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 #ifdef CONFIG_BTRFS_DEBUG
22 int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group)
24 struct btrfs_fs_info *fs_info = block_group->fs_info;
26 return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
27 block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
28 (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
29 block_group->flags & BTRFS_BLOCK_GROUP_DATA);
34 * Return target flags in extended format or 0 if restripe for this chunk_type
37 * Should be called with balance_lock held
39 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
41 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
47 if (flags & BTRFS_BLOCK_GROUP_DATA &&
48 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
49 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
50 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
51 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
52 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
53 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
54 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
55 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
62 * @flags: available profiles in extended format (see ctree.h)
64 * Return reduced profile in chunk format. If profile changing is in progress
65 * (either running or paused) picks the target profile (if it's already
66 * available), otherwise falls back to plain reducing.
68 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
70 u64 num_devices = fs_info->fs_devices->rw_devices;
76 * See if restripe for this chunk_type is in progress, if so try to
77 * reduce to the target profile
79 spin_lock(&fs_info->balance_lock);
80 target = get_restripe_target(fs_info, flags);
82 spin_unlock(&fs_info->balance_lock);
83 return extended_to_chunk(target);
85 spin_unlock(&fs_info->balance_lock);
87 /* First, mask out the RAID levels which aren't possible */
88 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
89 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
90 allowed |= btrfs_raid_array[raid_type].bg_flag;
94 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
95 allowed = BTRFS_BLOCK_GROUP_RAID6;
96 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
97 allowed = BTRFS_BLOCK_GROUP_RAID5;
98 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
99 allowed = BTRFS_BLOCK_GROUP_RAID10;
100 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
101 allowed = BTRFS_BLOCK_GROUP_RAID1;
102 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
103 allowed = BTRFS_BLOCK_GROUP_RAID0;
105 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
107 return extended_to_chunk(flags | allowed);
110 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
117 seq = read_seqbegin(&fs_info->profiles_lock);
119 if (flags & BTRFS_BLOCK_GROUP_DATA)
120 flags |= fs_info->avail_data_alloc_bits;
121 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
122 flags |= fs_info->avail_system_alloc_bits;
123 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
124 flags |= fs_info->avail_metadata_alloc_bits;
125 } while (read_seqretry(&fs_info->profiles_lock, seq));
127 return btrfs_reduce_alloc_profile(fs_info, flags);
130 void btrfs_get_block_group(struct btrfs_block_group *cache)
132 refcount_inc(&cache->refs);
135 void btrfs_put_block_group(struct btrfs_block_group *cache)
137 if (refcount_dec_and_test(&cache->refs)) {
138 WARN_ON(cache->pinned > 0);
140 * If there was a failure to cleanup a log tree, very likely due
141 * to an IO failure on a writeback attempt of one or more of its
142 * extent buffers, we could not do proper (and cheap) unaccounting
143 * of their reserved space, so don't warn on reserved > 0 in that
146 if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
147 !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
148 WARN_ON(cache->reserved > 0);
151 * A block_group shouldn't be on the discard_list anymore.
152 * Remove the block_group from the discard_list to prevent us
153 * from causing a panic due to NULL pointer dereference.
155 if (WARN_ON(!list_empty(&cache->discard_list)))
156 btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
160 * If not empty, someone is still holding mutex of
161 * full_stripe_lock, which can only be released by caller.
162 * And it will definitely cause use-after-free when caller
163 * tries to release full stripe lock.
165 * No better way to resolve, but only to warn.
167 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
168 kfree(cache->free_space_ctl);
169 kfree(cache->physical_map);
175 * This adds the block group to the fs_info rb tree for the block group cache
177 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
178 struct btrfs_block_group *block_group)
181 struct rb_node *parent = NULL;
182 struct btrfs_block_group *cache;
183 bool leftmost = true;
185 ASSERT(block_group->length != 0);
187 write_lock(&info->block_group_cache_lock);
188 p = &info->block_group_cache_tree.rb_root.rb_node;
192 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
193 if (block_group->start < cache->start) {
195 } else if (block_group->start > cache->start) {
199 write_unlock(&info->block_group_cache_lock);
204 rb_link_node(&block_group->cache_node, parent, p);
205 rb_insert_color_cached(&block_group->cache_node,
206 &info->block_group_cache_tree, leftmost);
208 write_unlock(&info->block_group_cache_lock);
214 * This will return the block group at or after bytenr if contains is 0, else
215 * it will return the block group that contains the bytenr
217 static struct btrfs_block_group *block_group_cache_tree_search(
218 struct btrfs_fs_info *info, u64 bytenr, int contains)
220 struct btrfs_block_group *cache, *ret = NULL;
224 read_lock(&info->block_group_cache_lock);
225 n = info->block_group_cache_tree.rb_root.rb_node;
228 cache = rb_entry(n, struct btrfs_block_group, cache_node);
229 end = cache->start + cache->length - 1;
230 start = cache->start;
232 if (bytenr < start) {
233 if (!contains && (!ret || start < ret->start))
236 } else if (bytenr > start) {
237 if (contains && bytenr <= end) {
248 btrfs_get_block_group(ret);
249 read_unlock(&info->block_group_cache_lock);
255 * Return the block group that starts at or after bytenr
257 struct btrfs_block_group *btrfs_lookup_first_block_group(
258 struct btrfs_fs_info *info, u64 bytenr)
260 return block_group_cache_tree_search(info, bytenr, 0);
264 * Return the block group that contains the given bytenr
266 struct btrfs_block_group *btrfs_lookup_block_group(
267 struct btrfs_fs_info *info, u64 bytenr)
269 return block_group_cache_tree_search(info, bytenr, 1);
272 struct btrfs_block_group *btrfs_next_block_group(
273 struct btrfs_block_group *cache)
275 struct btrfs_fs_info *fs_info = cache->fs_info;
276 struct rb_node *node;
278 read_lock(&fs_info->block_group_cache_lock);
280 /* If our block group was removed, we need a full search. */
281 if (RB_EMPTY_NODE(&cache->cache_node)) {
282 const u64 next_bytenr = cache->start + cache->length;
284 read_unlock(&fs_info->block_group_cache_lock);
285 btrfs_put_block_group(cache);
286 return btrfs_lookup_first_block_group(fs_info, next_bytenr);
288 node = rb_next(&cache->cache_node);
289 btrfs_put_block_group(cache);
291 cache = rb_entry(node, struct btrfs_block_group, cache_node);
292 btrfs_get_block_group(cache);
295 read_unlock(&fs_info->block_group_cache_lock);
300 * Check if we can do a NOCOW write for a given extent.
302 * @fs_info: The filesystem information object.
303 * @bytenr: Logical start address of the extent.
305 * Check if we can do a NOCOW write for the given extent, and increments the
306 * number of NOCOW writers in the block group that contains the extent, as long
307 * as the block group exists and it's currently not in read-only mode.
309 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
310 * is responsible for calling btrfs_dec_nocow_writers() later.
312 * Or NULL if we can not do a NOCOW write
314 struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
317 struct btrfs_block_group *bg;
318 bool can_nocow = true;
320 bg = btrfs_lookup_block_group(fs_info, bytenr);
324 spin_lock(&bg->lock);
328 atomic_inc(&bg->nocow_writers);
329 spin_unlock(&bg->lock);
332 btrfs_put_block_group(bg);
336 /* No put on block group, done by btrfs_dec_nocow_writers(). */
341 * Decrement the number of NOCOW writers in a block group.
343 * @bg: The block group.
345 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
346 * and on the block group returned by that call. Typically this is called after
347 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
350 * After this call, the caller should not use the block group anymore. It it wants
351 * to use it, then it should get a reference on it before calling this function.
353 void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
355 if (atomic_dec_and_test(&bg->nocow_writers))
356 wake_up_var(&bg->nocow_writers);
358 /* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
359 btrfs_put_block_group(bg);
362 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
364 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
367 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
370 struct btrfs_block_group *bg;
372 bg = btrfs_lookup_block_group(fs_info, start);
374 if (atomic_dec_and_test(&bg->reservations))
375 wake_up_var(&bg->reservations);
376 btrfs_put_block_group(bg);
379 void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
381 struct btrfs_space_info *space_info = bg->space_info;
385 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
389 * Our block group is read only but before we set it to read only,
390 * some task might have had allocated an extent from it already, but it
391 * has not yet created a respective ordered extent (and added it to a
392 * root's list of ordered extents).
393 * Therefore wait for any task currently allocating extents, since the
394 * block group's reservations counter is incremented while a read lock
395 * on the groups' semaphore is held and decremented after releasing
396 * the read access on that semaphore and creating the ordered extent.
398 down_write(&space_info->groups_sem);
399 up_write(&space_info->groups_sem);
401 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
404 struct btrfs_caching_control *btrfs_get_caching_control(
405 struct btrfs_block_group *cache)
407 struct btrfs_caching_control *ctl;
409 spin_lock(&cache->lock);
410 if (!cache->caching_ctl) {
411 spin_unlock(&cache->lock);
415 ctl = cache->caching_ctl;
416 refcount_inc(&ctl->count);
417 spin_unlock(&cache->lock);
421 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
423 if (refcount_dec_and_test(&ctl->count))
428 * When we wait for progress in the block group caching, its because our
429 * allocation attempt failed at least once. So, we must sleep and let some
430 * progress happen before we try again.
432 * This function will sleep at least once waiting for new free space to show
433 * up, and then it will check the block group free space numbers for our min
434 * num_bytes. Another option is to have it go ahead and look in the rbtree for
435 * a free extent of a given size, but this is a good start.
437 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
438 * any of the information in this block group.
440 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
443 struct btrfs_caching_control *caching_ctl;
445 caching_ctl = btrfs_get_caching_control(cache);
449 wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
450 (cache->free_space_ctl->free_space >= num_bytes));
452 btrfs_put_caching_control(caching_ctl);
455 static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
456 struct btrfs_caching_control *caching_ctl)
458 wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
459 return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
462 static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
464 struct btrfs_caching_control *caching_ctl;
467 caching_ctl = btrfs_get_caching_control(cache);
469 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
470 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
471 btrfs_put_caching_control(caching_ctl);
475 #ifdef CONFIG_BTRFS_DEBUG
476 static void fragment_free_space(struct btrfs_block_group *block_group)
478 struct btrfs_fs_info *fs_info = block_group->fs_info;
479 u64 start = block_group->start;
480 u64 len = block_group->length;
481 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
482 fs_info->nodesize : fs_info->sectorsize;
483 u64 step = chunk << 1;
485 while (len > chunk) {
486 btrfs_remove_free_space(block_group, start, chunk);
497 * This is only called by btrfs_cache_block_group, since we could have freed
498 * extents we need to check the pinned_extents for any extents that can't be
499 * used yet since their free space will be released as soon as the transaction
502 u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
504 struct btrfs_fs_info *info = block_group->fs_info;
505 u64 extent_start, extent_end, size, total_added = 0;
508 while (start < end) {
509 ret = find_first_extent_bit(&info->excluded_extents, start,
510 &extent_start, &extent_end,
511 EXTENT_DIRTY | EXTENT_UPTODATE,
516 if (extent_start <= start) {
517 start = extent_end + 1;
518 } else if (extent_start > start && extent_start < end) {
519 size = extent_start - start;
521 ret = btrfs_add_free_space_async_trimmed(block_group,
523 BUG_ON(ret); /* -ENOMEM or logic error */
524 start = extent_end + 1;
533 ret = btrfs_add_free_space_async_trimmed(block_group, start,
535 BUG_ON(ret); /* -ENOMEM or logic error */
541 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
543 struct btrfs_block_group *block_group = caching_ctl->block_group;
544 struct btrfs_fs_info *fs_info = block_group->fs_info;
545 struct btrfs_root *extent_root;
546 struct btrfs_path *path;
547 struct extent_buffer *leaf;
548 struct btrfs_key key;
555 path = btrfs_alloc_path();
559 last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
560 extent_root = btrfs_extent_root(fs_info, last);
562 #ifdef CONFIG_BTRFS_DEBUG
564 * If we're fragmenting we don't want to make anybody think we can
565 * allocate from this block group until we've had a chance to fragment
568 if (btrfs_should_fragment_free_space(block_group))
572 * We don't want to deadlock with somebody trying to allocate a new
573 * extent for the extent root while also trying to search the extent
574 * root to add free space. So we skip locking and search the commit
575 * root, since its read-only
577 path->skip_locking = 1;
578 path->search_commit_root = 1;
579 path->reada = READA_FORWARD;
583 key.type = BTRFS_EXTENT_ITEM_KEY;
586 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
590 leaf = path->nodes[0];
591 nritems = btrfs_header_nritems(leaf);
594 if (btrfs_fs_closing(fs_info) > 1) {
599 if (path->slots[0] < nritems) {
600 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
602 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
606 if (need_resched() ||
607 rwsem_is_contended(&fs_info->commit_root_sem)) {
608 btrfs_release_path(path);
609 up_read(&fs_info->commit_root_sem);
610 mutex_unlock(&caching_ctl->mutex);
612 mutex_lock(&caching_ctl->mutex);
613 down_read(&fs_info->commit_root_sem);
617 ret = btrfs_next_leaf(extent_root, path);
622 leaf = path->nodes[0];
623 nritems = btrfs_header_nritems(leaf);
627 if (key.objectid < last) {
630 key.type = BTRFS_EXTENT_ITEM_KEY;
631 btrfs_release_path(path);
635 if (key.objectid < block_group->start) {
640 if (key.objectid >= block_group->start + block_group->length)
643 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
644 key.type == BTRFS_METADATA_ITEM_KEY) {
645 total_found += add_new_free_space(block_group, last,
647 if (key.type == BTRFS_METADATA_ITEM_KEY)
648 last = key.objectid +
651 last = key.objectid + key.offset;
653 if (total_found > CACHING_CTL_WAKE_UP) {
656 wake_up(&caching_ctl->wait);
663 total_found += add_new_free_space(block_group, last,
664 block_group->start + block_group->length);
667 btrfs_free_path(path);
671 static noinline void caching_thread(struct btrfs_work *work)
673 struct btrfs_block_group *block_group;
674 struct btrfs_fs_info *fs_info;
675 struct btrfs_caching_control *caching_ctl;
678 caching_ctl = container_of(work, struct btrfs_caching_control, work);
679 block_group = caching_ctl->block_group;
680 fs_info = block_group->fs_info;
682 mutex_lock(&caching_ctl->mutex);
683 down_read(&fs_info->commit_root_sem);
685 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
686 ret = load_free_space_cache(block_group);
693 * We failed to load the space cache, set ourselves to
694 * CACHE_STARTED and carry on.
696 spin_lock(&block_group->lock);
697 block_group->cached = BTRFS_CACHE_STARTED;
698 spin_unlock(&block_group->lock);
699 wake_up(&caching_ctl->wait);
703 * If we are in the transaction that populated the free space tree we
704 * can't actually cache from the free space tree as our commit root and
705 * real root are the same, so we could change the contents of the blocks
706 * while caching. Instead do the slow caching in this case, and after
707 * the transaction has committed we will be safe.
709 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
710 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
711 ret = load_free_space_tree(caching_ctl);
713 ret = load_extent_tree_free(caching_ctl);
715 spin_lock(&block_group->lock);
716 block_group->caching_ctl = NULL;
717 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
718 spin_unlock(&block_group->lock);
720 #ifdef CONFIG_BTRFS_DEBUG
721 if (btrfs_should_fragment_free_space(block_group)) {
724 spin_lock(&block_group->space_info->lock);
725 spin_lock(&block_group->lock);
726 bytes_used = block_group->length - block_group->used;
727 block_group->space_info->bytes_used += bytes_used >> 1;
728 spin_unlock(&block_group->lock);
729 spin_unlock(&block_group->space_info->lock);
730 fragment_free_space(block_group);
734 up_read(&fs_info->commit_root_sem);
735 btrfs_free_excluded_extents(block_group);
736 mutex_unlock(&caching_ctl->mutex);
738 wake_up(&caching_ctl->wait);
740 btrfs_put_caching_control(caching_ctl);
741 btrfs_put_block_group(block_group);
744 int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
746 struct btrfs_fs_info *fs_info = cache->fs_info;
747 struct btrfs_caching_control *caching_ctl = NULL;
750 /* Allocator for zoned filesystems does not use the cache at all */
751 if (btrfs_is_zoned(fs_info))
754 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
758 INIT_LIST_HEAD(&caching_ctl->list);
759 mutex_init(&caching_ctl->mutex);
760 init_waitqueue_head(&caching_ctl->wait);
761 caching_ctl->block_group = cache;
762 refcount_set(&caching_ctl->count, 2);
763 btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
765 spin_lock(&cache->lock);
766 if (cache->cached != BTRFS_CACHE_NO) {
769 caching_ctl = cache->caching_ctl;
771 refcount_inc(&caching_ctl->count);
772 spin_unlock(&cache->lock);
775 WARN_ON(cache->caching_ctl);
776 cache->caching_ctl = caching_ctl;
777 cache->cached = BTRFS_CACHE_STARTED;
778 spin_unlock(&cache->lock);
780 write_lock(&fs_info->block_group_cache_lock);
781 refcount_inc(&caching_ctl->count);
782 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
783 write_unlock(&fs_info->block_group_cache_lock);
785 btrfs_get_block_group(cache);
787 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
789 if (wait && caching_ctl)
790 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
792 btrfs_put_caching_control(caching_ctl);
797 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
799 u64 extra_flags = chunk_to_extended(flags) &
800 BTRFS_EXTENDED_PROFILE_MASK;
802 write_seqlock(&fs_info->profiles_lock);
803 if (flags & BTRFS_BLOCK_GROUP_DATA)
804 fs_info->avail_data_alloc_bits &= ~extra_flags;
805 if (flags & BTRFS_BLOCK_GROUP_METADATA)
806 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
807 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
808 fs_info->avail_system_alloc_bits &= ~extra_flags;
809 write_sequnlock(&fs_info->profiles_lock);
813 * Clear incompat bits for the following feature(s):
815 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
816 * in the whole filesystem
818 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
820 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
822 bool found_raid56 = false;
823 bool found_raid1c34 = false;
825 if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
826 (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
827 (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
828 struct list_head *head = &fs_info->space_info;
829 struct btrfs_space_info *sinfo;
831 list_for_each_entry_rcu(sinfo, head, list) {
832 down_read(&sinfo->groups_sem);
833 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
835 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
837 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
838 found_raid1c34 = true;
839 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
840 found_raid1c34 = true;
841 up_read(&sinfo->groups_sem);
844 btrfs_clear_fs_incompat(fs_info, RAID56);
846 btrfs_clear_fs_incompat(fs_info, RAID1C34);
850 static int remove_block_group_item(struct btrfs_trans_handle *trans,
851 struct btrfs_path *path,
852 struct btrfs_block_group *block_group)
854 struct btrfs_fs_info *fs_info = trans->fs_info;
855 struct btrfs_root *root;
856 struct btrfs_key key;
859 root = btrfs_block_group_root(fs_info);
860 key.objectid = block_group->start;
861 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
862 key.offset = block_group->length;
864 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
870 ret = btrfs_del_item(trans, root, path);
874 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
875 u64 group_start, struct extent_map *em)
877 struct btrfs_fs_info *fs_info = trans->fs_info;
878 struct btrfs_path *path;
879 struct btrfs_block_group *block_group;
880 struct btrfs_free_cluster *cluster;
882 struct kobject *kobj = NULL;
886 struct btrfs_caching_control *caching_ctl = NULL;
888 bool remove_rsv = false;
890 block_group = btrfs_lookup_block_group(fs_info, group_start);
891 BUG_ON(!block_group);
892 BUG_ON(!block_group->ro);
894 trace_btrfs_remove_block_group(block_group);
896 * Free the reserved super bytes from this block group before
899 btrfs_free_excluded_extents(block_group);
900 btrfs_free_ref_tree_range(fs_info, block_group->start,
901 block_group->length);
903 index = btrfs_bg_flags_to_raid_index(block_group->flags);
904 factor = btrfs_bg_type_to_factor(block_group->flags);
906 /* make sure this block group isn't part of an allocation cluster */
907 cluster = &fs_info->data_alloc_cluster;
908 spin_lock(&cluster->refill_lock);
909 btrfs_return_cluster_to_free_space(block_group, cluster);
910 spin_unlock(&cluster->refill_lock);
913 * make sure this block group isn't part of a metadata
916 cluster = &fs_info->meta_alloc_cluster;
917 spin_lock(&cluster->refill_lock);
918 btrfs_return_cluster_to_free_space(block_group, cluster);
919 spin_unlock(&cluster->refill_lock);
921 btrfs_clear_treelog_bg(block_group);
922 btrfs_clear_data_reloc_bg(block_group);
924 path = btrfs_alloc_path();
931 * get the inode first so any iput calls done for the io_list
932 * aren't the final iput (no unlinks allowed now)
934 inode = lookup_free_space_inode(block_group, path);
936 mutex_lock(&trans->transaction->cache_write_mutex);
938 * Make sure our free space cache IO is done before removing the
941 spin_lock(&trans->transaction->dirty_bgs_lock);
942 if (!list_empty(&block_group->io_list)) {
943 list_del_init(&block_group->io_list);
945 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
947 spin_unlock(&trans->transaction->dirty_bgs_lock);
948 btrfs_wait_cache_io(trans, block_group, path);
949 btrfs_put_block_group(block_group);
950 spin_lock(&trans->transaction->dirty_bgs_lock);
953 if (!list_empty(&block_group->dirty_list)) {
954 list_del_init(&block_group->dirty_list);
956 btrfs_put_block_group(block_group);
958 spin_unlock(&trans->transaction->dirty_bgs_lock);
959 mutex_unlock(&trans->transaction->cache_write_mutex);
961 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
965 write_lock(&fs_info->block_group_cache_lock);
966 rb_erase_cached(&block_group->cache_node,
967 &fs_info->block_group_cache_tree);
968 RB_CLEAR_NODE(&block_group->cache_node);
970 /* Once for the block groups rbtree */
971 btrfs_put_block_group(block_group);
973 write_unlock(&fs_info->block_group_cache_lock);
975 down_write(&block_group->space_info->groups_sem);
977 * we must use list_del_init so people can check to see if they
978 * are still on the list after taking the semaphore
980 list_del_init(&block_group->list);
981 if (list_empty(&block_group->space_info->block_groups[index])) {
982 kobj = block_group->space_info->block_group_kobjs[index];
983 block_group->space_info->block_group_kobjs[index] = NULL;
984 clear_avail_alloc_bits(fs_info, block_group->flags);
986 up_write(&block_group->space_info->groups_sem);
987 clear_incompat_bg_bits(fs_info, block_group->flags);
993 if (block_group->cached == BTRFS_CACHE_STARTED)
994 btrfs_wait_block_group_cache_done(block_group);
996 write_lock(&fs_info->block_group_cache_lock);
997 caching_ctl = btrfs_get_caching_control(block_group);
999 struct btrfs_caching_control *ctl;
1001 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1002 if (ctl->block_group == block_group) {
1004 refcount_inc(&caching_ctl->count);
1010 list_del_init(&caching_ctl->list);
1011 write_unlock(&fs_info->block_group_cache_lock);
1014 /* Once for the caching bgs list and once for us. */
1015 btrfs_put_caching_control(caching_ctl);
1016 btrfs_put_caching_control(caching_ctl);
1019 spin_lock(&trans->transaction->dirty_bgs_lock);
1020 WARN_ON(!list_empty(&block_group->dirty_list));
1021 WARN_ON(!list_empty(&block_group->io_list));
1022 spin_unlock(&trans->transaction->dirty_bgs_lock);
1024 btrfs_remove_free_space_cache(block_group);
1026 spin_lock(&block_group->space_info->lock);
1027 list_del_init(&block_group->ro_list);
1029 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1030 WARN_ON(block_group->space_info->total_bytes
1031 < block_group->length);
1032 WARN_ON(block_group->space_info->bytes_readonly
1033 < block_group->length - block_group->zone_unusable);
1034 WARN_ON(block_group->space_info->bytes_zone_unusable
1035 < block_group->zone_unusable);
1036 WARN_ON(block_group->space_info->disk_total
1037 < block_group->length * factor);
1038 WARN_ON(test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
1039 &block_group->runtime_flags) &&
1040 block_group->space_info->active_total_bytes
1041 < block_group->length);
1043 block_group->space_info->total_bytes -= block_group->length;
1044 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
1045 block_group->space_info->active_total_bytes -= block_group->length;
1046 block_group->space_info->bytes_readonly -=
1047 (block_group->length - block_group->zone_unusable);
1048 block_group->space_info->bytes_zone_unusable -=
1049 block_group->zone_unusable;
1050 block_group->space_info->disk_total -= block_group->length * factor;
1052 spin_unlock(&block_group->space_info->lock);
1055 * Remove the free space for the block group from the free space tree
1056 * and the block group's item from the extent tree before marking the
1057 * block group as removed. This is to prevent races with tasks that
1058 * freeze and unfreeze a block group, this task and another task
1059 * allocating a new block group - the unfreeze task ends up removing
1060 * the block group's extent map before the task calling this function
1061 * deletes the block group item from the extent tree, allowing for
1062 * another task to attempt to create another block group with the same
1063 * item key (and failing with -EEXIST and a transaction abort).
1065 ret = remove_block_group_free_space(trans, block_group);
1069 ret = remove_block_group_item(trans, path, block_group);
1073 spin_lock(&block_group->lock);
1074 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1077 * At this point trimming or scrub can't start on this block group,
1078 * because we removed the block group from the rbtree
1079 * fs_info->block_group_cache_tree so no one can't find it anymore and
1080 * even if someone already got this block group before we removed it
1081 * from the rbtree, they have already incremented block_group->frozen -
1082 * if they didn't, for the trimming case they won't find any free space
1083 * entries because we already removed them all when we called
1084 * btrfs_remove_free_space_cache().
1086 * And we must not remove the extent map from the fs_info->mapping_tree
1087 * to prevent the same logical address range and physical device space
1088 * ranges from being reused for a new block group. This is needed to
1089 * avoid races with trimming and scrub.
1091 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1092 * completely transactionless, so while it is trimming a range the
1093 * currently running transaction might finish and a new one start,
1094 * allowing for new block groups to be created that can reuse the same
1095 * physical device locations unless we take this special care.
1097 * There may also be an implicit trim operation if the file system
1098 * is mounted with -odiscard. The same protections must remain
1099 * in place until the extents have been discarded completely when
1100 * the transaction commit has completed.
1102 remove_em = (atomic_read(&block_group->frozen) == 0);
1103 spin_unlock(&block_group->lock);
1106 struct extent_map_tree *em_tree;
1108 em_tree = &fs_info->mapping_tree;
1109 write_lock(&em_tree->lock);
1110 remove_extent_mapping(em_tree, em);
1111 write_unlock(&em_tree->lock);
1112 /* once for the tree */
1113 free_extent_map(em);
1117 /* Once for the lookup reference */
1118 btrfs_put_block_group(block_group);
1120 btrfs_delayed_refs_rsv_release(fs_info, 1);
1121 btrfs_free_path(path);
1125 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1126 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1128 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1129 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1130 struct extent_map *em;
1131 struct map_lookup *map;
1132 unsigned int num_items;
1134 read_lock(&em_tree->lock);
1135 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1136 read_unlock(&em_tree->lock);
1137 ASSERT(em && em->start == chunk_offset);
1140 * We need to reserve 3 + N units from the metadata space info in order
1141 * to remove a block group (done at btrfs_remove_chunk() and at
1142 * btrfs_remove_block_group()), which are used for:
1144 * 1 unit for adding the free space inode's orphan (located in the tree
1146 * 1 unit for deleting the block group item (located in the extent
1148 * 1 unit for deleting the free space item (located in tree of tree
1150 * N units for deleting N device extent items corresponding to each
1151 * stripe (located in the device tree).
1153 * In order to remove a block group we also need to reserve units in the
1154 * system space info in order to update the chunk tree (update one or
1155 * more device items and remove one chunk item), but this is done at
1156 * btrfs_remove_chunk() through a call to check_system_chunk().
1158 map = em->map_lookup;
1159 num_items = 3 + map->num_stripes;
1160 free_extent_map(em);
1162 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1166 * Mark block group @cache read-only, so later write won't happen to block
1169 * If @force is not set, this function will only mark the block group readonly
1170 * if we have enough free space (1M) in other metadata/system block groups.
1171 * If @force is not set, this function will mark the block group readonly
1172 * without checking free space.
1174 * NOTE: This function doesn't care if other block groups can contain all the
1175 * data in this block group. That check should be done by relocation routine,
1176 * not this function.
1178 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1180 struct btrfs_space_info *sinfo = cache->space_info;
1184 spin_lock(&sinfo->lock);
1185 spin_lock(&cache->lock);
1187 if (cache->swap_extents) {
1198 num_bytes = cache->length - cache->reserved - cache->pinned -
1199 cache->bytes_super - cache->zone_unusable - cache->used;
1202 * Data never overcommits, even in mixed mode, so do just the straight
1203 * check of left over space in how much we have allocated.
1207 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1208 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1211 * Here we make sure if we mark this bg RO, we still have enough
1212 * free space as buffer.
1214 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1218 * We overcommit metadata, so we need to do the
1219 * btrfs_can_overcommit check here, and we need to pass in
1220 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1221 * leeway to allow us to mark this block group as read only.
1223 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1224 BTRFS_RESERVE_NO_FLUSH))
1229 sinfo->bytes_readonly += num_bytes;
1230 if (btrfs_is_zoned(cache->fs_info)) {
1231 /* Migrate zone_unusable bytes to readonly */
1232 sinfo->bytes_readonly += cache->zone_unusable;
1233 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1234 cache->zone_unusable = 0;
1237 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1240 spin_unlock(&cache->lock);
1241 spin_unlock(&sinfo->lock);
1242 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1243 btrfs_info(cache->fs_info,
1244 "unable to make block group %llu ro", cache->start);
1245 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1250 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1251 struct btrfs_block_group *bg)
1253 struct btrfs_fs_info *fs_info = bg->fs_info;
1254 struct btrfs_transaction *prev_trans = NULL;
1255 const u64 start = bg->start;
1256 const u64 end = start + bg->length - 1;
1259 spin_lock(&fs_info->trans_lock);
1260 if (trans->transaction->list.prev != &fs_info->trans_list) {
1261 prev_trans = list_last_entry(&trans->transaction->list,
1262 struct btrfs_transaction, list);
1263 refcount_inc(&prev_trans->use_count);
1265 spin_unlock(&fs_info->trans_lock);
1268 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1269 * btrfs_finish_extent_commit(). If we are at transaction N, another
1270 * task might be running finish_extent_commit() for the previous
1271 * transaction N - 1, and have seen a range belonging to the block
1272 * group in pinned_extents before we were able to clear the whole block
1273 * group range from pinned_extents. This means that task can lookup for
1274 * the block group after we unpinned it from pinned_extents and removed
1275 * it, leading to a BUG_ON() at unpin_extent_range().
1277 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1279 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1285 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1288 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1290 btrfs_put_transaction(prev_trans);
1296 * Process the unused_bgs list and remove any that don't have any allocated
1297 * space inside of them.
1299 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1301 struct btrfs_block_group *block_group;
1302 struct btrfs_space_info *space_info;
1303 struct btrfs_trans_handle *trans;
1304 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1307 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1310 if (btrfs_fs_closing(fs_info))
1314 * Long running balances can keep us blocked here for eternity, so
1315 * simply skip deletion if we're unable to get the mutex.
1317 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1320 spin_lock(&fs_info->unused_bgs_lock);
1321 while (!list_empty(&fs_info->unused_bgs)) {
1324 block_group = list_first_entry(&fs_info->unused_bgs,
1325 struct btrfs_block_group,
1327 list_del_init(&block_group->bg_list);
1329 space_info = block_group->space_info;
1331 if (ret || btrfs_mixed_space_info(space_info)) {
1332 btrfs_put_block_group(block_group);
1335 spin_unlock(&fs_info->unused_bgs_lock);
1337 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1339 /* Don't want to race with allocators so take the groups_sem */
1340 down_write(&space_info->groups_sem);
1343 * Async discard moves the final block group discard to be prior
1344 * to the unused_bgs code path. Therefore, if it's not fully
1345 * trimmed, punt it back to the async discard lists.
1347 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1348 !btrfs_is_free_space_trimmed(block_group)) {
1349 trace_btrfs_skip_unused_block_group(block_group);
1350 up_write(&space_info->groups_sem);
1351 /* Requeue if we failed because of async discard */
1352 btrfs_discard_queue_work(&fs_info->discard_ctl,
1357 spin_lock(&block_group->lock);
1358 if (block_group->reserved || block_group->pinned ||
1359 block_group->used || block_group->ro ||
1360 list_is_singular(&block_group->list)) {
1362 * We want to bail if we made new allocations or have
1363 * outstanding allocations in this block group. We do
1364 * the ro check in case balance is currently acting on
1367 trace_btrfs_skip_unused_block_group(block_group);
1368 spin_unlock(&block_group->lock);
1369 up_write(&space_info->groups_sem);
1372 spin_unlock(&block_group->lock);
1374 /* We don't want to force the issue, only flip if it's ok. */
1375 ret = inc_block_group_ro(block_group, 0);
1376 up_write(&space_info->groups_sem);
1382 ret = btrfs_zone_finish(block_group);
1384 btrfs_dec_block_group_ro(block_group);
1391 * Want to do this before we do anything else so we can recover
1392 * properly if we fail to join the transaction.
1394 trans = btrfs_start_trans_remove_block_group(fs_info,
1395 block_group->start);
1396 if (IS_ERR(trans)) {
1397 btrfs_dec_block_group_ro(block_group);
1398 ret = PTR_ERR(trans);
1403 * We could have pending pinned extents for this block group,
1404 * just delete them, we don't care about them anymore.
1406 if (!clean_pinned_extents(trans, block_group)) {
1407 btrfs_dec_block_group_ro(block_group);
1412 * At this point, the block_group is read only and should fail
1413 * new allocations. However, btrfs_finish_extent_commit() can
1414 * cause this block_group to be placed back on the discard
1415 * lists because now the block_group isn't fully discarded.
1416 * Bail here and try again later after discarding everything.
1418 spin_lock(&fs_info->discard_ctl.lock);
1419 if (!list_empty(&block_group->discard_list)) {
1420 spin_unlock(&fs_info->discard_ctl.lock);
1421 btrfs_dec_block_group_ro(block_group);
1422 btrfs_discard_queue_work(&fs_info->discard_ctl,
1426 spin_unlock(&fs_info->discard_ctl.lock);
1428 /* Reset pinned so btrfs_put_block_group doesn't complain */
1429 spin_lock(&space_info->lock);
1430 spin_lock(&block_group->lock);
1432 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1433 -block_group->pinned);
1434 space_info->bytes_readonly += block_group->pinned;
1435 block_group->pinned = 0;
1437 spin_unlock(&block_group->lock);
1438 spin_unlock(&space_info->lock);
1441 * The normal path here is an unused block group is passed here,
1442 * then trimming is handled in the transaction commit path.
1443 * Async discard interposes before this to do the trimming
1444 * before coming down the unused block group path as trimming
1445 * will no longer be done later in the transaction commit path.
1447 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1451 * DISCARD can flip during remount. On zoned filesystems, we
1452 * need to reset sequential-required zones.
1454 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1455 btrfs_is_zoned(fs_info);
1457 /* Implicit trim during transaction commit. */
1459 btrfs_freeze_block_group(block_group);
1462 * Btrfs_remove_chunk will abort the transaction if things go
1465 ret = btrfs_remove_chunk(trans, block_group->start);
1469 btrfs_unfreeze_block_group(block_group);
1474 * If we're not mounted with -odiscard, we can just forget
1475 * about this block group. Otherwise we'll need to wait
1476 * until transaction commit to do the actual discard.
1479 spin_lock(&fs_info->unused_bgs_lock);
1481 * A concurrent scrub might have added us to the list
1482 * fs_info->unused_bgs, so use a list_move operation
1483 * to add the block group to the deleted_bgs list.
1485 list_move(&block_group->bg_list,
1486 &trans->transaction->deleted_bgs);
1487 spin_unlock(&fs_info->unused_bgs_lock);
1488 btrfs_get_block_group(block_group);
1491 btrfs_end_transaction(trans);
1493 btrfs_put_block_group(block_group);
1494 spin_lock(&fs_info->unused_bgs_lock);
1496 spin_unlock(&fs_info->unused_bgs_lock);
1497 mutex_unlock(&fs_info->reclaim_bgs_lock);
1501 btrfs_end_transaction(trans);
1502 mutex_unlock(&fs_info->reclaim_bgs_lock);
1503 btrfs_put_block_group(block_group);
1504 btrfs_discard_punt_unused_bgs_list(fs_info);
1507 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1509 struct btrfs_fs_info *fs_info = bg->fs_info;
1511 spin_lock(&fs_info->unused_bgs_lock);
1512 if (list_empty(&bg->bg_list)) {
1513 btrfs_get_block_group(bg);
1514 trace_btrfs_add_unused_block_group(bg);
1515 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1517 spin_unlock(&fs_info->unused_bgs_lock);
1521 * We want block groups with a low number of used bytes to be in the beginning
1522 * of the list, so they will get reclaimed first.
1524 static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1525 const struct list_head *b)
1527 const struct btrfs_block_group *bg1, *bg2;
1529 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1530 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1532 return bg1->used > bg2->used;
1535 static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1537 if (btrfs_is_zoned(fs_info))
1538 return btrfs_zoned_should_reclaim(fs_info);
1542 static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1544 const struct btrfs_space_info *space_info = bg->space_info;
1545 const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1546 const u64 new_val = bg->used;
1547 const u64 old_val = new_val + bytes_freed;
1550 if (reclaim_thresh == 0)
1553 thresh = div_factor_fine(bg->length, reclaim_thresh);
1556 * If we were below the threshold before don't reclaim, we are likely a
1557 * brand new block group and we don't want to relocate new block groups.
1559 if (old_val < thresh)
1561 if (new_val >= thresh)
1566 void btrfs_reclaim_bgs_work(struct work_struct *work)
1568 struct btrfs_fs_info *fs_info =
1569 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1570 struct btrfs_block_group *bg;
1571 struct btrfs_space_info *space_info;
1573 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1576 if (btrfs_fs_closing(fs_info))
1579 if (!btrfs_should_reclaim(fs_info))
1582 sb_start_write(fs_info->sb);
1584 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1585 sb_end_write(fs_info->sb);
1590 * Long running balances can keep us blocked here for eternity, so
1591 * simply skip reclaim if we're unable to get the mutex.
1593 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1594 btrfs_exclop_finish(fs_info);
1595 sb_end_write(fs_info->sb);
1599 spin_lock(&fs_info->unused_bgs_lock);
1601 * Sort happens under lock because we can't simply splice it and sort.
1602 * The block groups might still be in use and reachable via bg_list,
1603 * and their presence in the reclaim_bgs list must be preserved.
1605 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1606 while (!list_empty(&fs_info->reclaim_bgs)) {
1610 bg = list_first_entry(&fs_info->reclaim_bgs,
1611 struct btrfs_block_group,
1613 list_del_init(&bg->bg_list);
1615 space_info = bg->space_info;
1616 spin_unlock(&fs_info->unused_bgs_lock);
1618 /* Don't race with allocators so take the groups_sem */
1619 down_write(&space_info->groups_sem);
1621 spin_lock(&bg->lock);
1622 if (bg->reserved || bg->pinned || bg->ro) {
1624 * We want to bail if we made new allocations or have
1625 * outstanding allocations in this block group. We do
1626 * the ro check in case balance is currently acting on
1629 spin_unlock(&bg->lock);
1630 up_write(&space_info->groups_sem);
1633 if (bg->used == 0) {
1635 * It is possible that we trigger relocation on a block
1636 * group as its extents are deleted and it first goes
1637 * below the threshold, then shortly after goes empty.
1639 * In this case, relocating it does delete it, but has
1640 * some overhead in relocation specific metadata, looking
1641 * for the non-existent extents and running some extra
1642 * transactions, which we can avoid by using one of the
1643 * other mechanisms for dealing with empty block groups.
1645 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1646 btrfs_mark_bg_unused(bg);
1647 spin_unlock(&bg->lock);
1648 up_write(&space_info->groups_sem);
1653 * The block group might no longer meet the reclaim condition by
1654 * the time we get around to reclaiming it, so to avoid
1655 * reclaiming overly full block_groups, skip reclaiming them.
1657 * Since the decision making process also depends on the amount
1658 * being freed, pass in a fake giant value to skip that extra
1659 * check, which is more meaningful when adding to the list in
1662 if (!should_reclaim_block_group(bg, bg->length)) {
1663 spin_unlock(&bg->lock);
1664 up_write(&space_info->groups_sem);
1667 spin_unlock(&bg->lock);
1669 /* Get out fast, in case we're unmounting the filesystem */
1670 if (btrfs_fs_closing(fs_info)) {
1671 up_write(&space_info->groups_sem);
1676 * Cache the zone_unusable value before turning the block group
1677 * to read only. As soon as the blog group is read only it's
1678 * zone_unusable value gets moved to the block group's read-only
1679 * bytes and isn't available for calculations anymore.
1681 zone_unusable = bg->zone_unusable;
1682 ret = inc_block_group_ro(bg, 0);
1683 up_write(&space_info->groups_sem);
1688 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1689 bg->start, div_u64(bg->used * 100, bg->length),
1690 div64_u64(zone_unusable * 100, bg->length));
1691 trace_btrfs_reclaim_block_group(bg);
1692 ret = btrfs_relocate_chunk(fs_info, bg->start);
1694 btrfs_dec_block_group_ro(bg);
1695 btrfs_err(fs_info, "error relocating chunk %llu",
1700 btrfs_put_block_group(bg);
1701 spin_lock(&fs_info->unused_bgs_lock);
1703 spin_unlock(&fs_info->unused_bgs_lock);
1704 mutex_unlock(&fs_info->reclaim_bgs_lock);
1705 btrfs_exclop_finish(fs_info);
1706 sb_end_write(fs_info->sb);
1709 void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1711 spin_lock(&fs_info->unused_bgs_lock);
1712 if (!list_empty(&fs_info->reclaim_bgs))
1713 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1714 spin_unlock(&fs_info->unused_bgs_lock);
1717 void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1719 struct btrfs_fs_info *fs_info = bg->fs_info;
1721 spin_lock(&fs_info->unused_bgs_lock);
1722 if (list_empty(&bg->bg_list)) {
1723 btrfs_get_block_group(bg);
1724 trace_btrfs_add_reclaim_block_group(bg);
1725 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1727 spin_unlock(&fs_info->unused_bgs_lock);
1730 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1731 struct btrfs_path *path)
1733 struct extent_map_tree *em_tree;
1734 struct extent_map *em;
1735 struct btrfs_block_group_item bg;
1736 struct extent_buffer *leaf;
1741 slot = path->slots[0];
1742 leaf = path->nodes[0];
1744 em_tree = &fs_info->mapping_tree;
1745 read_lock(&em_tree->lock);
1746 em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1747 read_unlock(&em_tree->lock);
1750 "logical %llu len %llu found bg but no related chunk",
1751 key->objectid, key->offset);
1755 if (em->start != key->objectid || em->len != key->offset) {
1757 "block group %llu len %llu mismatch with chunk %llu len %llu",
1758 key->objectid, key->offset, em->start, em->len);
1763 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1765 flags = btrfs_stack_block_group_flags(&bg) &
1766 BTRFS_BLOCK_GROUP_TYPE_MASK;
1768 if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1770 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1771 key->objectid, key->offset, flags,
1772 (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1777 free_extent_map(em);
1781 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1782 struct btrfs_path *path,
1783 struct btrfs_key *key)
1785 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1787 struct btrfs_key found_key;
1789 btrfs_for_each_slot(root, key, &found_key, path, ret) {
1790 if (found_key.objectid >= key->objectid &&
1791 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1792 return read_bg_from_eb(fs_info, &found_key, path);
1798 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1800 u64 extra_flags = chunk_to_extended(flags) &
1801 BTRFS_EXTENDED_PROFILE_MASK;
1803 write_seqlock(&fs_info->profiles_lock);
1804 if (flags & BTRFS_BLOCK_GROUP_DATA)
1805 fs_info->avail_data_alloc_bits |= extra_flags;
1806 if (flags & BTRFS_BLOCK_GROUP_METADATA)
1807 fs_info->avail_metadata_alloc_bits |= extra_flags;
1808 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1809 fs_info->avail_system_alloc_bits |= extra_flags;
1810 write_sequnlock(&fs_info->profiles_lock);
1814 * Map a physical disk address to a list of logical addresses
1816 * @fs_info: the filesystem
1817 * @chunk_start: logical address of block group
1818 * @bdev: physical device to resolve, can be NULL to indicate any device
1819 * @physical: physical address to map to logical addresses
1820 * @logical: return array of logical addresses which map to @physical
1821 * @naddrs: length of @logical
1822 * @stripe_len: size of IO stripe for the given block group
1824 * Maps a particular @physical disk address to a list of @logical addresses.
1825 * Used primarily to exclude those portions of a block group that contain super
1828 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1829 struct block_device *bdev, u64 physical, u64 **logical,
1830 int *naddrs, int *stripe_len)
1832 struct extent_map *em;
1833 struct map_lookup *map;
1836 u64 data_stripe_length;
1841 em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1845 map = em->map_lookup;
1846 data_stripe_length = em->orig_block_len;
1847 io_stripe_size = map->stripe_len;
1848 chunk_start = em->start;
1850 /* For RAID5/6 adjust to a full IO stripe length */
1851 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1852 io_stripe_size = map->stripe_len * nr_data_stripes(map);
1854 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1860 for (i = 0; i < map->num_stripes; i++) {
1861 bool already_inserted = false;
1866 if (!in_range(physical, map->stripes[i].physical,
1867 data_stripe_length))
1870 if (bdev && map->stripes[i].dev->bdev != bdev)
1873 stripe_nr = physical - map->stripes[i].physical;
1874 stripe_nr = div64_u64_rem(stripe_nr, map->stripe_len, &offset);
1876 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
1877 BTRFS_BLOCK_GROUP_RAID10)) {
1878 stripe_nr = stripe_nr * map->num_stripes + i;
1879 stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1882 * The remaining case would be for RAID56, multiply by
1883 * nr_data_stripes(). Alternatively, just use rmap_len below
1884 * instead of map->stripe_len
1887 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
1889 /* Ensure we don't add duplicate addresses */
1890 for (j = 0; j < nr; j++) {
1891 if (buf[j] == bytenr) {
1892 already_inserted = true;
1897 if (!already_inserted)
1903 *stripe_len = io_stripe_size;
1905 free_extent_map(em);
1909 static int exclude_super_stripes(struct btrfs_block_group *cache)
1911 struct btrfs_fs_info *fs_info = cache->fs_info;
1912 const bool zoned = btrfs_is_zoned(fs_info);
1918 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1919 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1920 cache->bytes_super += stripe_len;
1921 ret = btrfs_add_excluded_extent(fs_info, cache->start,
1927 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1928 bytenr = btrfs_sb_offset(i);
1929 ret = btrfs_rmap_block(fs_info, cache->start, NULL,
1930 bytenr, &logical, &nr, &stripe_len);
1934 /* Shouldn't have super stripes in sequential zones */
1937 "zoned: block group %llu must not contain super block",
1943 u64 len = min_t(u64, stripe_len,
1944 cache->start + cache->length - logical[nr]);
1946 cache->bytes_super += len;
1947 ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1960 static struct btrfs_block_group *btrfs_create_block_group_cache(
1961 struct btrfs_fs_info *fs_info, u64 start)
1963 struct btrfs_block_group *cache;
1965 cache = kzalloc(sizeof(*cache), GFP_NOFS);
1969 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1971 if (!cache->free_space_ctl) {
1976 cache->start = start;
1978 cache->fs_info = fs_info;
1979 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1981 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1983 refcount_set(&cache->refs, 1);
1984 spin_lock_init(&cache->lock);
1985 init_rwsem(&cache->data_rwsem);
1986 INIT_LIST_HEAD(&cache->list);
1987 INIT_LIST_HEAD(&cache->cluster_list);
1988 INIT_LIST_HEAD(&cache->bg_list);
1989 INIT_LIST_HEAD(&cache->ro_list);
1990 INIT_LIST_HEAD(&cache->discard_list);
1991 INIT_LIST_HEAD(&cache->dirty_list);
1992 INIT_LIST_HEAD(&cache->io_list);
1993 INIT_LIST_HEAD(&cache->active_bg_list);
1994 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1995 atomic_set(&cache->frozen, 0);
1996 mutex_init(&cache->free_space_lock);
1997 cache->full_stripe_locks_root.root = RB_ROOT;
1998 mutex_init(&cache->full_stripe_locks_root.lock);
2004 * Iterate all chunks and verify that each of them has the corresponding block
2007 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2009 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
2010 struct extent_map *em;
2011 struct btrfs_block_group *bg;
2016 read_lock(&map_tree->lock);
2018 * lookup_extent_mapping will return the first extent map
2019 * intersecting the range, so setting @len to 1 is enough to
2020 * get the first chunk.
2022 em = lookup_extent_mapping(map_tree, start, 1);
2023 read_unlock(&map_tree->lock);
2027 bg = btrfs_lookup_block_group(fs_info, em->start);
2030 "chunk start=%llu len=%llu doesn't have corresponding block group",
2031 em->start, em->len);
2033 free_extent_map(em);
2036 if (bg->start != em->start || bg->length != em->len ||
2037 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2038 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2040 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2042 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2043 bg->start, bg->length,
2044 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2046 free_extent_map(em);
2047 btrfs_put_block_group(bg);
2050 start = em->start + em->len;
2051 free_extent_map(em);
2052 btrfs_put_block_group(bg);
2057 static int read_one_block_group(struct btrfs_fs_info *info,
2058 struct btrfs_block_group_item *bgi,
2059 const struct btrfs_key *key,
2062 struct btrfs_block_group *cache;
2063 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2066 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2068 cache = btrfs_create_block_group_cache(info, key->objectid);
2072 cache->length = key->offset;
2073 cache->used = btrfs_stack_block_group_used(bgi);
2074 cache->commit_used = cache->used;
2075 cache->flags = btrfs_stack_block_group_flags(bgi);
2076 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2078 set_free_space_tree_thresholds(cache);
2082 * When we mount with old space cache, we need to
2083 * set BTRFS_DC_CLEAR and set dirty flag.
2085 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2086 * truncate the old free space cache inode and
2088 * b) Setting 'dirty flag' makes sure that we flush
2089 * the new space cache info onto disk.
2091 if (btrfs_test_opt(info, SPACE_CACHE))
2092 cache->disk_cache_state = BTRFS_DC_CLEAR;
2094 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2095 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2097 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2103 ret = btrfs_load_block_group_zone_info(cache, false);
2105 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2111 * We need to exclude the super stripes now so that the space info has
2112 * super bytes accounted for, otherwise we'll think we have more space
2113 * than we actually do.
2115 ret = exclude_super_stripes(cache);
2117 /* We may have excluded something, so call this just in case. */
2118 btrfs_free_excluded_extents(cache);
2123 * For zoned filesystem, space after the allocation offset is the only
2124 * free space for a block group. So, we don't need any caching work.
2125 * btrfs_calc_zone_unusable() will set the amount of free space and
2126 * zone_unusable space.
2128 * For regular filesystem, check for two cases, either we are full, and
2129 * therefore don't need to bother with the caching work since we won't
2130 * find any space, or we are empty, and we can just add all the space
2131 * in and be done with it. This saves us _a_lot_ of time, particularly
2134 if (btrfs_is_zoned(info)) {
2135 btrfs_calc_zone_unusable(cache);
2136 /* Should not have any excluded extents. Just in case, though. */
2137 btrfs_free_excluded_extents(cache);
2138 } else if (cache->length == cache->used) {
2139 cache->cached = BTRFS_CACHE_FINISHED;
2140 btrfs_free_excluded_extents(cache);
2141 } else if (cache->used == 0) {
2142 cache->cached = BTRFS_CACHE_FINISHED;
2143 add_new_free_space(cache, cache->start,
2144 cache->start + cache->length);
2145 btrfs_free_excluded_extents(cache);
2148 ret = btrfs_add_block_group_cache(info, cache);
2150 btrfs_remove_free_space_cache(cache);
2153 trace_btrfs_add_block_group(info, cache, 0);
2154 btrfs_add_bg_to_space_info(info, cache);
2156 set_avail_alloc_bits(info, cache->flags);
2157 if (btrfs_chunk_writeable(info, cache->start)) {
2158 if (cache->used == 0) {
2159 ASSERT(list_empty(&cache->bg_list));
2160 if (btrfs_test_opt(info, DISCARD_ASYNC))
2161 btrfs_discard_queue_work(&info->discard_ctl, cache);
2163 btrfs_mark_bg_unused(cache);
2166 inc_block_group_ro(cache, 1);
2171 btrfs_put_block_group(cache);
2175 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2177 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
2178 struct rb_node *node;
2181 for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
2182 struct extent_map *em;
2183 struct map_lookup *map;
2184 struct btrfs_block_group *bg;
2186 em = rb_entry(node, struct extent_map, rb_node);
2187 map = em->map_lookup;
2188 bg = btrfs_create_block_group_cache(fs_info, em->start);
2194 /* Fill dummy cache as FULL */
2195 bg->length = em->len;
2196 bg->flags = map->type;
2197 bg->cached = BTRFS_CACHE_FINISHED;
2199 bg->flags = map->type;
2200 ret = btrfs_add_block_group_cache(fs_info, bg);
2202 * We may have some valid block group cache added already, in
2203 * that case we skip to the next one.
2205 if (ret == -EEXIST) {
2207 btrfs_put_block_group(bg);
2212 btrfs_remove_free_space_cache(bg);
2213 btrfs_put_block_group(bg);
2217 btrfs_add_bg_to_space_info(fs_info, bg);
2219 set_avail_alloc_bits(fs_info, bg->flags);
2222 btrfs_init_global_block_rsv(fs_info);
2226 int btrfs_read_block_groups(struct btrfs_fs_info *info)
2228 struct btrfs_root *root = btrfs_block_group_root(info);
2229 struct btrfs_path *path;
2231 struct btrfs_block_group *cache;
2232 struct btrfs_space_info *space_info;
2233 struct btrfs_key key;
2238 * Either no extent root (with ibadroots rescue option) or we have
2239 * unsupported RO options. The fs can never be mounted read-write, so no
2240 * need to waste time searching block group items.
2242 * This also allows new extent tree related changes to be RO compat,
2243 * no need for a full incompat flag.
2245 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2246 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2247 return fill_dummy_bgs(info);
2251 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2252 path = btrfs_alloc_path();
2256 cache_gen = btrfs_super_cache_generation(info->super_copy);
2257 if (btrfs_test_opt(info, SPACE_CACHE) &&
2258 btrfs_super_generation(info->super_copy) != cache_gen)
2260 if (btrfs_test_opt(info, CLEAR_CACHE))
2264 struct btrfs_block_group_item bgi;
2265 struct extent_buffer *leaf;
2268 ret = find_first_block_group(info, path, &key);
2274 leaf = path->nodes[0];
2275 slot = path->slots[0];
2277 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2280 btrfs_item_key_to_cpu(leaf, &key, slot);
2281 btrfs_release_path(path);
2282 ret = read_one_block_group(info, &bgi, &key, need_clear);
2285 key.objectid += key.offset;
2288 btrfs_release_path(path);
2290 list_for_each_entry(space_info, &info->space_info, list) {
2293 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2294 if (list_empty(&space_info->block_groups[i]))
2296 cache = list_first_entry(&space_info->block_groups[i],
2297 struct btrfs_block_group,
2299 btrfs_sysfs_add_block_group_type(cache);
2302 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2303 (BTRFS_BLOCK_GROUP_RAID10 |
2304 BTRFS_BLOCK_GROUP_RAID1_MASK |
2305 BTRFS_BLOCK_GROUP_RAID56_MASK |
2306 BTRFS_BLOCK_GROUP_DUP)))
2309 * Avoid allocating from un-mirrored block group if there are
2310 * mirrored block groups.
2312 list_for_each_entry(cache,
2313 &space_info->block_groups[BTRFS_RAID_RAID0],
2315 inc_block_group_ro(cache, 1);
2316 list_for_each_entry(cache,
2317 &space_info->block_groups[BTRFS_RAID_SINGLE],
2319 inc_block_group_ro(cache, 1);
2322 btrfs_init_global_block_rsv(info);
2323 ret = check_chunk_block_group_mappings(info);
2325 btrfs_free_path(path);
2327 * We've hit some error while reading the extent tree, and have
2328 * rescue=ibadroots mount option.
2329 * Try to fill the tree using dummy block groups so that the user can
2330 * continue to mount and grab their data.
2332 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2333 ret = fill_dummy_bgs(info);
2338 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2341 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2344 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2345 struct btrfs_block_group *block_group)
2347 struct btrfs_fs_info *fs_info = trans->fs_info;
2348 struct btrfs_block_group_item bgi;
2349 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2350 struct btrfs_key key;
2352 spin_lock(&block_group->lock);
2353 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2354 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2355 block_group->global_root_id);
2356 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2357 key.objectid = block_group->start;
2358 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2359 key.offset = block_group->length;
2360 spin_unlock(&block_group->lock);
2362 return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2365 static int insert_dev_extent(struct btrfs_trans_handle *trans,
2366 struct btrfs_device *device, u64 chunk_offset,
2367 u64 start, u64 num_bytes)
2369 struct btrfs_fs_info *fs_info = device->fs_info;
2370 struct btrfs_root *root = fs_info->dev_root;
2371 struct btrfs_path *path;
2372 struct btrfs_dev_extent *extent;
2373 struct extent_buffer *leaf;
2374 struct btrfs_key key;
2377 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2378 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2379 path = btrfs_alloc_path();
2383 key.objectid = device->devid;
2384 key.type = BTRFS_DEV_EXTENT_KEY;
2386 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2390 leaf = path->nodes[0];
2391 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2392 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2393 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2394 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2395 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2397 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2398 btrfs_mark_buffer_dirty(leaf);
2400 btrfs_free_path(path);
2405 * This function belongs to phase 2.
2407 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2410 static int insert_dev_extents(struct btrfs_trans_handle *trans,
2411 u64 chunk_offset, u64 chunk_size)
2413 struct btrfs_fs_info *fs_info = trans->fs_info;
2414 struct btrfs_device *device;
2415 struct extent_map *em;
2416 struct map_lookup *map;
2422 em = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2426 map = em->map_lookup;
2427 stripe_size = em->orig_block_len;
2430 * Take the device list mutex to prevent races with the final phase of
2431 * a device replace operation that replaces the device object associated
2432 * with the map's stripes, because the device object's id can change
2433 * at any time during that final phase of the device replace operation
2434 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2435 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2436 * resulting in persisting a device extent item with such ID.
2438 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2439 for (i = 0; i < map->num_stripes; i++) {
2440 device = map->stripes[i].dev;
2441 dev_offset = map->stripes[i].physical;
2443 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2448 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2450 free_extent_map(em);
2455 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2458 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2461 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2463 struct btrfs_fs_info *fs_info = trans->fs_info;
2464 struct btrfs_block_group *block_group;
2467 while (!list_empty(&trans->new_bgs)) {
2470 block_group = list_first_entry(&trans->new_bgs,
2471 struct btrfs_block_group,
2476 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2478 ret = insert_block_group_item(trans, block_group);
2480 btrfs_abort_transaction(trans, ret);
2481 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2482 &block_group->runtime_flags)) {
2483 mutex_lock(&fs_info->chunk_mutex);
2484 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2485 mutex_unlock(&fs_info->chunk_mutex);
2487 btrfs_abort_transaction(trans, ret);
2489 ret = insert_dev_extents(trans, block_group->start,
2490 block_group->length);
2492 btrfs_abort_transaction(trans, ret);
2493 add_block_group_free_space(trans, block_group);
2496 * If we restriped during balance, we may have added a new raid
2497 * type, so now add the sysfs entries when it is safe to do so.
2498 * We don't have to worry about locking here as it's handled in
2499 * btrfs_sysfs_add_block_group_type.
2501 if (block_group->space_info->block_group_kobjs[index] == NULL)
2502 btrfs_sysfs_add_block_group_type(block_group);
2504 /* Already aborted the transaction if it failed. */
2506 btrfs_delayed_refs_rsv_release(fs_info, 1);
2507 list_del_init(&block_group->bg_list);
2509 btrfs_trans_release_chunk_metadata(trans);
2513 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2514 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2516 static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2521 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2522 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2524 /* If we have a smaller fs index based on 128MiB. */
2525 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2528 offset = div64_u64(offset, div);
2529 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2533 struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2534 u64 bytes_used, u64 type,
2535 u64 chunk_offset, u64 size)
2537 struct btrfs_fs_info *fs_info = trans->fs_info;
2538 struct btrfs_block_group *cache;
2541 btrfs_set_log_full_commit(trans);
2543 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2545 return ERR_PTR(-ENOMEM);
2547 cache->length = size;
2548 set_free_space_tree_thresholds(cache);
2549 cache->used = bytes_used;
2550 cache->flags = type;
2551 cache->cached = BTRFS_CACHE_FINISHED;
2552 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2554 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2555 cache->needs_free_space = 1;
2557 ret = btrfs_load_block_group_zone_info(cache, true);
2559 btrfs_put_block_group(cache);
2560 return ERR_PTR(ret);
2563 ret = exclude_super_stripes(cache);
2565 /* We may have excluded something, so call this just in case */
2566 btrfs_free_excluded_extents(cache);
2567 btrfs_put_block_group(cache);
2568 return ERR_PTR(ret);
2571 add_new_free_space(cache, chunk_offset, chunk_offset + size);
2573 btrfs_free_excluded_extents(cache);
2576 * Ensure the corresponding space_info object is created and
2577 * assigned to our block group. We want our bg to be added to the rbtree
2578 * with its ->space_info set.
2580 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2581 ASSERT(cache->space_info);
2583 ret = btrfs_add_block_group_cache(fs_info, cache);
2585 btrfs_remove_free_space_cache(cache);
2586 btrfs_put_block_group(cache);
2587 return ERR_PTR(ret);
2591 * Now that our block group has its ->space_info set and is inserted in
2592 * the rbtree, update the space info's counters.
2594 trace_btrfs_add_block_group(fs_info, cache, 1);
2595 btrfs_add_bg_to_space_info(fs_info, cache);
2596 btrfs_update_global_block_rsv(fs_info);
2598 #ifdef CONFIG_BTRFS_DEBUG
2599 if (btrfs_should_fragment_free_space(cache)) {
2600 u64 new_bytes_used = size - bytes_used;
2602 cache->space_info->bytes_used += new_bytes_used >> 1;
2603 fragment_free_space(cache);
2607 list_add_tail(&cache->bg_list, &trans->new_bgs);
2608 trans->delayed_ref_updates++;
2609 btrfs_update_delayed_refs_rsv(trans);
2611 set_avail_alloc_bits(fs_info, type);
2616 * Mark one block group RO, can be called several times for the same block
2619 * @cache: the destination block group
2620 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2621 * ensure we still have some free space after marking this
2624 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2625 bool do_chunk_alloc)
2627 struct btrfs_fs_info *fs_info = cache->fs_info;
2628 struct btrfs_trans_handle *trans;
2629 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2632 bool dirty_bg_running;
2635 * This can only happen when we are doing read-only scrub on read-only
2637 * In that case we should not start a new transaction on read-only fs.
2638 * Thus here we skip all chunk allocations.
2640 if (sb_rdonly(fs_info->sb)) {
2641 mutex_lock(&fs_info->ro_block_group_mutex);
2642 ret = inc_block_group_ro(cache, 0);
2643 mutex_unlock(&fs_info->ro_block_group_mutex);
2648 trans = btrfs_join_transaction(root);
2650 return PTR_ERR(trans);
2652 dirty_bg_running = false;
2655 * We're not allowed to set block groups readonly after the dirty
2656 * block group cache has started writing. If it already started,
2657 * back off and let this transaction commit.
2659 mutex_lock(&fs_info->ro_block_group_mutex);
2660 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2661 u64 transid = trans->transid;
2663 mutex_unlock(&fs_info->ro_block_group_mutex);
2664 btrfs_end_transaction(trans);
2666 ret = btrfs_wait_for_commit(fs_info, transid);
2669 dirty_bg_running = true;
2671 } while (dirty_bg_running);
2673 if (do_chunk_alloc) {
2675 * If we are changing raid levels, try to allocate a
2676 * corresponding block group with the new raid level.
2678 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2679 if (alloc_flags != cache->flags) {
2680 ret = btrfs_chunk_alloc(trans, alloc_flags,
2683 * ENOSPC is allowed here, we may have enough space
2684 * already allocated at the new raid level to carry on
2693 ret = inc_block_group_ro(cache, 0);
2694 if (!do_chunk_alloc || ret == -ETXTBSY)
2698 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2699 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2703 * We have allocated a new chunk. We also need to activate that chunk to
2704 * grant metadata tickets for zoned filesystem.
2706 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2710 ret = inc_block_group_ro(cache, 0);
2711 if (ret == -ETXTBSY)
2714 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2715 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2716 mutex_lock(&fs_info->chunk_mutex);
2717 check_system_chunk(trans, alloc_flags);
2718 mutex_unlock(&fs_info->chunk_mutex);
2721 mutex_unlock(&fs_info->ro_block_group_mutex);
2723 btrfs_end_transaction(trans);
2727 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2729 struct btrfs_space_info *sinfo = cache->space_info;
2734 spin_lock(&sinfo->lock);
2735 spin_lock(&cache->lock);
2737 if (btrfs_is_zoned(cache->fs_info)) {
2738 /* Migrate zone_unusable bytes back */
2739 cache->zone_unusable =
2740 (cache->alloc_offset - cache->used) +
2741 (cache->length - cache->zone_capacity);
2742 sinfo->bytes_zone_unusable += cache->zone_unusable;
2743 sinfo->bytes_readonly -= cache->zone_unusable;
2745 num_bytes = cache->length - cache->reserved -
2746 cache->pinned - cache->bytes_super -
2747 cache->zone_unusable - cache->used;
2748 sinfo->bytes_readonly -= num_bytes;
2749 list_del_init(&cache->ro_list);
2751 spin_unlock(&cache->lock);
2752 spin_unlock(&sinfo->lock);
2755 static int update_block_group_item(struct btrfs_trans_handle *trans,
2756 struct btrfs_path *path,
2757 struct btrfs_block_group *cache)
2759 struct btrfs_fs_info *fs_info = trans->fs_info;
2761 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2763 struct extent_buffer *leaf;
2764 struct btrfs_block_group_item bgi;
2765 struct btrfs_key key;
2766 u64 old_commit_used;
2770 * Block group items update can be triggered out of commit transaction
2771 * critical section, thus we need a consistent view of used bytes.
2772 * We cannot use cache->used directly outside of the spin lock, as it
2775 spin_lock(&cache->lock);
2776 old_commit_used = cache->commit_used;
2778 /* No change in used bytes, can safely skip it. */
2779 if (cache->commit_used == used) {
2780 spin_unlock(&cache->lock);
2783 cache->commit_used = used;
2784 spin_unlock(&cache->lock);
2786 key.objectid = cache->start;
2787 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2788 key.offset = cache->length;
2790 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2797 leaf = path->nodes[0];
2798 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2799 btrfs_set_stack_block_group_used(&bgi, used);
2800 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2801 cache->global_root_id);
2802 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2803 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2804 btrfs_mark_buffer_dirty(leaf);
2806 btrfs_release_path(path);
2807 /* We didn't update the block group item, need to revert @commit_used. */
2809 spin_lock(&cache->lock);
2810 cache->commit_used = old_commit_used;
2811 spin_unlock(&cache->lock);
2817 static int cache_save_setup(struct btrfs_block_group *block_group,
2818 struct btrfs_trans_handle *trans,
2819 struct btrfs_path *path)
2821 struct btrfs_fs_info *fs_info = block_group->fs_info;
2822 struct btrfs_root *root = fs_info->tree_root;
2823 struct inode *inode = NULL;
2824 struct extent_changeset *data_reserved = NULL;
2826 int dcs = BTRFS_DC_ERROR;
2831 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2835 * If this block group is smaller than 100 megs don't bother caching the
2838 if (block_group->length < (100 * SZ_1M)) {
2839 spin_lock(&block_group->lock);
2840 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2841 spin_unlock(&block_group->lock);
2845 if (TRANS_ABORTED(trans))
2848 inode = lookup_free_space_inode(block_group, path);
2849 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2850 ret = PTR_ERR(inode);
2851 btrfs_release_path(path);
2855 if (IS_ERR(inode)) {
2859 if (block_group->ro)
2862 ret = create_free_space_inode(trans, block_group, path);
2869 * We want to set the generation to 0, that way if anything goes wrong
2870 * from here on out we know not to trust this cache when we load up next
2873 BTRFS_I(inode)->generation = 0;
2874 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2877 * So theoretically we could recover from this, simply set the
2878 * super cache generation to 0 so we know to invalidate the
2879 * cache, but then we'd have to keep track of the block groups
2880 * that fail this way so we know we _have_ to reset this cache
2881 * before the next commit or risk reading stale cache. So to
2882 * limit our exposure to horrible edge cases lets just abort the
2883 * transaction, this only happens in really bad situations
2886 btrfs_abort_transaction(trans, ret);
2891 /* We've already setup this transaction, go ahead and exit */
2892 if (block_group->cache_generation == trans->transid &&
2893 i_size_read(inode)) {
2894 dcs = BTRFS_DC_SETUP;
2898 if (i_size_read(inode) > 0) {
2899 ret = btrfs_check_trunc_cache_free_space(fs_info,
2900 &fs_info->global_block_rsv);
2904 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2909 spin_lock(&block_group->lock);
2910 if (block_group->cached != BTRFS_CACHE_FINISHED ||
2911 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2913 * don't bother trying to write stuff out _if_
2914 * a) we're not cached,
2915 * b) we're with nospace_cache mount option,
2916 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2918 dcs = BTRFS_DC_WRITTEN;
2919 spin_unlock(&block_group->lock);
2922 spin_unlock(&block_group->lock);
2925 * We hit an ENOSPC when setting up the cache in this transaction, just
2926 * skip doing the setup, we've already cleared the cache so we're safe.
2928 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2934 * Try to preallocate enough space based on how big the block group is.
2935 * Keep in mind this has to include any pinned space which could end up
2936 * taking up quite a bit since it's not folded into the other space
2939 cache_size = div_u64(block_group->length, SZ_256M);
2944 cache_size *= fs_info->sectorsize;
2946 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2951 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
2952 cache_size, cache_size,
2955 * Our cache requires contiguous chunks so that we don't modify a bunch
2956 * of metadata or split extents when writing the cache out, which means
2957 * we can enospc if we are heavily fragmented in addition to just normal
2958 * out of space conditions. So if we hit this just skip setting up any
2959 * other block groups for this transaction, maybe we'll unpin enough
2960 * space the next time around.
2963 dcs = BTRFS_DC_SETUP;
2964 else if (ret == -ENOSPC)
2965 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2970 btrfs_release_path(path);
2972 spin_lock(&block_group->lock);
2973 if (!ret && dcs == BTRFS_DC_SETUP)
2974 block_group->cache_generation = trans->transid;
2975 block_group->disk_cache_state = dcs;
2976 spin_unlock(&block_group->lock);
2978 extent_changeset_free(data_reserved);
2982 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2984 struct btrfs_fs_info *fs_info = trans->fs_info;
2985 struct btrfs_block_group *cache, *tmp;
2986 struct btrfs_transaction *cur_trans = trans->transaction;
2987 struct btrfs_path *path;
2989 if (list_empty(&cur_trans->dirty_bgs) ||
2990 !btrfs_test_opt(fs_info, SPACE_CACHE))
2993 path = btrfs_alloc_path();
2997 /* Could add new block groups, use _safe just in case */
2998 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3000 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3001 cache_save_setup(cache, trans, path);
3004 btrfs_free_path(path);
3009 * Transaction commit does final block group cache writeback during a critical
3010 * section where nothing is allowed to change the FS. This is required in
3011 * order for the cache to actually match the block group, but can introduce a
3012 * lot of latency into the commit.
3014 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3015 * There's a chance we'll have to redo some of it if the block group changes
3016 * again during the commit, but it greatly reduces the commit latency by
3017 * getting rid of the easy block groups while we're still allowing others to
3020 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3022 struct btrfs_fs_info *fs_info = trans->fs_info;
3023 struct btrfs_block_group *cache;
3024 struct btrfs_transaction *cur_trans = trans->transaction;
3027 struct btrfs_path *path = NULL;
3029 struct list_head *io = &cur_trans->io_bgs;
3032 spin_lock(&cur_trans->dirty_bgs_lock);
3033 if (list_empty(&cur_trans->dirty_bgs)) {
3034 spin_unlock(&cur_trans->dirty_bgs_lock);
3037 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3038 spin_unlock(&cur_trans->dirty_bgs_lock);
3041 /* Make sure all the block groups on our dirty list actually exist */
3042 btrfs_create_pending_block_groups(trans);
3045 path = btrfs_alloc_path();
3053 * cache_write_mutex is here only to save us from balance or automatic
3054 * removal of empty block groups deleting this block group while we are
3055 * writing out the cache
3057 mutex_lock(&trans->transaction->cache_write_mutex);
3058 while (!list_empty(&dirty)) {
3059 bool drop_reserve = true;
3061 cache = list_first_entry(&dirty, struct btrfs_block_group,
3064 * This can happen if something re-dirties a block group that
3065 * is already under IO. Just wait for it to finish and then do
3068 if (!list_empty(&cache->io_list)) {
3069 list_del_init(&cache->io_list);
3070 btrfs_wait_cache_io(trans, cache, path);
3071 btrfs_put_block_group(cache);
3076 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3077 * it should update the cache_state. Don't delete until after
3080 * Since we're not running in the commit critical section
3081 * we need the dirty_bgs_lock to protect from update_block_group
3083 spin_lock(&cur_trans->dirty_bgs_lock);
3084 list_del_init(&cache->dirty_list);
3085 spin_unlock(&cur_trans->dirty_bgs_lock);
3089 cache_save_setup(cache, trans, path);
3091 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3092 cache->io_ctl.inode = NULL;
3093 ret = btrfs_write_out_cache(trans, cache, path);
3094 if (ret == 0 && cache->io_ctl.inode) {
3098 * The cache_write_mutex is protecting the
3099 * io_list, also refer to the definition of
3100 * btrfs_transaction::io_bgs for more details
3102 list_add_tail(&cache->io_list, io);
3105 * If we failed to write the cache, the
3106 * generation will be bad and life goes on
3112 ret = update_block_group_item(trans, path, cache);
3114 * Our block group might still be attached to the list
3115 * of new block groups in the transaction handle of some
3116 * other task (struct btrfs_trans_handle->new_bgs). This
3117 * means its block group item isn't yet in the extent
3118 * tree. If this happens ignore the error, as we will
3119 * try again later in the critical section of the
3120 * transaction commit.
3122 if (ret == -ENOENT) {
3124 spin_lock(&cur_trans->dirty_bgs_lock);
3125 if (list_empty(&cache->dirty_list)) {
3126 list_add_tail(&cache->dirty_list,
3127 &cur_trans->dirty_bgs);
3128 btrfs_get_block_group(cache);
3129 drop_reserve = false;
3131 spin_unlock(&cur_trans->dirty_bgs_lock);
3133 btrfs_abort_transaction(trans, ret);
3137 /* If it's not on the io list, we need to put the block group */
3139 btrfs_put_block_group(cache);
3141 btrfs_delayed_refs_rsv_release(fs_info, 1);
3143 * Avoid blocking other tasks for too long. It might even save
3144 * us from writing caches for block groups that are going to be
3147 mutex_unlock(&trans->transaction->cache_write_mutex);
3150 mutex_lock(&trans->transaction->cache_write_mutex);
3152 mutex_unlock(&trans->transaction->cache_write_mutex);
3155 * Go through delayed refs for all the stuff we've just kicked off
3156 * and then loop back (just once)
3159 ret = btrfs_run_delayed_refs(trans, 0);
3160 if (!ret && loops == 0) {
3162 spin_lock(&cur_trans->dirty_bgs_lock);
3163 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3165 * dirty_bgs_lock protects us from concurrent block group
3166 * deletes too (not just cache_write_mutex).
3168 if (!list_empty(&dirty)) {
3169 spin_unlock(&cur_trans->dirty_bgs_lock);
3172 spin_unlock(&cur_trans->dirty_bgs_lock);
3176 spin_lock(&cur_trans->dirty_bgs_lock);
3177 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3178 spin_unlock(&cur_trans->dirty_bgs_lock);
3179 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3182 btrfs_free_path(path);
3186 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3188 struct btrfs_fs_info *fs_info = trans->fs_info;
3189 struct btrfs_block_group *cache;
3190 struct btrfs_transaction *cur_trans = trans->transaction;
3193 struct btrfs_path *path;
3194 struct list_head *io = &cur_trans->io_bgs;
3196 path = btrfs_alloc_path();
3201 * Even though we are in the critical section of the transaction commit,
3202 * we can still have concurrent tasks adding elements to this
3203 * transaction's list of dirty block groups. These tasks correspond to
3204 * endio free space workers started when writeback finishes for a
3205 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3206 * allocate new block groups as a result of COWing nodes of the root
3207 * tree when updating the free space inode. The writeback for the space
3208 * caches is triggered by an earlier call to
3209 * btrfs_start_dirty_block_groups() and iterations of the following
3211 * Also we want to do the cache_save_setup first and then run the
3212 * delayed refs to make sure we have the best chance at doing this all
3215 spin_lock(&cur_trans->dirty_bgs_lock);
3216 while (!list_empty(&cur_trans->dirty_bgs)) {
3217 cache = list_first_entry(&cur_trans->dirty_bgs,
3218 struct btrfs_block_group,
3222 * This can happen if cache_save_setup re-dirties a block group
3223 * that is already under IO. Just wait for it to finish and
3224 * then do it all again
3226 if (!list_empty(&cache->io_list)) {
3227 spin_unlock(&cur_trans->dirty_bgs_lock);
3228 list_del_init(&cache->io_list);
3229 btrfs_wait_cache_io(trans, cache, path);
3230 btrfs_put_block_group(cache);
3231 spin_lock(&cur_trans->dirty_bgs_lock);
3235 * Don't remove from the dirty list until after we've waited on
3238 list_del_init(&cache->dirty_list);
3239 spin_unlock(&cur_trans->dirty_bgs_lock);
3242 cache_save_setup(cache, trans, path);
3245 ret = btrfs_run_delayed_refs(trans,
3246 (unsigned long) -1);
3248 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3249 cache->io_ctl.inode = NULL;
3250 ret = btrfs_write_out_cache(trans, cache, path);
3251 if (ret == 0 && cache->io_ctl.inode) {
3253 list_add_tail(&cache->io_list, io);
3256 * If we failed to write the cache, the
3257 * generation will be bad and life goes on
3263 ret = update_block_group_item(trans, path, cache);
3265 * One of the free space endio workers might have
3266 * created a new block group while updating a free space
3267 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3268 * and hasn't released its transaction handle yet, in
3269 * which case the new block group is still attached to
3270 * its transaction handle and its creation has not
3271 * finished yet (no block group item in the extent tree
3272 * yet, etc). If this is the case, wait for all free
3273 * space endio workers to finish and retry. This is a
3274 * very rare case so no need for a more efficient and
3277 if (ret == -ENOENT) {
3278 wait_event(cur_trans->writer_wait,
3279 atomic_read(&cur_trans->num_writers) == 1);
3280 ret = update_block_group_item(trans, path, cache);
3283 btrfs_abort_transaction(trans, ret);
3286 /* If its not on the io list, we need to put the block group */
3288 btrfs_put_block_group(cache);
3289 btrfs_delayed_refs_rsv_release(fs_info, 1);
3290 spin_lock(&cur_trans->dirty_bgs_lock);
3292 spin_unlock(&cur_trans->dirty_bgs_lock);
3295 * Refer to the definition of io_bgs member for details why it's safe
3296 * to use it without any locking
3298 while (!list_empty(io)) {
3299 cache = list_first_entry(io, struct btrfs_block_group,
3301 list_del_init(&cache->io_list);
3302 btrfs_wait_cache_io(trans, cache, path);
3303 btrfs_put_block_group(cache);
3306 btrfs_free_path(path);
3310 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3311 u64 bytenr, u64 num_bytes, bool alloc)
3313 struct btrfs_fs_info *info = trans->fs_info;
3314 struct btrfs_block_group *cache = NULL;
3315 u64 total = num_bytes;
3321 /* Block accounting for super block */
3322 spin_lock(&info->delalloc_root_lock);
3323 old_val = btrfs_super_bytes_used(info->super_copy);
3325 old_val += num_bytes;
3327 old_val -= num_bytes;
3328 btrfs_set_super_bytes_used(info->super_copy, old_val);
3329 spin_unlock(&info->delalloc_root_lock);
3334 cache = btrfs_lookup_block_group(info, bytenr);
3339 factor = btrfs_bg_type_to_factor(cache->flags);
3342 * If this block group has free space cache written out, we
3343 * need to make sure to load it if we are removing space. This
3344 * is because we need the unpinning stage to actually add the
3345 * space back to the block group, otherwise we will leak space.
3347 if (!alloc && !btrfs_block_group_done(cache))
3348 btrfs_cache_block_group(cache, true);
3350 byte_in_group = bytenr - cache->start;
3351 WARN_ON(byte_in_group > cache->length);
3353 spin_lock(&cache->space_info->lock);
3354 spin_lock(&cache->lock);
3356 if (btrfs_test_opt(info, SPACE_CACHE) &&
3357 cache->disk_cache_state < BTRFS_DC_CLEAR)
3358 cache->disk_cache_state = BTRFS_DC_CLEAR;
3360 old_val = cache->used;
3361 num_bytes = min(total, cache->length - byte_in_group);
3363 old_val += num_bytes;
3364 cache->used = old_val;
3365 cache->reserved -= num_bytes;
3366 cache->space_info->bytes_reserved -= num_bytes;
3367 cache->space_info->bytes_used += num_bytes;
3368 cache->space_info->disk_used += num_bytes * factor;
3369 spin_unlock(&cache->lock);
3370 spin_unlock(&cache->space_info->lock);
3372 old_val -= num_bytes;
3373 cache->used = old_val;
3374 cache->pinned += num_bytes;
3375 btrfs_space_info_update_bytes_pinned(info,
3376 cache->space_info, num_bytes);
3377 cache->space_info->bytes_used -= num_bytes;
3378 cache->space_info->disk_used -= num_bytes * factor;
3380 reclaim = should_reclaim_block_group(cache, num_bytes);
3381 spin_unlock(&cache->lock);
3382 spin_unlock(&cache->space_info->lock);
3384 set_extent_dirty(&trans->transaction->pinned_extents,
3385 bytenr, bytenr + num_bytes - 1,
3386 GFP_NOFS | __GFP_NOFAIL);
3389 spin_lock(&trans->transaction->dirty_bgs_lock);
3390 if (list_empty(&cache->dirty_list)) {
3391 list_add_tail(&cache->dirty_list,
3392 &trans->transaction->dirty_bgs);
3393 trans->delayed_ref_updates++;
3394 btrfs_get_block_group(cache);
3396 spin_unlock(&trans->transaction->dirty_bgs_lock);
3399 * No longer have used bytes in this block group, queue it for
3400 * deletion. We do this after adding the block group to the
3401 * dirty list to avoid races between cleaner kthread and space
3404 if (!alloc && old_val == 0) {
3405 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3406 btrfs_mark_bg_unused(cache);
3407 } else if (!alloc && reclaim) {
3408 btrfs_mark_bg_to_reclaim(cache);
3411 btrfs_put_block_group(cache);
3413 bytenr += num_bytes;
3416 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3417 btrfs_update_delayed_refs_rsv(trans);
3422 * btrfs_add_reserved_bytes - update the block_group and space info counters
3423 * @cache: The cache we are manipulating
3424 * @ram_bytes: The number of bytes of file content, and will be same to
3425 * @num_bytes except for the compress path.
3426 * @num_bytes: The number of bytes in question
3427 * @delalloc: The blocks are allocated for the delalloc write
3429 * This is called by the allocator when it reserves space. If this is a
3430 * reservation and the block group has become read only we cannot make the
3431 * reservation and return -EAGAIN, otherwise this function always succeeds.
3433 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3434 u64 ram_bytes, u64 num_bytes, int delalloc)
3436 struct btrfs_space_info *space_info = cache->space_info;
3439 spin_lock(&space_info->lock);
3440 spin_lock(&cache->lock);
3444 cache->reserved += num_bytes;
3445 space_info->bytes_reserved += num_bytes;
3446 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3447 space_info->flags, num_bytes, 1);
3448 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3449 space_info, -ram_bytes);
3451 cache->delalloc_bytes += num_bytes;
3454 * Compression can use less space than we reserved, so wake
3455 * tickets if that happens
3457 if (num_bytes < ram_bytes)
3458 btrfs_try_granting_tickets(cache->fs_info, space_info);
3460 spin_unlock(&cache->lock);
3461 spin_unlock(&space_info->lock);
3466 * btrfs_free_reserved_bytes - update the block_group and space info counters
3467 * @cache: The cache we are manipulating
3468 * @num_bytes: The number of bytes in question
3469 * @delalloc: The blocks are allocated for the delalloc write
3471 * This is called by somebody who is freeing space that was never actually used
3472 * on disk. For example if you reserve some space for a new leaf in transaction
3473 * A and before transaction A commits you free that leaf, you call this with
3474 * reserve set to 0 in order to clear the reservation.
3476 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3477 u64 num_bytes, int delalloc)
3479 struct btrfs_space_info *space_info = cache->space_info;
3481 spin_lock(&space_info->lock);
3482 spin_lock(&cache->lock);
3484 space_info->bytes_readonly += num_bytes;
3485 cache->reserved -= num_bytes;
3486 space_info->bytes_reserved -= num_bytes;
3487 space_info->max_extent_size = 0;
3490 cache->delalloc_bytes -= num_bytes;
3491 spin_unlock(&cache->lock);
3493 btrfs_try_granting_tickets(cache->fs_info, space_info);
3494 spin_unlock(&space_info->lock);
3497 static void force_metadata_allocation(struct btrfs_fs_info *info)
3499 struct list_head *head = &info->space_info;
3500 struct btrfs_space_info *found;
3502 list_for_each_entry(found, head, list) {
3503 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3504 found->force_alloc = CHUNK_ALLOC_FORCE;
3508 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3509 struct btrfs_space_info *sinfo, int force)
3511 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3514 if (force == CHUNK_ALLOC_FORCE)
3518 * in limited mode, we want to have some free space up to
3519 * about 1% of the FS size.
3521 if (force == CHUNK_ALLOC_LIMITED) {
3522 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3523 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3525 if (sinfo->total_bytes - bytes_used < thresh)
3529 if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3534 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3536 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3538 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3541 static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3543 struct btrfs_block_group *bg;
3547 * Check if we have enough space in the system space info because we
3548 * will need to update device items in the chunk btree and insert a new
3549 * chunk item in the chunk btree as well. This will allocate a new
3550 * system block group if needed.
3552 check_system_chunk(trans, flags);
3554 bg = btrfs_create_chunk(trans, flags);
3560 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3562 * Normally we are not expected to fail with -ENOSPC here, since we have
3563 * previously reserved space in the system space_info and allocated one
3564 * new system chunk if necessary. However there are three exceptions:
3566 * 1) We may have enough free space in the system space_info but all the
3567 * existing system block groups have a profile which can not be used
3568 * for extent allocation.
3570 * This happens when mounting in degraded mode. For example we have a
3571 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3572 * using the other device in degraded mode. If we then allocate a chunk,
3573 * we may have enough free space in the existing system space_info, but
3574 * none of the block groups can be used for extent allocation since they
3575 * have a RAID1 profile, and because we are in degraded mode with a
3576 * single device, we are forced to allocate a new system chunk with a
3577 * SINGLE profile. Making check_system_chunk() iterate over all system
3578 * block groups and check if they have a usable profile and enough space
3579 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3580 * try again after forcing allocation of a new system chunk. Like this
3581 * we avoid paying the cost of that search in normal circumstances, when
3582 * we were not mounted in degraded mode;
3584 * 2) We had enough free space info the system space_info, and one suitable
3585 * block group to allocate from when we called check_system_chunk()
3586 * above. However right after we called it, the only system block group
3587 * with enough free space got turned into RO mode by a running scrub,
3588 * and in this case we have to allocate a new one and retry. We only
3589 * need do this allocate and retry once, since we have a transaction
3590 * handle and scrub uses the commit root to search for block groups;
3592 * 3) We had one system block group with enough free space when we called
3593 * check_system_chunk(), but after that, right before we tried to
3594 * allocate the last extent buffer we needed, a discard operation came
3595 * in and it temporarily removed the last free space entry from the
3596 * block group (discard removes a free space entry, discards it, and
3597 * then adds back the entry to the block group cache).
3599 if (ret == -ENOSPC) {
3600 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3601 struct btrfs_block_group *sys_bg;
3603 sys_bg = btrfs_create_chunk(trans, sys_flags);
3604 if (IS_ERR(sys_bg)) {
3605 ret = PTR_ERR(sys_bg);
3606 btrfs_abort_transaction(trans, ret);
3610 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3612 btrfs_abort_transaction(trans, ret);
3616 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3618 btrfs_abort_transaction(trans, ret);
3622 btrfs_abort_transaction(trans, ret);
3626 btrfs_trans_release_chunk_metadata(trans);
3629 return ERR_PTR(ret);
3631 btrfs_get_block_group(bg);
3636 * Chunk allocation is done in 2 phases:
3638 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3639 * the chunk, the chunk mapping, create its block group and add the items
3640 * that belong in the chunk btree to it - more specifically, we need to
3641 * update device items in the chunk btree and add a new chunk item to it.
3643 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3644 * group item to the extent btree and the device extent items to the devices
3647 * This is done to prevent deadlocks. For example when COWing a node from the
3648 * extent btree we are holding a write lock on the node's parent and if we
3649 * trigger chunk allocation and attempted to insert the new block group item
3650 * in the extent btree right way, we could deadlock because the path for the
3651 * insertion can include that parent node. At first glance it seems impossible
3652 * to trigger chunk allocation after starting a transaction since tasks should
3653 * reserve enough transaction units (metadata space), however while that is true
3654 * most of the time, chunk allocation may still be triggered for several reasons:
3656 * 1) When reserving metadata, we check if there is enough free space in the
3657 * metadata space_info and therefore don't trigger allocation of a new chunk.
3658 * However later when the task actually tries to COW an extent buffer from
3659 * the extent btree or from the device btree for example, it is forced to
3660 * allocate a new block group (chunk) because the only one that had enough
3661 * free space was just turned to RO mode by a running scrub for example (or
3662 * device replace, block group reclaim thread, etc), so we can not use it
3663 * for allocating an extent and end up being forced to allocate a new one;
3665 * 2) Because we only check that the metadata space_info has enough free bytes,
3666 * we end up not allocating a new metadata chunk in that case. However if
3667 * the filesystem was mounted in degraded mode, none of the existing block
3668 * groups might be suitable for extent allocation due to their incompatible
3669 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3670 * use a RAID1 profile, in degraded mode using a single device). In this case
3671 * when the task attempts to COW some extent buffer of the extent btree for
3672 * example, it will trigger allocation of a new metadata block group with a
3673 * suitable profile (SINGLE profile in the example of the degraded mount of
3674 * the RAID1 filesystem);
3676 * 3) The task has reserved enough transaction units / metadata space, but when
3677 * it attempts to COW an extent buffer from the extent or device btree for
3678 * example, it does not find any free extent in any metadata block group,
3679 * therefore forced to try to allocate a new metadata block group.
3680 * This is because some other task allocated all available extents in the
3681 * meanwhile - this typically happens with tasks that don't reserve space
3682 * properly, either intentionally or as a bug. One example where this is
3683 * done intentionally is fsync, as it does not reserve any transaction units
3684 * and ends up allocating a variable number of metadata extents for log
3685 * tree extent buffers;
3687 * 4) The task has reserved enough transaction units / metadata space, but right
3688 * before it tries to allocate the last extent buffer it needs, a discard
3689 * operation comes in and, temporarily, removes the last free space entry from
3690 * the only metadata block group that had free space (discard starts by
3691 * removing a free space entry from a block group, then does the discard
3692 * operation and, once it's done, it adds back the free space entry to the
3695 * We also need this 2 phases setup when adding a device to a filesystem with
3696 * a seed device - we must create new metadata and system chunks without adding
3697 * any of the block group items to the chunk, extent and device btrees. If we
3698 * did not do it this way, we would get ENOSPC when attempting to update those
3699 * btrees, since all the chunks from the seed device are read-only.
3701 * Phase 1 does the updates and insertions to the chunk btree because if we had
3702 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3703 * parallel, we risk having too many system chunks allocated by many tasks if
3704 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3705 * extreme case this leads to exhaustion of the system chunk array in the
3706 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3707 * and with RAID filesystems (so we have more device items in the chunk btree).
3708 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3709 * the system chunk array due to concurrent allocations") provides more details.
3711 * Allocation of system chunks does not happen through this function. A task that
3712 * needs to update the chunk btree (the only btree that uses system chunks), must
3713 * preallocate chunk space by calling either check_system_chunk() or
3714 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
3715 * metadata chunk or when removing a chunk, while the later is used before doing
3716 * a modification to the chunk btree - use cases for the later are adding,
3717 * removing and resizing a device as well as relocation of a system chunk.
3718 * See the comment below for more details.
3720 * The reservation of system space, done through check_system_chunk(), as well
3721 * as all the updates and insertions into the chunk btree must be done while
3722 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
3723 * an extent buffer from the chunks btree we never trigger allocation of a new
3724 * system chunk, which would result in a deadlock (trying to lock twice an
3725 * extent buffer of the chunk btree, first time before triggering the chunk
3726 * allocation and the second time during chunk allocation while attempting to
3727 * update the chunks btree). The system chunk array is also updated while holding
3728 * that mutex. The same logic applies to removing chunks - we must reserve system
3729 * space, update the chunk btree and the system chunk array in the superblock
3730 * while holding fs_info->chunk_mutex.
3732 * This function, btrfs_chunk_alloc(), belongs to phase 1.
3734 * If @force is CHUNK_ALLOC_FORCE:
3735 * - return 1 if it successfully allocates a chunk,
3736 * - return errors including -ENOSPC otherwise.
3737 * If @force is NOT CHUNK_ALLOC_FORCE:
3738 * - return 0 if it doesn't need to allocate a new chunk,
3739 * - return 1 if it successfully allocates a chunk,
3740 * - return errors including -ENOSPC otherwise.
3742 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3743 enum btrfs_chunk_alloc_enum force)
3745 struct btrfs_fs_info *fs_info = trans->fs_info;
3746 struct btrfs_space_info *space_info;
3747 struct btrfs_block_group *ret_bg;
3748 bool wait_for_alloc = false;
3749 bool should_alloc = false;
3750 bool from_extent_allocation = false;
3753 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
3754 from_extent_allocation = true;
3755 force = CHUNK_ALLOC_FORCE;
3758 /* Don't re-enter if we're already allocating a chunk */
3759 if (trans->allocating_chunk)
3762 * Allocation of system chunks can not happen through this path, as we
3763 * could end up in a deadlock if we are allocating a data or metadata
3764 * chunk and there is another task modifying the chunk btree.
3766 * This is because while we are holding the chunk mutex, we will attempt
3767 * to add the new chunk item to the chunk btree or update an existing
3768 * device item in the chunk btree, while the other task that is modifying
3769 * the chunk btree is attempting to COW an extent buffer while holding a
3770 * lock on it and on its parent - if the COW operation triggers a system
3771 * chunk allocation, then we can deadlock because we are holding the
3772 * chunk mutex and we may need to access that extent buffer or its parent
3773 * in order to add the chunk item or update a device item.
3775 * Tasks that want to modify the chunk tree should reserve system space
3776 * before updating the chunk btree, by calling either
3777 * btrfs_reserve_chunk_metadata() or check_system_chunk().
3778 * It's possible that after a task reserves the space, it still ends up
3779 * here - this happens in the cases described above at do_chunk_alloc().
3780 * The task will have to either retry or fail.
3782 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3785 space_info = btrfs_find_space_info(fs_info, flags);
3789 spin_lock(&space_info->lock);
3790 if (force < space_info->force_alloc)
3791 force = space_info->force_alloc;
3792 should_alloc = should_alloc_chunk(fs_info, space_info, force);
3793 if (space_info->full) {
3794 /* No more free physical space */
3799 spin_unlock(&space_info->lock);
3801 } else if (!should_alloc) {
3802 spin_unlock(&space_info->lock);
3804 } else if (space_info->chunk_alloc) {
3806 * Someone is already allocating, so we need to block
3807 * until this someone is finished and then loop to
3808 * recheck if we should continue with our allocation
3811 wait_for_alloc = true;
3812 force = CHUNK_ALLOC_NO_FORCE;
3813 spin_unlock(&space_info->lock);
3814 mutex_lock(&fs_info->chunk_mutex);
3815 mutex_unlock(&fs_info->chunk_mutex);
3817 /* Proceed with allocation */
3818 space_info->chunk_alloc = 1;
3819 wait_for_alloc = false;
3820 spin_unlock(&space_info->lock);
3824 } while (wait_for_alloc);
3826 mutex_lock(&fs_info->chunk_mutex);
3827 trans->allocating_chunk = true;
3830 * If we have mixed data/metadata chunks we want to make sure we keep
3831 * allocating mixed chunks instead of individual chunks.
3833 if (btrfs_mixed_space_info(space_info))
3834 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3837 * if we're doing a data chunk, go ahead and make sure that
3838 * we keep a reasonable number of metadata chunks allocated in the
3841 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3842 fs_info->data_chunk_allocations++;
3843 if (!(fs_info->data_chunk_allocations %
3844 fs_info->metadata_ratio))
3845 force_metadata_allocation(fs_info);
3848 ret_bg = do_chunk_alloc(trans, flags);
3849 trans->allocating_chunk = false;
3851 if (IS_ERR(ret_bg)) {
3852 ret = PTR_ERR(ret_bg);
3853 } else if (from_extent_allocation) {
3855 * New block group is likely to be used soon. Try to activate
3856 * it now. Failure is OK for now.
3858 btrfs_zone_activate(ret_bg);
3862 btrfs_put_block_group(ret_bg);
3864 spin_lock(&space_info->lock);
3867 space_info->full = 1;
3872 space_info->max_extent_size = 0;
3875 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3877 space_info->chunk_alloc = 0;
3878 spin_unlock(&space_info->lock);
3879 mutex_unlock(&fs_info->chunk_mutex);
3884 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3888 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3890 num_dev = fs_info->fs_devices->rw_devices;
3895 static void reserve_chunk_space(struct btrfs_trans_handle *trans,
3899 struct btrfs_fs_info *fs_info = trans->fs_info;
3900 struct btrfs_space_info *info;
3905 * Needed because we can end up allocating a system chunk and for an
3906 * atomic and race free space reservation in the chunk block reserve.
3908 lockdep_assert_held(&fs_info->chunk_mutex);
3910 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3911 spin_lock(&info->lock);
3912 left = info->total_bytes - btrfs_space_info_used(info, true);
3913 spin_unlock(&info->lock);
3915 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3916 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3918 btrfs_dump_space_info(fs_info, info, 0, 0);
3922 u64 flags = btrfs_system_alloc_profile(fs_info);
3923 struct btrfs_block_group *bg;
3926 * Ignore failure to create system chunk. We might end up not
3927 * needing it, as we might not need to COW all nodes/leafs from
3928 * the paths we visit in the chunk tree (they were already COWed
3929 * or created in the current transaction for example).
3931 bg = btrfs_create_chunk(trans, flags);
3936 * We have a new chunk. We also need to activate it for
3939 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
3944 * If we fail to add the chunk item here, we end up
3945 * trying again at phase 2 of chunk allocation, at
3946 * btrfs_create_pending_block_groups(). So ignore
3947 * any error here. An ENOSPC here could happen, due to
3948 * the cases described at do_chunk_alloc() - the system
3949 * block group we just created was just turned into RO
3950 * mode by a scrub for example, or a running discard
3951 * temporarily removed its free space entries, etc.
3953 btrfs_chunk_alloc_add_chunk_item(trans, bg);
3958 ret = btrfs_block_rsv_add(fs_info,
3959 &fs_info->chunk_block_rsv,
3960 bytes, BTRFS_RESERVE_NO_FLUSH);
3962 trans->chunk_bytes_reserved += bytes;
3967 * Reserve space in the system space for allocating or removing a chunk.
3968 * The caller must be holding fs_info->chunk_mutex.
3970 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3972 struct btrfs_fs_info *fs_info = trans->fs_info;
3973 const u64 num_devs = get_profile_num_devs(fs_info, type);
3976 /* num_devs device items to update and 1 chunk item to add or remove. */
3977 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
3978 btrfs_calc_insert_metadata_size(fs_info, 1);
3980 reserve_chunk_space(trans, bytes, type);
3984 * Reserve space in the system space, if needed, for doing a modification to the
3987 * @trans: A transaction handle.
3988 * @is_item_insertion: Indicate if the modification is for inserting a new item
3989 * in the chunk btree or if it's for the deletion or update
3990 * of an existing item.
3992 * This is used in a context where we need to update the chunk btree outside
3993 * block group allocation and removal, to avoid a deadlock with a concurrent
3994 * task that is allocating a metadata or data block group and therefore needs to
3995 * update the chunk btree while holding the chunk mutex. After the update to the
3996 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
3999 void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4000 bool is_item_insertion)
4002 struct btrfs_fs_info *fs_info = trans->fs_info;
4005 if (is_item_insertion)
4006 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4008 bytes = btrfs_calc_metadata_size(fs_info, 1);
4010 mutex_lock(&fs_info->chunk_mutex);
4011 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4012 mutex_unlock(&fs_info->chunk_mutex);
4015 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4017 struct btrfs_block_group *block_group;
4019 block_group = btrfs_lookup_first_block_group(info, 0);
4020 while (block_group) {
4021 btrfs_wait_block_group_cache_done(block_group);
4022 spin_lock(&block_group->lock);
4023 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4024 &block_group->runtime_flags)) {
4025 struct inode *inode = block_group->inode;
4027 block_group->inode = NULL;
4028 spin_unlock(&block_group->lock);
4030 ASSERT(block_group->io_ctl.inode == NULL);
4033 spin_unlock(&block_group->lock);
4035 block_group = btrfs_next_block_group(block_group);
4040 * Must be called only after stopping all workers, since we could have block
4041 * group caching kthreads running, and therefore they could race with us if we
4042 * freed the block groups before stopping them.
4044 int btrfs_free_block_groups(struct btrfs_fs_info *info)
4046 struct btrfs_block_group *block_group;
4047 struct btrfs_space_info *space_info;
4048 struct btrfs_caching_control *caching_ctl;
4051 write_lock(&info->block_group_cache_lock);
4052 while (!list_empty(&info->caching_block_groups)) {
4053 caching_ctl = list_entry(info->caching_block_groups.next,
4054 struct btrfs_caching_control, list);
4055 list_del(&caching_ctl->list);
4056 btrfs_put_caching_control(caching_ctl);
4058 write_unlock(&info->block_group_cache_lock);
4060 spin_lock(&info->unused_bgs_lock);
4061 while (!list_empty(&info->unused_bgs)) {
4062 block_group = list_first_entry(&info->unused_bgs,
4063 struct btrfs_block_group,
4065 list_del_init(&block_group->bg_list);
4066 btrfs_put_block_group(block_group);
4069 while (!list_empty(&info->reclaim_bgs)) {
4070 block_group = list_first_entry(&info->reclaim_bgs,
4071 struct btrfs_block_group,
4073 list_del_init(&block_group->bg_list);
4074 btrfs_put_block_group(block_group);
4076 spin_unlock(&info->unused_bgs_lock);
4078 spin_lock(&info->zone_active_bgs_lock);
4079 while (!list_empty(&info->zone_active_bgs)) {
4080 block_group = list_first_entry(&info->zone_active_bgs,
4081 struct btrfs_block_group,
4083 list_del_init(&block_group->active_bg_list);
4084 btrfs_put_block_group(block_group);
4086 spin_unlock(&info->zone_active_bgs_lock);
4088 write_lock(&info->block_group_cache_lock);
4089 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4090 block_group = rb_entry(n, struct btrfs_block_group,
4092 rb_erase_cached(&block_group->cache_node,
4093 &info->block_group_cache_tree);
4094 RB_CLEAR_NODE(&block_group->cache_node);
4095 write_unlock(&info->block_group_cache_lock);
4097 down_write(&block_group->space_info->groups_sem);
4098 list_del(&block_group->list);
4099 up_write(&block_group->space_info->groups_sem);
4102 * We haven't cached this block group, which means we could
4103 * possibly have excluded extents on this block group.
4105 if (block_group->cached == BTRFS_CACHE_NO ||
4106 block_group->cached == BTRFS_CACHE_ERROR)
4107 btrfs_free_excluded_extents(block_group);
4109 btrfs_remove_free_space_cache(block_group);
4110 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4111 ASSERT(list_empty(&block_group->dirty_list));
4112 ASSERT(list_empty(&block_group->io_list));
4113 ASSERT(list_empty(&block_group->bg_list));
4114 ASSERT(refcount_read(&block_group->refs) == 1);
4115 ASSERT(block_group->swap_extents == 0);
4116 btrfs_put_block_group(block_group);
4118 write_lock(&info->block_group_cache_lock);
4120 write_unlock(&info->block_group_cache_lock);
4122 btrfs_release_global_block_rsv(info);
4124 while (!list_empty(&info->space_info)) {
4125 space_info = list_entry(info->space_info.next,
4126 struct btrfs_space_info,
4130 * Do not hide this behind enospc_debug, this is actually
4131 * important and indicates a real bug if this happens.
4133 if (WARN_ON(space_info->bytes_pinned > 0 ||
4134 space_info->bytes_may_use > 0))
4135 btrfs_dump_space_info(info, space_info, 0, 0);
4138 * If there was a failure to cleanup a log tree, very likely due
4139 * to an IO failure on a writeback attempt of one or more of its
4140 * extent buffers, we could not do proper (and cheap) unaccounting
4141 * of their reserved space, so don't warn on bytes_reserved > 0 in
4144 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4145 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4146 if (WARN_ON(space_info->bytes_reserved > 0))
4147 btrfs_dump_space_info(info, space_info, 0, 0);
4150 WARN_ON(space_info->reclaim_size > 0);
4151 list_del(&space_info->list);
4152 btrfs_sysfs_remove_space_info(space_info);
4157 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4159 atomic_inc(&cache->frozen);
4162 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4164 struct btrfs_fs_info *fs_info = block_group->fs_info;
4165 struct extent_map_tree *em_tree;
4166 struct extent_map *em;
4169 spin_lock(&block_group->lock);
4170 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4171 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4172 spin_unlock(&block_group->lock);
4175 em_tree = &fs_info->mapping_tree;
4176 write_lock(&em_tree->lock);
4177 em = lookup_extent_mapping(em_tree, block_group->start,
4179 BUG_ON(!em); /* logic error, can't happen */
4180 remove_extent_mapping(em_tree, em);
4181 write_unlock(&em_tree->lock);
4183 /* once for us and once for the tree */
4184 free_extent_map(em);
4185 free_extent_map(em);
4188 * We may have left one free space entry and other possible
4189 * tasks trimming this block group have left 1 entry each one.
4192 btrfs_remove_free_space_cache(block_group);
4196 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4200 spin_lock(&bg->lock);
4205 spin_unlock(&bg->lock);
4210 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4212 spin_lock(&bg->lock);
4214 ASSERT(bg->swap_extents >= amount);
4215 bg->swap_extents -= amount;
4216 spin_unlock(&bg->lock);