs_info;
515 struct btrfs_root *extent_root = fs_info->extent_root;
516 struct btrfs_path *path;
517 struct extent_buffer *leaf;
518 struct btrfs_key key;
525 path = btrfs_alloc_path();
529 last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
531 #ifdef CONFIG_BTRFS_DEBUG
533 * If we're fragmenting we don't want to make anybody think we can
534 * allocate from this block group until we've had a chance to fragment
537 if (btrfs_should_fragment_free_space(block_group))
541 * We don't want to deadlock with somebody trying to allocate a new
542 * extent for the extent root while also trying to search the extent
543 * root to add free space. So we skip locking and search the commit
544 * root, since its read-only
546 path->skip_locking = 1;
547 path->search_commit_root = 1;
548 path->reada = READA_FORWARD;
552 key.type = BTRFS_EXTENT_ITEM_KEY;
555 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
559 leaf = path->nodes[0];
560 nritems = btrfs_header_nritems(leaf);
563 if (btrfs_fs_closing(fs_info) > 1) {
568 if (path->slots[0] < nritems) {
569 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
571 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
575 if (need_resched() ||
576 rwsem_is_contended(&fs_info->commit_root_sem)) {
578 caching_ctl->progress = last;
579 btrfs_release_path(path);
580 up_read(&fs_info->commit_root_sem);
581 mutex_unlock(&caching_ctl->mutex);
583 mutex_lock(&caching_ctl->mutex);
584 down_read(&fs_info->commit_root_sem);
588 ret = btrfs_next_leaf(extent_root, path);
593 leaf = path->nodes[0];
594 nritems = btrfs_header_nritems(leaf);
598 if (key.objectid < last) {
601 key.type = BTRFS_EXTENT_ITEM_KEY;
604 caching_ctl->progress = last;
605 btrfs_release_path(path);
609 if (key.objectid < block_group->start) {
614 if (key.objectid >= block_group->start + block_group->length)
617 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
618 key.type == BTRFS_METADATA_ITEM_KEY) {
619 total_found += add_new_free_space(block_group, last,
621 if (key.type == BTRFS_METADATA_ITEM_KEY)
622 last = key.objectid +
625 last = key.objectid + key.offset;
627 if (total_found > CACHING_CTL_WAKE_UP) {
630 wake_up(&caching_ctl->wait);
637 total_found += add_new_free_space(block_group, last,
638 block_group->start + block_group->length);
639 caching_ctl->progress = (u64)-1;
642 btrfs_free_path(path);
646 static noinline void caching_thread(struct btrfs_work *work)
648 struct btrfs_block_group *block_group;
649 struct btrfs_fs_info *fs_info;
650 struct btrfs_caching_control *caching_ctl;
653 caching_ctl = container_of(work, struct btrfs_caching_control, work);
654 block_group = caching_ctl->block_group;
655 fs_info = block_group->fs_info;
657 mutex_lock(&caching_ctl->mutex);
658 down_read(&fs_info->commit_root_sem);
660 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
661 ret = load_free_space_cache(block_group);
668 * We failed to load the space cache, set ourselves to
669 * CACHE_STARTED and carry on.
671 spin_lock(&block_group->lock);
672 block_group->cached = BTRFS_CACHE_STARTED;
673 spin_unlock(&block_group->lock);
674 wake_up(&caching_ctl->wait);
678 * If we are in the transaction that populated the free space tree we
679 * can't actually cache from the free space tree as our commit root and
680 * real root are the same, so we could change the contents of the blocks
681 * while caching. Instead do the slow caching in this case, and after
682 * the transaction has committed we will be safe.
684 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
685 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
686 ret = load_free_space_tree(caching_ctl);
688 ret = load_extent_tree_free(caching_ctl);
690 spin_lock(&block_group->lock);
691 block_group->caching_ctl = NULL;
692 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
693 spin_unlock(&block_group->lock);
695 #ifdef CONFIG_BTRFS_DEBUG
696 if (btrfs_should_fragment_free_space(block_group)) {
699 spin_lock(&block_group->space_info->lock);
700 spin_lock(&block_group->lock);
701 bytes_used = block_group->length - block_group->used;
702 block_group->space_info->bytes_used += bytes_used >> 1;
703 spin_unlock(&block_group->lock);
704 spin_unlock(&block_group->space_info->lock);
705 fragment_free_space(block_group);
709 caching_ctl->progress = (u64)-1;
711 up_read(&fs_info->commit_root_sem);
712 btrfs_free_excluded_extents(block_group);
713 mutex_unlock(&caching_ctl->mutex);
715 wake_up(&caching_ctl->wait);
717 btrfs_put_caching_control(caching_ctl);
718 btrfs_put_block_group(block_group);
721 int btrfs_cache_block_group(struct btrfs_block_group *cache, int load_cache_only)
724 struct btrfs_fs_info *fs_info = cache->fs_info;
725 struct btrfs_caching_control *caching_ctl = NULL;
728 /* Allocator for zoned filesystems does not use the cache at all */
729 if (btrfs_is_zoned(fs_info))
732 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
736 INIT_LIST_HEAD(&caching_ctl->list);
737 mutex_init(&caching_ctl->mutex);
738 init_waitqueue_head(&caching_ctl->wait);
739 caching_ctl->block_group = cache;
740 caching_ctl->progress = cache->start;
741 refcount_set(&caching_ctl->count, 2);
742 btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
744 spin_lock(&cache->lock);
745 if (cache->cached != BTRFS_CACHE_NO) {
748 caching_ctl = cache->caching_ctl;
750 refcount_inc(&caching_ctl->count);
751 spin_unlock(&cache->lock);
754 WARN_ON(cache->caching_ctl);
755 cache->caching_ctl = caching_ctl;
756 if (btrfs_test_opt(fs_info, SPACE_CACHE))
757 cache->cached = BTRFS_CACHE_FAST;
759 cache->cached = BTRFS_CACHE_STARTED;
760 cache->has_caching_ctl = 1;
761 spin_unlock(&cache->lock);
763 spin_lock(&fs_info->block_group_cache_lock);
764 refcount_inc(&caching_ctl->count);
765 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
766 spin_unlock(&fs_info->block_group_cache_lock);
768 btrfs_get_block_group(cache);
770 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
772 if (load_cache_only && caching_ctl)
773 btrfs_wait_space_cache_v1_finished(cache, caching_ctl);
775 btrfs_put_caching_control(caching_ctl);
780 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
782 u64 extra_flags = chunk_to_extended(flags) &
783 BTRFS_EXTENDED_PROFILE_MASK;
785 write_seqlock(&fs_info->profiles_lock);
786 if (flags & BTRFS_BLOCK_GROUP_DATA)
787 fs_info->avail_data_alloc_bits &= ~extra_flags;
788 if (flags & BTRFS_BLOCK_GROUP_METADATA)
789 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
790 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
791 fs_info->avail_system_alloc_bits &= ~extra_flags;
792 write_sequnlock(&fs_info->profiles_lock);
796 * Clear incompat bits for the following feature(s):
798 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
799 * in the whole filesystem
801 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
803 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
805 bool found_raid56 = false;
806 bool found_raid1c34 = false;
808 if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
809 (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
810 (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
811 struct list_head *head = &fs_info->space_info;
812 struct btrfs_space_info *sinfo;
814 list_for_each_entry_rcu(sinfo, head, list) {
815 down_read(&sinfo->groups_sem);
816 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
818 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
820 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
821 found_raid1c34 = true;
822 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
823 found_raid1c34 = true;
824 up_read(&sinfo->groups_sem);
827 btrfs_clear_fs_incompat(fs_info, RAID56);
829 btrfs_clear_fs_incompat(fs_info, RAID1C34);
833 static int remove_block_group_item(struct btrfs_trans_handle *trans,
834 struct btrfs_path *path,
835 struct btrfs_block_group *block_group)
837 struct btrfs_fs_info *fs_info = trans->fs_info;
838 struct btrfs_root *root;
839 struct btrfs_key key;
842 root = fs_info->extent_root;
843 key.objectid = block_group->start;
844 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
845 key.offset = block_group->length;
847 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
853 ret = btrfs_del_item(trans, root, path);
857 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
858 u64 group_start, struct extent_map *em)
860 struct btrfs_fs_info *fs_info = trans->fs_info;
861 struct btrfs_path *path;
862 struct btrfs_block_group *block_group;
863 struct btrfs_free_cluster *cluster;
865 struct kobject *kobj = NULL;
869 struct btrfs_caching_control *caching_ctl = NULL;
871 bool remove_rsv = false;
873 block_group = btrfs_lookup_block_group(fs_info, group_start);
874 BUG_ON(!block_group);
875 BUG_ON(!block_group->ro);
877 trace_btrfs_remove_block_group(block_group);
879 * Free the reserved super bytes from this block group before
882 btrfs_free_excluded_extents(block_group);
883 btrfs_free_ref_tree_range(fs_info, block_group->start,
884 block_group->length);
886 index = btrfs_bg_flags_to_raid_index(block_group->flags);
887 factor = btrfs_bg_type_to_factor(block_group->flags);
889 /* make sure this block group isn't part of an allocation cluster */
890 cluster = &fs_info->data_alloc_cluster;
891 spin_lock(&cluster->refill_lock);
892 btrfs_return_cluster_to_free_space(block_group, cluster);
893 spin_unlock(&cluster->refill_lock);
896 * make sure this block group isn't part of a metadata
899 cluster = &fs_info->meta_alloc_cluster;
900 spin_lock(&cluster->refill_lock);
901 btrfs_return_cluster_to_free_space(block_group, cluster);
902 spin_unlock(&cluster->refill_lock);
904 btrfs_clear_treelog_bg(block_group);
906 path = btrfs_alloc_path();
913 * get the inode first so any iput calls done for the io_list
914 * aren't the final iput (no unlinks allowed now)
916 inode = lookup_free_space_inode(block_group, path);
918 mutex_lock(&trans->transaction->cache_write_mutex);
920 * Make sure our free space cache IO is done before removing the
923 spin_lock(&trans->transaction->dirty_bgs_lock);
924 if (!list_empty(&block_group->io_list)) {
925 list_del_init(&block_group->io_list);
927 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
929 spin_unlock(&trans->transaction->dirty_bgs_lock);
930 btrfs_wait_cache_io(trans, block_group, path);
931 btrfs_put_block_group(block_group);
932 spin_lock(&trans->transaction->dirty_bgs_lock);
935 if (!list_empty(&block_group->dirty_list)) {
936 list_del_init(&block_group->dirty_list);
938 btrfs_put_block_group(block_group);
940 spin_unlock(&trans->transaction->dirty_bgs_lock);
941 mutex_unlock(&trans->transaction->cache_write_mutex);
943 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
947 spin_lock(&fs_info->block_group_cache_lock);
948 rb_erase(&block_group->cache_node,
949 &fs_info->block_group_cache_tree);
950 RB_CLEAR_NODE(&block_group->cache_node);
952 /* Once for the block groups rbtree */
953 btrfs_put_block_group(block_group);
955 if (fs_info->first_logical_byte == block_group->start)
956 fs_info->first_logical_byte = (u64)-1;
957 spin_unlock(&fs_info->block_group_cache_lock);
959 down_write(&block_group->space_info->groups_sem);
961 * we must use list_del_init so people can check to see if they
962 * are still on the list after taking the semaphore
964 list_del_init(&block_group->list);
965 if (list_empty(&block_group->space_info->block_groups[index])) {
966 kobj = block_group->space_info->block_group_kobjs[index];
967 block_group->space_info->block_group_kobjs[index] = NULL;
968 clear_avail_alloc_bits(fs_info, block_group->flags);
970 up_write(&block_group->space_info->groups_sem);
971 clear_incompat_bg_bits(fs_info, block_group->flags);
977 if (block_group->has_caching_ctl)
978 caching_ctl = btrfs_get_caching_control(block_group);
979 if (block_group->cached == BTRFS_CACHE_STARTED)
980 btrfs_wait_block_group_cache_done(block_group);
981 if (block_group->has_caching_ctl) {
982 spin_lock(&fs_info->block_group_cache_lock);
984 struct btrfs_caching_control *ctl;
986 list_for_each_entry(ctl,
987 &fs_info->caching_block_groups, list)
988 if (ctl->block_group == block_group) {
990 refcount_inc(&caching_ctl->count);
995 list_del_init(&caching_ctl->list);
996 spin_unlock(&fs_info->block_group_cache_lock);
998 /* Once for the caching bgs list and once for us. */
999 btrfs_put_caching_control(caching_ctl);
1000 btrfs_put_caching_control(caching_ctl);
1004 spin_lock(&trans->transaction->dirty_bgs_lock);
1005 WARN_ON(!list_empty(&block_group->dirty_list));
1006 WARN_ON(!list_empty(&block_group->io_list));
1007 spin_unlock(&trans->transaction->dirty_bgs_lock);
1009 btrfs_remove_free_space_cache(block_group);
1011 spin_lock(&block_group->space_info->lock);
1012 list_del_init(&block_group->ro_list);
1014 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1015 WARN_ON(block_group->space_info->total_bytes
1016 < block_group->length);
1017 WARN_ON(block_group->space_info->bytes_readonly
1018 < block_group->length - block_group->zone_unusable);
1019 WARN_ON(block_group->space_info->bytes_zone_unusable
1020 < block_group->zone_unusable);
1021 WARN_ON(block_group->space_info->disk_total
1022 < block_group->length * factor);
1024 block_group->space_info->total_bytes -= block_group->length;
1025 block_group->space_info->bytes_readonly -=
1026 (block_group->length - block_group->zone_unusable);
1027 block_group->space_info->bytes_zone_unusable -=
1028 block_group->zone_unusable;
1029 block_group->space_info->disk_total -= block_group->length * factor;
1031 spin_unlock(&block_group->space_info->lock);
1034 * Remove the free space for the block group from the free space tree
1035 * and the block group's item from the extent tree before marking the
1036 * block group as removed. This is to prevent races with tasks that
1037 * freeze and unfreeze a block group, this task and another task
1038 * allocating a new block group - the unfreeze task ends up removing
1039 * the block group's extent map before the task calling this function
1040 * deletes the block group item from the extent tree, allowing for
1041 * another task to attempt to create another block group with the same
1042 * item key (and failing with -EEXIST and a transaction abort).
1044 ret = remove_block_group_free_space(trans, block_group);
1048 ret = remove_block_group_item(trans, path, block_group);
1052 spin_lock(&block_group->lock);
1053 block_group->removed = 1;
1055 * At this point trimming or scrub can't start on this block group,
1056 * because we removed the block group from the rbtree
1057 * fs_info->block_group_cache_tree so no one can't find it anymore and
1058 * even if someone already got this block group before we removed it
1059 * from the rbtree, they have already incremented block_group->frozen -
1060 * if they didn't, for the trimming case they won't find any free space
1061 * entries because we already removed them all when we called
1062 * btrfs_remove_free_space_cache().
1064 * And we must not remove the extent map from the fs_info->mapping_tree
1065 * to prevent the same logical address range and physical device space
1066 * ranges from being reused for a new block group. This is needed to
1067 * avoid races with trimming and scrub.
1069 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1070 * completely transactionless, so while it is trimming a range the
1071 * currently running transaction might finish and a new one start,
1072 * allowing for new block groups to be created that can reuse the same
1073 * physical device locations unless we take this special care.
1075 * There may also be an implicit trim operation if the file system
1076 * is mounted with -odiscard. The same protections must remain
1077 * in place until the extents have been discarded completely when
1078 * the transaction commit has completed.
1080 remove_em = (atomic_read(&block_group->frozen) == 0);
1081 spin_unlock(&block_group->lock);
1084 struct extent_map_tree *em_tree;
1086 em_tree = &fs_info->mapping_tree;
1087 write_lock(&em_tree->lock);
1088 remove_extent_mapping(em_tree, em);
1089 write_unlock(&em_tree->lock);
1090 /* once for the tree */
1091 free_extent_map(em);
1095 /* Once for the lookup reference */
1096 btrfs_put_block_group(block_group);
1098 btrfs_delayed_refs_rsv_release(fs_info, 1);
1099 btrfs_free_path(path);
1103 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1104 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1106 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1107 struct extent_map *em;
1108 struct map_lookup *map;
1109 unsigned int num_items;
1111 read_lock(&em_tree->lock);
1112 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1113 read_unlock(&em_tree->lock);
1114 ASSERT(em && em->start == chunk_offset);
1117 * We need to reserve 3 + N units from the metadata space info in order
1118 * to remove a block group (done at btrfs_remove_chunk() and at
1119 * btrfs_remove_block_group()), which are used for:
1121 * 1 unit for adding the free space inode's orphan (located in the tree
1123 * 1 unit for deleting the block group item (located in the extent
1125 * 1 unit for deleting the free space item (located in tree of tree
1127 * N units for deleting N device extent items corresponding to each
1128 * stripe (located in the device tree).
1130 * In order to remove a block group we also need to reserve units in the
1131 * system space info in order to update the chunk tree (update one or
1132 * more device items and remove one chunk item), but this is done at
1133 * btrfs_remove_chunk() through a call to check_system_chunk().
1135 map = em->map_lookup;
1136 num_items = 3 + map->num_stripes;
1137 free_extent_map(em);
1139 return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
1144 * Mark block group @cache read-only, so later write won't happen to block
1147 * If @force is not set, this function will only mark the block group readonly
1148 * if we have enough free space (1M) in other metadata/system block groups.
1149 * If @force is not set, this function will mark the block group readonly
1150 * without checking free space.
1152 * NOTE: This function doesn't care if other block groups can contain all the
1153 * data in this block group. That check should be done by relocation routine,
1154 * not this function.
1156 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1158 struct btrfs_space_info *sinfo = cache->space_info;
1162 spin_lock(&sinfo->lock);
1163 spin_lock(&cache->lock);
1165 if (cache->swap_extents) {
1176 num_bytes = cache->length - cache->reserved - cache->pinned -
1177 cache->bytes_super - cache->zone_unusable - cache->used;
1180 * Data never overcommits, even in mixed mode, so do just the straight
1181 * check of left over space in how much we have allocated.
1185 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1186 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1189 * Here we make sure if we mark this bg RO, we still have enough
1190 * free space as buffer.
1192 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1196 * We overcommit metadata, so we need to do the
1197 * btrfs_can_overcommit check here, and we need to pass in
1198 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1199 * leeway to allow us to mark this block group as read only.
1201 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1202 BTRFS_RESERVE_NO_FLUSH))
1207 sinfo->bytes_readonly += num_bytes;
1208 if (btrfs_is_zoned(cache->fs_info)) {
1209 /* Migrate zone_unusable bytes to readonly */
1210 sinfo->bytes_readonly += cache->zone_unusable;
1211 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1212 cache->zone_unusable = 0;
1215 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1218 spin_unlock(&cache->lock);
1219 spin_unlock(&sinfo->lock);
1220 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1221 btrfs_info(cache->fs_info,
1222 "unable to make block group %llu ro", cache->start);
1223 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1228 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1229 struct btrfs_block_group *bg)
1231 struct btrfs_fs_info *fs_info = bg->fs_info;
1232 struct btrfs_transaction *prev_trans = NULL;
1233 const u64 start = bg->start;
1234 const u64 end = start + bg->length - 1;
1237 spin_lock(&fs_info->trans_lock);
1238 if (trans->transaction->list.prev != &fs_info->trans_list) {
1239 prev_trans = list_last_entry(&trans->transaction->list,
1240 struct btrfs_transaction, list);
1241 refcount_inc(&prev_trans->use_count);
1243 spin_unlock(&fs_info->trans_lock);
1246 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1247 * btrfs_finish_extent_commit(). If we are at transaction N, another
1248 * task might be running finish_extent_commit() for the previous
1249 * transaction N - 1, and have seen a range belonging to the block
1250 * group in pinned_extents before we were able to clear the whole block
1251 * group range from pinned_extents. This means that task can lookup for
1252 * the block group after we unpinned it from pinned_extents and removed
1253 * it, leading to a BUG_ON() at unpin_extent_range().
1255 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1257 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1263 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1266 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1268 btrfs_put_transaction(prev_trans);
1274 * Process the unused_bgs list and remove any that don't have any allocated
1275 * space inside of them.
1277 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1279 struct btrfs_block_group *block_group;
1280 struct btrfs_space_info *space_info;
1281 struct btrfs_trans_handle *trans;
1282 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1285 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1289 * Long running balances can keep us blocked here for eternity, so
1290 * simply skip deletion if we're unable to get the mutex.
1292 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1295 spin_lock(&fs_info->unused_bgs_lock);
1296 while (!list_empty(&fs_info->unused_bgs)) {
1299 block_group = list_first_entry(&fs_info->unused_bgs,
1300 struct btrfs_block_group,
1302 list_del_init(&block_group->bg_list);
1304 space_info = block_group->space_info;
1306 if (ret || btrfs_mixed_space_info(space_info)) {
1307 btrfs_put_block_group(block_group);
1310 spin_unlock(&fs_info->unused_bgs_lock);
1312 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1314 /* Don't want to race with allocators so take the groups_sem */
1315 down_write(&space_info->groups_sem);
1318 * Async discard moves the final block group discard to be prior
1319 * to the unused_bgs code path. Therefore, if it's not fully
1320 * trimmed, punt it back to the async discard lists.
1322 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1323 !btrfs_is_free_space_trimmed(block_group)) {
1324 trace_btrfs_skip_unused_block_group(block_group);
1325 up_write(&space_info->groups_sem);
1326 /* Requeue if we failed because of async discard */
1327 btrfs_discard_queue_work(&fs_info->discard_ctl,
1332 spin_lock(&block_group->lock);
1333 if (block_group->reserved || block_group->pinned ||
1334 block_group->used || block_group->ro ||
1335 list_is_singular(&block_group->list)) {
1337 * We want to bail if we made new allocations or have
1338 * outstanding allocations in this block group. We do
1339 * the ro check in case balance is currently acting on
1342 trace_btrfs_skip_unused_block_group(block_group);
1343 spin_unlock(&block_group->lock);
1344 up_write(&space_info->groups_sem);
1347 spin_unlock(&block_group->lock);
1349 /* We don't want to force the issue, only flip if it's ok. */
1350 ret = inc_block_group_ro(block_group, 0);
1351 up_write(&space_info->groups_sem);
1358 * Want to do this before we do anything else so we can recover
1359 * properly if we fail to join the transaction.
1361 trans = btrfs_start_trans_remove_block_group(fs_info,
1362 block_group->start);
1363 if (IS_ERR(trans)) {
1364 btrfs_dec_block_group_ro(block_group);
1365 ret = PTR_ERR(trans);
1370 * We could have pending pinned extents for this block group,
1371 * just delete them, we don't care about them anymore.
1373 if (!clean_pinned_extents(trans, block_group)) {
1374 btrfs_dec_block_group_ro(block_group);
1379 * At this point, the block_group is read only and should fail
1380 * new allocations. However, btrfs_finish_extent_commit() can
1381 * cause this block_group to be placed back on the discard
1382 * lists because now the block_group isn't fully discarded.
1383 * Bail here and try again later after discarding everything.
1385 spin_lock(&fs_info->discard_ctl.lock);
1386 if (!list_empty(&block_group->discard_list)) {
1387 spin_unlock(&fs_info->discard_ctl.lock);
1388 btrfs_dec_block_group_ro(block_group);
1389 btrfs_discard_queue_work(&fs_info->discard_ctl,
1393 spin_unlock(&fs_info->discard_ctl.lock);
1395 /* Reset pinned so btrfs_put_block_group doesn't complain */
1396 spin_lock(&space_info->lock);
1397 spin_lock(&block_group->lock);
1399 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1400 -block_group->pinned);
1401 space_info->bytes_readonly += block_group->pinned;
1402 block_group->pinned = 0;
1404 spin_unlock(&block_group->lock);
1405 spin_unlock(&space_info->lock);
1408 * The normal path here is an unused block group is passed here,
1409 * then trimming is handled in the transaction commit path.
1410 * Async discard interposes before this to do the trimming
1411 * before coming down the unused block group path as trimming
1412 * will no longer be done later in the transaction commit path.
1414 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1418 * DISCARD can flip during remount. On zoned filesystems, we
1419 * need to reset sequential-required zones.
1421 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1422 btrfs_is_zoned(fs_info);
1424 /* Implicit trim during transaction commit. */
1426 btrfs_freeze_block_group(block_group);
1429 * Btrfs_remove_chunk will abort the transaction if things go
1432 ret = btrfs_remove_chunk(trans, block_group->start);
1436 btrfs_unfreeze_block_group(block_group);
1441 * If we're not mounted with -odiscard, we can just forget
1442 * about this block group. Otherwise we'll need to wait
1443 * until transaction commit to do the actual discard.
1446 spin_lock(&fs_info->unused_bgs_lock);
1448 * A concurrent scrub might have added us to the list
1449 * fs_info->unused_bgs, so use a list_move operation
1450 * to add the block group to the deleted_bgs list.
1452 list_move(&block_group->bg_list,
1453 &trans->transaction->deleted_bgs);
1454 spin_unlock(&fs_info->unused_bgs_lock);
1455 btrfs_get_block_group(block_group);
1458 btrfs_end_transaction(trans);
1460 btrfs_put_block_group(block_group);
1461 spin_lock(&fs_info->unused_bgs_lock);
1463 spin_unlock(&fs_info->unused_bgs_lock);
1464 mutex_unlock(&fs_info->reclaim_bgs_lock);
1468 btrfs_end_transaction(trans);
1469 mutex_unlock(&fs_info->reclaim_bgs_lock);
1470 btrfs_put_block_group(block_group);
1471 btrfs_discard_punt_unused_bgs_list(fs_info);
1474 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1476 struct btrfs_fs_info *fs_info = bg->fs_info;
1478 spin_lock(&fs_info->unused_bgs_lock);
1479 if (list_empty(&bg->bg_list)) {
1480 btrfs_get_block_group(bg);
1481 trace_btrfs_add_unused_block_group(bg);
1482 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1484 spin_unlock(&fs_info->unused_bgs_lock);
1487 void btrfs_reclaim_bgs_work(struct work_struct *work)
1489 struct btrfs_fs_info *fs_info =
1490 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1491 struct btrfs_block_group *bg;
1492 struct btrfs_space_info *space_info;
1493 LIST_HEAD(again_list);
1495 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1498 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE))
1502 * Long running balances can keep us blocked here for eternity, so
1503 * simply skip reclaim if we're unable to get the mutex.
1505 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1506 btrfs_exclop_finish(fs_info);
1510 spin_lock(&fs_info->unused_bgs_lock);
1511 while (!list_empty(&fs_info->reclaim_bgs)) {
1515 bg = list_first_entry(&fs_info->reclaim_bgs,
1516 struct btrfs_block_group,
1518 list_del_init(&bg->bg_list);
1520 space_info = bg->space_info;
1521 spin_unlock(&fs_info->unused_bgs_lock);
1523 /* Don't race with allocators so take the groups_sem */
1524 down_write(&space_info->groups_sem);
1526 spin_lock(&bg->lock);
1527 if (bg->reserved || bg->pinned || bg->ro) {
1529 * We want to bail if we made new allocations or have
1530 * outstanding allocations in this block group. We do
1531 * the ro check in case balance is currently acting on
1534 spin_unlock(&bg->lock);
1535 up_write(&space_info->groups_sem);
1538 spin_unlock(&bg->lock);
1540 /* Get out fast, in case we're unmounting the filesystem */
1541 if (btrfs_fs_closing(fs_info)) {
1542 up_write(&space_info->groups_sem);
1547 * Cache the zone_unusable value before turning the block group
1548 * to read only. As soon as the blog group is read only it's
1549 * zone_unusable value gets moved to the block group's read-only
1550 * bytes and isn't available for calculations anymore.
1552 zone_unusable = bg->zone_unusable;
1553 ret = inc_block_group_ro(bg, 0);
1554 up_write(&space_info->groups_sem);
1559 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1560 bg->start, div_u64(bg->used * 100, bg->length),
1561 div64_u64(zone_unusable * 100, bg->length));
1562 trace_btrfs_reclaim_block_group(bg);
1563 ret = btrfs_relocate_chunk(fs_info, bg->start);
1565 btrfs_err(fs_info, "error relocating chunk %llu",
1569 spin_lock(&fs_info->unused_bgs_lock);
1570 if (ret == -EAGAIN && list_empty(&bg->bg_list))
1571 list_add_tail(&bg->bg_list, &again_list);
1573 btrfs_put_block_group(bg);
1575 list_splice_tail(&again_list, &fs_info->reclaim_bgs);
1576 spin_unlock(&fs_info->unused_bgs_lock);
1577 mutex_unlock(&fs_info->reclaim_bgs_lock);
1578 btrfs_exclop_finish(fs_info);
1581 void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1583 spin_lock(&fs_info->unused_bgs_lock);
1584 if (!list_empty(&fs_info->reclaim_bgs))
1585 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1586 spin_unlock(&fs_info->unused_bgs_lock);
1589 void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1591 struct btrfs_fs_info *fs_info = bg->fs_info;
1593 spin_lock(&fs_info->unused_bgs_lock);
1594 if (list_empty(&bg->bg_list)) {
1595 btrfs_get_block_group(bg);
1596 trace_btrfs_add_reclaim_block_group(bg);
1597 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1599 spin_unlock(&fs_info->unused_bgs_lock);
1602 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1603 struct btrfs_path *path)
1605 struct extent_map_tree *em_tree;
1606 struct extent_map *em;
1607 struct btrfs_block_group_item bg;
1608 struct extent_buffer *leaf;
1613 slot = path->slots[0];
1614 leaf = path->nodes[0];
1616 em_tree = &fs_info->mapping_tree;
1617 read_lock(&em_tree->lock);
1618 em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1619 read_unlock(&em_tree->lock);
1622 "logical %llu len %llu found bg but no related chunk",
1623 key->objectid, key->offset);
1627 if (em->start != key->objectid || em->len != key->offset) {
1629 "block group %llu len %llu mismatch with chunk %llu len %llu",
1630 key->objectid, key->offset, em->start, em->len);
1635 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1637 flags = btrfs_stack_block_group_flags(&bg) &
1638 BTRFS_BLOCK_GROUP_TYPE_MASK;
1640 if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1642 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1643 key->objectid, key->offset, flags,
1644 (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1649 free_extent_map(em);
1653 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1654 struct btrfs_path *path,
1655 struct btrfs_key *key)
1657 struct btrfs_root *root = fs_info->extent_root;
1659 struct btrfs_key found_key;
1660 struct extent_buffer *leaf;
1663 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1668 slot = path->slots[0];
1669 leaf = path->nodes[0];
1670 if (slot >= btrfs_header_nritems(leaf)) {
1671 ret = btrfs_next_leaf(root, path);
1678 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1680 if (found_key.objectid >= key->objectid &&
1681 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1682 ret = read_bg_from_eb(fs_info, &found_key, path);
1692 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1694 u64 extra_flags = chunk_to_extended(flags) &
1695 BTRFS_EXTENDED_PROFILE_MASK;
1697 write_seqlock(&fs_info->profiles_lock);
1698 if (flags & BTRFS_BLOCK_GROUP_DATA)
1699 fs_info->avail_data_alloc_bits |= extra_flags;
1700 if (flags & BTRFS_BLOCK_GROUP_METADATA)
1701 fs_info->avail_metadata_alloc_bits |= extra_flags;
1702 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1703 fs_info->avail_system_alloc_bits |= extra_flags;
1704 write_sequnlock(&fs_info->profiles_lock);
1708 * Map a physical disk address to a list of logical addresses
1710 * @fs_info: the filesystem
1711 * @chunk_start: logical address of block group
1712 * @bdev: physical device to resolve, can be NULL to indicate any device
1713 * @physical: physical address to map to logical addresses
1714 * @logical: return array of logical addresses which map to @physical
1715 * @naddrs: length of @logical
1716 * @stripe_len: size of IO stripe for the given block group
1718 * Maps a particular @physical disk address to a list of @logical addresses.
1719 * Used primarily to exclude those portions of a block group that contain super
1722 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1723 struct block_device *bdev, u64 physical, u64 **logical,
1724 int *naddrs, int *stripe_len)
1726 struct extent_map *em;
1727 struct map_lookup *map;
1730 u64 data_stripe_length;
1735 em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1739 map = em->map_lookup;
1740 data_stripe_length = em->orig_block_len;
1741 io_stripe_size = map->stripe_len;
1742 chunk_start = em->start;
1744 /* For RAID5/6 adjust to a full IO stripe length */
1745 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1746 io_stripe_size = map->stripe_len * nr_data_stripes(map);
1748 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1754 for (i = 0; i < map->num_stripes; i++) {
1755 bool already_inserted = false;
1760 if (!in_range(physical, map->stripes[i].physical,
1761 data_stripe_length))
1764 if (bdev && map->stripes[i].dev->bdev != bdev)
1767 stripe_nr = physical - map->stripes[i].physical;
1768 stripe_nr = div64_u64_rem(stripe_nr, map->stripe_len, &offset);
1770 if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1771 stripe_nr = stripe_nr * map->num_stripes + i;
1772 stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1773 } else if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1774 stripe_nr = stripe_nr * map->num_stripes + i;
1777 * The remaining case would be for RAID56, multiply by
1778 * nr_data_stripes(). Alternatively, just use rmap_len below
1779 * instead of map->stripe_len
1782 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
1784 /* Ensure we don't add duplicate addresses */
1785 for (j = 0; j < nr; j++) {
1786 if (buf[j] == bytenr) {
1787 already_inserted = true;
1792 if (!already_inserted)
1798 *stripe_len = io_stripe_size;
1800 free_extent_map(em);
1804 static int exclude_super_stripes(struct btrfs_block_group *cache)
1806 struct btrfs_fs_info *fs_info = cache->fs_info;
1807 const bool zoned = btrfs_is_zoned(fs_info);
1813 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1814 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1815 cache->bytes_super += stripe_len;
1816 ret = btrfs_add_excluded_extent(fs_info, cache->start,
1822 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1823 bytenr = btrfs_sb_offset(i);
1824 ret = btrfs_rmap_block(fs_info, cache->start, NULL,
1825 bytenr, &logical, &nr, &stripe_len);
1829 /* Shouldn't have super stripes in sequential zones */
1832 "zoned: block group %llu must not contain super block",
1838 u64 len = min_t(u64, stripe_len,
1839 cache->start + cache->length - logical[nr]);
1841 cache->bytes_super += len;
1842 ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1855 static void link_block_group(struct btrfs_block_group *cache)
1857 struct btrfs_space_info *space_info = cache->space_info;
1858 int index = btrfs_bg_flags_to_raid_index(cache->flags);
1860 down_write(&space_info->groups_sem);
1861 list_add_tail(&cache->list, &space_info->block_groups[index]);
1862 up_write(&space_info->groups_sem);
1865 static struct btrfs_block_group *btrfs_create_block_group_cache(
1866 struct btrfs_fs_info *fs_info, u64 start)
1868 struct btrfs_block_group *cache;
1870 cache = kzalloc(sizeof(*cache), GFP_NOFS);
1874 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1876 if (!cache->free_space_ctl) {
1881 cache->start = start;
1883 cache->fs_info = fs_info;
1884 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1886 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1888 refcount_set(&cache->refs, 1);
1889 spin_lock_init(&cache->lock);
1890 init_rwsem(&cache->data_rwsem);
1891 INIT_LIST_HEAD(&cache->list);
1892 INIT_LIST_HEAD(&cache->cluster_list);
1893 INIT_LIST_HEAD(&cache->bg_list);
1894 INIT_LIST_HEAD(&cache->ro_list);
1895 INIT_LIST_HEAD(&cache->discard_list);
1896 INIT_LIST_HEAD(&cache->dirty_list);
1897 INIT_LIST_HEAD(&cache->io_list);
1898 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1899 atomic_set(&cache->frozen, 0);
1900 mutex_init(&cache->free_space_lock);
1901 btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
1907 * Iterate all chunks and verify that each of them has the corresponding block
1910 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
1912 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
1913 struct extent_map *em;
1914 struct btrfs_block_group *bg;
1919 read_lock(&map_tree->lock);
1921 * lookup_extent_mapping will return the first extent map
1922 * intersecting the range, so setting @len to 1 is enough to
1923 * get the first chunk.
1925 em = lookup_extent_mapping(map_tree, start, 1);
1926 read_unlock(&map_tree->lock);
1930 bg = btrfs_lookup_block_group(fs_info, em->start);
1933 "chunk start=%llu len=%llu doesn't have corresponding block group",
1934 em->start, em->len);
1936 free_extent_map(em);
1939 if (bg->start != em->start || bg->length != em->len ||
1940 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
1941 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1943 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1945 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
1946 bg->start, bg->length,
1947 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
1949 free_extent_map(em);
1950 btrfs_put_block_group(bg);
1953 start = em->start + em->len;
1954 free_extent_map(em);
1955 btrfs_put_block_group(bg);
1960 static int read_one_block_group(struct btrfs_fs_info *info,
1961 struct btrfs_block_group_item *bgi,
1962 const struct btrfs_key *key,
1965 struct btrfs_block_group *cache;
1966 struct btrfs_space_info *space_info;
1967 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
1970 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
1972 cache = btrfs_create_block_group_cache(info, key->objectid);
1976 cache->length = key->offset;
1977 cache->used = btrfs_stack_block_group_used(bgi);
1978 cache->flags = btrfs_stack_block_group_flags(bgi);
1980 set_free_space_tree_thresholds(cache);
1984 * When we mount with old space cache, we need to
1985 * set BTRFS_DC_CLEAR and set dirty flag.
1987 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
1988 * truncate the old free space cache inode and
1990 * b) Setting 'dirty flag' makes sure that we flush
1991 * the new space cache info onto disk.
1993 if (btrfs_test_opt(info, SPACE_CACHE))
1994 cache->disk_cache_state = BTRFS_DC_CLEAR;
1996 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
1997 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
1999 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2005 ret = btrfs_load_block_group_zone_info(cache, false);
2007 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2013 * We need to exclude the super stripes now so that the space info has
2014 * super bytes accounted for, otherwise we'll think we have more space
2015 * than we actually do.
2017 ret = exclude_super_stripes(cache);
2019 /* We may have excluded something, so call this just in case. */
2020 btrfs_free_excluded_extents(cache);
2025 * For zoned filesystem, space after the allocation offset is the only
2026 * free space for a block group. So, we don't need any caching work.
2027 * btrfs_calc_zone_unusable() will set the amount of free space and
2028 * zone_unusable space.
2030 * For regular filesystem, check for two cases, either we are full, and
2031 * therefore don't need to bother with the caching work since we won't
2032 * find any space, or we are empty, and we can just add all the space
2033 * in and be done with it. This saves us _a_lot_ of time, particularly
2036 if (btrfs_is_zoned(info)) {
2037 btrfs_calc_zone_unusable(cache);
2038 } else if (cache->length == cache->used) {
2039 cache->last_byte_to_unpin = (u64)-1;
2040 cache->cached = BTRFS_CACHE_FINISHED;
2041 btrfs_free_excluded_extents(cache);
2042 } else if (cache->used == 0) {
2043 cache->last_byte_to_unpin = (u64)-1;
2044 cache->cached = BTRFS_CACHE_FINISHED;
2045 add_new_free_space(cache, cache->start,
2046 cache->start + cache->length);
2047 btrfs_free_excluded_extents(cache);
2050 ret = btrfs_add_block_group_cache(info, cache);
2052 btrfs_remove_free_space_cache(cache);
2055 trace_btrfs_add_block_group(info, cache, 0);
2056 btrfs_update_space_info(info, cache->flags, cache->length,
2057 cache->used, cache->bytes_super,
2058 cache->zone_unusable, &space_info);
2060 cache->space_info = space_info;
2062 link_block_group(cache);
2064 set_avail_alloc_bits(info, cache->flags);
2065 if (btrfs_chunk_readonly(info, cache->start)) {
2066 inc_block_group_ro(cache, 1);
2067 } else if (cache->used == 0) {
2068 ASSERT(list_empty(&cache->bg_list));
2069 if (btrfs_test_opt(info, DISCARD_ASYNC))
2070 btrfs_discard_queue_work(&info->discard_ctl, cache);
2072 btrfs_mark_bg_unused(cache);
2076 btrfs_put_block_group(cache);
2080 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2082 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
2083 struct btrfs_space_info *space_info;
2084 struct rb_node *node;
2087 for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
2088 struct extent_map *em;
2089 struct map_lookup *map;
2090 struct btrfs_block_group *bg;
2092 em = rb_entry(node, struct extent_map, rb_node);
2093 map = em->map_lookup;
2094 bg = btrfs_create_block_group_cache(fs_info, em->start);
2100 /* Fill dummy cache as FULL */
2101 bg->length = em->len;
2102 bg->flags = map->type;
2103 bg->last_byte_to_unpin = (u64)-1;
2104 bg->cached = BTRFS_CACHE_FINISHED;
2106 bg->flags = map->type;
2107 ret = btrfs_add_block_group_cache(fs_info, bg);
2109 btrfs_remove_free_space_cache(bg);
2110 btrfs_put_block_group(bg);
2113 btrfs_update_space_info(fs_info, bg->flags, em->len, em->len,
2115 bg->space_info = space_info;
2116 link_block_group(bg);
2118 set_avail_alloc_bits(fs_info, bg->flags);
2121 btrfs_init_global_block_rsv(fs_info);
2125 int btrfs_read_block_groups(struct btrfs_fs_info *info)
2127 struct btrfs_path *path;
2129 struct btrfs_block_group *cache;
2130 struct btrfs_space_info *space_info;
2131 struct btrfs_key key;
2135 if (!info->extent_root)
2136 return fill_dummy_bgs(info);
2140 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2141 path = btrfs_alloc_path();
2145 cache_gen = btrfs_super_cache_generation(info->super_copy);
2146 if (btrfs_test_opt(info, SPACE_CACHE) &&
2147 btrfs_super_generation(info->super_copy) != cache_gen)
2149 if (btrfs_test_opt(info, CLEAR_CACHE))
2153 struct btrfs_block_group_item bgi;
2154 struct extent_buffer *leaf;
2157 ret = find_first_block_group(info, path, &key);
2163 leaf = path->nodes[0];
2164 slot = path->slots[0];
2166 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2169 btrfs_item_key_to_cpu(leaf, &key, slot);
2170 btrfs_release_path(path);
2171 ret = read_one_block_group(info, &bgi, &key, need_clear);
2174 key.objectid += key.offset;
2177 btrfs_release_path(path);
2179 list_for_each_entry(space_info, &info->space_info, list) {
2182 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2183 if (list_empty(&space_info->block_groups[i]))
2185 cache = list_first_entry(&space_info->block_groups[i],
2186 struct btrfs_block_group,
2188 btrfs_sysfs_add_block_group_type(cache);
2191 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2192 (BTRFS_BLOCK_GROUP_RAID10 |
2193 BTRFS_BLOCK_GROUP_RAID1_MASK |
2194 BTRFS_BLOCK_GROUP_RAID56_MASK |
2195 BTRFS_BLOCK_GROUP_DUP)))
2198 * Avoid allocating from un-mirrored block group if there are
2199 * mirrored block groups.
2201 list_for_each_entry(cache,
2202 &space_info->block_groups[BTRFS_RAID_RAID0],
2204 inc_block_group_ro(cache, 1);
2205 list_for_each_entry(cache,
2206 &space_info->block_groups[BTRFS_RAID_SINGLE],
2208 inc_block_group_ro(cache, 1);
2211 btrfs_init_global_block_rsv(info);
2212 ret = check_chunk_block_group_mappings(info);
2214 btrfs_free_path(path);
2219 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2222 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2225 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2226 struct btrfs_block_group *block_group)
2228 struct btrfs_fs_info *fs_info = trans->fs_info;
2229 struct btrfs_block_group_item bgi;
2230 struct btrfs_root *root;
2231 struct btrfs_key key;
2233 spin_lock(&block_group->lock);
2234 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2235 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2236 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2237 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2238 key.objectid = block_group->start;
2239 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2240 key.offset = block_group->length;
2241 spin_unlock(&block_group->lock);
2243 root = fs_info->extent_root;
2244 return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2248 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2251 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2254 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2256 struct btrfs_fs_info *fs_info = trans->fs_info;
2257 struct btrfs_block_group *block_group;
2260 while (!list_empty(&trans->new_bgs)) {
2263 block_group = list_first_entry(&trans->new_bgs,
2264 struct btrfs_block_group,
2269 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2271 ret = insert_block_group_item(trans, block_group);
2273 btrfs_abort_transaction(trans, ret);
2274 if (!block_group->chunk_item_inserted) {
2275 mutex_lock(&fs_info->chunk_mutex);
2276 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2277 mutex_unlock(&fs_info->chunk_mutex);
2279 btrfs_abort_transaction(trans, ret);
2281 ret = btrfs_finish_chunk_alloc(trans, block_group->start,
2282 block_group->length);
2284 btrfs_abort_transaction(trans, ret);
2285 add_block_group_free_space(trans, block_group);
2288 * If we restriped during balance, we may have added a new raid
2289 * type, so now add the sysfs entries when it is safe to do so.
2290 * We don't have to worry about locking here as it's handled in
2291 * btrfs_sysfs_add_block_group_type.
2293 if (block_group->space_info->block_group_kobjs[index] == NULL)
2294 btrfs_sysfs_add_block_group_type(block_group);
2296 /* Already aborted the transaction if it failed. */
2298 btrfs_delayed_refs_rsv_release(fs_info, 1);
2299 list_del_init(&block_group->bg_list);
2301 btrfs_trans_release_chunk_metadata(trans);
2304 struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2305 u64 bytes_used, u64 type,
2306 u64 chunk_offset, u64 size)
2308 struct btrfs_fs_info *fs_info = trans->fs_info;
2309 struct btrfs_block_group *cache;
2312 btrfs_set_log_full_commit(trans);
2314 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2316 return ERR_PTR(-ENOMEM);
2318 cache->length = size;
2319 set_free_space_tree_thresholds(cache);
2320 cache->used = bytes_used;
2321 cache->flags = type;
2322 cache->last_byte_to_unpin = (u64)-1;
2323 cache->cached = BTRFS_CACHE_FINISHED;
2324 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2325 cache->needs_free_space = 1;
2327 ret = btrfs_load_block_group_zone_info(cache, true);
2329 btrfs_put_block_group(cache);
2330 return ERR_PTR(ret);
2333 ret = exclude_super_stripes(cache);
2335 /* We may have excluded something, so call this just in case */
2336 btrfs_free_excluded_extents(cache);
2337 btrfs_put_block_group(cache);
2338 return ERR_PTR(ret);
2341 add_new_free_space(cache, chunk_offset, chunk_offset + size);
2343 btrfs_free_excluded_extents(cache);
2345 #ifdef CONFIG_BTRFS_DEBUG
2346 if (btrfs_should_fragment_free_space(cache)) {
2347 u64 new_bytes_used = size - bytes_used;
2349 bytes_used += new_bytes_used >> 1;
2350 fragment_free_space(cache);
2354 * Ensure the corresponding space_info object is created and
2355 * assigned to our block group. We want our bg to be added to the rbtree
2356 * with its ->space_info set.
2358 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2359 ASSERT(cache->space_info);
2361 ret = btrfs_add_block_group_cache(fs_info, cache);
2363 btrfs_remove_free_space_cache(cache);
2364 btrfs_put_block_group(cache);
2365 return ERR_PTR(ret);
2369 * Now that our block group has its ->space_info set and is inserted in
2370 * the rbtree, update the space info's counters.
2372 trace_btrfs_add_block_group(fs_info, cache, 1);
2373 btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
2374 cache->bytes_super, 0, &cache->space_info);
2375 btrfs_update_global_block_rsv(fs_info);
2377 link_block_group(cache);
2379 list_add_tail(&cache->bg_list, &trans->new_bgs);
2380 trans->delayed_ref_updates++;
2381 btrfs_update_delayed_refs_rsv(trans);
2383 set_avail_alloc_bits(fs_info, type);
2388 * Mark one block group RO, can be called several times for the same block
2391 * @cache: the destination block group
2392 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2393 * ensure we still have some free space after marking this
2396 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2397 bool do_chunk_alloc)
2399 struct btrfs_fs_info *fs_info = cache->fs_info;
2400 struct btrfs_trans_handle *trans;
2403 bool dirty_bg_running;
2406 trans = btrfs_join_transaction(fs_info->extent_root);
2408 return PTR_ERR(trans);
2410 dirty_bg_running = false;
2413 * We're not allowed to set block groups readonly after the dirty
2414 * block group cache has started writing. If it already started,
2415 * back off and let this transaction commit.
2417 mutex_lock(&fs_info->ro_block_group_mutex);
2418 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2419 u64 transid = trans->transid;
2421 mutex_unlock(&fs_info->ro_block_group_mutex);
2422 btrfs_end_transaction(trans);
2424 ret = btrfs_wait_for_commit(fs_info, transid);
2427 dirty_bg_running = true;
2429 } while (dirty_bg_running);
2431 if (do_chunk_alloc) {
2433 * If we are changing raid levels, try to allocate a
2434 * corresponding block group with the new raid level.
2436 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2437 if (alloc_flags != cache->flags) {
2438 ret = btrfs_chunk_alloc(trans, alloc_flags,
2441 * ENOSPC is allowed here, we may have enough space
2442 * already allocated at the new raid level to carry on
2451 ret = inc_block_group_ro(cache, 0);
2452 if (!do_chunk_alloc || ret == -ETXTBSY)
2456 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2457 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2460 ret = inc_block_group_ro(cache, 0);
2461 if (ret == -ETXTBSY)
2464 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2465 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2466 mutex_lock(&fs_info->chunk_mutex);
2467 check_system_chunk(trans, alloc_flags);
2468 mutex_unlock(&fs_info->chunk_mutex);
2471 mutex_unlock(&fs_info->ro_block_group_mutex);
2473 btrfs_end_transaction(trans);
2477 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2479 struct btrfs_space_info *sinfo = cache->space_info;
2484 spin_lock(&sinfo->lock);
2485 spin_lock(&cache->lock);
2487 if (btrfs_is_zoned(cache->fs_info)) {
2488 /* Migrate zone_unusable bytes back */
2489 cache->zone_unusable = cache->alloc_offset - cache->used;
2490 sinfo->bytes_zone_unusable += cache->zone_unusable;
2491 sinfo->bytes_readonly -= cache->zone_unusable;
2493 num_bytes = cache->length - cache->reserved -
2494 cache->pinned - cache->bytes_super -
2495 cache->zone_unusable - cache->used;
2496 sinfo->bytes_readonly -= num_bytes;
2497 list_del_init(&cache->ro_list);
2499 spin_unlock(&cache->lock);
2500 spin_unlock(&sinfo->lock);
2503 static int update_block_group_item(struct btrfs_trans_handle *trans,
2504 struct btrfs_path *path,
2505 struct btrfs_block_group *cache)
2507 struct btrfs_fs_info *fs_info = trans->fs_info;
2509 struct btrfs_root *root = fs_info->extent_root;
2511 struct extent_buffer *leaf;
2512 struct btrfs_block_group_item bgi;
2513 struct btrfs_key key;
2515 key.objectid = cache->start;
2516 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2517 key.offset = cache->length;
2519 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2526 leaf = path->nodes[0];
2527 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2528 btrfs_set_stack_block_group_used(&bgi, cache->used);
2529 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2530 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2531 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2532 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2533 btrfs_mark_buffer_dirty(leaf);
2535 btrfs_release_path(path);
2540 static int cache_save_setup(struct btrfs_block_group *block_group,
2541 struct btrfs_trans_handle *trans,
2542 struct btrfs_path *path)
2544 struct btrfs_fs_info *fs_info = block_group->fs_info;
2545 struct btrfs_root *root = fs_info->tree_root;
2546 struct inode *inode = NULL;
2547 struct extent_changeset *data_reserved = NULL;
2549 int dcs = BTRFS_DC_ERROR;
2554 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2558 * If this block group is smaller than 100 megs don't bother caching the
2561 if (block_group->length < (100 * SZ_1M)) {
2562 spin_lock(&block_group->lock);
2563 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2564 spin_unlock(&block_group->lock);
2568 if (TRANS_ABORTED(trans))
2571 inode = lookup_free_space_inode(block_group, path);
2572 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2573 ret = PTR_ERR(inode);
2574 btrfs_release_path(path);
2578 if (IS_ERR(inode)) {
2582 if (block_group->ro)
2585 ret = create_free_space_inode(trans, block_group, path);
2592 * We want to set the generation to 0, that way if anything goes wrong
2593 * from here on out we know not to trust this cache when we load up next
2596 BTRFS_I(inode)->generation = 0;
2597 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2600 * So theoretically we could recover from this, simply set the
2601 * super cache generation to 0 so we know to invalidate the
2602 * cache, but then we'd have to keep track of the block groups
2603 * that fail this way so we know we _have_ to reset this cache
2604 * before the next commit or risk reading stale cache. So to
2605 * limit our exposure to horrible edge cases lets just abort the
2606 * transaction, this only happens in really bad situations
2609 btrfs_abort_transaction(trans, ret);
2614 /* We've already setup this transaction, go ahead and exit */
2615 if (block_group->cache_generation == trans->transid &&
2616 i_size_read(inode)) {
2617 dcs = BTRFS_DC_SETUP;
2621 if (i_size_read(inode) > 0) {
2622 ret = btrfs_check_trunc_cache_free_space(fs_info,
2623 &fs_info->global_block_rsv);
2627 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2632 spin_lock(&block_group->lock);
2633 if (block_group->cached != BTRFS_CACHE_FINISHED ||
2634 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2636 * don't bother trying to write stuff out _if_
2637 * a) we're not cached,
2638 * b) we're with nospace_cache mount option,
2639 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2641 dcs = BTRFS_DC_WRITTEN;
2642 spin_unlock(&block_group->lock);
2645 spin_unlock(&block_group->lock);
2648 * We hit an ENOSPC when setting up the cache in this transaction, just
2649 * skip doing the setup, we've already cleared the cache so we're safe.
2651 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2657 * Try to preallocate enough space based on how big the block group is.
2658 * Keep in mind this has to include any pinned space which could end up
2659 * taking up quite a bit since it's not folded into the other space
2662 cache_size = div_u64(block_group->length, SZ_256M);
2667 cache_size *= fs_info->sectorsize;
2669 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2674 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
2675 cache_size, cache_size,
2678 * Our cache requires contiguous chunks so that we don't modify a bunch
2679 * of metadata or split extents when writing the cache out, which means
2680 * we can enospc if we are heavily fragmented in addition to just normal
2681 * out of space conditions. So if we hit this just skip setting up any
2682 * other block groups for this transaction, maybe we'll unpin enough
2683 * space the next time around.
2686 dcs = BTRFS_DC_SETUP;
2687 else if (ret == -ENOSPC)
2688 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2693 btrfs_release_path(path);
2695 spin_lock(&block_group->lock);
2696 if (!ret && dcs == BTRFS_DC_SETUP)
2697 block_group->cache_generation = trans->transid;
2698 block_group->disk_cache_state = dcs;
2699 spin_unlock(&block_group->lock);
2701 extent_changeset_free(data_reserved);
2705 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2707 struct btrfs_fs_info *fs_info = trans->fs_info;
2708 struct btrfs_block_group *cache, *tmp;
2709 struct btrfs_transaction *cur_trans = trans->transaction;
2710 struct btrfs_path *path;
2712 if (list_empty(&cur_trans->dirty_bgs) ||
2713 !btrfs_test_opt(fs_info, SPACE_CACHE))
2716 path = btrfs_alloc_path();
2720 /* Could add new block groups, use _safe just in case */
2721 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
2723 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2724 cache_save_setup(cache, trans, path);
2727 btrfs_free_path(path);
2732 * Transaction commit does final block group cache writeback during a critical
2733 * section where nothing is allowed to change the FS. This is required in
2734 * order for the cache to actually match the block group, but can introduce a
2735 * lot of latency into the commit.
2737 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2738 * There's a chance we'll have to redo some of it if the block group changes
2739 * again during the commit, but it greatly reduces the commit latency by
2740 * getting rid of the easy block groups while we're still allowing others to
2743 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
2745 struct btrfs_fs_info *fs_info = trans->fs_info;
2746 struct btrfs_block_group *cache;
2747 struct btrfs_transaction *cur_trans = trans->transaction;
2750 struct btrfs_path *path = NULL;
2752 struct list_head *io = &cur_trans->io_bgs;
2753 int num_started = 0;
2756 spin_lock(&cur_trans->dirty_bgs_lock);
2757 if (list_empty(&cur_trans->dirty_bgs)) {
2758 spin_unlock(&cur_trans->dirty_bgs_lock);
2761 list_splice_init(&cur_trans->dirty_bgs, &dirty);
2762 spin_unlock(&cur_trans->dirty_bgs_lock);
2765 /* Make sure all the block groups on our dirty list actually exist */
2766 btrfs_create_pending_block_groups(trans);
2769 path = btrfs_alloc_path();
2777 * cache_write_mutex is here only to save us from balance or automatic
2778 * removal of empty block groups deleting this block group while we are
2779 * writing out the cache
2781 mutex_lock(&trans->transaction->cache_write_mutex);
2782 while (!list_empty(&dirty)) {
2783 bool drop_reserve = true;
2785 cache = list_first_entry(&dirty, struct btrfs_block_group,
2788 * This can happen if something re-dirties a block group that
2789 * is already under IO. Just wait for it to finish and then do
2792 if (!list_empty(&cache->io_list)) {
2793 list_del_init(&cache->io_list);
2794 btrfs_wait_cache_io(trans, cache, path);
2795 btrfs_put_block_group(cache);
2800 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2801 * it should update the cache_state. Don't delete until after
2804 * Since we're not running in the commit critical section
2805 * we need the dirty_bgs_lock to protect from update_block_group
2807 spin_lock(&cur_trans->dirty_bgs_lock);
2808 list_del_init(&cache->dirty_list);
2809 spin_unlock(&cur_trans->dirty_bgs_lock);
2813 cache_save_setup(cache, trans, path);
2815 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
2816 cache->io_ctl.inode = NULL;
2817 ret = btrfs_write_out_cache(trans, cache, path);
2818 if (ret == 0 && cache->io_ctl.inode) {
2823 * The cache_write_mutex is protecting the
2824 * io_list, also refer to the definition of
2825 * btrfs_transaction::io_bgs for more details
2827 list_add_tail(&cache->io_list, io);
2830 * If we failed to write the cache, the
2831 * generation will be bad and life goes on
2837 ret = update_block_group_item(trans, path, cache);
2839 * Our block group might still be attached to the list
2840 * of new block groups in the transaction handle of some
2841 * other task (struct btrfs_trans_handle->new_bgs). This
2842 * means its block group item isn't yet in the extent
2843 * tree. If this happens ignore the error, as we will
2844 * try again later in the critical section of the
2845 * transaction commit.
2847 if (ret == -ENOENT) {
2849 spin_lock(&cur_trans->dirty_bgs_lock);
2850 if (list_empty(&cache->dirty_list)) {
2851 list_add_tail(&cache->dirty_list,
2852 &cur_trans->dirty_bgs);
2853 btrfs_get_block_group(cache);
2854 drop_reserve = false;
2856 spin_unlock(&cur_trans->dirty_bgs_lock);
2858 btrfs_abort_transaction(trans, ret);
2862 /* If it's not on the io list, we need to put the block group */
2864 btrfs_put_block_group(cache);
2866 btrfs_delayed_refs_rsv_release(fs_info, 1);
2868 * Avoid blocking other tasks for too long. It might even save
2869 * us from writing caches for block groups that are going to be
2872 mutex_unlock(&trans->transaction->cache_write_mutex);
2875 mutex_lock(&trans->transaction->cache_write_mutex);
2877 mutex_unlock(&trans->transaction->cache_write_mutex);
2880 * Go through delayed refs for all the stuff we've just kicked off
2881 * and then loop back (just once)
2884 ret = btrfs_run_delayed_refs(trans, 0);
2885 if (!ret && loops == 0) {
2887 spin_lock(&cur_trans->dirty_bgs_lock);
2888 list_splice_init(&cur_trans->dirty_bgs, &dirty);
2890 * dirty_bgs_lock protects us from concurrent block group
2891 * deletes too (not just cache_write_mutex).
2893 if (!list_empty(&dirty)) {
2894 spin_unlock(&cur_trans->dirty_bgs_lock);
2897 spin_unlock(&cur_trans->dirty_bgs_lock);
2901 spin_lock(&cur_trans->dirty_bgs_lock);
2902 list_splice_init(&dirty, &cur_trans->dirty_bgs);
2903 spin_unlock(&cur_trans->dirty_bgs_lock);
2904 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
2907 btrfs_free_path(path);
2911 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
2913 struct btrfs_fs_info *fs_info = trans->fs_info;
2914 struct btrfs_block_group *cache;
2915 struct btrfs_transaction *cur_trans = trans->transaction;
2918 struct btrfs_path *path;
2919 struct list_head *io = &cur_trans->io_bgs;
2920 int num_started = 0;
2922 path = btrfs_alloc_path();
2927 * Even though we are in the critical section of the transaction commit,
2928 * we can still have concurrent tasks adding elements to this
2929 * transaction's list of dirty block groups. These tasks correspond to
2930 * endio free space workers started when writeback finishes for a
2931 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
2932 * allocate new block groups as a result of COWing nodes of the root
2933 * tree when updating the free space inode. The writeback for the space
2934 * caches is triggered by an earlier call to
2935 * btrfs_start_dirty_block_groups() and iterations of the following
2937 * Also we want to do the cache_save_setup first and then run the
2938 * delayed refs to make sure we have the best chance at doing this all
2941 spin_lock(&cur_trans->dirty_bgs_lock);
2942 while (!list_empty(&cur_trans->dirty_bgs)) {
2943 cache = list_first_entry(&cur_trans->dirty_bgs,
2944 struct btrfs_block_group,
2948 * This can happen if cache_save_setup re-dirties a block group
2949 * that is already under IO. Just wait for it to finish and
2950 * then do it all again
2952 if (!list_empty(&cache->io_list)) {
2953 spin_unlock(&cur_trans->dirty_bgs_lock);
2954 list_del_init(&cache->io_list);
2955 btrfs_wait_cache_io(trans, cache, path);
2956 btrfs_put_block_group(cache);
2957 spin_lock(&cur_trans->dirty_bgs_lock);
2961 * Don't remove from the dirty list until after we've waited on
2964 list_del_init(&cache->dirty_list);
2965 spin_unlock(&cur_trans->dirty_bgs_lock);
2968 cache_save_setup(cache, trans, path);
2971 ret = btrfs_run_delayed_refs(trans,
2972 (unsigned long) -1);
2974 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
2975 cache->io_ctl.inode = NULL;
2976 ret = btrfs_write_out_cache(trans, cache, path);
2977 if (ret == 0 && cache->io_ctl.inode) {
2980 list_add_tail(&cache->io_list, io);
2983 * If we failed to write the cache, the
2984 * generation will be bad and life goes on
2990 ret = update_block_group_item(trans, path, cache);
2992 * One of the free space endio workers might have
2993 * created a new block group while updating a free space
2994 * cache's inode (at inode.c:btrfs_finish_ordered_io())
2995 * and hasn't released its transaction handle yet, in
2996 * which case the new block group is still attached to
2997 * its transaction handle and its creation has not
2998 * finished yet (no block group item in the extent tree
2999 * yet, etc). If this is the case, wait for all free
3000 * space endio workers to finish and retry. This is a
3001 * very rare case so no need for a more efficient and
3004 if (ret == -ENOENT) {
3005 wait_event(cur_trans->writer_wait,
3006 atomic_read(&cur_trans->num_writers) == 1);
3007 ret = update_block_group_item(trans, path, cache);
3010 btrfs_abort_transaction(trans, ret);
3013 /* If its not on the io list, we need to put the block group */
3015 btrfs_put_block_group(cache);
3016 btrfs_delayed_refs_rsv_release(fs_info, 1);
3017 spin_lock(&cur_trans->dirty_bgs_lock);
3019 spin_unlock(&cur_trans->dirty_bgs_lock);
3022 * Refer to the definition of io_bgs member for details why it's safe
3023 * to use it without any locking
3025 while (!list_empty(io)) {
3026 cache = list_first_entry(io, struct btrfs_block_group,
3028 list_del_init(&cache->io_list);
3029 btrfs_wait_cache_io(trans, cache, path);
3030 btrfs_put_block_group(cache);
3033 btrfs_free_path(path);
3037 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3038 u64 bytenr, u64 num_bytes, int alloc)
3040 struct btrfs_fs_info *info = trans->fs_info;
3041 struct btrfs_block_group *cache = NULL;
3042 u64 total = num_bytes;
3048 /* Block accounting for super block */
3049 spin_lock(&info->delalloc_root_lock);
3050 old_val = btrfs_super_bytes_used(info->super_copy);
3052 old_val += num_bytes;
3054 old_val -= num_bytes;
3055 btrfs_set_super_bytes_used(info->super_copy, old_val);
3056 spin_unlock(&info->delalloc_root_lock);
3059 cache = btrfs_lookup_block_group(info, bytenr);
3064 factor = btrfs_bg_type_to_factor(cache->flags);
3067 * If this block group has free space cache written out, we
3068 * need to make sure to load it if we are removing space. This
3069 * is because we need the unpinning stage to actually add the
3070 * space back to the block group, otherwise we will leak space.
3072 if (!alloc && !btrfs_block_group_done(cache))
3073 btrfs_cache_block_group(cache, 1);
3075 byte_in_group = bytenr - cache->start;
3076 WARN_ON(byte_in_group > cache->length);
3078 spin_lock(&cache->space_info->lock);
3079 spin_lock(&cache->lock);
3081 if (btrfs_test_opt(info, SPACE_CACHE) &&
3082 cache->disk_cache_state < BTRFS_DC_CLEAR)
3083 cache->disk_cache_state = BTRFS_DC_CLEAR;
3085 old_val = cache->used;
3086 num_bytes = min(total, cache->length - byte_in_group);
3088 old_val += num_bytes;
3089 cache->used = old_val;
3090 cache->reserved -= num_bytes;
3091 cache->space_info->bytes_reserved -= num_bytes;
3092 cache->space_info->bytes_used += num_bytes;
3093 cache->space_info->disk_used += num_bytes * factor;
3094 spin_unlock(&cache->lock);
3095 spin_unlock(&cache->space_info->lock);
3097 old_val -= num_bytes;
3098 cache->used = old_val;
3099 cache->pinned += num_bytes;
3100 btrfs_space_info_update_bytes_pinned(info,
3101 cache->space_info, num_bytes);
3102 cache->space_info->bytes_used -= num_bytes;
3103 cache->space_info->disk_used -= num_bytes * factor;
3104 spin_unlock(&cache->lock);
3105 spin_unlock(&cache->space_info->lock);
3107 set_extent_dirty(&trans->transaction->pinned_extents,
3108 bytenr, bytenr + num_bytes - 1,
3109 GFP_NOFS | __GFP_NOFAIL);
3112 spin_lock(&trans->transaction->dirty_bgs_lock);
3113 if (list_empty(&cache->dirty_list)) {
3114 list_add_tail(&cache->dirty_list,
3115 &trans->transaction->dirty_bgs);
3116 trans->delayed_ref_updates++;
3117 btrfs_get_block_group(cache);
3119 spin_unlock(&trans->transaction->dirty_bgs_lock);
3122 * No longer have used bytes in this block group, queue it for
3123 * deletion. We do this after adding the block group to the
3124 * dirty list to avoid races between cleaner kthread and space
3127 if (!alloc && old_val == 0) {
3128 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3129 btrfs_mark_bg_unused(cache);
3132 btrfs_put_block_group(cache);
3134 bytenr += num_bytes;
3137 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3138 btrfs_update_delayed_refs_rsv(trans);
3143 * btrfs_add_reserved_bytes - update the block_group and space info counters
3144 * @cache: The cache we are manipulating
3145 * @ram_bytes: The number of bytes of file content, and will be same to
3146 * @num_bytes except for the compress path.
3147 * @num_bytes: The number of bytes in question
3148 * @delalloc: The blocks are allocated for the delalloc write
3150 * This is called by the allocator when it reserves space. If this is a
3151 * reservation and the block group has become read only we cannot make the
3152 * reservation and return -EAGAIN, otherwise this function always succeeds.
3154 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3155 u64 ram_bytes, u64 num_bytes, int delalloc)
3157 struct btrfs_space_info *space_info = cache->space_info;
3160 spin_lock(&space_info->lock);
3161 spin_lock(&cache->lock);
3165 cache->reserved += num_bytes;
3166 space_info->bytes_reserved += num_bytes;
3167 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3168 space_info->flags, num_bytes, 1);
3169 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3170 space_info, -ram_bytes);
3172 cache->delalloc_bytes += num_bytes;
3175 * Compression can use less space than we reserved, so wake
3176 * tickets if that happens
3178 if (num_bytes < ram_bytes)
3179 btrfs_try_granting_tickets(cache->fs_info, space_info);
3181 spin_unlock(&cache->lock);
3182 spin_unlock(&space_info->lock);
3187 * btrfs_free_reserved_bytes - update the block_group and space info counters
3188 * @cache: The cache we are manipulating
3189 * @num_bytes: The number of bytes in question
3190 * @delalloc: The blocks are allocated for the delalloc write
3192 * This is called by somebody who is freeing space that was never actually used
3193 * on disk. For example if you reserve some space for a new leaf in transaction
3194 * A and before transaction A commits you free that leaf, you call this with
3195 * reserve set to 0 in order to clear the reservation.
3197 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3198 u64 num_bytes, int delalloc)
3200 struct btrfs_space_info *space_info = cache->space_info;
3202 spin_lock(&space_info->lock);
3203 spin_lock(&cache->lock);
3205 space_info->bytes_readonly += num_bytes;
3206 cache->reserved -= num_bytes;
3207 space_info->bytes_reserved -= num_bytes;
3208 space_info->max_extent_size = 0;
3211 cache->delalloc_bytes -= num_bytes;
3212 spin_unlock(&cache->lock);
3214 btrfs_try_granting_tickets(cache->fs_info, space_info);
3215 spin_unlock(&space_info->lock);
3218 static void force_metadata_allocation(struct btrfs_fs_info *info)
3220 struct list_head *head = &info->space_info;
3221 struct btrfs_space_info *found;
3223 list_for_each_entry(found, head, list) {
3224 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3225 found->force_alloc = CHUNK_ALLOC_FORCE;
3229 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3230 struct btrfs_space_info *sinfo, int force)
3232 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3235 if (force == CHUNK_ALLOC_FORCE)
3239 * in limited mode, we want to have some free space up to
3240 * about 1% of the FS size.
3242 if (force == CHUNK_ALLOC_LIMITED) {
3243 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3244 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3246 if (sinfo->total_bytes - bytes_used < thresh)
3250 if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3255 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3257 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3259 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3262 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3264 struct btrfs_block_group *bg;
3268 * Check if we have enough space in the system space info because we
3269 * will need to update device items in the chunk btree and insert a new
3270 * chunk item in the chunk btree as well. This will allocate a new
3271 * system block group if needed.
3273 check_system_chunk(trans, flags);
3275 bg = btrfs_alloc_chunk(trans, flags);
3282 * If this is a system chunk allocation then stop right here and do not
3283 * add the chunk item to the chunk btree. This is to prevent a deadlock
3284 * because this system chunk allocation can be triggered while COWing
3285 * some extent buffer of the chunk btree and while holding a lock on a
3286 * parent extent buffer, in which case attempting to insert the chunk
3287 * item (or update the device item) would result in a deadlock on that
3288 * parent extent buffer. In this case defer the chunk btree updates to
3289 * the second phase of chunk allocation and keep our reservation until
3290 * the second phase completes.
3292 * This is a rare case and can only be triggered by the very few cases
3293 * we have where we need to touch the chunk btree outside chunk allocation
3294 * and chunk removal. These cases are basically adding a device, removing
3295 * a device or resizing a device.
3297 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3300 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3302 * Normally we are not expected to fail with -ENOSPC here, since we have
3303 * previously reserved space in the system space_info and allocated one
3304 * new system chunk if necessary. However there are two exceptions:
3306 * 1) We may have enough free space in the system space_info but all the
3307 * existing system block groups have a profile which can not be used
3308 * for extent allocation.
3310 * This happens when mounting in degraded mode. For example we have a
3311 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3312 * using the other device in degraded mode. If we then allocate a chunk,
3313 * we may have enough free space in the existing system space_info, but
3314 * none of the block groups can be used for extent allocation since they
3315 * have a RAID1 profile, and because we are in degraded mode with a
3316 * single device, we are forced to allocate a new system chunk with a
3317 * SINGLE profile. Making check_system_chunk() iterate over all system
3318 * block groups and check if they have a usable profile and enough space
3319 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3320 * try again after forcing allocation of a new system chunk. Like this
3321 * we avoid paying the cost of that search in normal circumstances, when
3322 * we were not mounted in degraded mode;
3324 * 2) We had enough free space info the system space_info, and one suitable
3325 * block group to allocate from when we called check_system_chunk()
3326 * above. However right after we called it, the only system block group
3327 * with enough free space got turned into RO mode by a running scrub,
3328 * and in this case we have to allocate a new one and retry. We only
3329 * need do this allocate and retry once, since we have a transaction
3330 * handle and scrub uses the commit root to search for block groups.
3332 if (ret == -ENOSPC) {
3333 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3334 struct btrfs_block_group *sys_bg;
3336 sys_bg = btrfs_alloc_chunk(trans, sys_flags);
3337 if (IS_ERR(sys_bg)) {
3338 ret = PTR_ERR(sys_bg);
3339 btrfs_abort_transaction(trans, ret);
3343 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3345 btrfs_abort_transaction(trans, ret);
3349 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3351 btrfs_abort_transaction(trans, ret);
3355 btrfs_abort_transaction(trans, ret);
3359 btrfs_trans_release_chunk_metadata(trans);
3365 * Chunk allocation is done in 2 phases:
3367 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3368 * the chunk, the chunk mapping, create its block group and add the items
3369 * that belong in the chunk btree to it - more specifically, we need to
3370 * update device items in the chunk btree and add a new chunk item to it.
3372 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3373 * group item to the extent btree and the device extent items to the devices
3376 * This is done to prevent deadlocks. For example when COWing a node from the
3377 * extent btree we are holding a write lock on the node's parent and if we
3378 * trigger chunk allocation and attempted to insert the new block group item
3379 * in the extent btree right way, we could deadlock because the path for the
3380 * insertion can include that parent node. At first glance it seems impossible
3381 * to trigger chunk allocation after starting a transaction since tasks should
3382 * reserve enough transaction units (metadata space), however while that is true
3383 * most of the time, chunk allocation may still be triggered for several reasons:
3385 * 1) When reserving metadata, we check if there is enough free space in the
3386 * metadata space_info and therefore don't trigger allocation of a new chunk.
3387 * However later when the task actually tries to COW an extent buffer from
3388 * the extent btree or from the device btree for example, it is forced to
3389 * allocate a new block group (chunk) because the only one that had enough
3390 * free space was just turned to RO mode by a running scrub for example (or
3391 * device replace, block group reclaim thread, etc), so we can not use it
3392 * for allocating an extent and end up being forced to allocate a new one;
3394 * 2) Because we only check that the metadata space_info has enough free bytes,
3395 * we end up not allocating a new metadata chunk in that case. However if
3396 * the filesystem was mounted in degraded mode, none of the existing block
3397 * groups might be suitable for extent allocation due to their incompatible
3398 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3399 * use a RAID1 profile, in degraded mode using a single device). In this case
3400 * when the task attempts to COW some extent buffer of the extent btree for
3401 * example, it will trigger allocation of a new metadata block group with a
3402 * suitable profile (SINGLE profile in the example of the degraded mount of
3403 * the RAID1 filesystem);
3405 * 3) The task has reserved enough transaction units / metadata space, but when
3406 * it attempts to COW an extent buffer from the extent or device btree for
3407 * example, it does not find any free extent in any metadata block group,
3408 * therefore forced to try to allocate a new metadata block group.
3409 * This is because some other task allocated all available extents in the
3410 * meanwhile - this typically happens with tasks that don't reserve space
3411 * properly, either intentionally or as a bug. One example where this is
3412 * done intentionally is fsync, as it does not reserve any transaction units
3413 * and ends up allocating a variable number of metadata extents for log
3414 * tree extent buffers.
3416 * We also need this 2 phases setup when adding a device to a filesystem with
3417 * a seed device - we must create new metadata and system chunks without adding
3418 * any of the block group items to the chunk, extent and device btrees. If we
3419 * did not do it this way, we would get ENOSPC when attempting to update those
3420 * btrees, since all the chunks from the seed device are read-only.
3422 * Phase 1 does the updates and insertions to the chunk btree because if we had
3423 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3424 * parallel, we risk having too many system chunks allocated by many tasks if
3425 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3426 * extreme case this leads to exhaustion of the system chunk array in the
3427 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3428 * and with RAID filesystems (so we have more device items in the chunk btree).
3429 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3430 * the system chunk array due to concurrent allocations") provides more details.
3432 * For allocation of system chunks, we defer the updates and insertions into the
3433 * chunk btree to phase 2. This is to prevent deadlocks on extent buffers because
3434 * if the chunk allocation is triggered while COWing an extent buffer of the
3435 * chunk btree, we are holding a lock on the parent of that extent buffer and
3436 * doing the chunk btree updates and insertions can require locking that parent.
3437 * This is for the very few and rare cases where we update the chunk btree that
3438 * are not chunk allocation or chunk removal: adding a device, removing a device
3439 * or resizing a device.
3441 * The reservation of system space, done through check_system_chunk(), as well
3442 * as all the updates and insertions into the chunk btree must be done while
3443 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
3444 * an extent buffer from the chunks btree we never trigger allocation of a new
3445 * system chunk, which would result in a deadlock (trying to lock twice an
3446 * extent buffer of the chunk btree, first time before triggering the chunk
3447 * allocation and the second time during chunk allocation while attempting to
3448 * update the chunks btree). The system chunk array is also updated while holding
3449 * that mutex. The same logic applies to removing chunks - we must reserve system
3450 * space, update the chunk btree and the system chunk array in the superblock
3451 * while holding fs_info->chunk_mutex.
3453 * This function, btrfs_chunk_alloc(), belongs to phase 1.
3455 * If @force is CHUNK_ALLOC_FORCE:
3456 * - return 1 if it successfully allocates a chunk,
3457 * - return errors including -ENOSPC otherwise.
3458 * If @force is NOT CHUNK_ALLOC_FORCE:
3459 * - return 0 if it doesn't need to allocate a new chunk,
3460 * - return 1 if it successfully allocates a chunk,
3461 * - return errors including -ENOSPC otherwise.
3463 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3464 enum btrfs_chunk_alloc_enum force)
3466 struct btrfs_fs_info *fs_info = trans->fs_info;
3467 struct btrfs_space_info *space_info;
3468 bool wait_for_alloc = false;
3469 bool should_alloc = false;
3472 /* Don't re-enter if we're already allocating a chunk */
3473 if (trans->allocating_chunk)
3476 * If we are removing a chunk, don't re-enter or we would deadlock.
3477 * System space reservation and system chunk allocation is done by the
3478 * chunk remove operation (btrfs_remove_chunk()).
3480 if (trans->removing_chunk)
3483 space_info = btrfs_find_space_info(fs_info, flags);
3487 spin_lock(&space_info->lock);
3488 if (force < space_info->force_alloc)
3489 force = space_info->force_alloc;
3490 should_alloc = should_alloc_chunk(fs_info, space_info, force);
3491 if (space_info->full) {
3492 /* No more free physical space */
3497 spin_unlock(&space_info->lock);
3499 } else if (!should_alloc) {
3500 spin_unlock(&space_info->lock);
3502 } else if (space_info->chunk_alloc) {
3504 * Someone is already allocating, so we need to block
3505 * until this someone is finished and then loop to
3506 * recheck if we should continue with our allocation
3509 wait_for_alloc = true;
3510 spin_unlock(&space_info->lock);
3511 mutex_lock(&fs_info->chunk_mutex);
3512 mutex_unlock(&fs_info->chunk_mutex);
3514 /* Proceed with allocation */
3515 space_info->chunk_alloc = 1;
3516 wait_for_alloc = false;
3517 spin_unlock(&space_info->lock);
3521 } while (wait_for_alloc);
3523 mutex_lock(&fs_info->chunk_mutex);
3524 trans->allocating_chunk = true;
3527 * If we have mixed data/metadata chunks we want to make sure we keep
3528 * allocating mixed chunks instead of individual chunks.
3530 if (btrfs_mixed_space_info(space_info))
3531 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3534 * if we're doing a data chunk, go ahead and make sure that
3535 * we keep a reasonable number of metadata chunks allocated in the
3538 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3539 fs_info->data_chunk_allocations++;
3540 if (!(fs_info->data_chunk_allocations %
3541 fs_info->metadata_ratio))
3542 force_metadata_allocation(fs_info);
3545 ret = do_chunk_alloc(trans, flags);
3546 trans->allocating_chunk = false;
3548 spin_lock(&space_info->lock);
3551 space_info->full = 1;
3556 space_info->max_extent_size = 0;
3559 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3561 space_info->chunk_alloc = 0;
3562 spin_unlock(&space_info->lock);
3563 mutex_unlock(&fs_info->chunk_mutex);
3568 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3572 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3574 num_dev = fs_info->fs_devices->rw_devices;
3580 * Reserve space in the system space for allocating or removing a chunk
3582 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3584 struct btrfs_fs_info *fs_info = trans->fs_info;
3585 struct btrfs_space_info *info;
3592 * Needed because we can end up allocating a system chunk and for an
3593 * atomic and race free space reservation in the chunk block reserve.
3595 lockdep_assert_held(&fs_info->chunk_mutex);
3597 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3598 spin_lock(&info->lock);
3599 left = info->total_bytes - btrfs_space_info_used(info, true);
3600 spin_unlock(&info->lock);
3602 num_devs = get_profile_num_devs(fs_info, type);
3604 /* num_devs device items to update and 1 chunk item to add or remove */
3605 thresh = btrfs_calc_metadata_size(fs_info, num_devs) +
3606 btrfs_calc_insert_metadata_size(fs_info, 1);
3608 if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3609 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3610 left, thresh, type);
3611 btrfs_dump_space_info(fs_info, info, 0, 0);
3614 if (left < thresh) {
3615 u64 flags = btrfs_system_alloc_profile(fs_info);
3616 struct btrfs_block_group *bg;
3619 * Ignore failure to create system chunk. We might end up not
3620 * needing it, as we might not need to COW all nodes/leafs from
3621 * the paths we visit in the chunk tree (they were already COWed
3622 * or created in the current transaction for example).
3624 * Also, if our caller is allocating a system chunk, do not
3625 * attempt to insert the chunk item in the chunk btree, as we
3626 * could deadlock on an extent buffer since our caller may be
3627 * COWing an extent buffer from the chunk btree.
3629 bg = btrfs_alloc_chunk(trans, flags);
3632 } else if (!(type & BTRFS_BLOCK_GROUP_SYSTEM)) {
3634 * If we fail to add the chunk item here, we end up
3635 * trying again at phase 2 of chunk allocation, at
3636 * btrfs_create_pending_block_groups(). So ignore
3639 btrfs_chunk_alloc_add_chunk_item(trans, bg);
3644 ret = btrfs_block_rsv_add(fs_info->chunk_root,
3645 &fs_info->chunk_block_rsv,
3646 thresh, BTRFS_RESERVE_NO_FLUSH);
3648 trans->chunk_bytes_reserved += thresh;
3652 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
3654 struct btrfs_block_group *block_group;
3658 struct inode *inode;
3660 block_group = btrfs_lookup_first_block_group(info, last);
3661 while (block_group) {
3662 btrfs_wait_block_group_cache_done(block_group);
3663 spin_lock(&block_group->lock);
3664 if (block_group->iref)
3666 spin_unlock(&block_group->lock);
3667 block_group = btrfs_next_block_group(block_group);
3676 inode = block_group->inode;
3677 block_group->iref = 0;
3678 block_group->inode = NULL;
3679 spin_unlock(&block_group->lock);
3680 ASSERT(block_group->io_ctl.inode == NULL);
3682 last = block_group->start + block_group->length;
3683 btrfs_put_block_group(block_group);
3688 * Must be called only after stopping all workers, since we could have block
3689 * group caching kthreads running, and therefore they could race with us if we
3690 * freed the block groups before stopping them.
3692 int btrfs_free_block_groups(struct btrfs_fs_info *info)
3694 struct btrfs_block_group *block_group;
3695 struct btrfs_space_info *space_info;
3696 struct btrfs_caching_control *caching_ctl;
3699 spin_lock(&info->block_group_cache_lock);
3700 while (!list_empty(&info->caching_block_groups)) {
3701 caching_ctl = list_entry(info->caching_block_groups.next,
3702 struct btrfs_caching_control, list);
3703 list_del(&caching_ctl->list);
3704 btrfs_put_caching_control(caching_ctl);
3706 spin_unlock(&info->block_group_cache_lock);
3708 spin_lock(&info->unused_bgs_lock);
3709 while (!list_empty(&info->unused_bgs)) {
3710 block_group = list_first_entry(&info->unused_bgs,
3711 struct btrfs_block_group,
3713 list_del_init(&block_group->bg_list);
3714 btrfs_put_block_group(block_group);
3716 spin_unlock(&info->unused_bgs_lock);
3718 spin_lock(&info->unused_bgs_lock);
3719 while (!list_empty(&info->reclaim_bgs)) {
3720 block_group = list_first_entry(&info->reclaim_bgs,
3721 struct btrfs_block_group,
3723 list_del_init(&block_group->bg_list);
3724 btrfs_put_block_group(block_group);
3726 spin_unlock(&info->unused_bgs_lock);
3728 spin_lock(&info->block_group_cache_lock);
3729 while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
3730 block_group = rb_entry(n, struct btrfs_block_group,
3732 rb_erase(&block_group->cache_node,
3733 &info->block_group_cache_tree);
3734 RB_CLEAR_NODE(&block_group->cache_node);
3735 spin_unlock(&info->block_group_cache_lock);
3737 down_write(&block_group->space_info->groups_sem);
3738 list_del(&block_group->list);
3739 up_write(&block_group->space_info->groups_sem);
3742 * We haven't cached this block group, which means we could
3743 * possibly have excluded extents on this block group.
3745 if (block_group->cached == BTRFS_CACHE_NO ||
3746 block_group->cached == BTRFS_CACHE_ERROR)
3747 btrfs_free_excluded_extents(block_group);
3749 btrfs_remove_free_space_cache(block_group);
3750 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
3751 ASSERT(list_empty(&block_group->dirty_list));
3752 ASSERT(list_empty(&block_group->io_list));
3753 ASSERT(list_empty(&block_group->bg_list));
3754 ASSERT(refcount_read(&block_group->refs) == 1);
3755 ASSERT(block_group->swap_extents == 0);
3756 btrfs_put_block_group(block_group);
3758 spin_lock(&info->block_group_cache_lock);
3760 spin_unlock(&info->block_group_cache_lock);
3762 btrfs_release_global_block_rsv(info);
3764 while (!list_empty(&info->space_info)) {
3765 space_info = list_entry(info->space_info.next,
3766 struct btrfs_space_info,
3770 * Do not hide this behind enospc_debug, this is actually
3771 * important and indicates a real bug if this happens.
3773 if (WARN_ON(space_info->bytes_pinned > 0 ||
3774 space_info->bytes_reserved > 0 ||
3775 space_info->bytes_may_use > 0))
3776 btrfs_dump_space_info(info, space_info, 0, 0);
3777 WARN_ON(space_info->reclaim_size > 0);
3778 list_del(&space_info->list);
3779 btrfs_sysfs_remove_space_info(space_info);
3784 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
3786 atomic_inc(&cache->frozen);
3789 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
3791 struct btrfs_fs_info *fs_info = block_group->fs_info;
3792 struct extent_map_tree *em_tree;
3793 struct extent_map *em;
3796 spin_lock(&block_group->lock);
3797 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
3798 block_group->removed);
3799 spin_unlock(&block_group->lock);
3802 em_tree = &fs_info->mapping_tree;
3803 write_lock(&em_tree->lock);
3804 em = lookup_extent_mapping(em_tree, block_group->start,
3806 BUG_ON(!em); /* logic error, can't happen */
3807 remove_extent_mapping(em_tree, em);
3808 write_unlock(&em_tree->lock);
3810 /* once for us and once for the tree */
3811 free_extent_map(em);
3812 free_extent_map(em);
3815 * We may have left one free space entry and other possible
3816 * tasks trimming this block group have left 1 entry each one.
3819 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3823 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
3827 spin_lock(&bg->lock);
3832 spin_unlock(&bg->lock);
3837 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
3839 spin_lock(&bg->lock);
3841 ASSERT(bg->swap_extents >= amount);
3842 bg->swap_extents -= amount;
3843 spin_unlock(&bg->lock);