1 // SPDX-License-Identifier: GPL-2.0
3 #include <linux/list_sort.h>
6 #include "block-group.h"
7 #include "space-info.h"
9 #include "free-space-cache.h"
10 #include "free-space-tree.h"
12 #include "transaction.h"
13 #include "ref-verify.h"
16 #include "delalloc-space.h"
21 #include "accessors.h"
22 #include "extent-tree.h"
24 #ifdef CONFIG_BTRFS_DEBUG
25 int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group)
27 struct btrfs_fs_info *fs_info = block_group->fs_info;
29 return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
30 block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
31 (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
32 block_group->flags & BTRFS_BLOCK_GROUP_DATA);
37 * Return target flags in extended format or 0 if restripe for this chunk_type
40 * Should be called with balance_lock held
42 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
44 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
50 if (flags & BTRFS_BLOCK_GROUP_DATA &&
51 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
52 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
53 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
54 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
55 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
56 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
57 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
58 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
65 * @flags: available profiles in extended format (see ctree.h)
67 * Return reduced profile in chunk format. If profile changing is in progress
68 * (either running or paused) picks the target profile (if it's already
69 * available), otherwise falls back to plain reducing.
71 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
73 u64 num_devices = fs_info->fs_devices->rw_devices;
79 * See if restripe for this chunk_type is in progress, if so try to
80 * reduce to the target profile
82 spin_lock(&fs_info->balance_lock);
83 target = get_restripe_target(fs_info, flags);
85 spin_unlock(&fs_info->balance_lock);
86 return extended_to_chunk(target);
88 spin_unlock(&fs_info->balance_lock);
90 /* First, mask out the RAID levels which aren't possible */
91 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
92 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
93 allowed |= btrfs_raid_array[raid_type].bg_flag;
97 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
98 allowed = BTRFS_BLOCK_GROUP_RAID6;
99 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
100 allowed = BTRFS_BLOCK_GROUP_RAID5;
101 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
102 allowed = BTRFS_BLOCK_GROUP_RAID10;
103 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
104 allowed = BTRFS_BLOCK_GROUP_RAID1;
105 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
106 allowed = BTRFS_BLOCK_GROUP_RAID0;
108 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
110 return extended_to_chunk(flags | allowed);
113 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
120 seq = read_seqbegin(&fs_info->profiles_lock);
122 if (flags & BTRFS_BLOCK_GROUP_DATA)
123 flags |= fs_info->avail_data_alloc_bits;
124 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
125 flags |= fs_info->avail_system_alloc_bits;
126 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
127 flags |= fs_info->avail_metadata_alloc_bits;
128 } while (read_seqretry(&fs_info->profiles_lock, seq));
130 return btrfs_reduce_alloc_profile(fs_info, flags);
133 void btrfs_get_block_group(struct btrfs_block_group *cache)
135 refcount_inc(&cache->refs);
138 void btrfs_put_block_group(struct btrfs_block_group *cache)
140 if (refcount_dec_and_test(&cache->refs)) {
141 WARN_ON(cache->pinned > 0);
143 * If there was a failure to cleanup a log tree, very likely due
144 * to an IO failure on a writeback attempt of one or more of its
145 * extent buffers, we could not do proper (and cheap) unaccounting
146 * of their reserved space, so don't warn on reserved > 0 in that
149 if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
150 !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
151 WARN_ON(cache->reserved > 0);
154 * A block_group shouldn't be on the discard_list anymore.
155 * Remove the block_group from the discard_list to prevent us
156 * from causing a panic due to NULL pointer dereference.
158 if (WARN_ON(!list_empty(&cache->discard_list)))
159 btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
163 * If not empty, someone is still holding mutex of
164 * full_stripe_lock, which can only be released by caller.
165 * And it will definitely cause use-after-free when caller
166 * tries to release full stripe lock.
168 * No better way to resolve, but only to warn.
170 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
171 kfree(cache->free_space_ctl);
172 kfree(cache->physical_map);
178 * This adds the block group to the fs_info rb tree for the block group cache
180 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
181 struct btrfs_block_group *block_group)
184 struct rb_node *parent = NULL;
185 struct btrfs_block_group *cache;
186 bool leftmost = true;
188 ASSERT(block_group->length != 0);
190 write_lock(&info->block_group_cache_lock);
191 p = &info->block_group_cache_tree.rb_root.rb_node;
195 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
196 if (block_group->start < cache->start) {
198 } else if (block_group->start > cache->start) {
202 write_unlock(&info->block_group_cache_lock);
207 rb_link_node(&block_group->cache_node, parent, p);
208 rb_insert_color_cached(&block_group->cache_node,
209 &info->block_group_cache_tree, leftmost);
211 write_unlock(&info->block_group_cache_lock);
217 * This will return the block group at or after bytenr if contains is 0, else
218 * it will return the block group that contains the bytenr
220 static struct btrfs_block_group *block_group_cache_tree_search(
221 struct btrfs_fs_info *info, u64 bytenr, int contains)
223 struct btrfs_block_group *cache, *ret = NULL;
227 read_lock(&info->block_group_cache_lock);
228 n = info->block_group_cache_tree.rb_root.rb_node;
231 cache = rb_entry(n, struct btrfs_block_group, cache_node);
232 end = cache->start + cache->length - 1;
233 start = cache->start;
235 if (bytenr < start) {
236 if (!contains && (!ret || start < ret->start))
239 } else if (bytenr > start) {
240 if (contains && bytenr <= end) {
251 btrfs_get_block_group(ret);
252 read_unlock(&info->block_group_cache_lock);
258 * Return the block group that starts at or after bytenr
260 struct btrfs_block_group *btrfs_lookup_first_block_group(
261 struct btrfs_fs_info *info, u64 bytenr)
263 return block_group_cache_tree_search(info, bytenr, 0);
267 * Return the block group that contains the given bytenr
269 struct btrfs_block_group *btrfs_lookup_block_group(
270 struct btrfs_fs_info *info, u64 bytenr)
272 return block_group_cache_tree_search(info, bytenr, 1);
275 struct btrfs_block_group *btrfs_next_block_group(
276 struct btrfs_block_group *cache)
278 struct btrfs_fs_info *fs_info = cache->fs_info;
279 struct rb_node *node;
281 read_lock(&fs_info->block_group_cache_lock);
283 /* If our block group was removed, we need a full search. */
284 if (RB_EMPTY_NODE(&cache->cache_node)) {
285 const u64 next_bytenr = cache->start + cache->length;
287 read_unlock(&fs_info->block_group_cache_lock);
288 btrfs_put_block_group(cache);
289 return btrfs_lookup_first_block_group(fs_info, next_bytenr);
291 node = rb_next(&cache->cache_node);
292 btrfs_put_block_group(cache);
294 cache = rb_entry(node, struct btrfs_block_group, cache_node);
295 btrfs_get_block_group(cache);
298 read_unlock(&fs_info->block_group_cache_lock);
303 * Check if we can do a NOCOW write for a given extent.
305 * @fs_info: The filesystem information object.
306 * @bytenr: Logical start address of the extent.
308 * Check if we can do a NOCOW write for the given extent, and increments the
309 * number of NOCOW writers in the block group that contains the extent, as long
310 * as the block group exists and it's currently not in read-only mode.
312 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
313 * is responsible for calling btrfs_dec_nocow_writers() later.
315 * Or NULL if we can not do a NOCOW write
317 struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
320 struct btrfs_block_group *bg;
321 bool can_nocow = true;
323 bg = btrfs_lookup_block_group(fs_info, bytenr);
327 spin_lock(&bg->lock);
331 atomic_inc(&bg->nocow_writers);
332 spin_unlock(&bg->lock);
335 btrfs_put_block_group(bg);
339 /* No put on block group, done by btrfs_dec_nocow_writers(). */
344 * Decrement the number of NOCOW writers in a block group.
346 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
347 * and on the block group returned by that call. Typically this is called after
348 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
351 * After this call, the caller should not use the block group anymore. It it wants
352 * to use it, then it should get a reference on it before calling this function.
354 void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
356 if (atomic_dec_and_test(&bg->nocow_writers))
357 wake_up_var(&bg->nocow_writers);
359 /* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
360 btrfs_put_block_group(bg);
363 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
365 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
368 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
371 struct btrfs_block_group *bg;
373 bg = btrfs_lookup_block_group(fs_info, start);
375 if (atomic_dec_and_test(&bg->reservations))
376 wake_up_var(&bg->reservations);
377 btrfs_put_block_group(bg);
380 void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
382 struct btrfs_space_info *space_info = bg->space_info;
386 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
390 * Our block group is read only but before we set it to read only,
391 * some task might have had allocated an extent from it already, but it
392 * has not yet created a respective ordered extent (and added it to a
393 * root's list of ordered extents).
394 * Therefore wait for any task currently allocating extents, since the
395 * block group's reservations counter is incremented while a read lock
396 * on the groups' semaphore is held and decremented after releasing
397 * the read access on that semaphore and creating the ordered extent.
399 down_write(&space_info->groups_sem);
400 up_write(&space_info->groups_sem);
402 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
405 struct btrfs_caching_control *btrfs_get_caching_control(
406 struct btrfs_block_group *cache)
408 struct btrfs_caching_control *ctl;
410 spin_lock(&cache->lock);
411 if (!cache->caching_ctl) {
412 spin_unlock(&cache->lock);
416 ctl = cache->caching_ctl;
417 refcount_inc(&ctl->count);
418 spin_unlock(&cache->lock);
422 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
424 if (refcount_dec_and_test(&ctl->count))
429 * When we wait for progress in the block group caching, its because our
430 * allocation attempt failed at least once. So, we must sleep and let some
431 * progress happen before we try again.
433 * This function will sleep at least once waiting for new free space to show
434 * up, and then it will check the block group free space numbers for our min
435 * num_bytes. Another option is to have it go ahead and look in the rbtree for
436 * a free extent of a given size, but this is a good start.
438 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
439 * any of the information in this block group.
441 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
444 struct btrfs_caching_control *caching_ctl;
446 caching_ctl = btrfs_get_caching_control(cache);
450 wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
451 (cache->free_space_ctl->free_space >= num_bytes));
453 btrfs_put_caching_control(caching_ctl);
456 static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
457 struct btrfs_caching_control *caching_ctl)
459 wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
460 return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
463 static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
465 struct btrfs_caching_control *caching_ctl;
468 caching_ctl = btrfs_get_caching_control(cache);
470 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
471 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
472 btrfs_put_caching_control(caching_ctl);
476 #ifdef CONFIG_BTRFS_DEBUG
477 static void fragment_free_space(struct btrfs_block_group *block_group)
479 struct btrfs_fs_info *fs_info = block_group->fs_info;
480 u64 start = block_group->start;
481 u64 len = block_group->length;
482 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
483 fs_info->nodesize : fs_info->sectorsize;
484 u64 step = chunk << 1;
486 while (len > chunk) {
487 btrfs_remove_free_space(block_group, start, chunk);
498 * This is only called by btrfs_cache_block_group, since we could have freed
499 * extents we need to check the pinned_extents for any extents that can't be
500 * used yet since their free space will be released as soon as the transaction
503 u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
505 struct btrfs_fs_info *info = block_group->fs_info;
506 u64 extent_start, extent_end, size, total_added = 0;
509 while (start < end) {
510 ret = find_first_extent_bit(&info->excluded_extents, start,
511 &extent_start, &extent_end,
512 EXTENT_DIRTY | EXTENT_UPTODATE,
517 if (extent_start <= start) {
518 start = extent_end + 1;
519 } else if (extent_start > start && extent_start < end) {
520 size = extent_start - start;
522 ret = btrfs_add_free_space_async_trimmed(block_group,
524 BUG_ON(ret); /* -ENOMEM or logic error */
525 start = extent_end + 1;
534 ret = btrfs_add_free_space_async_trimmed(block_group, start,
536 BUG_ON(ret); /* -ENOMEM or logic error */
542 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
544 struct btrfs_block_group *block_group = caching_ctl->block_group;
545 struct btrfs_fs_info *fs_info = block_group->fs_info;
546 struct btrfs_root *extent_root;
547 struct btrfs_path *path;
548 struct extent_buffer *leaf;
549 struct btrfs_key key;
556 path = btrfs_alloc_path();
560 last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
561 extent_root = btrfs_extent_root(fs_info, last);
563 #ifdef CONFIG_BTRFS_DEBUG
565 * If we're fragmenting we don't want to make anybody think we can
566 * allocate from this block group until we've had a chance to fragment
569 if (btrfs_should_fragment_free_space(block_group))
573 * We don't want to deadlock with somebody trying to allocate a new
574 * extent for the extent root while also trying to search the extent
575 * root to add free space. So we skip locking and search the commit
576 * root, since its read-only
578 path->skip_locking = 1;
579 path->search_commit_root = 1;
580 path->reada = READA_FORWARD;
584 key.type = BTRFS_EXTENT_ITEM_KEY;
587 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
591 leaf = path->nodes[0];
592 nritems = btrfs_header_nritems(leaf);
595 if (btrfs_fs_closing(fs_info) > 1) {
600 if (path->slots[0] < nritems) {
601 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
603 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
607 if (need_resched() ||
608 rwsem_is_contended(&fs_info->commit_root_sem)) {
609 btrfs_release_path(path);
610 up_read(&fs_info->commit_root_sem);
611 mutex_unlock(&caching_ctl->mutex);
613 mutex_lock(&caching_ctl->mutex);
614 down_read(&fs_info->commit_root_sem);
618 ret = btrfs_next_leaf(extent_root, path);
623 leaf = path->nodes[0];
624 nritems = btrfs_header_nritems(leaf);
628 if (key.objectid < last) {
631 key.type = BTRFS_EXTENT_ITEM_KEY;
632 btrfs_release_path(path);
636 if (key.objectid < block_group->start) {
641 if (key.objectid >= block_group->start + block_group->length)
644 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
645 key.type == BTRFS_METADATA_ITEM_KEY) {
646 total_found += add_new_free_space(block_group, last,
648 if (key.type == BTRFS_METADATA_ITEM_KEY)
649 last = key.objectid +
652 last = key.objectid + key.offset;
654 if (total_found > CACHING_CTL_WAKE_UP) {
657 wake_up(&caching_ctl->wait);
664 total_found += add_new_free_space(block_group, last,
665 block_group->start + block_group->length);
668 btrfs_free_path(path);
672 static noinline void caching_thread(struct btrfs_work *work)
674 struct btrfs_block_group *block_group;
675 struct btrfs_fs_info *fs_info;
676 struct btrfs_caching_control *caching_ctl;
679 caching_ctl = container_of(work, struct btrfs_caching_control, work);
680 block_group = caching_ctl->block_group;
681 fs_info = block_group->fs_info;
683 mutex_lock(&caching_ctl->mutex);
684 down_read(&fs_info->commit_root_sem);
686 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
687 ret = load_free_space_cache(block_group);
694 * We failed to load the space cache, set ourselves to
695 * CACHE_STARTED and carry on.
697 spin_lock(&block_group->lock);
698 block_group->cached = BTRFS_CACHE_STARTED;
699 spin_unlock(&block_group->lock);
700 wake_up(&caching_ctl->wait);
704 * If we are in the transaction that populated the free space tree we
705 * can't actually cache from the free space tree as our commit root and
706 * real root are the same, so we could change the contents of the blocks
707 * while caching. Instead do the slow caching in this case, and after
708 * the transaction has committed we will be safe.
710 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
711 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
712 ret = load_free_space_tree(caching_ctl);
714 ret = load_extent_tree_free(caching_ctl);
716 spin_lock(&block_group->lock);
717 block_group->caching_ctl = NULL;
718 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
719 spin_unlock(&block_group->lock);
721 #ifdef CONFIG_BTRFS_DEBUG
722 if (btrfs_should_fragment_free_space(block_group)) {
725 spin_lock(&block_group->space_info->lock);
726 spin_lock(&block_group->lock);
727 bytes_used = block_group->length - block_group->used;
728 block_group->space_info->bytes_used += bytes_used >> 1;
729 spin_unlock(&block_group->lock);
730 spin_unlock(&block_group->space_info->lock);
731 fragment_free_space(block_group);
735 up_read(&fs_info->commit_root_sem);
736 btrfs_free_excluded_extents(block_group);
737 mutex_unlock(&caching_ctl->mutex);
739 wake_up(&caching_ctl->wait);
741 btrfs_put_caching_control(caching_ctl);
742 btrfs_put_block_group(block_group);
745 int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
747 struct btrfs_fs_info *fs_info = cache->fs_info;
748 struct btrfs_caching_control *caching_ctl = NULL;
751 /* Allocator for zoned filesystems does not use the cache at all */
752 if (btrfs_is_zoned(fs_info))
755 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
759 INIT_LIST_HEAD(&caching_ctl->list);
760 mutex_init(&caching_ctl->mutex);
761 init_waitqueue_head(&caching_ctl->wait);
762 caching_ctl->block_group = cache;
763 refcount_set(&caching_ctl->count, 2);
764 btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
766 spin_lock(&cache->lock);
767 if (cache->cached != BTRFS_CACHE_NO) {
770 caching_ctl = cache->caching_ctl;
772 refcount_inc(&caching_ctl->count);
773 spin_unlock(&cache->lock);
776 WARN_ON(cache->caching_ctl);
777 cache->caching_ctl = caching_ctl;
778 cache->cached = BTRFS_CACHE_STARTED;
779 spin_unlock(&cache->lock);
781 write_lock(&fs_info->block_group_cache_lock);
782 refcount_inc(&caching_ctl->count);
783 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
784 write_unlock(&fs_info->block_group_cache_lock);
786 btrfs_get_block_group(cache);
788 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
790 if (wait && caching_ctl)
791 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
793 btrfs_put_caching_control(caching_ctl);
798 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
800 u64 extra_flags = chunk_to_extended(flags) &
801 BTRFS_EXTENDED_PROFILE_MASK;
803 write_seqlock(&fs_info->profiles_lock);
804 if (flags & BTRFS_BLOCK_GROUP_DATA)
805 fs_info->avail_data_alloc_bits &= ~extra_flags;
806 if (flags & BTRFS_BLOCK_GROUP_METADATA)
807 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
808 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
809 fs_info->avail_system_alloc_bits &= ~extra_flags;
810 write_sequnlock(&fs_info->profiles_lock);
814 * Clear incompat bits for the following feature(s):
816 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
817 * in the whole filesystem
819 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
821 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
823 bool found_raid56 = false;
824 bool found_raid1c34 = false;
826 if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
827 (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
828 (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
829 struct list_head *head = &fs_info->space_info;
830 struct btrfs_space_info *sinfo;
832 list_for_each_entry_rcu(sinfo, head, list) {
833 down_read(&sinfo->groups_sem);
834 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
836 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
838 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
839 found_raid1c34 = true;
840 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
841 found_raid1c34 = true;
842 up_read(&sinfo->groups_sem);
845 btrfs_clear_fs_incompat(fs_info, RAID56);
847 btrfs_clear_fs_incompat(fs_info, RAID1C34);
851 static int remove_block_group_item(struct btrfs_trans_handle *trans,
852 struct btrfs_path *path,
853 struct btrfs_block_group *block_group)
855 struct btrfs_fs_info *fs_info = trans->fs_info;
856 struct btrfs_root *root;
857 struct btrfs_key key;
860 root = btrfs_block_group_root(fs_info);
861 key.objectid = block_group->start;
862 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
863 key.offset = block_group->length;
865 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
871 ret = btrfs_del_item(trans, root, path);
875 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
876 u64 group_start, struct extent_map *em)
878 struct btrfs_fs_info *fs_info = trans->fs_info;
879 struct btrfs_path *path;
880 struct btrfs_block_group *block_group;
881 struct btrfs_free_cluster *cluster;
883 struct kobject *kobj = NULL;
887 struct btrfs_caching_control *caching_ctl = NULL;
889 bool remove_rsv = false;
891 block_group = btrfs_lookup_block_group(fs_info, group_start);
892 BUG_ON(!block_group);
893 BUG_ON(!block_group->ro);
895 trace_btrfs_remove_block_group(block_group);
897 * Free the reserved super bytes from this block group before
900 btrfs_free_excluded_extents(block_group);
901 btrfs_free_ref_tree_range(fs_info, block_group->start,
902 block_group->length);
904 index = btrfs_bg_flags_to_raid_index(block_group->flags);
905 factor = btrfs_bg_type_to_factor(block_group->flags);
907 /* make sure this block group isn't part of an allocation cluster */
908 cluster = &fs_info->data_alloc_cluster;
909 spin_lock(&cluster->refill_lock);
910 btrfs_return_cluster_to_free_space(block_group, cluster);
911 spin_unlock(&cluster->refill_lock);
914 * make sure this block group isn't part of a metadata
917 cluster = &fs_info->meta_alloc_cluster;
918 spin_lock(&cluster->refill_lock);
919 btrfs_return_cluster_to_free_space(block_group, cluster);
920 spin_unlock(&cluster->refill_lock);
922 btrfs_clear_treelog_bg(block_group);
923 btrfs_clear_data_reloc_bg(block_group);
925 path = btrfs_alloc_path();
932 * get the inode first so any iput calls done for the io_list
933 * aren't the final iput (no unlinks allowed now)
935 inode = lookup_free_space_inode(block_group, path);
937 mutex_lock(&trans->transaction->cache_write_mutex);
939 * Make sure our free space cache IO is done before removing the
942 spin_lock(&trans->transaction->dirty_bgs_lock);
943 if (!list_empty(&block_group->io_list)) {
944 list_del_init(&block_group->io_list);
946 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
948 spin_unlock(&trans->transaction->dirty_bgs_lock);
949 btrfs_wait_cache_io(trans, block_group, path);
950 btrfs_put_block_group(block_group);
951 spin_lock(&trans->transaction->dirty_bgs_lock);
954 if (!list_empty(&block_group->dirty_list)) {
955 list_del_init(&block_group->dirty_list);
957 btrfs_put_block_group(block_group);
959 spin_unlock(&trans->transaction->dirty_bgs_lock);
960 mutex_unlock(&trans->transaction->cache_write_mutex);
962 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
966 write_lock(&fs_info->block_group_cache_lock);
967 rb_erase_cached(&block_group->cache_node,
968 &fs_info->block_group_cache_tree);
969 RB_CLEAR_NODE(&block_group->cache_node);
971 /* Once for the block groups rbtree */
972 btrfs_put_block_group(block_group);
974 write_unlock(&fs_info->block_group_cache_lock);
976 down_write(&block_group->space_info->groups_sem);
978 * we must use list_del_init so people can check to see if they
979 * are still on the list after taking the semaphore
981 list_del_init(&block_group->list);
982 if (list_empty(&block_group->space_info->block_groups[index])) {
983 kobj = block_group->space_info->block_group_kobjs[index];
984 block_group->space_info->block_group_kobjs[index] = NULL;
985 clear_avail_alloc_bits(fs_info, block_group->flags);
987 up_write(&block_group->space_info->groups_sem);
988 clear_incompat_bg_bits(fs_info, block_group->flags);
994 if (block_group->cached == BTRFS_CACHE_STARTED)
995 btrfs_wait_block_group_cache_done(block_group);
997 write_lock(&fs_info->block_group_cache_lock);
998 caching_ctl = btrfs_get_caching_control(block_group);
1000 struct btrfs_caching_control *ctl;
1002 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1003 if (ctl->block_group == block_group) {
1005 refcount_inc(&caching_ctl->count);
1011 list_del_init(&caching_ctl->list);
1012 write_unlock(&fs_info->block_group_cache_lock);
1015 /* Once for the caching bgs list and once for us. */
1016 btrfs_put_caching_control(caching_ctl);
1017 btrfs_put_caching_control(caching_ctl);
1020 spin_lock(&trans->transaction->dirty_bgs_lock);
1021 WARN_ON(!list_empty(&block_group->dirty_list));
1022 WARN_ON(!list_empty(&block_group->io_list));
1023 spin_unlock(&trans->transaction->dirty_bgs_lock);
1025 btrfs_remove_free_space_cache(block_group);
1027 spin_lock(&block_group->space_info->lock);
1028 list_del_init(&block_group->ro_list);
1030 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1031 WARN_ON(block_group->space_info->total_bytes
1032 < block_group->length);
1033 WARN_ON(block_group->space_info->bytes_readonly
1034 < block_group->length - block_group->zone_unusable);
1035 WARN_ON(block_group->space_info->bytes_zone_unusable
1036 < block_group->zone_unusable);
1037 WARN_ON(block_group->space_info->disk_total
1038 < block_group->length * factor);
1039 WARN_ON(test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
1040 &block_group->runtime_flags) &&
1041 block_group->space_info->active_total_bytes
1042 < block_group->length);
1044 block_group->space_info->total_bytes -= block_group->length;
1045 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
1046 block_group->space_info->active_total_bytes -= block_group->length;
1047 block_group->space_info->bytes_readonly -=
1048 (block_group->length - block_group->zone_unusable);
1049 block_group->space_info->bytes_zone_unusable -=
1050 block_group->zone_unusable;
1051 block_group->space_info->disk_total -= block_group->length * factor;
1053 spin_unlock(&block_group->space_info->lock);
1056 * Remove the free space for the block group from the free space tree
1057 * and the block group's item from the extent tree before marking the
1058 * block group as removed. This is to prevent races with tasks that
1059 * freeze and unfreeze a block group, this task and another task
1060 * allocating a new block group - the unfreeze task ends up removing
1061 * the block group's extent map before the task calling this function
1062 * deletes the block group item from the extent tree, allowing for
1063 * another task to attempt to create another block group with the same
1064 * item key (and failing with -EEXIST and a transaction abort).
1066 ret = remove_block_group_free_space(trans, block_group);
1070 ret = remove_block_group_item(trans, path, block_group);
1074 spin_lock(&block_group->lock);
1075 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1078 * At this point trimming or scrub can't start on this block group,
1079 * because we removed the block group from the rbtree
1080 * fs_info->block_group_cache_tree so no one can't find it anymore and
1081 * even if someone already got this block group before we removed it
1082 * from the rbtree, they have already incremented block_group->frozen -
1083 * if they didn't, for the trimming case they won't find any free space
1084 * entries because we already removed them all when we called
1085 * btrfs_remove_free_space_cache().
1087 * And we must not remove the extent map from the fs_info->mapping_tree
1088 * to prevent the same logical address range and physical device space
1089 * ranges from being reused for a new block group. This is needed to
1090 * avoid races with trimming and scrub.
1092 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1093 * completely transactionless, so while it is trimming a range the
1094 * currently running transaction might finish and a new one start,
1095 * allowing for new block groups to be created that can reuse the same
1096 * physical device locations unless we take this special care.
1098 * There may also be an implicit trim operation if the file system
1099 * is mounted with -odiscard. The same protections must remain
1100 * in place until the extents have been discarded completely when
1101 * the transaction commit has completed.
1103 remove_em = (atomic_read(&block_group->frozen) == 0);
1104 spin_unlock(&block_group->lock);
1107 struct extent_map_tree *em_tree;
1109 em_tree = &fs_info->mapping_tree;
1110 write_lock(&em_tree->lock);
1111 remove_extent_mapping(em_tree, em);
1112 write_unlock(&em_tree->lock);
1113 /* once for the tree */
1114 free_extent_map(em);
1118 /* Once for the lookup reference */
1119 btrfs_put_block_group(block_group);
1121 btrfs_delayed_refs_rsv_release(fs_info, 1);
1122 btrfs_free_path(path);
1126 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1127 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1129 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1130 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1131 struct extent_map *em;
1132 struct map_lookup *map;
1133 unsigned int num_items;
1135 read_lock(&em_tree->lock);
1136 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1137 read_unlock(&em_tree->lock);
1138 ASSERT(em && em->start == chunk_offset);
1141 * We need to reserve 3 + N units from the metadata space info in order
1142 * to remove a block group (done at btrfs_remove_chunk() and at
1143 * btrfs_remove_block_group()), which are used for:
1145 * 1 unit for adding the free space inode's orphan (located in the tree
1147 * 1 unit for deleting the block group item (located in the extent
1149 * 1 unit for deleting the free space item (located in tree of tree
1151 * N units for deleting N device extent items corresponding to each
1152 * stripe (located in the device tree).
1154 * In order to remove a block group we also need to reserve units in the
1155 * system space info in order to update the chunk tree (update one or
1156 * more device items and remove one chunk item), but this is done at
1157 * btrfs_remove_chunk() through a call to check_system_chunk().
1159 map = em->map_lookup;
1160 num_items = 3 + map->num_stripes;
1161 free_extent_map(em);
1163 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1167 * Mark block group @cache read-only, so later write won't happen to block
1170 * If @force is not set, this function will only mark the block group readonly
1171 * if we have enough free space (1M) in other metadata/system block groups.
1172 * If @force is not set, this function will mark the block group readonly
1173 * without checking free space.
1175 * NOTE: This function doesn't care if other block groups can contain all the
1176 * data in this block group. That check should be done by relocation routine,
1177 * not this function.
1179 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1181 struct btrfs_space_info *sinfo = cache->space_info;
1185 spin_lock(&sinfo->lock);
1186 spin_lock(&cache->lock);
1188 if (cache->swap_extents) {
1199 num_bytes = cache->length - cache->reserved - cache->pinned -
1200 cache->bytes_super - cache->zone_unusable - cache->used;
1203 * Data never overcommits, even in mixed mode, so do just the straight
1204 * check of left over space in how much we have allocated.
1208 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1209 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1212 * Here we make sure if we mark this bg RO, we still have enough
1213 * free space as buffer.
1215 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1219 * We overcommit metadata, so we need to do the
1220 * btrfs_can_overcommit check here, and we need to pass in
1221 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1222 * leeway to allow us to mark this block group as read only.
1224 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1225 BTRFS_RESERVE_NO_FLUSH))
1230 sinfo->bytes_readonly += num_bytes;
1231 if (btrfs_is_zoned(cache->fs_info)) {
1232 /* Migrate zone_unusable bytes to readonly */
1233 sinfo->bytes_readonly += cache->zone_unusable;
1234 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1235 cache->zone_unusable = 0;
1238 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1241 spin_unlock(&cache->lock);
1242 spin_unlock(&sinfo->lock);
1243 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1244 btrfs_info(cache->fs_info,
1245 "unable to make block group %llu ro", cache->start);
1246 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1251 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1252 struct btrfs_block_group *bg)
1254 struct btrfs_fs_info *fs_info = bg->fs_info;
1255 struct btrfs_transaction *prev_trans = NULL;
1256 const u64 start = bg->start;
1257 const u64 end = start + bg->length - 1;
1260 spin_lock(&fs_info->trans_lock);
1261 if (trans->transaction->list.prev != &fs_info->trans_list) {
1262 prev_trans = list_last_entry(&trans->transaction->list,
1263 struct btrfs_transaction, list);
1264 refcount_inc(&prev_trans->use_count);
1266 spin_unlock(&fs_info->trans_lock);
1269 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1270 * btrfs_finish_extent_commit(). If we are at transaction N, another
1271 * task might be running finish_extent_commit() for the previous
1272 * transaction N - 1, and have seen a range belonging to the block
1273 * group in pinned_extents before we were able to clear the whole block
1274 * group range from pinned_extents. This means that task can lookup for
1275 * the block group after we unpinned it from pinned_extents and removed
1276 * it, leading to a BUG_ON() at unpin_extent_range().
1278 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1280 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1286 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1289 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1291 btrfs_put_transaction(prev_trans);
1297 * Process the unused_bgs list and remove any that don't have any allocated
1298 * space inside of them.
1300 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1302 struct btrfs_block_group *block_group;
1303 struct btrfs_space_info *space_info;
1304 struct btrfs_trans_handle *trans;
1305 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1308 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1311 if (btrfs_fs_closing(fs_info))
1315 * Long running balances can keep us blocked here for eternity, so
1316 * simply skip deletion if we're unable to get the mutex.
1318 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1321 spin_lock(&fs_info->unused_bgs_lock);
1322 while (!list_empty(&fs_info->unused_bgs)) {
1325 block_group = list_first_entry(&fs_info->unused_bgs,
1326 struct btrfs_block_group,
1328 list_del_init(&block_group->bg_list);
1330 space_info = block_group->space_info;
1332 if (ret || btrfs_mixed_space_info(space_info)) {
1333 btrfs_put_block_group(block_group);
1336 spin_unlock(&fs_info->unused_bgs_lock);
1338 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1340 /* Don't want to race with allocators so take the groups_sem */
1341 down_write(&space_info->groups_sem);
1344 * Async discard moves the final block group discard to be prior
1345 * to the unused_bgs code path. Therefore, if it's not fully
1346 * trimmed, punt it back to the async discard lists.
1348 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1349 !btrfs_is_free_space_trimmed(block_group)) {
1350 trace_btrfs_skip_unused_block_group(block_group);
1351 up_write(&space_info->groups_sem);
1352 /* Requeue if we failed because of async discard */
1353 btrfs_discard_queue_work(&fs_info->discard_ctl,
1358 spin_lock(&block_group->lock);
1359 if (block_group->reserved || block_group->pinned ||
1360 block_group->used || block_group->ro ||
1361 list_is_singular(&block_group->list)) {
1363 * We want to bail if we made new allocations or have
1364 * outstanding allocations in this block group. We do
1365 * the ro check in case balance is currently acting on
1368 trace_btrfs_skip_unused_block_group(block_group);
1369 spin_unlock(&block_group->lock);
1370 up_write(&space_info->groups_sem);
1373 spin_unlock(&block_group->lock);
1375 /* We don't want to force the issue, only flip if it's ok. */
1376 ret = inc_block_group_ro(block_group, 0);
1377 up_write(&space_info->groups_sem);
1383 ret = btrfs_zone_finish(block_group);
1385 btrfs_dec_block_group_ro(block_group);
1392 * Want to do this before we do anything else so we can recover
1393 * properly if we fail to join the transaction.
1395 trans = btrfs_start_trans_remove_block_group(fs_info,
1396 block_group->start);
1397 if (IS_ERR(trans)) {
1398 btrfs_dec_block_group_ro(block_group);
1399 ret = PTR_ERR(trans);
1404 * We could have pending pinned extents for this block group,
1405 * just delete them, we don't care about them anymore.
1407 if (!clean_pinned_extents(trans, block_group)) {
1408 btrfs_dec_block_group_ro(block_group);
1413 * At this point, the block_group is read only and should fail
1414 * new allocations. However, btrfs_finish_extent_commit() can
1415 * cause this block_group to be placed back on the discard
1416 * lists because now the block_group isn't fully discarded.
1417 * Bail here and try again later after discarding everything.
1419 spin_lock(&fs_info->discard_ctl.lock);
1420 if (!list_empty(&block_group->discard_list)) {
1421 spin_unlock(&fs_info->discard_ctl.lock);
1422 btrfs_dec_block_group_ro(block_group);
1423 btrfs_discard_queue_work(&fs_info->discard_ctl,
1427 spin_unlock(&fs_info->discard_ctl.lock);
1429 /* Reset pinned so btrfs_put_block_group doesn't complain */
1430 spin_lock(&space_info->lock);
1431 spin_lock(&block_group->lock);
1433 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1434 -block_group->pinned);
1435 space_info->bytes_readonly += block_group->pinned;
1436 block_group->pinned = 0;
1438 spin_unlock(&block_group->lock);
1439 spin_unlock(&space_info->lock);
1442 * The normal path here is an unused block group is passed here,
1443 * then trimming is handled in the transaction commit path.
1444 * Async discard interposes before this to do the trimming
1445 * before coming down the unused block group path as trimming
1446 * will no longer be done later in the transaction commit path.
1448 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1452 * DISCARD can flip during remount. On zoned filesystems, we
1453 * need to reset sequential-required zones.
1455 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1456 btrfs_is_zoned(fs_info);
1458 /* Implicit trim during transaction commit. */
1460 btrfs_freeze_block_group(block_group);
1463 * Btrfs_remove_chunk will abort the transaction if things go
1466 ret = btrfs_remove_chunk(trans, block_group->start);
1470 btrfs_unfreeze_block_group(block_group);
1475 * If we're not mounted with -odiscard, we can just forget
1476 * about this block group. Otherwise we'll need to wait
1477 * until transaction commit to do the actual discard.
1480 spin_lock(&fs_info->unused_bgs_lock);
1482 * A concurrent scrub might have added us to the list
1483 * fs_info->unused_bgs, so use a list_move operation
1484 * to add the block group to the deleted_bgs list.
1486 list_move(&block_group->bg_list,
1487 &trans->transaction->deleted_bgs);
1488 spin_unlock(&fs_info->unused_bgs_lock);
1489 btrfs_get_block_group(block_group);
1492 btrfs_end_transaction(trans);
1494 btrfs_put_block_group(block_group);
1495 spin_lock(&fs_info->unused_bgs_lock);
1497 spin_unlock(&fs_info->unused_bgs_lock);
1498 mutex_unlock(&fs_info->reclaim_bgs_lock);
1502 btrfs_end_transaction(trans);
1503 mutex_unlock(&fs_info->reclaim_bgs_lock);
1504 btrfs_put_block_group(block_group);
1505 btrfs_discard_punt_unused_bgs_list(fs_info);
1508 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1510 struct btrfs_fs_info *fs_info = bg->fs_info;
1512 spin_lock(&fs_info->unused_bgs_lock);
1513 if (list_empty(&bg->bg_list)) {
1514 btrfs_get_block_group(bg);
1515 trace_btrfs_add_unused_block_group(bg);
1516 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1518 spin_unlock(&fs_info->unused_bgs_lock);
1522 * We want block groups with a low number of used bytes to be in the beginning
1523 * of the list, so they will get reclaimed first.
1525 static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1526 const struct list_head *b)
1528 const struct btrfs_block_group *bg1, *bg2;
1530 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1531 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1533 return bg1->used > bg2->used;
1536 static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1538 if (btrfs_is_zoned(fs_info))
1539 return btrfs_zoned_should_reclaim(fs_info);
1543 static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1545 const struct btrfs_space_info *space_info = bg->space_info;
1546 const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1547 const u64 new_val = bg->used;
1548 const u64 old_val = new_val + bytes_freed;
1551 if (reclaim_thresh == 0)
1554 thresh = mult_perc(bg->length, reclaim_thresh);
1557 * If we were below the threshold before don't reclaim, we are likely a
1558 * brand new block group and we don't want to relocate new block groups.
1560 if (old_val < thresh)
1562 if (new_val >= thresh)
1567 void btrfs_reclaim_bgs_work(struct work_struct *work)
1569 struct btrfs_fs_info *fs_info =
1570 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1571 struct btrfs_block_group *bg;
1572 struct btrfs_space_info *space_info;
1574 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1577 if (btrfs_fs_closing(fs_info))
1580 if (!btrfs_should_reclaim(fs_info))
1583 sb_start_write(fs_info->sb);
1585 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1586 sb_end_write(fs_info->sb);
1591 * Long running balances can keep us blocked here for eternity, so
1592 * simply skip reclaim if we're unable to get the mutex.
1594 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1595 btrfs_exclop_finish(fs_info);
1596 sb_end_write(fs_info->sb);
1600 spin_lock(&fs_info->unused_bgs_lock);
1602 * Sort happens under lock because we can't simply splice it and sort.
1603 * The block groups might still be in use and reachable via bg_list,
1604 * and their presence in the reclaim_bgs list must be preserved.
1606 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1607 while (!list_empty(&fs_info->reclaim_bgs)) {
1611 bg = list_first_entry(&fs_info->reclaim_bgs,
1612 struct btrfs_block_group,
1614 list_del_init(&bg->bg_list);
1616 space_info = bg->space_info;
1617 spin_unlock(&fs_info->unused_bgs_lock);
1619 /* Don't race with allocators so take the groups_sem */
1620 down_write(&space_info->groups_sem);
1622 spin_lock(&bg->lock);
1623 if (bg->reserved || bg->pinned || bg->ro) {
1625 * We want to bail if we made new allocations or have
1626 * outstanding allocations in this block group. We do
1627 * the ro check in case balance is currently acting on
1630 spin_unlock(&bg->lock);
1631 up_write(&space_info->groups_sem);
1634 if (bg->used == 0) {
1636 * It is possible that we trigger relocation on a block
1637 * group as its extents are deleted and it first goes
1638 * below the threshold, then shortly after goes empty.
1640 * In this case, relocating it does delete it, but has
1641 * some overhead in relocation specific metadata, looking
1642 * for the non-existent extents and running some extra
1643 * transactions, which we can avoid by using one of the
1644 * other mechanisms for dealing with empty block groups.
1646 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1647 btrfs_mark_bg_unused(bg);
1648 spin_unlock(&bg->lock);
1649 up_write(&space_info->groups_sem);
1654 * The block group might no longer meet the reclaim condition by
1655 * the time we get around to reclaiming it, so to avoid
1656 * reclaiming overly full block_groups, skip reclaiming them.
1658 * Since the decision making process also depends on the amount
1659 * being freed, pass in a fake giant value to skip that extra
1660 * check, which is more meaningful when adding to the list in
1663 if (!should_reclaim_block_group(bg, bg->length)) {
1664 spin_unlock(&bg->lock);
1665 up_write(&space_info->groups_sem);
1668 spin_unlock(&bg->lock);
1670 /* Get out fast, in case we're unmounting the filesystem */
1671 if (btrfs_fs_closing(fs_info)) {
1672 up_write(&space_info->groups_sem);
1677 * Cache the zone_unusable value before turning the block group
1678 * to read only. As soon as the blog group is read only it's
1679 * zone_unusable value gets moved to the block group's read-only
1680 * bytes and isn't available for calculations anymore.
1682 zone_unusable = bg->zone_unusable;
1683 ret = inc_block_group_ro(bg, 0);
1684 up_write(&space_info->groups_sem);
1689 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1690 bg->start, div_u64(bg->used * 100, bg->length),
1691 div64_u64(zone_unusable * 100, bg->length));
1692 trace_btrfs_reclaim_block_group(bg);
1693 ret = btrfs_relocate_chunk(fs_info, bg->start);
1695 btrfs_dec_block_group_ro(bg);
1696 btrfs_err(fs_info, "error relocating chunk %llu",
1701 btrfs_put_block_group(bg);
1702 spin_lock(&fs_info->unused_bgs_lock);
1704 spin_unlock(&fs_info->unused_bgs_lock);
1705 mutex_unlock(&fs_info->reclaim_bgs_lock);
1706 btrfs_exclop_finish(fs_info);
1707 sb_end_write(fs_info->sb);
1710 void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1712 spin_lock(&fs_info->unused_bgs_lock);
1713 if (!list_empty(&fs_info->reclaim_bgs))
1714 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1715 spin_unlock(&fs_info->unused_bgs_lock);
1718 void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1720 struct btrfs_fs_info *fs_info = bg->fs_info;
1722 spin_lock(&fs_info->unused_bgs_lock);
1723 if (list_empty(&bg->bg_list)) {
1724 btrfs_get_block_group(bg);
1725 trace_btrfs_add_reclaim_block_group(bg);
1726 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1728 spin_unlock(&fs_info->unused_bgs_lock);
1731 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1732 struct btrfs_path *path)
1734 struct extent_map_tree *em_tree;
1735 struct extent_map *em;
1736 struct btrfs_block_group_item bg;
1737 struct extent_buffer *leaf;
1742 slot = path->slots[0];
1743 leaf = path->nodes[0];
1745 em_tree = &fs_info->mapping_tree;
1746 read_lock(&em_tree->lock);
1747 em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1748 read_unlock(&em_tree->lock);
1751 "logical %llu len %llu found bg but no related chunk",
1752 key->objectid, key->offset);
1756 if (em->start != key->objectid || em->len != key->offset) {
1758 "block group %llu len %llu mismatch with chunk %llu len %llu",
1759 key->objectid, key->offset, em->start, em->len);
1764 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1766 flags = btrfs_stack_block_group_flags(&bg) &
1767 BTRFS_BLOCK_GROUP_TYPE_MASK;
1769 if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1771 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1772 key->objectid, key->offset, flags,
1773 (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1778 free_extent_map(em);
1782 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1783 struct btrfs_path *path,
1784 struct btrfs_key *key)
1786 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1788 struct btrfs_key found_key;
1790 btrfs_for_each_slot(root, key, &found_key, path, ret) {
1791 if (found_key.objectid >= key->objectid &&
1792 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1793 return read_bg_from_eb(fs_info, &found_key, path);
1799 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1801 u64 extra_flags = chunk_to_extended(flags) &
1802 BTRFS_EXTENDED_PROFILE_MASK;
1804 write_seqlock(&fs_info->profiles_lock);
1805 if (flags & BTRFS_BLOCK_GROUP_DATA)
1806 fs_info->avail_data_alloc_bits |= extra_flags;
1807 if (flags & BTRFS_BLOCK_GROUP_METADATA)
1808 fs_info->avail_metadata_alloc_bits |= extra_flags;
1809 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1810 fs_info->avail_system_alloc_bits |= extra_flags;
1811 write_sequnlock(&fs_info->profiles_lock);
1815 * Map a physical disk address to a list of logical addresses.
1817 * @fs_info: the filesystem
1818 * @chunk_start: logical address of block group
1819 * @bdev: physical device to resolve, can be NULL to indicate any device
1820 * @physical: physical address to map to logical addresses
1821 * @logical: return array of logical addresses which map to @physical
1822 * @naddrs: length of @logical
1823 * @stripe_len: size of IO stripe for the given block group
1825 * Maps a particular @physical disk address to a list of @logical addresses.
1826 * Used primarily to exclude those portions of a block group that contain super
1829 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1830 struct block_device *bdev, u64 physical, u64 **logical,
1831 int *naddrs, int *stripe_len)
1833 struct extent_map *em;
1834 struct map_lookup *map;
1837 u64 data_stripe_length;
1842 em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1846 map = em->map_lookup;
1847 data_stripe_length = em->orig_block_len;
1848 io_stripe_size = map->stripe_len;
1849 chunk_start = em->start;
1851 /* For RAID5/6 adjust to a full IO stripe length */
1852 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1853 io_stripe_size = map->stripe_len * nr_data_stripes(map);
1855 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1861 for (i = 0; i < map->num_stripes; i++) {
1862 bool already_inserted = false;
1867 if (!in_range(physical, map->stripes[i].physical,
1868 data_stripe_length))
1871 if (bdev && map->stripes[i].dev->bdev != bdev)
1874 stripe_nr = physical - map->stripes[i].physical;
1875 stripe_nr = div64_u64_rem(stripe_nr, map->stripe_len, &offset);
1877 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
1878 BTRFS_BLOCK_GROUP_RAID10)) {
1879 stripe_nr = stripe_nr * map->num_stripes + i;
1880 stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1883 * The remaining case would be for RAID56, multiply by
1884 * nr_data_stripes(). Alternatively, just use rmap_len below
1885 * instead of map->stripe_len
1888 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
1890 /* Ensure we don't add duplicate addresses */
1891 for (j = 0; j < nr; j++) {
1892 if (buf[j] == bytenr) {
1893 already_inserted = true;
1898 if (!already_inserted)
1904 *stripe_len = io_stripe_size;
1906 free_extent_map(em);
1910 static int exclude_super_stripes(struct btrfs_block_group *cache)
1912 struct btrfs_fs_info *fs_info = cache->fs_info;
1913 const bool zoned = btrfs_is_zoned(fs_info);
1919 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1920 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1921 cache->bytes_super += stripe_len;
1922 ret = btrfs_add_excluded_extent(fs_info, cache->start,
1928 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1929 bytenr = btrfs_sb_offset(i);
1930 ret = btrfs_rmap_block(fs_info, cache->start, NULL,
1931 bytenr, &logical, &nr, &stripe_len);
1935 /* Shouldn't have super stripes in sequential zones */
1938 "zoned: block group %llu must not contain super block",
1944 u64 len = min_t(u64, stripe_len,
1945 cache->start + cache->length - logical[nr]);
1947 cache->bytes_super += len;
1948 ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1961 static struct btrfs_block_group *btrfs_create_block_group_cache(
1962 struct btrfs_fs_info *fs_info, u64 start)
1964 struct btrfs_block_group *cache;
1966 cache = kzalloc(sizeof(*cache), GFP_NOFS);
1970 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1972 if (!cache->free_space_ctl) {
1977 cache->start = start;
1979 cache->fs_info = fs_info;
1980 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1982 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1984 refcount_set(&cache->refs, 1);
1985 spin_lock_init(&cache->lock);
1986 init_rwsem(&cache->data_rwsem);
1987 INIT_LIST_HEAD(&cache->list);
1988 INIT_LIST_HEAD(&cache->cluster_list);
1989 INIT_LIST_HEAD(&cache->bg_list);
1990 INIT_LIST_HEAD(&cache->ro_list);
1991 INIT_LIST_HEAD(&cache->discard_list);
1992 INIT_LIST_HEAD(&cache->dirty_list);
1993 INIT_LIST_HEAD(&cache->io_list);
1994 INIT_LIST_HEAD(&cache->active_bg_list);
1995 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1996 atomic_set(&cache->frozen, 0);
1997 mutex_init(&cache->free_space_lock);
1998 cache->full_stripe_locks_root.root = RB_ROOT;
1999 mutex_init(&cache->full_stripe_locks_root.lock);
2005 * Iterate all chunks and verify that each of them has the corresponding block
2008 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2010 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
2011 struct extent_map *em;
2012 struct btrfs_block_group *bg;
2017 read_lock(&map_tree->lock);
2019 * lookup_extent_mapping will return the first extent map
2020 * intersecting the range, so setting @len to 1 is enough to
2021 * get the first chunk.
2023 em = lookup_extent_mapping(map_tree, start, 1);
2024 read_unlock(&map_tree->lock);
2028 bg = btrfs_lookup_block_group(fs_info, em->start);
2031 "chunk start=%llu len=%llu doesn't have corresponding block group",
2032 em->start, em->len);
2034 free_extent_map(em);
2037 if (bg->start != em->start || bg->length != em->len ||
2038 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2039 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2041 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2043 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2044 bg->start, bg->length,
2045 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2047 free_extent_map(em);
2048 btrfs_put_block_group(bg);
2051 start = em->start + em->len;
2052 free_extent_map(em);
2053 btrfs_put_block_group(bg);
2058 static int read_one_block_group(struct btrfs_fs_info *info,
2059 struct btrfs_block_group_item *bgi,
2060 const struct btrfs_key *key,
2063 struct btrfs_block_group *cache;
2064 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2067 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2069 cache = btrfs_create_block_group_cache(info, key->objectid);
2073 cache->length = key->offset;
2074 cache->used = btrfs_stack_block_group_used(bgi);
2075 cache->commit_used = cache->used;
2076 cache->flags = btrfs_stack_block_group_flags(bgi);
2077 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2079 set_free_space_tree_thresholds(cache);
2083 * When we mount with old space cache, we need to
2084 * set BTRFS_DC_CLEAR and set dirty flag.
2086 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2087 * truncate the old free space cache inode and
2089 * b) Setting 'dirty flag' makes sure that we flush
2090 * the new space cache info onto disk.
2092 if (btrfs_test_opt(info, SPACE_CACHE))
2093 cache->disk_cache_state = BTRFS_DC_CLEAR;
2095 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2096 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2098 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2104 ret = btrfs_load_block_group_zone_info(cache, false);
2106 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2112 * We need to exclude the super stripes now so that the space info has
2113 * super bytes accounted for, otherwise we'll think we have more space
2114 * than we actually do.
2116 ret = exclude_super_stripes(cache);
2118 /* We may have excluded something, so call this just in case. */
2119 btrfs_free_excluded_extents(cache);
2124 * For zoned filesystem, space after the allocation offset is the only
2125 * free space for a block group. So, we don't need any caching work.
2126 * btrfs_calc_zone_unusable() will set the amount of free space and
2127 * zone_unusable space.
2129 * For regular filesystem, check for two cases, either we are full, and
2130 * therefore don't need to bother with the caching work since we won't
2131 * find any space, or we are empty, and we can just add all the space
2132 * in and be done with it. This saves us _a_lot_ of time, particularly
2135 if (btrfs_is_zoned(info)) {
2136 btrfs_calc_zone_unusable(cache);
2137 /* Should not have any excluded extents. Just in case, though. */
2138 btrfs_free_excluded_extents(cache);
2139 } else if (cache->length == cache->used) {
2140 cache->cached = BTRFS_CACHE_FINISHED;
2141 btrfs_free_excluded_extents(cache);
2142 } else if (cache->used == 0) {
2143 cache->cached = BTRFS_CACHE_FINISHED;
2144 add_new_free_space(cache, cache->start,
2145 cache->start + cache->length);
2146 btrfs_free_excluded_extents(cache);
2149 ret = btrfs_add_block_group_cache(info, cache);
2151 btrfs_remove_free_space_cache(cache);
2154 trace_btrfs_add_block_group(info, cache, 0);
2155 btrfs_add_bg_to_space_info(info, cache);
2157 set_avail_alloc_bits(info, cache->flags);
2158 if (btrfs_chunk_writeable(info, cache->start)) {
2159 if (cache->used == 0) {
2160 ASSERT(list_empty(&cache->bg_list));
2161 if (btrfs_test_opt(info, DISCARD_ASYNC))
2162 btrfs_discard_queue_work(&info->discard_ctl, cache);
2164 btrfs_mark_bg_unused(cache);
2167 inc_block_group_ro(cache, 1);
2172 btrfs_put_block_group(cache);
2176 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2178 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
2179 struct rb_node *node;
2182 for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
2183 struct extent_map *em;
2184 struct map_lookup *map;
2185 struct btrfs_block_group *bg;
2187 em = rb_entry(node, struct extent_map, rb_node);
2188 map = em->map_lookup;
2189 bg = btrfs_create_block_group_cache(fs_info, em->start);
2195 /* Fill dummy cache as FULL */
2196 bg->length = em->len;
2197 bg->flags = map->type;
2198 bg->cached = BTRFS_CACHE_FINISHED;
2200 bg->flags = map->type;
2201 ret = btrfs_add_block_group_cache(fs_info, bg);
2203 * We may have some valid block group cache added already, in
2204 * that case we skip to the next one.
2206 if (ret == -EEXIST) {
2208 btrfs_put_block_group(bg);
2213 btrfs_remove_free_space_cache(bg);
2214 btrfs_put_block_group(bg);
2218 btrfs_add_bg_to_space_info(fs_info, bg);
2220 set_avail_alloc_bits(fs_info, bg->flags);
2223 btrfs_init_global_block_rsv(fs_info);
2227 int btrfs_read_block_groups(struct btrfs_fs_info *info)
2229 struct btrfs_root *root = btrfs_block_group_root(info);
2230 struct btrfs_path *path;
2232 struct btrfs_block_group *cache;
2233 struct btrfs_space_info *space_info;
2234 struct btrfs_key key;
2239 * Either no extent root (with ibadroots rescue option) or we have
2240 * unsupported RO options. The fs can never be mounted read-write, so no
2241 * need to waste time searching block group items.
2243 * This also allows new extent tree related changes to be RO compat,
2244 * no need for a full incompat flag.
2246 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2247 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2248 return fill_dummy_bgs(info);
2252 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2253 path = btrfs_alloc_path();
2257 cache_gen = btrfs_super_cache_generation(info->super_copy);
2258 if (btrfs_test_opt(info, SPACE_CACHE) &&
2259 btrfs_super_generation(info->super_copy) != cache_gen)
2261 if (btrfs_test_opt(info, CLEAR_CACHE))
2265 struct btrfs_block_group_item bgi;
2266 struct extent_buffer *leaf;
2269 ret = find_first_block_group(info, path, &key);
2275 leaf = path->nodes[0];
2276 slot = path->slots[0];
2278 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2281 btrfs_item_key_to_cpu(leaf, &key, slot);
2282 btrfs_release_path(path);
2283 ret = read_one_block_group(info, &bgi, &key, need_clear);
2286 key.objectid += key.offset;
2289 btrfs_release_path(path);
2291 list_for_each_entry(space_info, &info->space_info, list) {
2294 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2295 if (list_empty(&space_info->block_groups[i]))
2297 cache = list_first_entry(&space_info->block_groups[i],
2298 struct btrfs_block_group,
2300 btrfs_sysfs_add_block_group_type(cache);
2303 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2304 (BTRFS_BLOCK_GROUP_RAID10 |
2305 BTRFS_BLOCK_GROUP_RAID1_MASK |
2306 BTRFS_BLOCK_GROUP_RAID56_MASK |
2307 BTRFS_BLOCK_GROUP_DUP)))
2310 * Avoid allocating from un-mirrored block group if there are
2311 * mirrored block groups.
2313 list_for_each_entry(cache,
2314 &space_info->block_groups[BTRFS_RAID_RAID0],
2316 inc_block_group_ro(cache, 1);
2317 list_for_each_entry(cache,
2318 &space_info->block_groups[BTRFS_RAID_SINGLE],
2320 inc_block_group_ro(cache, 1);
2323 btrfs_init_global_block_rsv(info);
2324 ret = check_chunk_block_group_mappings(info);
2326 btrfs_free_path(path);
2328 * We've hit some error while reading the extent tree, and have
2329 * rescue=ibadroots mount option.
2330 * Try to fill the tree using dummy block groups so that the user can
2331 * continue to mount and grab their data.
2333 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2334 ret = fill_dummy_bgs(info);
2339 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2342 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2345 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2346 struct btrfs_block_group *block_group)
2348 struct btrfs_fs_info *fs_info = trans->fs_info;
2349 struct btrfs_block_group_item bgi;
2350 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2351 struct btrfs_key key;
2353 spin_lock(&block_group->lock);
2354 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2355 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2356 block_group->global_root_id);
2357 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2358 key.objectid = block_group->start;
2359 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2360 key.offset = block_group->length;
2361 spin_unlock(&block_group->lock);
2363 return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2366 static int insert_dev_extent(struct btrfs_trans_handle *trans,
2367 struct btrfs_device *device, u64 chunk_offset,
2368 u64 start, u64 num_bytes)
2370 struct btrfs_fs_info *fs_info = device->fs_info;
2371 struct btrfs_root *root = fs_info->dev_root;
2372 struct btrfs_path *path;
2373 struct btrfs_dev_extent *extent;
2374 struct extent_buffer *leaf;
2375 struct btrfs_key key;
2378 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2379 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2380 path = btrfs_alloc_path();
2384 key.objectid = device->devid;
2385 key.type = BTRFS_DEV_EXTENT_KEY;
2387 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2391 leaf = path->nodes[0];
2392 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2393 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2394 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2395 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2396 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2398 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2399 btrfs_mark_buffer_dirty(leaf);
2401 btrfs_free_path(path);
2406 * This function belongs to phase 2.
2408 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2411 static int insert_dev_extents(struct btrfs_trans_handle *trans,
2412 u64 chunk_offset, u64 chunk_size)
2414 struct btrfs_fs_info *fs_info = trans->fs_info;
2415 struct btrfs_device *device;
2416 struct extent_map *em;
2417 struct map_lookup *map;
2423 em = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2427 map = em->map_lookup;
2428 stripe_size = em->orig_block_len;
2431 * Take the device list mutex to prevent races with the final phase of
2432 * a device replace operation that replaces the device object associated
2433 * with the map's stripes, because the device object's id can change
2434 * at any time during that final phase of the device replace operation
2435 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2436 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2437 * resulting in persisting a device extent item with such ID.
2439 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2440 for (i = 0; i < map->num_stripes; i++) {
2441 device = map->stripes[i].dev;
2442 dev_offset = map->stripes[i].physical;
2444 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2449 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2451 free_extent_map(em);
2456 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2459 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2462 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2464 struct btrfs_fs_info *fs_info = trans->fs_info;
2465 struct btrfs_block_group *block_group;
2468 while (!list_empty(&trans->new_bgs)) {
2471 block_group = list_first_entry(&trans->new_bgs,
2472 struct btrfs_block_group,
2477 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2479 ret = insert_block_group_item(trans, block_group);
2481 btrfs_abort_transaction(trans, ret);
2482 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2483 &block_group->runtime_flags)) {
2484 mutex_lock(&fs_info->chunk_mutex);
2485 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2486 mutex_unlock(&fs_info->chunk_mutex);
2488 btrfs_abort_transaction(trans, ret);
2490 ret = insert_dev_extents(trans, block_group->start,
2491 block_group->length);
2493 btrfs_abort_transaction(trans, ret);
2494 add_block_group_free_space(trans, block_group);
2497 * If we restriped during balance, we may have added a new raid
2498 * type, so now add the sysfs entries when it is safe to do so.
2499 * We don't have to worry about locking here as it's handled in
2500 * btrfs_sysfs_add_block_group_type.
2502 if (block_group->space_info->block_group_kobjs[index] == NULL)
2503 btrfs_sysfs_add_block_group_type(block_group);
2505 /* Already aborted the transaction if it failed. */
2507 btrfs_delayed_refs_rsv_release(fs_info, 1);
2508 list_del_init(&block_group->bg_list);
2510 btrfs_trans_release_chunk_metadata(trans);
2514 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2515 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2517 static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2522 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2523 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2525 /* If we have a smaller fs index based on 128MiB. */
2526 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2529 offset = div64_u64(offset, div);
2530 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2534 struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2535 u64 bytes_used, u64 type,
2536 u64 chunk_offset, u64 size)
2538 struct btrfs_fs_info *fs_info = trans->fs_info;
2539 struct btrfs_block_group *cache;
2542 btrfs_set_log_full_commit(trans);
2544 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2546 return ERR_PTR(-ENOMEM);
2548 cache->length = size;
2549 set_free_space_tree_thresholds(cache);
2550 cache->used = bytes_used;
2551 cache->flags = type;
2552 cache->cached = BTRFS_CACHE_FINISHED;
2553 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2555 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2556 set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2558 ret = btrfs_load_block_group_zone_info(cache, true);
2560 btrfs_put_block_group(cache);
2561 return ERR_PTR(ret);
2564 ret = exclude_super_stripes(cache);
2566 /* We may have excluded something, so call this just in case */
2567 btrfs_free_excluded_extents(cache);
2568 btrfs_put_block_group(cache);
2569 return ERR_PTR(ret);
2572 add_new_free_space(cache, chunk_offset, chunk_offset + size);
2574 btrfs_free_excluded_extents(cache);
2577 * Ensure the corresponding space_info object is created and
2578 * assigned to our block group. We want our bg to be added to the rbtree
2579 * with its ->space_info set.
2581 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2582 ASSERT(cache->space_info);
2584 ret = btrfs_add_block_group_cache(fs_info, cache);
2586 btrfs_remove_free_space_cache(cache);
2587 btrfs_put_block_group(cache);
2588 return ERR_PTR(ret);
2592 * Now that our block group has its ->space_info set and is inserted in
2593 * the rbtree, update the space info's counters.
2595 trace_btrfs_add_block_group(fs_info, cache, 1);
2596 btrfs_add_bg_to_space_info(fs_info, cache);
2597 btrfs_update_global_block_rsv(fs_info);
2599 #ifdef CONFIG_BTRFS_DEBUG
2600 if (btrfs_should_fragment_free_space(cache)) {
2601 u64 new_bytes_used = size - bytes_used;
2603 cache->space_info->bytes_used += new_bytes_used >> 1;
2604 fragment_free_space(cache);
2608 list_add_tail(&cache->bg_list, &trans->new_bgs);
2609 trans->delayed_ref_updates++;
2610 btrfs_update_delayed_refs_rsv(trans);
2612 set_avail_alloc_bits(fs_info, type);
2617 * Mark one block group RO, can be called several times for the same block
2620 * @cache: the destination block group
2621 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2622 * ensure we still have some free space after marking this
2625 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2626 bool do_chunk_alloc)
2628 struct btrfs_fs_info *fs_info = cache->fs_info;
2629 struct btrfs_trans_handle *trans;
2630 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2633 bool dirty_bg_running;
2636 * This can only happen when we are doing read-only scrub on read-only
2638 * In that case we should not start a new transaction on read-only fs.
2639 * Thus here we skip all chunk allocations.
2641 if (sb_rdonly(fs_info->sb)) {
2642 mutex_lock(&fs_info->ro_block_group_mutex);
2643 ret = inc_block_group_ro(cache, 0);
2644 mutex_unlock(&fs_info->ro_block_group_mutex);
2649 trans = btrfs_join_transaction(root);
2651 return PTR_ERR(trans);
2653 dirty_bg_running = false;
2656 * We're not allowed to set block groups readonly after the dirty
2657 * block group cache has started writing. If it already started,
2658 * back off and let this transaction commit.
2660 mutex_lock(&fs_info->ro_block_group_mutex);
2661 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2662 u64 transid = trans->transid;
2664 mutex_unlock(&fs_info->ro_block_group_mutex);
2665 btrfs_end_transaction(trans);
2667 ret = btrfs_wait_for_commit(fs_info, transid);
2670 dirty_bg_running = true;
2672 } while (dirty_bg_running);
2674 if (do_chunk_alloc) {
2676 * If we are changing raid levels, try to allocate a
2677 * corresponding block group with the new raid level.
2679 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2680 if (alloc_flags != cache->flags) {
2681 ret = btrfs_chunk_alloc(trans, alloc_flags,
2684 * ENOSPC is allowed here, we may have enough space
2685 * already allocated at the new raid level to carry on
2694 ret = inc_block_group_ro(cache, 0);
2695 if (!do_chunk_alloc || ret == -ETXTBSY)
2699 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2700 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2704 * We have allocated a new chunk. We also need to activate that chunk to
2705 * grant metadata tickets for zoned filesystem.
2707 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2711 ret = inc_block_group_ro(cache, 0);
2712 if (ret == -ETXTBSY)
2715 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2716 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2717 mutex_lock(&fs_info->chunk_mutex);
2718 check_system_chunk(trans, alloc_flags);
2719 mutex_unlock(&fs_info->chunk_mutex);
2722 mutex_unlock(&fs_info->ro_block_group_mutex);
2724 btrfs_end_transaction(trans);
2728 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2730 struct btrfs_space_info *sinfo = cache->space_info;
2735 spin_lock(&sinfo->lock);
2736 spin_lock(&cache->lock);
2738 if (btrfs_is_zoned(cache->fs_info)) {
2739 /* Migrate zone_unusable bytes back */
2740 cache->zone_unusable =
2741 (cache->alloc_offset - cache->used) +
2742 (cache->length - cache->zone_capacity);
2743 sinfo->bytes_zone_unusable += cache->zone_unusable;
2744 sinfo->bytes_readonly -= cache->zone_unusable;
2746 num_bytes = cache->length - cache->reserved -
2747 cache->pinned - cache->bytes_super -
2748 cache->zone_unusable - cache->used;
2749 sinfo->bytes_readonly -= num_bytes;
2750 list_del_init(&cache->ro_list);
2752 spin_unlock(&cache->lock);
2753 spin_unlock(&sinfo->lock);
2756 static int update_block_group_item(struct btrfs_trans_handle *trans,
2757 struct btrfs_path *path,
2758 struct btrfs_block_group *cache)
2760 struct btrfs_fs_info *fs_info = trans->fs_info;
2762 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2764 struct extent_buffer *leaf;
2765 struct btrfs_block_group_item bgi;
2766 struct btrfs_key key;
2767 u64 old_commit_used;
2771 * Block group items update can be triggered out of commit transaction
2772 * critical section, thus we need a consistent view of used bytes.
2773 * We cannot use cache->used directly outside of the spin lock, as it
2776 spin_lock(&cache->lock);
2777 old_commit_used = cache->commit_used;
2779 /* No change in used bytes, can safely skip it. */
2780 if (cache->commit_used == used) {
2781 spin_unlock(&cache->lock);
2784 cache->commit_used = used;
2785 spin_unlock(&cache->lock);
2787 key.objectid = cache->start;
2788 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2789 key.offset = cache->length;
2791 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2798 leaf = path->nodes[0];
2799 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2800 btrfs_set_stack_block_group_used(&bgi, used);
2801 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2802 cache->global_root_id);
2803 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2804 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2805 btrfs_mark_buffer_dirty(leaf);
2807 btrfs_release_path(path);
2808 /* We didn't update the block group item, need to revert @commit_used. */
2810 spin_lock(&cache->lock);
2811 cache->commit_used = old_commit_used;
2812 spin_unlock(&cache->lock);
2818 static int cache_save_setup(struct btrfs_block_group *block_group,
2819 struct btrfs_trans_handle *trans,
2820 struct btrfs_path *path)
2822 struct btrfs_fs_info *fs_info = block_group->fs_info;
2823 struct btrfs_root *root = fs_info->tree_root;
2824 struct inode *inode = NULL;
2825 struct extent_changeset *data_reserved = NULL;
2827 int dcs = BTRFS_DC_ERROR;
2832 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2836 * If this block group is smaller than 100 megs don't bother caching the
2839 if (block_group->length < (100 * SZ_1M)) {
2840 spin_lock(&block_group->lock);
2841 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2842 spin_unlock(&block_group->lock);
2846 if (TRANS_ABORTED(trans))
2849 inode = lookup_free_space_inode(block_group, path);
2850 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2851 ret = PTR_ERR(inode);
2852 btrfs_release_path(path);
2856 if (IS_ERR(inode)) {
2860 if (block_group->ro)
2863 ret = create_free_space_inode(trans, block_group, path);
2870 * We want to set the generation to 0, that way if anything goes wrong
2871 * from here on out we know not to trust this cache when we load up next
2874 BTRFS_I(inode)->generation = 0;
2875 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2878 * So theoretically we could recover from this, simply set the
2879 * super cache generation to 0 so we know to invalidate the
2880 * cache, but then we'd have to keep track of the block groups
2881 * that fail this way so we know we _have_ to reset this cache
2882 * before the next commit or risk reading stale cache. So to
2883 * limit our exposure to horrible edge cases lets just abort the
2884 * transaction, this only happens in really bad situations
2887 btrfs_abort_transaction(trans, ret);
2892 /* We've already setup this transaction, go ahead and exit */
2893 if (block_group->cache_generation == trans->transid &&
2894 i_size_read(inode)) {
2895 dcs = BTRFS_DC_SETUP;
2899 if (i_size_read(inode) > 0) {
2900 ret = btrfs_check_trunc_cache_free_space(fs_info,
2901 &fs_info->global_block_rsv);
2905 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2910 spin_lock(&block_group->lock);
2911 if (block_group->cached != BTRFS_CACHE_FINISHED ||
2912 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2914 * don't bother trying to write stuff out _if_
2915 * a) we're not cached,
2916 * b) we're with nospace_cache mount option,
2917 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2919 dcs = BTRFS_DC_WRITTEN;
2920 spin_unlock(&block_group->lock);
2923 spin_unlock(&block_group->lock);
2926 * We hit an ENOSPC when setting up the cache in this transaction, just
2927 * skip doing the setup, we've already cleared the cache so we're safe.
2929 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2935 * Try to preallocate enough space based on how big the block group is.
2936 * Keep in mind this has to include any pinned space which could end up
2937 * taking up quite a bit since it's not folded into the other space
2940 cache_size = div_u64(block_group->length, SZ_256M);
2945 cache_size *= fs_info->sectorsize;
2947 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2952 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
2953 cache_size, cache_size,
2956 * Our cache requires contiguous chunks so that we don't modify a bunch
2957 * of metadata or split extents when writing the cache out, which means
2958 * we can enospc if we are heavily fragmented in addition to just normal
2959 * out of space conditions. So if we hit this just skip setting up any
2960 * other block groups for this transaction, maybe we'll unpin enough
2961 * space the next time around.
2964 dcs = BTRFS_DC_SETUP;
2965 else if (ret == -ENOSPC)
2966 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2971 btrfs_release_path(path);
2973 spin_lock(&block_group->lock);
2974 if (!ret && dcs == BTRFS_DC_SETUP)
2975 block_group->cache_generation = trans->transid;
2976 block_group->disk_cache_state = dcs;
2977 spin_unlock(&block_group->lock);
2979 extent_changeset_free(data_reserved);
2983 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2985 struct btrfs_fs_info *fs_info = trans->fs_info;
2986 struct btrfs_block_group *cache, *tmp;
2987 struct btrfs_transaction *cur_trans = trans->transaction;
2988 struct btrfs_path *path;
2990 if (list_empty(&cur_trans->dirty_bgs) ||
2991 !btrfs_test_opt(fs_info, SPACE_CACHE))
2994 path = btrfs_alloc_path();
2998 /* Could add new block groups, use _safe just in case */
2999 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3001 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3002 cache_save_setup(cache, trans, path);
3005 btrfs_free_path(path);
3010 * Transaction commit does final block group cache writeback during a critical
3011 * section where nothing is allowed to change the FS. This is required in
3012 * order for the cache to actually match the block group, but can introduce a
3013 * lot of latency into the commit.
3015 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3016 * There's a chance we'll have to redo some of it if the block group changes
3017 * again during the commit, but it greatly reduces the commit latency by
3018 * getting rid of the easy block groups while we're still allowing others to
3021 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3023 struct btrfs_fs_info *fs_info = trans->fs_info;
3024 struct btrfs_block_group *cache;
3025 struct btrfs_transaction *cur_trans = trans->transaction;
3028 struct btrfs_path *path = NULL;
3030 struct list_head *io = &cur_trans->io_bgs;
3033 spin_lock(&cur_trans->dirty_bgs_lock);
3034 if (list_empty(&cur_trans->dirty_bgs)) {
3035 spin_unlock(&cur_trans->dirty_bgs_lock);
3038 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3039 spin_unlock(&cur_trans->dirty_bgs_lock);
3042 /* Make sure all the block groups on our dirty list actually exist */
3043 btrfs_create_pending_block_groups(trans);
3046 path = btrfs_alloc_path();
3054 * cache_write_mutex is here only to save us from balance or automatic
3055 * removal of empty block groups deleting this block group while we are
3056 * writing out the cache
3058 mutex_lock(&trans->transaction->cache_write_mutex);
3059 while (!list_empty(&dirty)) {
3060 bool drop_reserve = true;
3062 cache = list_first_entry(&dirty, struct btrfs_block_group,
3065 * This can happen if something re-dirties a block group that
3066 * is already under IO. Just wait for it to finish and then do
3069 if (!list_empty(&cache->io_list)) {
3070 list_del_init(&cache->io_list);
3071 btrfs_wait_cache_io(trans, cache, path);
3072 btrfs_put_block_group(cache);
3077 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3078 * it should update the cache_state. Don't delete until after
3081 * Since we're not running in the commit critical section
3082 * we need the dirty_bgs_lock to protect from update_block_group
3084 spin_lock(&cur_trans->dirty_bgs_lock);
3085 list_del_init(&cache->dirty_list);
3086 spin_unlock(&cur_trans->dirty_bgs_lock);
3090 cache_save_setup(cache, trans, path);
3092 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3093 cache->io_ctl.inode = NULL;
3094 ret = btrfs_write_out_cache(trans, cache, path);
3095 if (ret == 0 && cache->io_ctl.inode) {
3099 * The cache_write_mutex is protecting the
3100 * io_list, also refer to the definition of
3101 * btrfs_transaction::io_bgs for more details
3103 list_add_tail(&cache->io_list, io);
3106 * If we failed to write the cache, the
3107 * generation will be bad and life goes on
3113 ret = update_block_group_item(trans, path, cache);
3115 * Our block group might still be attached to the list
3116 * of new block groups in the transaction handle of some
3117 * other task (struct btrfs_trans_handle->new_bgs). This
3118 * means its block group item isn't yet in the extent
3119 * tree. If this happens ignore the error, as we will
3120 * try again later in the critical section of the
3121 * transaction commit.
3123 if (ret == -ENOENT) {
3125 spin_lock(&cur_trans->dirty_bgs_lock);
3126 if (list_empty(&cache->dirty_list)) {
3127 list_add_tail(&cache->dirty_list,
3128 &cur_trans->dirty_bgs);
3129 btrfs_get_block_group(cache);
3130 drop_reserve = false;
3132 spin_unlock(&cur_trans->dirty_bgs_lock);
3134 btrfs_abort_transaction(trans, ret);
3138 /* If it's not on the io list, we need to put the block group */
3140 btrfs_put_block_group(cache);
3142 btrfs_delayed_refs_rsv_release(fs_info, 1);
3144 * Avoid blocking other tasks for too long. It might even save
3145 * us from writing caches for block groups that are going to be
3148 mutex_unlock(&trans->transaction->cache_write_mutex);
3151 mutex_lock(&trans->transaction->cache_write_mutex);
3153 mutex_unlock(&trans->transaction->cache_write_mutex);
3156 * Go through delayed refs for all the stuff we've just kicked off
3157 * and then loop back (just once)
3160 ret = btrfs_run_delayed_refs(trans, 0);
3161 if (!ret && loops == 0) {
3163 spin_lock(&cur_trans->dirty_bgs_lock);
3164 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3166 * dirty_bgs_lock protects us from concurrent block group
3167 * deletes too (not just cache_write_mutex).
3169 if (!list_empty(&dirty)) {
3170 spin_unlock(&cur_trans->dirty_bgs_lock);
3173 spin_unlock(&cur_trans->dirty_bgs_lock);
3177 spin_lock(&cur_trans->dirty_bgs_lock);
3178 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3179 spin_unlock(&cur_trans->dirty_bgs_lock);
3180 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3183 btrfs_free_path(path);
3187 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3189 struct btrfs_fs_info *fs_info = trans->fs_info;
3190 struct btrfs_block_group *cache;
3191 struct btrfs_transaction *cur_trans = trans->transaction;
3194 struct btrfs_path *path;
3195 struct list_head *io = &cur_trans->io_bgs;
3197 path = btrfs_alloc_path();
3202 * Even though we are in the critical section of the transaction commit,
3203 * we can still have concurrent tasks adding elements to this
3204 * transaction's list of dirty block groups. These tasks correspond to
3205 * endio free space workers started when writeback finishes for a
3206 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3207 * allocate new block groups as a result of COWing nodes of the root
3208 * tree when updating the free space inode. The writeback for the space
3209 * caches is triggered by an earlier call to
3210 * btrfs_start_dirty_block_groups() and iterations of the following
3212 * Also we want to do the cache_save_setup first and then run the
3213 * delayed refs to make sure we have the best chance at doing this all
3216 spin_lock(&cur_trans->dirty_bgs_lock);
3217 while (!list_empty(&cur_trans->dirty_bgs)) {
3218 cache = list_first_entry(&cur_trans->dirty_bgs,
3219 struct btrfs_block_group,
3223 * This can happen if cache_save_setup re-dirties a block group
3224 * that is already under IO. Just wait for it to finish and
3225 * then do it all again
3227 if (!list_empty(&cache->io_list)) {
3228 spin_unlock(&cur_trans->dirty_bgs_lock);
3229 list_del_init(&cache->io_list);
3230 btrfs_wait_cache_io(trans, cache, path);
3231 btrfs_put_block_group(cache);
3232 spin_lock(&cur_trans->dirty_bgs_lock);
3236 * Don't remove from the dirty list until after we've waited on
3239 list_del_init(&cache->dirty_list);
3240 spin_unlock(&cur_trans->dirty_bgs_lock);
3243 cache_save_setup(cache, trans, path);
3246 ret = btrfs_run_delayed_refs(trans,
3247 (unsigned long) -1);
3249 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3250 cache->io_ctl.inode = NULL;
3251 ret = btrfs_write_out_cache(trans, cache, path);
3252 if (ret == 0 && cache->io_ctl.inode) {
3254 list_add_tail(&cache->io_list, io);
3257 * If we failed to write the cache, the
3258 * generation will be bad and life goes on
3264 ret = update_block_group_item(trans, path, cache);
3266 * One of the free space endio workers might have
3267 * created a new block group while updating a free space
3268 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3269 * and hasn't released its transaction handle yet, in
3270 * which case the new block group is still attached to
3271 * its transaction handle and its creation has not
3272 * finished yet (no block group item in the extent tree
3273 * yet, etc). If this is the case, wait for all free
3274 * space endio workers to finish and retry. This is a
3275 * very rare case so no need for a more efficient and
3278 if (ret == -ENOENT) {
3279 wait_event(cur_trans->writer_wait,
3280 atomic_read(&cur_trans->num_writers) == 1);
3281 ret = update_block_group_item(trans, path, cache);
3284 btrfs_abort_transaction(trans, ret);
3287 /* If its not on the io list, we need to put the block group */
3289 btrfs_put_block_group(cache);
3290 btrfs_delayed_refs_rsv_release(fs_info, 1);
3291 spin_lock(&cur_trans->dirty_bgs_lock);
3293 spin_unlock(&cur_trans->dirty_bgs_lock);
3296 * Refer to the definition of io_bgs member for details why it's safe
3297 * to use it without any locking
3299 while (!list_empty(io)) {
3300 cache = list_first_entry(io, struct btrfs_block_group,
3302 list_del_init(&cache->io_list);
3303 btrfs_wait_cache_io(trans, cache, path);
3304 btrfs_put_block_group(cache);
3307 btrfs_free_path(path);
3311 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3312 u64 bytenr, u64 num_bytes, bool alloc)
3314 struct btrfs_fs_info *info = trans->fs_info;
3315 struct btrfs_block_group *cache = NULL;
3316 u64 total = num_bytes;
3322 /* Block accounting for super block */
3323 spin_lock(&info->delalloc_root_lock);
3324 old_val = btrfs_super_bytes_used(info->super_copy);
3326 old_val += num_bytes;
3328 old_val -= num_bytes;
3329 btrfs_set_super_bytes_used(info->super_copy, old_val);
3330 spin_unlock(&info->delalloc_root_lock);
3335 cache = btrfs_lookup_block_group(info, bytenr);
3340 factor = btrfs_bg_type_to_factor(cache->flags);
3343 * If this block group has free space cache written out, we
3344 * need to make sure to load it if we are removing space. This
3345 * is because we need the unpinning stage to actually add the
3346 * space back to the block group, otherwise we will leak space.
3348 if (!alloc && !btrfs_block_group_done(cache))
3349 btrfs_cache_block_group(cache, true);
3351 byte_in_group = bytenr - cache->start;
3352 WARN_ON(byte_in_group > cache->length);
3354 spin_lock(&cache->space_info->lock);
3355 spin_lock(&cache->lock);
3357 if (btrfs_test_opt(info, SPACE_CACHE) &&
3358 cache->disk_cache_state < BTRFS_DC_CLEAR)
3359 cache->disk_cache_state = BTRFS_DC_CLEAR;
3361 old_val = cache->used;
3362 num_bytes = min(total, cache->length - byte_in_group);
3364 old_val += num_bytes;
3365 cache->used = old_val;
3366 cache->reserved -= num_bytes;
3367 cache->space_info->bytes_reserved -= num_bytes;
3368 cache->space_info->bytes_used += num_bytes;
3369 cache->space_info->disk_used += num_bytes * factor;
3370 spin_unlock(&cache->lock);
3371 spin_unlock(&cache->space_info->lock);
3373 old_val -= num_bytes;
3374 cache->used = old_val;
3375 cache->pinned += num_bytes;
3376 btrfs_space_info_update_bytes_pinned(info,
3377 cache->space_info, num_bytes);
3378 cache->space_info->bytes_used -= num_bytes;
3379 cache->space_info->disk_used -= num_bytes * factor;
3381 reclaim = should_reclaim_block_group(cache, num_bytes);
3382 spin_unlock(&cache->lock);
3383 spin_unlock(&cache->space_info->lock);
3385 set_extent_dirty(&trans->transaction->pinned_extents,
3386 bytenr, bytenr + num_bytes - 1,
3387 GFP_NOFS | __GFP_NOFAIL);
3390 spin_lock(&trans->transaction->dirty_bgs_lock);
3391 if (list_empty(&cache->dirty_list)) {
3392 list_add_tail(&cache->dirty_list,
3393 &trans->transaction->dirty_bgs);
3394 trans->delayed_ref_updates++;
3395 btrfs_get_block_group(cache);
3397 spin_unlock(&trans->transaction->dirty_bgs_lock);
3400 * No longer have used bytes in this block group, queue it for
3401 * deletion. We do this after adding the block group to the
3402 * dirty list to avoid races between cleaner kthread and space
3405 if (!alloc && old_val == 0) {
3406 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3407 btrfs_mark_bg_unused(cache);
3408 } else if (!alloc && reclaim) {
3409 btrfs_mark_bg_to_reclaim(cache);
3412 btrfs_put_block_group(cache);
3414 bytenr += num_bytes;
3417 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3418 btrfs_update_delayed_refs_rsv(trans);
3423 * Update the block_group and space info counters.
3425 * @cache: The cache we are manipulating
3426 * @ram_bytes: The number of bytes of file content, and will be same to
3427 * @num_bytes except for the compress path.
3428 * @num_bytes: The number of bytes in question
3429 * @delalloc: The blocks are allocated for the delalloc write
3431 * This is called by the allocator when it reserves space. If this is a
3432 * reservation and the block group has become read only we cannot make the
3433 * reservation and return -EAGAIN, otherwise this function always succeeds.
3435 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3436 u64 ram_bytes, u64 num_bytes, int delalloc)
3438 struct btrfs_space_info *space_info = cache->space_info;
3441 spin_lock(&space_info->lock);
3442 spin_lock(&cache->lock);
3446 cache->reserved += num_bytes;
3447 space_info->bytes_reserved += num_bytes;
3448 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3449 space_info->flags, num_bytes, 1);
3450 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3451 space_info, -ram_bytes);
3453 cache->delalloc_bytes += num_bytes;
3456 * Compression can use less space than we reserved, so wake
3457 * tickets if that happens
3459 if (num_bytes < ram_bytes)
3460 btrfs_try_granting_tickets(cache->fs_info, space_info);
3462 spin_unlock(&cache->lock);
3463 spin_unlock(&space_info->lock);
3468 * Update the block_group and space info counters.
3470 * @cache: The cache we are manipulating
3471 * @num_bytes: The number of bytes in question
3472 * @delalloc: The blocks are allocated for the delalloc write
3474 * This is called by somebody who is freeing space that was never actually used
3475 * on disk. For example if you reserve some space for a new leaf in transaction
3476 * A and before transaction A commits you free that leaf, you call this with
3477 * reserve set to 0 in order to clear the reservation.
3479 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3480 u64 num_bytes, int delalloc)
3482 struct btrfs_space_info *space_info = cache->space_info;
3484 spin_lock(&space_info->lock);
3485 spin_lock(&cache->lock);
3487 space_info->bytes_readonly += num_bytes;
3488 cache->reserved -= num_bytes;
3489 space_info->bytes_reserved -= num_bytes;
3490 space_info->max_extent_size = 0;
3493 cache->delalloc_bytes -= num_bytes;
3494 spin_unlock(&cache->lock);
3496 btrfs_try_granting_tickets(cache->fs_info, space_info);
3497 spin_unlock(&space_info->lock);
3500 static void force_metadata_allocation(struct btrfs_fs_info *info)
3502 struct list_head *head = &info->space_info;
3503 struct btrfs_space_info *found;
3505 list_for_each_entry(found, head, list) {
3506 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3507 found->force_alloc = CHUNK_ALLOC_FORCE;
3511 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3512 struct btrfs_space_info *sinfo, int force)
3514 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3517 if (force == CHUNK_ALLOC_FORCE)
3521 * in limited mode, we want to have some free space up to
3522 * about 1% of the FS size.
3524 if (force == CHUNK_ALLOC_LIMITED) {
3525 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3526 thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3528 if (sinfo->total_bytes - bytes_used < thresh)
3532 if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3537 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3539 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3541 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3544 static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3546 struct btrfs_block_group *bg;
3550 * Check if we have enough space in the system space info because we
3551 * will need to update device items in the chunk btree and insert a new
3552 * chunk item in the chunk btree as well. This will allocate a new
3553 * system block group if needed.
3555 check_system_chunk(trans, flags);
3557 bg = btrfs_create_chunk(trans, flags);
3563 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3565 * Normally we are not expected to fail with -ENOSPC here, since we have
3566 * previously reserved space in the system space_info and allocated one
3567 * new system chunk if necessary. However there are three exceptions:
3569 * 1) We may have enough free space in the system space_info but all the
3570 * existing system block groups have a profile which can not be used
3571 * for extent allocation.
3573 * This happens when mounting in degraded mode. For example we have a
3574 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3575 * using the other device in degraded mode. If we then allocate a chunk,
3576 * we may have enough free space in the existing system space_info, but
3577 * none of the block groups can be used for extent allocation since they
3578 * have a RAID1 profile, and because we are in degraded mode with a
3579 * single device, we are forced to allocate a new system chunk with a
3580 * SINGLE profile. Making check_system_chunk() iterate over all system
3581 * block groups and check if they have a usable profile and enough space
3582 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3583 * try again after forcing allocation of a new system chunk. Like this
3584 * we avoid paying the cost of that search in normal circumstances, when
3585 * we were not mounted in degraded mode;
3587 * 2) We had enough free space info the system space_info, and one suitable
3588 * block group to allocate from when we called check_system_chunk()
3589 * above. However right after we called it, the only system block group
3590 * with enough free space got turned into RO mode by a running scrub,
3591 * and in this case we have to allocate a new one and retry. We only
3592 * need do this allocate and retry once, since we have a transaction
3593 * handle and scrub uses the commit root to search for block groups;
3595 * 3) We had one system block group with enough free space when we called
3596 * check_system_chunk(), but after that, right before we tried to
3597 * allocate the last extent buffer we needed, a discard operation came
3598 * in and it temporarily removed the last free space entry from the
3599 * block group (discard removes a free space entry, discards it, and
3600 * then adds back the entry to the block group cache).
3602 if (ret == -ENOSPC) {
3603 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3604 struct btrfs_block_group *sys_bg;
3606 sys_bg = btrfs_create_chunk(trans, sys_flags);
3607 if (IS_ERR(sys_bg)) {
3608 ret = PTR_ERR(sys_bg);
3609 btrfs_abort_transaction(trans, ret);
3613 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3615 btrfs_abort_transaction(trans, ret);
3619 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3621 btrfs_abort_transaction(trans, ret);
3625 btrfs_abort_transaction(trans, ret);
3629 btrfs_trans_release_chunk_metadata(trans);
3632 return ERR_PTR(ret);
3634 btrfs_get_block_group(bg);
3639 * Chunk allocation is done in 2 phases:
3641 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3642 * the chunk, the chunk mapping, create its block group and add the items
3643 * that belong in the chunk btree to it - more specifically, we need to
3644 * update device items in the chunk btree and add a new chunk item to it.
3646 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3647 * group item to the extent btree and the device extent items to the devices
3650 * This is done to prevent deadlocks. For example when COWing a node from the
3651 * extent btree we are holding a write lock on the node's parent and if we
3652 * trigger chunk allocation and attempted to insert the new block group item
3653 * in the extent btree right way, we could deadlock because the path for the
3654 * insertion can include that parent node. At first glance it seems impossible
3655 * to trigger chunk allocation after starting a transaction since tasks should
3656 * reserve enough transaction units (metadata space), however while that is true
3657 * most of the time, chunk allocation may still be triggered for several reasons:
3659 * 1) When reserving metadata, we check if there is enough free space in the
3660 * metadata space_info and therefore don't trigger allocation of a new chunk.
3661 * However later when the task actually tries to COW an extent buffer from
3662 * the extent btree or from the device btree for example, it is forced to
3663 * allocate a new block group (chunk) because the only one that had enough
3664 * free space was just turned to RO mode by a running scrub for example (or
3665 * device replace, block group reclaim thread, etc), so we can not use it
3666 * for allocating an extent and end up being forced to allocate a new one;
3668 * 2) Because we only check that the metadata space_info has enough free bytes,
3669 * we end up not allocating a new metadata chunk in that case. However if
3670 * the filesystem was mounted in degraded mode, none of the existing block
3671 * groups might be suitable for extent allocation due to their incompatible
3672 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3673 * use a RAID1 profile, in degraded mode using a single device). In this case
3674 * when the task attempts to COW some extent buffer of the extent btree for
3675 * example, it will trigger allocation of a new metadata block group with a
3676 * suitable profile (SINGLE profile in the example of the degraded mount of
3677 * the RAID1 filesystem);
3679 * 3) The task has reserved enough transaction units / metadata space, but when
3680 * it attempts to COW an extent buffer from the extent or device btree for
3681 * example, it does not find any free extent in any metadata block group,
3682 * therefore forced to try to allocate a new metadata block group.
3683 * This is because some other task allocated all available extents in the
3684 * meanwhile - this typically happens with tasks that don't reserve space
3685 * properly, either intentionally or as a bug. One example where this is
3686 * done intentionally is fsync, as it does not reserve any transaction units
3687 * and ends up allocating a variable number of metadata extents for log
3688 * tree extent buffers;
3690 * 4) The task has reserved enough transaction units / metadata space, but right
3691 * before it tries to allocate the last extent buffer it needs, a discard
3692 * operation comes in and, temporarily, removes the last free space entry from
3693 * the only metadata block group that had free space (discard starts by
3694 * removing a free space entry from a block group, then does the discard
3695 * operation and, once it's done, it adds back the free space entry to the
3698 * We also need this 2 phases setup when adding a device to a filesystem with
3699 * a seed device - we must create new metadata and system chunks without adding
3700 * any of the block group items to the chunk, extent and device btrees. If we
3701 * did not do it this way, we would get ENOSPC when attempting to update those
3702 * btrees, since all the chunks from the seed device are read-only.
3704 * Phase 1 does the updates and insertions to the chunk btree because if we had
3705 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3706 * parallel, we risk having too many system chunks allocated by many tasks if
3707 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3708 * extreme case this leads to exhaustion of the system chunk array in the
3709 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3710 * and with RAID filesystems (so we have more device items in the chunk btree).
3711 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3712 * the system chunk array due to concurrent allocations") provides more details.
3714 * Allocation of system chunks does not happen through this function. A task that
3715 * needs to update the chunk btree (the only btree that uses system chunks), must
3716 * preallocate chunk space by calling either check_system_chunk() or
3717 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
3718 * metadata chunk or when removing a chunk, while the later is used before doing
3719 * a modification to the chunk btree - use cases for the later are adding,
3720 * removing and resizing a device as well as relocation of a system chunk.
3721 * See the comment below for more details.
3723 * The reservation of system space, done through check_system_chunk(), as well
3724 * as all the updates and insertions into the chunk btree must be done while
3725 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
3726 * an extent buffer from the chunks btree we never trigger allocation of a new
3727 * system chunk, which would result in a deadlock (trying to lock twice an
3728 * extent buffer of the chunk btree, first time before triggering the chunk
3729 * allocation and the second time during chunk allocation while attempting to
3730 * update the chunks btree). The system chunk array is also updated while holding
3731 * that mutex. The same logic applies to removing chunks - we must reserve system
3732 * space, update the chunk btree and the system chunk array in the superblock
3733 * while holding fs_info->chunk_mutex.
3735 * This function, btrfs_chunk_alloc(), belongs to phase 1.
3737 * If @force is CHUNK_ALLOC_FORCE:
3738 * - return 1 if it successfully allocates a chunk,
3739 * - return errors including -ENOSPC otherwise.
3740 * If @force is NOT CHUNK_ALLOC_FORCE:
3741 * - return 0 if it doesn't need to allocate a new chunk,
3742 * - return 1 if it successfully allocates a chunk,
3743 * - return errors including -ENOSPC otherwise.
3745 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3746 enum btrfs_chunk_alloc_enum force)
3748 struct btrfs_fs_info *fs_info = trans->fs_info;
3749 struct btrfs_space_info *space_info;
3750 struct btrfs_block_group *ret_bg;
3751 bool wait_for_alloc = false;
3752 bool should_alloc = false;
3753 bool from_extent_allocation = false;
3756 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
3757 from_extent_allocation = true;
3758 force = CHUNK_ALLOC_FORCE;
3761 /* Don't re-enter if we're already allocating a chunk */
3762 if (trans->allocating_chunk)
3765 * Allocation of system chunks can not happen through this path, as we
3766 * could end up in a deadlock if we are allocating a data or metadata
3767 * chunk and there is another task modifying the chunk btree.
3769 * This is because while we are holding the chunk mutex, we will attempt
3770 * to add the new chunk item to the chunk btree or update an existing
3771 * device item in the chunk btree, while the other task that is modifying
3772 * the chunk btree is attempting to COW an extent buffer while holding a
3773 * lock on it and on its parent - if the COW operation triggers a system
3774 * chunk allocation, then we can deadlock because we are holding the
3775 * chunk mutex and we may need to access that extent buffer or its parent
3776 * in order to add the chunk item or update a device item.
3778 * Tasks that want to modify the chunk tree should reserve system space
3779 * before updating the chunk btree, by calling either
3780 * btrfs_reserve_chunk_metadata() or check_system_chunk().
3781 * It's possible that after a task reserves the space, it still ends up
3782 * here - this happens in the cases described above at do_chunk_alloc().
3783 * The task will have to either retry or fail.
3785 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3788 space_info = btrfs_find_space_info(fs_info, flags);
3792 spin_lock(&space_info->lock);
3793 if (force < space_info->force_alloc)
3794 force = space_info->force_alloc;
3795 should_alloc = should_alloc_chunk(fs_info, space_info, force);
3796 if (space_info->full) {
3797 /* No more free physical space */
3802 spin_unlock(&space_info->lock);
3804 } else if (!should_alloc) {
3805 spin_unlock(&space_info->lock);
3807 } else if (space_info->chunk_alloc) {
3809 * Someone is already allocating, so we need to block
3810 * until this someone is finished and then loop to
3811 * recheck if we should continue with our allocation
3814 wait_for_alloc = true;
3815 force = CHUNK_ALLOC_NO_FORCE;
3816 spin_unlock(&space_info->lock);
3817 mutex_lock(&fs_info->chunk_mutex);
3818 mutex_unlock(&fs_info->chunk_mutex);
3820 /* Proceed with allocation */
3821 space_info->chunk_alloc = 1;
3822 wait_for_alloc = false;
3823 spin_unlock(&space_info->lock);
3827 } while (wait_for_alloc);
3829 mutex_lock(&fs_info->chunk_mutex);
3830 trans->allocating_chunk = true;
3833 * If we have mixed data/metadata chunks we want to make sure we keep
3834 * allocating mixed chunks instead of individual chunks.
3836 if (btrfs_mixed_space_info(space_info))
3837 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3840 * if we're doing a data chunk, go ahead and make sure that
3841 * we keep a reasonable number of metadata chunks allocated in the
3844 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3845 fs_info->data_chunk_allocations++;
3846 if (!(fs_info->data_chunk_allocations %
3847 fs_info->metadata_ratio))
3848 force_metadata_allocation(fs_info);
3851 ret_bg = do_chunk_alloc(trans, flags);
3852 trans->allocating_chunk = false;
3854 if (IS_ERR(ret_bg)) {
3855 ret = PTR_ERR(ret_bg);
3856 } else if (from_extent_allocation) {
3858 * New block group is likely to be used soon. Try to activate
3859 * it now. Failure is OK for now.
3861 btrfs_zone_activate(ret_bg);
3865 btrfs_put_block_group(ret_bg);
3867 spin_lock(&space_info->lock);
3870 space_info->full = 1;
3875 space_info->max_extent_size = 0;
3878 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3880 space_info->chunk_alloc = 0;
3881 spin_unlock(&space_info->lock);
3882 mutex_unlock(&fs_info->chunk_mutex);
3887 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3891 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3893 num_dev = fs_info->fs_devices->rw_devices;
3898 static void reserve_chunk_space(struct btrfs_trans_handle *trans,
3902 struct btrfs_fs_info *fs_info = trans->fs_info;
3903 struct btrfs_space_info *info;
3908 * Needed because we can end up allocating a system chunk and for an
3909 * atomic and race free space reservation in the chunk block reserve.
3911 lockdep_assert_held(&fs_info->chunk_mutex);
3913 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3914 spin_lock(&info->lock);
3915 left = info->total_bytes - btrfs_space_info_used(info, true);
3916 spin_unlock(&info->lock);
3918 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3919 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3921 btrfs_dump_space_info(fs_info, info, 0, 0);
3925 u64 flags = btrfs_system_alloc_profile(fs_info);
3926 struct btrfs_block_group *bg;
3929 * Ignore failure to create system chunk. We might end up not
3930 * needing it, as we might not need to COW all nodes/leafs from
3931 * the paths we visit in the chunk tree (they were already COWed
3932 * or created in the current transaction for example).
3934 bg = btrfs_create_chunk(trans, flags);
3939 * We have a new chunk. We also need to activate it for
3942 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
3947 * If we fail to add the chunk item here, we end up
3948 * trying again at phase 2 of chunk allocation, at
3949 * btrfs_create_pending_block_groups(). So ignore
3950 * any error here. An ENOSPC here could happen, due to
3951 * the cases described at do_chunk_alloc() - the system
3952 * block group we just created was just turned into RO
3953 * mode by a scrub for example, or a running discard
3954 * temporarily removed its free space entries, etc.
3956 btrfs_chunk_alloc_add_chunk_item(trans, bg);
3961 ret = btrfs_block_rsv_add(fs_info,
3962 &fs_info->chunk_block_rsv,
3963 bytes, BTRFS_RESERVE_NO_FLUSH);
3965 trans->chunk_bytes_reserved += bytes;
3970 * Reserve space in the system space for allocating or removing a chunk.
3971 * The caller must be holding fs_info->chunk_mutex.
3973 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3975 struct btrfs_fs_info *fs_info = trans->fs_info;
3976 const u64 num_devs = get_profile_num_devs(fs_info, type);
3979 /* num_devs device items to update and 1 chunk item to add or remove. */
3980 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
3981 btrfs_calc_insert_metadata_size(fs_info, 1);
3983 reserve_chunk_space(trans, bytes, type);
3987 * Reserve space in the system space, if needed, for doing a modification to the
3990 * @trans: A transaction handle.
3991 * @is_item_insertion: Indicate if the modification is for inserting a new item
3992 * in the chunk btree or if it's for the deletion or update
3993 * of an existing item.
3995 * This is used in a context where we need to update the chunk btree outside
3996 * block group allocation and removal, to avoid a deadlock with a concurrent
3997 * task that is allocating a metadata or data block group and therefore needs to
3998 * update the chunk btree while holding the chunk mutex. After the update to the
3999 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4002 void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4003 bool is_item_insertion)
4005 struct btrfs_fs_info *fs_info = trans->fs_info;
4008 if (is_item_insertion)
4009 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4011 bytes = btrfs_calc_metadata_size(fs_info, 1);
4013 mutex_lock(&fs_info->chunk_mutex);
4014 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4015 mutex_unlock(&fs_info->chunk_mutex);
4018 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4020 struct btrfs_block_group *block_group;
4022 block_group = btrfs_lookup_first_block_group(info, 0);
4023 while (block_group) {
4024 btrfs_wait_block_group_cache_done(block_group);
4025 spin_lock(&block_group->lock);
4026 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4027 &block_group->runtime_flags)) {
4028 struct inode *inode = block_group->inode;
4030 block_group->inode = NULL;
4031 spin_unlock(&block_group->lock);
4033 ASSERT(block_group->io_ctl.inode == NULL);
4036 spin_unlock(&block_group->lock);
4038 block_group = btrfs_next_block_group(block_group);
4043 * Must be called only after stopping all workers, since we could have block
4044 * group caching kthreads running, and therefore they could race with us if we
4045 * freed the block groups before stopping them.
4047 int btrfs_free_block_groups(struct btrfs_fs_info *info)
4049 struct btrfs_block_group *block_group;
4050 struct btrfs_space_info *space_info;
4051 struct btrfs_caching_control *caching_ctl;
4054 write_lock(&info->block_group_cache_lock);
4055 while (!list_empty(&info->caching_block_groups)) {
4056 caching_ctl = list_entry(info->caching_block_groups.next,
4057 struct btrfs_caching_control, list);
4058 list_del(&caching_ctl->list);
4059 btrfs_put_caching_control(caching_ctl);
4061 write_unlock(&info->block_group_cache_lock);
4063 spin_lock(&info->unused_bgs_lock);
4064 while (!list_empty(&info->unused_bgs)) {
4065 block_group = list_first_entry(&info->unused_bgs,
4066 struct btrfs_block_group,
4068 list_del_init(&block_group->bg_list);
4069 btrfs_put_block_group(block_group);
4072 while (!list_empty(&info->reclaim_bgs)) {
4073 block_group = list_first_entry(&info->reclaim_bgs,
4074 struct btrfs_block_group,
4076 list_del_init(&block_group->bg_list);
4077 btrfs_put_block_group(block_group);
4079 spin_unlock(&info->unused_bgs_lock);
4081 spin_lock(&info->zone_active_bgs_lock);
4082 while (!list_empty(&info->zone_active_bgs)) {
4083 block_group = list_first_entry(&info->zone_active_bgs,
4084 struct btrfs_block_group,
4086 list_del_init(&block_group->active_bg_list);
4087 btrfs_put_block_group(block_group);
4089 spin_unlock(&info->zone_active_bgs_lock);
4091 write_lock(&info->block_group_cache_lock);
4092 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4093 block_group = rb_entry(n, struct btrfs_block_group,
4095 rb_erase_cached(&block_group->cache_node,
4096 &info->block_group_cache_tree);
4097 RB_CLEAR_NODE(&block_group->cache_node);
4098 write_unlock(&info->block_group_cache_lock);
4100 down_write(&block_group->space_info->groups_sem);
4101 list_del(&block_group->list);
4102 up_write(&block_group->space_info->groups_sem);
4105 * We haven't cached this block group, which means we could
4106 * possibly have excluded extents on this block group.
4108 if (block_group->cached == BTRFS_CACHE_NO ||
4109 block_group->cached == BTRFS_CACHE_ERROR)
4110 btrfs_free_excluded_extents(block_group);
4112 btrfs_remove_free_space_cache(block_group);
4113 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4114 ASSERT(list_empty(&block_group->dirty_list));
4115 ASSERT(list_empty(&block_group->io_list));
4116 ASSERT(list_empty(&block_group->bg_list));
4117 ASSERT(refcount_read(&block_group->refs) == 1);
4118 ASSERT(block_group->swap_extents == 0);
4119 btrfs_put_block_group(block_group);
4121 write_lock(&info->block_group_cache_lock);
4123 write_unlock(&info->block_group_cache_lock);
4125 btrfs_release_global_block_rsv(info);
4127 while (!list_empty(&info->space_info)) {
4128 space_info = list_entry(info->space_info.next,
4129 struct btrfs_space_info,
4133 * Do not hide this behind enospc_debug, this is actually
4134 * important and indicates a real bug if this happens.
4136 if (WARN_ON(space_info->bytes_pinned > 0 ||
4137 space_info->bytes_may_use > 0))
4138 btrfs_dump_space_info(info, space_info, 0, 0);
4141 * If there was a failure to cleanup a log tree, very likely due
4142 * to an IO failure on a writeback attempt of one or more of its
4143 * extent buffers, we could not do proper (and cheap) unaccounting
4144 * of their reserved space, so don't warn on bytes_reserved > 0 in
4147 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4148 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4149 if (WARN_ON(space_info->bytes_reserved > 0))
4150 btrfs_dump_space_info(info, space_info, 0, 0);
4153 WARN_ON(space_info->reclaim_size > 0);
4154 list_del(&space_info->list);
4155 btrfs_sysfs_remove_space_info(space_info);
4160 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4162 atomic_inc(&cache->frozen);
4165 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4167 struct btrfs_fs_info *fs_info = block_group->fs_info;
4168 struct extent_map_tree *em_tree;
4169 struct extent_map *em;
4172 spin_lock(&block_group->lock);
4173 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4174 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4175 spin_unlock(&block_group->lock);
4178 em_tree = &fs_info->mapping_tree;
4179 write_lock(&em_tree->lock);
4180 em = lookup_extent_mapping(em_tree, block_group->start,
4182 BUG_ON(!em); /* logic error, can't happen */
4183 remove_extent_mapping(em_tree, em);
4184 write_unlock(&em_tree->lock);
4186 /* once for us and once for the tree */
4187 free_extent_map(em);
4188 free_extent_map(em);
4191 * We may have left one free space entry and other possible
4192 * tasks trimming this block group have left 1 entry each one.
4195 btrfs_remove_free_space_cache(block_group);
4199 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4203 spin_lock(&bg->lock);
4208 spin_unlock(&bg->lock);
4213 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4215 spin_lock(&bg->lock);
4217 ASSERT(bg->swap_extents >= amount);
4218 bg->swap_extents -= amount;
4219 spin_unlock(&bg->lock);