1 // SPDX-License-Identifier: GPL-2.0
3 #include <linux/sizes.h>
4 #include <linux/list_sort.h>
7 #include "block-group.h"
8 #include "space-info.h"
10 #include "free-space-cache.h"
11 #include "free-space-tree.h"
13 #include "transaction.h"
14 #include "ref-verify.h"
17 #include "delalloc-space.h"
22 #include "accessors.h"
23 #include "extent-tree.h"
25 #ifdef CONFIG_BTRFS_DEBUG
26 int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group)
28 struct btrfs_fs_info *fs_info = block_group->fs_info;
30 return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
31 block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
32 (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
33 block_group->flags & BTRFS_BLOCK_GROUP_DATA);
38 * Return target flags in extended format or 0 if restripe for this chunk_type
41 * Should be called with balance_lock held
43 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
45 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
51 if (flags & BTRFS_BLOCK_GROUP_DATA &&
52 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
53 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
54 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
55 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
56 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
57 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
58 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
59 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
66 * @flags: available profiles in extended format (see ctree.h)
68 * Return reduced profile in chunk format. If profile changing is in progress
69 * (either running or paused) picks the target profile (if it's already
70 * available), otherwise falls back to plain reducing.
72 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
74 u64 num_devices = fs_info->fs_devices->rw_devices;
80 * See if restripe for this chunk_type is in progress, if so try to
81 * reduce to the target profile
83 spin_lock(&fs_info->balance_lock);
84 target = get_restripe_target(fs_info, flags);
86 spin_unlock(&fs_info->balance_lock);
87 return extended_to_chunk(target);
89 spin_unlock(&fs_info->balance_lock);
91 /* First, mask out the RAID levels which aren't possible */
92 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
93 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
94 allowed |= btrfs_raid_array[raid_type].bg_flag;
98 /* Select the highest-redundancy RAID level. */
99 if (allowed & BTRFS_BLOCK_GROUP_RAID1C4)
100 allowed = BTRFS_BLOCK_GROUP_RAID1C4;
101 else if (allowed & BTRFS_BLOCK_GROUP_RAID6)
102 allowed = BTRFS_BLOCK_GROUP_RAID6;
103 else if (allowed & BTRFS_BLOCK_GROUP_RAID1C3)
104 allowed = BTRFS_BLOCK_GROUP_RAID1C3;
105 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
106 allowed = BTRFS_BLOCK_GROUP_RAID5;
107 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
108 allowed = BTRFS_BLOCK_GROUP_RAID10;
109 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
110 allowed = BTRFS_BLOCK_GROUP_RAID1;
111 else if (allowed & BTRFS_BLOCK_GROUP_DUP)
112 allowed = BTRFS_BLOCK_GROUP_DUP;
113 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
114 allowed = BTRFS_BLOCK_GROUP_RAID0;
116 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
118 return extended_to_chunk(flags | allowed);
121 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
128 seq = read_seqbegin(&fs_info->profiles_lock);
130 if (flags & BTRFS_BLOCK_GROUP_DATA)
131 flags |= fs_info->avail_data_alloc_bits;
132 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
133 flags |= fs_info->avail_system_alloc_bits;
134 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
135 flags |= fs_info->avail_metadata_alloc_bits;
136 } while (read_seqretry(&fs_info->profiles_lock, seq));
138 return btrfs_reduce_alloc_profile(fs_info, flags);
141 void btrfs_get_block_group(struct btrfs_block_group *cache)
143 refcount_inc(&cache->refs);
146 void btrfs_put_block_group(struct btrfs_block_group *cache)
148 if (refcount_dec_and_test(&cache->refs)) {
149 WARN_ON(cache->pinned > 0);
151 * If there was a failure to cleanup a log tree, very likely due
152 * to an IO failure on a writeback attempt of one or more of its
153 * extent buffers, we could not do proper (and cheap) unaccounting
154 * of their reserved space, so don't warn on reserved > 0 in that
157 if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
158 !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
159 WARN_ON(cache->reserved > 0);
162 * A block_group shouldn't be on the discard_list anymore.
163 * Remove the block_group from the discard_list to prevent us
164 * from causing a panic due to NULL pointer dereference.
166 if (WARN_ON(!list_empty(&cache->discard_list)))
167 btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
170 kfree(cache->free_space_ctl);
171 btrfs_free_chunk_map(cache->physical_map);
177 * This adds the block group to the fs_info rb tree for the block group cache
179 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
180 struct btrfs_block_group *block_group)
183 struct rb_node *parent = NULL;
184 struct btrfs_block_group *cache;
185 bool leftmost = true;
187 ASSERT(block_group->length != 0);
189 write_lock(&info->block_group_cache_lock);
190 p = &info->block_group_cache_tree.rb_root.rb_node;
194 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
195 if (block_group->start < cache->start) {
197 } else if (block_group->start > cache->start) {
201 write_unlock(&info->block_group_cache_lock);
206 rb_link_node(&block_group->cache_node, parent, p);
207 rb_insert_color_cached(&block_group->cache_node,
208 &info->block_group_cache_tree, leftmost);
210 write_unlock(&info->block_group_cache_lock);
216 * This will return the block group at or after bytenr if contains is 0, else
217 * it will return the block group that contains the bytenr
219 static struct btrfs_block_group *block_group_cache_tree_search(
220 struct btrfs_fs_info *info, u64 bytenr, int contains)
222 struct btrfs_block_group *cache, *ret = NULL;
226 read_lock(&info->block_group_cache_lock);
227 n = info->block_group_cache_tree.rb_root.rb_node;
230 cache = rb_entry(n, struct btrfs_block_group, cache_node);
231 end = cache->start + cache->length - 1;
232 start = cache->start;
234 if (bytenr < start) {
235 if (!contains && (!ret || start < ret->start))
238 } else if (bytenr > start) {
239 if (contains && bytenr <= end) {
250 btrfs_get_block_group(ret);
251 read_unlock(&info->block_group_cache_lock);
257 * Return the block group that starts at or after bytenr
259 struct btrfs_block_group *btrfs_lookup_first_block_group(
260 struct btrfs_fs_info *info, u64 bytenr)
262 return block_group_cache_tree_search(info, bytenr, 0);
266 * Return the block group that contains the given bytenr
268 struct btrfs_block_group *btrfs_lookup_block_group(
269 struct btrfs_fs_info *info, u64 bytenr)
271 return block_group_cache_tree_search(info, bytenr, 1);
274 struct btrfs_block_group *btrfs_next_block_group(
275 struct btrfs_block_group *cache)
277 struct btrfs_fs_info *fs_info = cache->fs_info;
278 struct rb_node *node;
280 read_lock(&fs_info->block_group_cache_lock);
282 /* If our block group was removed, we need a full search. */
283 if (RB_EMPTY_NODE(&cache->cache_node)) {
284 const u64 next_bytenr = cache->start + cache->length;
286 read_unlock(&fs_info->block_group_cache_lock);
287 btrfs_put_block_group(cache);
288 return btrfs_lookup_first_block_group(fs_info, next_bytenr);
290 node = rb_next(&cache->cache_node);
291 btrfs_put_block_group(cache);
293 cache = rb_entry(node, struct btrfs_block_group, cache_node);
294 btrfs_get_block_group(cache);
297 read_unlock(&fs_info->block_group_cache_lock);
302 * Check if we can do a NOCOW write for a given extent.
304 * @fs_info: The filesystem information object.
305 * @bytenr: Logical start address of the extent.
307 * Check if we can do a NOCOW write for the given extent, and increments the
308 * number of NOCOW writers in the block group that contains the extent, as long
309 * as the block group exists and it's currently not in read-only mode.
311 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
312 * is responsible for calling btrfs_dec_nocow_writers() later.
314 * Or NULL if we can not do a NOCOW write
316 struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
319 struct btrfs_block_group *bg;
320 bool can_nocow = true;
322 bg = btrfs_lookup_block_group(fs_info, bytenr);
326 spin_lock(&bg->lock);
330 atomic_inc(&bg->nocow_writers);
331 spin_unlock(&bg->lock);
334 btrfs_put_block_group(bg);
338 /* No put on block group, done by btrfs_dec_nocow_writers(). */
343 * Decrement the number of NOCOW writers in a 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 static 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;
446 caching_ctl = btrfs_get_caching_control(cache);
451 * We've already failed to allocate from this block group, so even if
452 * there's enough space in the block group it isn't contiguous enough to
453 * allow for an allocation, so wait for at least the next wakeup tick,
454 * or for the thing to be done.
456 progress = atomic_read(&caching_ctl->progress);
458 wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
459 (progress != atomic_read(&caching_ctl->progress) &&
460 (cache->free_space_ctl->free_space >= num_bytes)));
462 btrfs_put_caching_control(caching_ctl);
465 static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
466 struct btrfs_caching_control *caching_ctl)
468 wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
469 return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
472 static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
474 struct btrfs_caching_control *caching_ctl;
477 caching_ctl = btrfs_get_caching_control(cache);
479 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
480 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
481 btrfs_put_caching_control(caching_ctl);
485 #ifdef CONFIG_BTRFS_DEBUG
486 static void fragment_free_space(struct btrfs_block_group *block_group)
488 struct btrfs_fs_info *fs_info = block_group->fs_info;
489 u64 start = block_group->start;
490 u64 len = block_group->length;
491 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
492 fs_info->nodesize : fs_info->sectorsize;
493 u64 step = chunk << 1;
495 while (len > chunk) {
496 btrfs_remove_free_space(block_group, start, chunk);
507 * Add a free space range to the in memory free space cache of a block group.
508 * This checks if the range contains super block locations and any such
509 * locations are not added to the free space cache.
511 * @block_group: The target block group.
512 * @start: Start offset of the range.
513 * @end: End offset of the range (exclusive).
514 * @total_added_ret: Optional pointer to return the total amount of space
515 * added to the block group's free space cache.
517 * Returns 0 on success or < 0 on error.
519 int btrfs_add_new_free_space(struct btrfs_block_group *block_group, u64 start,
520 u64 end, u64 *total_added_ret)
522 struct btrfs_fs_info *info = block_group->fs_info;
523 u64 extent_start, extent_end, size;
527 *total_added_ret = 0;
529 while (start < end) {
530 if (!find_first_extent_bit(&info->excluded_extents, start,
531 &extent_start, &extent_end,
532 EXTENT_DIRTY | EXTENT_UPTODATE,
536 if (extent_start <= start) {
537 start = extent_end + 1;
538 } else if (extent_start > start && extent_start < end) {
539 size = extent_start - start;
540 ret = btrfs_add_free_space_async_trimmed(block_group,
545 *total_added_ret += size;
546 start = extent_end + 1;
554 ret = btrfs_add_free_space_async_trimmed(block_group, start,
559 *total_added_ret += size;
566 * Get an arbitrary extent item index / max_index through the block group
568 * @block_group the block group to sample from
569 * @index: the integral step through the block group to grab from
570 * @max_index: the granularity of the sampling
571 * @key: return value parameter for the item we find
573 * Pre-conditions on indices:
574 * 0 <= index <= max_index
577 * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
578 * error code on error.
580 static int sample_block_group_extent_item(struct btrfs_caching_control *caching_ctl,
581 struct btrfs_block_group *block_group,
582 int index, int max_index,
583 struct btrfs_key *found_key)
585 struct btrfs_fs_info *fs_info = block_group->fs_info;
586 struct btrfs_root *extent_root;
588 u64 search_end = block_group->start + block_group->length;
589 struct btrfs_path *path;
590 struct btrfs_key search_key;
594 ASSERT(index <= max_index);
595 ASSERT(max_index > 0);
596 lockdep_assert_held(&caching_ctl->mutex);
597 lockdep_assert_held_read(&fs_info->commit_root_sem);
599 path = btrfs_alloc_path();
603 extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
604 BTRFS_SUPER_INFO_OFFSET));
606 path->skip_locking = 1;
607 path->search_commit_root = 1;
608 path->reada = READA_FORWARD;
610 search_offset = index * div_u64(block_group->length, max_index);
611 search_key.objectid = block_group->start + search_offset;
612 search_key.type = BTRFS_EXTENT_ITEM_KEY;
613 search_key.offset = 0;
615 btrfs_for_each_slot(extent_root, &search_key, found_key, path, ret) {
616 /* Success; sampled an extent item in the block group */
617 if (found_key->type == BTRFS_EXTENT_ITEM_KEY &&
618 found_key->objectid >= block_group->start &&
619 found_key->objectid + found_key->offset <= search_end)
622 /* We can't possibly find a valid extent item anymore */
623 if (found_key->objectid >= search_end) {
629 lockdep_assert_held(&caching_ctl->mutex);
630 lockdep_assert_held_read(&fs_info->commit_root_sem);
631 btrfs_free_path(path);
636 * Best effort attempt to compute a block group's size class while caching it.
638 * @block_group: the block group we are caching
640 * We cannot infer the size class while adding free space extents, because that
641 * logic doesn't care about contiguous file extents (it doesn't differentiate
642 * between a 100M extent and 100 contiguous 1M extents). So we need to read the
643 * file extent items. Reading all of them is quite wasteful, because usually
644 * only a handful are enough to give a good answer. Therefore, we just grab 5 of
645 * them at even steps through the block group and pick the smallest size class
646 * we see. Since size class is best effort, and not guaranteed in general,
647 * inaccuracy is acceptable.
649 * To be more explicit about why this algorithm makes sense:
651 * If we are caching in a block group from disk, then there are three major cases
653 * 1. the block group is well behaved and all extents in it are the same size
655 * 2. the block group is mostly one size class with rare exceptions for last
657 * 3. the block group was populated before size classes and can have a totally
658 * arbitrary mix of size classes.
660 * In case 1, looking at any extent in the block group will yield the correct
661 * result. For the mixed cases, taking the minimum size class seems like a good
662 * approximation, since gaps from frees will be usable to the size class. For
663 * 2., a small handful of file extents is likely to yield the right answer. For
664 * 3, we can either read every file extent, or admit that this is best effort
665 * anyway and try to stay fast.
667 * Returns: 0 on success, negative error code on error.
669 static int load_block_group_size_class(struct btrfs_caching_control *caching_ctl,
670 struct btrfs_block_group *block_group)
672 struct btrfs_fs_info *fs_info = block_group->fs_info;
673 struct btrfs_key key;
675 u64 min_size = block_group->length;
676 enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
679 if (!btrfs_block_group_should_use_size_class(block_group))
682 lockdep_assert_held(&caching_ctl->mutex);
683 lockdep_assert_held_read(&fs_info->commit_root_sem);
684 for (i = 0; i < 5; ++i) {
685 ret = sample_block_group_extent_item(caching_ctl, block_group, i, 5, &key);
690 min_size = min_t(u64, min_size, key.offset);
691 size_class = btrfs_calc_block_group_size_class(min_size);
693 if (size_class != BTRFS_BG_SZ_NONE) {
694 spin_lock(&block_group->lock);
695 block_group->size_class = size_class;
696 spin_unlock(&block_group->lock);
702 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
704 struct btrfs_block_group *block_group = caching_ctl->block_group;
705 struct btrfs_fs_info *fs_info = block_group->fs_info;
706 struct btrfs_root *extent_root;
707 struct btrfs_path *path;
708 struct extent_buffer *leaf;
709 struct btrfs_key key;
716 path = btrfs_alloc_path();
720 last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
721 extent_root = btrfs_extent_root(fs_info, last);
723 #ifdef CONFIG_BTRFS_DEBUG
725 * If we're fragmenting we don't want to make anybody think we can
726 * allocate from this block group until we've had a chance to fragment
729 if (btrfs_should_fragment_free_space(block_group))
733 * We don't want to deadlock with somebody trying to allocate a new
734 * extent for the extent root while also trying to search the extent
735 * root to add free space. So we skip locking and search the commit
736 * root, since its read-only
738 path->skip_locking = 1;
739 path->search_commit_root = 1;
740 path->reada = READA_FORWARD;
744 key.type = BTRFS_EXTENT_ITEM_KEY;
747 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
751 leaf = path->nodes[0];
752 nritems = btrfs_header_nritems(leaf);
755 if (btrfs_fs_closing(fs_info) > 1) {
760 if (path->slots[0] < nritems) {
761 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
763 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
767 if (need_resched() ||
768 rwsem_is_contended(&fs_info->commit_root_sem)) {
769 btrfs_release_path(path);
770 up_read(&fs_info->commit_root_sem);
771 mutex_unlock(&caching_ctl->mutex);
773 mutex_lock(&caching_ctl->mutex);
774 down_read(&fs_info->commit_root_sem);
778 ret = btrfs_next_leaf(extent_root, path);
783 leaf = path->nodes[0];
784 nritems = btrfs_header_nritems(leaf);
788 if (key.objectid < last) {
791 key.type = BTRFS_EXTENT_ITEM_KEY;
792 btrfs_release_path(path);
796 if (key.objectid < block_group->start) {
801 if (key.objectid >= block_group->start + block_group->length)
804 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
805 key.type == BTRFS_METADATA_ITEM_KEY) {
808 ret = btrfs_add_new_free_space(block_group, last,
809 key.objectid, &space_added);
812 total_found += space_added;
813 if (key.type == BTRFS_METADATA_ITEM_KEY)
814 last = key.objectid +
817 last = key.objectid + key.offset;
819 if (total_found > CACHING_CTL_WAKE_UP) {
822 atomic_inc(&caching_ctl->progress);
823 wake_up(&caching_ctl->wait);
830 ret = btrfs_add_new_free_space(block_group, last,
831 block_group->start + block_group->length,
834 btrfs_free_path(path);
838 static inline void btrfs_free_excluded_extents(const struct btrfs_block_group *bg)
840 clear_extent_bits(&bg->fs_info->excluded_extents, bg->start,
841 bg->start + bg->length - 1, EXTENT_UPTODATE);
844 static noinline void caching_thread(struct btrfs_work *work)
846 struct btrfs_block_group *block_group;
847 struct btrfs_fs_info *fs_info;
848 struct btrfs_caching_control *caching_ctl;
851 caching_ctl = container_of(work, struct btrfs_caching_control, work);
852 block_group = caching_ctl->block_group;
853 fs_info = block_group->fs_info;
855 mutex_lock(&caching_ctl->mutex);
856 down_read(&fs_info->commit_root_sem);
858 load_block_group_size_class(caching_ctl, block_group);
859 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
860 ret = load_free_space_cache(block_group);
867 * We failed to load the space cache, set ourselves to
868 * CACHE_STARTED and carry on.
870 spin_lock(&block_group->lock);
871 block_group->cached = BTRFS_CACHE_STARTED;
872 spin_unlock(&block_group->lock);
873 wake_up(&caching_ctl->wait);
877 * If we are in the transaction that populated the free space tree we
878 * can't actually cache from the free space tree as our commit root and
879 * real root are the same, so we could change the contents of the blocks
880 * while caching. Instead do the slow caching in this case, and after
881 * the transaction has committed we will be safe.
883 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
884 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
885 ret = load_free_space_tree(caching_ctl);
887 ret = load_extent_tree_free(caching_ctl);
889 spin_lock(&block_group->lock);
890 block_group->caching_ctl = NULL;
891 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
892 spin_unlock(&block_group->lock);
894 #ifdef CONFIG_BTRFS_DEBUG
895 if (btrfs_should_fragment_free_space(block_group)) {
898 spin_lock(&block_group->space_info->lock);
899 spin_lock(&block_group->lock);
900 bytes_used = block_group->length - block_group->used;
901 block_group->space_info->bytes_used += bytes_used >> 1;
902 spin_unlock(&block_group->lock);
903 spin_unlock(&block_group->space_info->lock);
904 fragment_free_space(block_group);
908 up_read(&fs_info->commit_root_sem);
909 btrfs_free_excluded_extents(block_group);
910 mutex_unlock(&caching_ctl->mutex);
912 wake_up(&caching_ctl->wait);
914 btrfs_put_caching_control(caching_ctl);
915 btrfs_put_block_group(block_group);
918 int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
920 struct btrfs_fs_info *fs_info = cache->fs_info;
921 struct btrfs_caching_control *caching_ctl = NULL;
924 /* Allocator for zoned filesystems does not use the cache at all */
925 if (btrfs_is_zoned(fs_info))
928 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
932 INIT_LIST_HEAD(&caching_ctl->list);
933 mutex_init(&caching_ctl->mutex);
934 init_waitqueue_head(&caching_ctl->wait);
935 caching_ctl->block_group = cache;
936 refcount_set(&caching_ctl->count, 2);
937 atomic_set(&caching_ctl->progress, 0);
938 btrfs_init_work(&caching_ctl->work, caching_thread, NULL);
940 spin_lock(&cache->lock);
941 if (cache->cached != BTRFS_CACHE_NO) {
944 caching_ctl = cache->caching_ctl;
946 refcount_inc(&caching_ctl->count);
947 spin_unlock(&cache->lock);
950 WARN_ON(cache->caching_ctl);
951 cache->caching_ctl = caching_ctl;
952 cache->cached = BTRFS_CACHE_STARTED;
953 spin_unlock(&cache->lock);
955 write_lock(&fs_info->block_group_cache_lock);
956 refcount_inc(&caching_ctl->count);
957 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
958 write_unlock(&fs_info->block_group_cache_lock);
960 btrfs_get_block_group(cache);
962 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
964 if (wait && caching_ctl)
965 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
967 btrfs_put_caching_control(caching_ctl);
972 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
974 u64 extra_flags = chunk_to_extended(flags) &
975 BTRFS_EXTENDED_PROFILE_MASK;
977 write_seqlock(&fs_info->profiles_lock);
978 if (flags & BTRFS_BLOCK_GROUP_DATA)
979 fs_info->avail_data_alloc_bits &= ~extra_flags;
980 if (flags & BTRFS_BLOCK_GROUP_METADATA)
981 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
982 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
983 fs_info->avail_system_alloc_bits &= ~extra_flags;
984 write_sequnlock(&fs_info->profiles_lock);
988 * Clear incompat bits for the following feature(s):
990 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
991 * in the whole filesystem
993 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
995 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
997 bool found_raid56 = false;
998 bool found_raid1c34 = false;
1000 if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
1001 (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
1002 (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
1003 struct list_head *head = &fs_info->space_info;
1004 struct btrfs_space_info *sinfo;
1006 list_for_each_entry_rcu(sinfo, head, list) {
1007 down_read(&sinfo->groups_sem);
1008 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
1009 found_raid56 = true;
1010 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
1011 found_raid56 = true;
1012 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
1013 found_raid1c34 = true;
1014 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
1015 found_raid1c34 = true;
1016 up_read(&sinfo->groups_sem);
1019 btrfs_clear_fs_incompat(fs_info, RAID56);
1020 if (!found_raid1c34)
1021 btrfs_clear_fs_incompat(fs_info, RAID1C34);
1025 static int remove_block_group_item(struct btrfs_trans_handle *trans,
1026 struct btrfs_path *path,
1027 struct btrfs_block_group *block_group)
1029 struct btrfs_fs_info *fs_info = trans->fs_info;
1030 struct btrfs_root *root;
1031 struct btrfs_key key;
1034 root = btrfs_block_group_root(fs_info);
1035 key.objectid = block_group->start;
1036 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1037 key.offset = block_group->length;
1039 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1045 ret = btrfs_del_item(trans, root, path);
1049 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1050 struct btrfs_chunk_map *map)
1052 struct btrfs_fs_info *fs_info = trans->fs_info;
1053 struct btrfs_path *path;
1054 struct btrfs_block_group *block_group;
1055 struct btrfs_free_cluster *cluster;
1056 struct inode *inode;
1057 struct kobject *kobj = NULL;
1061 struct btrfs_caching_control *caching_ctl = NULL;
1063 bool remove_rsv = false;
1065 block_group = btrfs_lookup_block_group(fs_info, map->start);
1069 BUG_ON(!block_group->ro);
1071 trace_btrfs_remove_block_group(block_group);
1073 * Free the reserved super bytes from this block group before
1076 btrfs_free_excluded_extents(block_group);
1077 btrfs_free_ref_tree_range(fs_info, block_group->start,
1078 block_group->length);
1080 index = btrfs_bg_flags_to_raid_index(block_group->flags);
1081 factor = btrfs_bg_type_to_factor(block_group->flags);
1083 /* make sure this block group isn't part of an allocation cluster */
1084 cluster = &fs_info->data_alloc_cluster;
1085 spin_lock(&cluster->refill_lock);
1086 btrfs_return_cluster_to_free_space(block_group, cluster);
1087 spin_unlock(&cluster->refill_lock);
1090 * make sure this block group isn't part of a metadata
1091 * allocation cluster
1093 cluster = &fs_info->meta_alloc_cluster;
1094 spin_lock(&cluster->refill_lock);
1095 btrfs_return_cluster_to_free_space(block_group, cluster);
1096 spin_unlock(&cluster->refill_lock);
1098 btrfs_clear_treelog_bg(block_group);
1099 btrfs_clear_data_reloc_bg(block_group);
1101 path = btrfs_alloc_path();
1108 * get the inode first so any iput calls done for the io_list
1109 * aren't the final iput (no unlinks allowed now)
1111 inode = lookup_free_space_inode(block_group, path);
1113 mutex_lock(&trans->transaction->cache_write_mutex);
1115 * Make sure our free space cache IO is done before removing the
1118 spin_lock(&trans->transaction->dirty_bgs_lock);
1119 if (!list_empty(&block_group->io_list)) {
1120 list_del_init(&block_group->io_list);
1122 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1124 spin_unlock(&trans->transaction->dirty_bgs_lock);
1125 btrfs_wait_cache_io(trans, block_group, path);
1126 btrfs_put_block_group(block_group);
1127 spin_lock(&trans->transaction->dirty_bgs_lock);
1130 if (!list_empty(&block_group->dirty_list)) {
1131 list_del_init(&block_group->dirty_list);
1133 btrfs_put_block_group(block_group);
1135 spin_unlock(&trans->transaction->dirty_bgs_lock);
1136 mutex_unlock(&trans->transaction->cache_write_mutex);
1138 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1142 write_lock(&fs_info->block_group_cache_lock);
1143 rb_erase_cached(&block_group->cache_node,
1144 &fs_info->block_group_cache_tree);
1145 RB_CLEAR_NODE(&block_group->cache_node);
1147 /* Once for the block groups rbtree */
1148 btrfs_put_block_group(block_group);
1150 write_unlock(&fs_info->block_group_cache_lock);
1152 down_write(&block_group->space_info->groups_sem);
1154 * we must use list_del_init so people can check to see if they
1155 * are still on the list after taking the semaphore
1157 list_del_init(&block_group->list);
1158 if (list_empty(&block_group->space_info->block_groups[index])) {
1159 kobj = block_group->space_info->block_group_kobjs[index];
1160 block_group->space_info->block_group_kobjs[index] = NULL;
1161 clear_avail_alloc_bits(fs_info, block_group->flags);
1163 up_write(&block_group->space_info->groups_sem);
1164 clear_incompat_bg_bits(fs_info, block_group->flags);
1170 if (block_group->cached == BTRFS_CACHE_STARTED)
1171 btrfs_wait_block_group_cache_done(block_group);
1173 write_lock(&fs_info->block_group_cache_lock);
1174 caching_ctl = btrfs_get_caching_control(block_group);
1176 struct btrfs_caching_control *ctl;
1178 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1179 if (ctl->block_group == block_group) {
1181 refcount_inc(&caching_ctl->count);
1187 list_del_init(&caching_ctl->list);
1188 write_unlock(&fs_info->block_group_cache_lock);
1191 /* Once for the caching bgs list and once for us. */
1192 btrfs_put_caching_control(caching_ctl);
1193 btrfs_put_caching_control(caching_ctl);
1196 spin_lock(&trans->transaction->dirty_bgs_lock);
1197 WARN_ON(!list_empty(&block_group->dirty_list));
1198 WARN_ON(!list_empty(&block_group->io_list));
1199 spin_unlock(&trans->transaction->dirty_bgs_lock);
1201 btrfs_remove_free_space_cache(block_group);
1203 spin_lock(&block_group->space_info->lock);
1204 list_del_init(&block_group->ro_list);
1206 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1207 WARN_ON(block_group->space_info->total_bytes
1208 < block_group->length);
1209 WARN_ON(block_group->space_info->bytes_readonly
1210 < block_group->length - block_group->zone_unusable);
1211 WARN_ON(block_group->space_info->bytes_zone_unusable
1212 < block_group->zone_unusable);
1213 WARN_ON(block_group->space_info->disk_total
1214 < block_group->length * factor);
1216 block_group->space_info->total_bytes -= block_group->length;
1217 block_group->space_info->bytes_readonly -=
1218 (block_group->length - block_group->zone_unusable);
1219 block_group->space_info->bytes_zone_unusable -=
1220 block_group->zone_unusable;
1221 block_group->space_info->disk_total -= block_group->length * factor;
1223 spin_unlock(&block_group->space_info->lock);
1226 * Remove the free space for the block group from the free space tree
1227 * and the block group's item from the extent tree before marking the
1228 * block group as removed. This is to prevent races with tasks that
1229 * freeze and unfreeze a block group, this task and another task
1230 * allocating a new block group - the unfreeze task ends up removing
1231 * the block group's extent map before the task calling this function
1232 * deletes the block group item from the extent tree, allowing for
1233 * another task to attempt to create another block group with the same
1234 * item key (and failing with -EEXIST and a transaction abort).
1236 ret = remove_block_group_free_space(trans, block_group);
1240 ret = remove_block_group_item(trans, path, block_group);
1244 spin_lock(&block_group->lock);
1245 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1248 * At this point trimming or scrub can't start on this block group,
1249 * because we removed the block group from the rbtree
1250 * fs_info->block_group_cache_tree so no one can't find it anymore and
1251 * even if someone already got this block group before we removed it
1252 * from the rbtree, they have already incremented block_group->frozen -
1253 * if they didn't, for the trimming case they won't find any free space
1254 * entries because we already removed them all when we called
1255 * btrfs_remove_free_space_cache().
1257 * And we must not remove the chunk map from the fs_info->mapping_tree
1258 * to prevent the same logical address range and physical device space
1259 * ranges from being reused for a new block group. This is needed to
1260 * avoid races with trimming and scrub.
1262 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1263 * completely transactionless, so while it is trimming a range the
1264 * currently running transaction might finish and a new one start,
1265 * allowing for new block groups to be created that can reuse the same
1266 * physical device locations unless we take this special care.
1268 * There may also be an implicit trim operation if the file system
1269 * is mounted with -odiscard. The same protections must remain
1270 * in place until the extents have been discarded completely when
1271 * the transaction commit has completed.
1273 remove_map = (atomic_read(&block_group->frozen) == 0);
1274 spin_unlock(&block_group->lock);
1277 btrfs_remove_chunk_map(fs_info, map);
1280 /* Once for the lookup reference */
1281 btrfs_put_block_group(block_group);
1283 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
1284 btrfs_free_path(path);
1288 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1289 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1291 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1292 struct btrfs_chunk_map *map;
1293 unsigned int num_items;
1295 map = btrfs_find_chunk_map(fs_info, chunk_offset, 1);
1296 ASSERT(map != NULL);
1297 ASSERT(map->start == chunk_offset);
1300 * We need to reserve 3 + N units from the metadata space info in order
1301 * to remove a block group (done at btrfs_remove_chunk() and at
1302 * btrfs_remove_block_group()), which are used for:
1304 * 1 unit for adding the free space inode's orphan (located in the tree
1306 * 1 unit for deleting the block group item (located in the extent
1308 * 1 unit for deleting the free space item (located in tree of tree
1310 * N units for deleting N device extent items corresponding to each
1311 * stripe (located in the device tree).
1313 * In order to remove a block group we also need to reserve units in the
1314 * system space info in order to update the chunk tree (update one or
1315 * more device items and remove one chunk item), but this is done at
1316 * btrfs_remove_chunk() through a call to check_system_chunk().
1318 num_items = 3 + map->num_stripes;
1319 btrfs_free_chunk_map(map);
1321 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1325 * Mark block group @cache read-only, so later write won't happen to block
1328 * If @force is not set, this function will only mark the block group readonly
1329 * if we have enough free space (1M) in other metadata/system block groups.
1330 * If @force is not set, this function will mark the block group readonly
1331 * without checking free space.
1333 * NOTE: This function doesn't care if other block groups can contain all the
1334 * data in this block group. That check should be done by relocation routine,
1335 * not this function.
1337 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1339 struct btrfs_space_info *sinfo = cache->space_info;
1343 spin_lock(&sinfo->lock);
1344 spin_lock(&cache->lock);
1346 if (cache->swap_extents) {
1357 num_bytes = cache->length - cache->reserved - cache->pinned -
1358 cache->bytes_super - cache->zone_unusable - cache->used;
1361 * Data never overcommits, even in mixed mode, so do just the straight
1362 * check of left over space in how much we have allocated.
1366 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1367 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1370 * Here we make sure if we mark this bg RO, we still have enough
1371 * free space as buffer.
1373 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1377 * We overcommit metadata, so we need to do the
1378 * btrfs_can_overcommit check here, and we need to pass in
1379 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1380 * leeway to allow us to mark this block group as read only.
1382 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1383 BTRFS_RESERVE_NO_FLUSH))
1388 sinfo->bytes_readonly += num_bytes;
1389 if (btrfs_is_zoned(cache->fs_info)) {
1390 /* Migrate zone_unusable bytes to readonly */
1391 sinfo->bytes_readonly += cache->zone_unusable;
1392 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1393 cache->zone_unusable = 0;
1396 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1399 spin_unlock(&cache->lock);
1400 spin_unlock(&sinfo->lock);
1401 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1402 btrfs_info(cache->fs_info,
1403 "unable to make block group %llu ro", cache->start);
1404 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1409 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1410 struct btrfs_block_group *bg)
1412 struct btrfs_fs_info *fs_info = bg->fs_info;
1413 struct btrfs_transaction *prev_trans = NULL;
1414 const u64 start = bg->start;
1415 const u64 end = start + bg->length - 1;
1418 spin_lock(&fs_info->trans_lock);
1419 if (trans->transaction->list.prev != &fs_info->trans_list) {
1420 prev_trans = list_last_entry(&trans->transaction->list,
1421 struct btrfs_transaction, list);
1422 refcount_inc(&prev_trans->use_count);
1424 spin_unlock(&fs_info->trans_lock);
1427 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1428 * btrfs_finish_extent_commit(). If we are at transaction N, another
1429 * task might be running finish_extent_commit() for the previous
1430 * transaction N - 1, and have seen a range belonging to the block
1431 * group in pinned_extents before we were able to clear the whole block
1432 * group range from pinned_extents. This means that task can lookup for
1433 * the block group after we unpinned it from pinned_extents and removed
1434 * it, leading to an error at unpin_extent_range().
1436 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1438 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1444 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1447 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1449 btrfs_put_transaction(prev_trans);
1455 * Process the unused_bgs list and remove any that don't have any allocated
1456 * space inside of them.
1458 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1460 LIST_HEAD(retry_list);
1461 struct btrfs_block_group *block_group;
1462 struct btrfs_space_info *space_info;
1463 struct btrfs_trans_handle *trans;
1464 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1467 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1470 if (btrfs_fs_closing(fs_info))
1474 * Long running balances can keep us blocked here for eternity, so
1475 * simply skip deletion if we're unable to get the mutex.
1477 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1480 spin_lock(&fs_info->unused_bgs_lock);
1481 while (!list_empty(&fs_info->unused_bgs)) {
1485 block_group = list_first_entry(&fs_info->unused_bgs,
1486 struct btrfs_block_group,
1488 list_del_init(&block_group->bg_list);
1490 space_info = block_group->space_info;
1492 if (ret || btrfs_mixed_space_info(space_info)) {
1493 btrfs_put_block_group(block_group);
1496 spin_unlock(&fs_info->unused_bgs_lock);
1498 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1500 /* Don't want to race with allocators so take the groups_sem */
1501 down_write(&space_info->groups_sem);
1504 * Async discard moves the final block group discard to be prior
1505 * to the unused_bgs code path. Therefore, if it's not fully
1506 * trimmed, punt it back to the async discard lists.
1508 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1509 !btrfs_is_free_space_trimmed(block_group)) {
1510 trace_btrfs_skip_unused_block_group(block_group);
1511 up_write(&space_info->groups_sem);
1512 /* Requeue if we failed because of async discard */
1513 btrfs_discard_queue_work(&fs_info->discard_ctl,
1518 spin_lock(&space_info->lock);
1519 spin_lock(&block_group->lock);
1520 if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1521 list_is_singular(&block_group->list)) {
1523 * We want to bail if we made new allocations or have
1524 * outstanding allocations in this block group. We do
1525 * the ro check in case balance is currently acting on
1528 * Also bail out if this is the only block group for its
1529 * type, because otherwise we would lose profile
1530 * information from fs_info->avail_*_alloc_bits and the
1531 * next block group of this type would be created with a
1532 * "single" profile (even if we're in a raid fs) because
1533 * fs_info->avail_*_alloc_bits would be 0.
1535 trace_btrfs_skip_unused_block_group(block_group);
1536 spin_unlock(&block_group->lock);
1537 spin_unlock(&space_info->lock);
1538 up_write(&space_info->groups_sem);
1543 * The block group may be unused but there may be space reserved
1544 * accounting with the existence of that block group, that is,
1545 * space_info->bytes_may_use was incremented by a task but no
1546 * space was yet allocated from the block group by the task.
1547 * That space may or may not be allocated, as we are generally
1548 * pessimistic about space reservation for metadata as well as
1549 * for data when using compression (as we reserve space based on
1550 * the worst case, when data can't be compressed, and before
1551 * actually attempting compression, before starting writeback).
1553 * So check if the total space of the space_info minus the size
1554 * of this block group is less than the used space of the
1555 * space_info - if that's the case, then it means we have tasks
1556 * that might be relying on the block group in order to allocate
1557 * extents, and add back the block group to the unused list when
1558 * we finish, so that we retry later in case no tasks ended up
1559 * needing to allocate extents from the block group.
1561 used = btrfs_space_info_used(space_info, true);
1562 if (space_info->total_bytes - block_group->length < used &&
1563 block_group->zone_unusable < block_group->length) {
1565 * Add a reference for the list, compensate for the ref
1566 * drop under the "next" label for the
1567 * fs_info->unused_bgs list.
1569 btrfs_get_block_group(block_group);
1570 list_add_tail(&block_group->bg_list, &retry_list);
1572 trace_btrfs_skip_unused_block_group(block_group);
1573 spin_unlock(&block_group->lock);
1574 spin_unlock(&space_info->lock);
1575 up_write(&space_info->groups_sem);
1579 spin_unlock(&block_group->lock);
1580 spin_unlock(&space_info->lock);
1582 /* We don't want to force the issue, only flip if it's ok. */
1583 ret = inc_block_group_ro(block_group, 0);
1584 up_write(&space_info->groups_sem);
1590 ret = btrfs_zone_finish(block_group);
1592 btrfs_dec_block_group_ro(block_group);
1599 * Want to do this before we do anything else so we can recover
1600 * properly if we fail to join the transaction.
1602 trans = btrfs_start_trans_remove_block_group(fs_info,
1603 block_group->start);
1604 if (IS_ERR(trans)) {
1605 btrfs_dec_block_group_ro(block_group);
1606 ret = PTR_ERR(trans);
1611 * We could have pending pinned extents for this block group,
1612 * just delete them, we don't care about them anymore.
1614 if (!clean_pinned_extents(trans, block_group)) {
1615 btrfs_dec_block_group_ro(block_group);
1620 * At this point, the block_group is read only and should fail
1621 * new allocations. However, btrfs_finish_extent_commit() can
1622 * cause this block_group to be placed back on the discard
1623 * lists because now the block_group isn't fully discarded.
1624 * Bail here and try again later after discarding everything.
1626 spin_lock(&fs_info->discard_ctl.lock);
1627 if (!list_empty(&block_group->discard_list)) {
1628 spin_unlock(&fs_info->discard_ctl.lock);
1629 btrfs_dec_block_group_ro(block_group);
1630 btrfs_discard_queue_work(&fs_info->discard_ctl,
1634 spin_unlock(&fs_info->discard_ctl.lock);
1636 /* Reset pinned so btrfs_put_block_group doesn't complain */
1637 spin_lock(&space_info->lock);
1638 spin_lock(&block_group->lock);
1640 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1641 -block_group->pinned);
1642 space_info->bytes_readonly += block_group->pinned;
1643 block_group->pinned = 0;
1645 spin_unlock(&block_group->lock);
1646 spin_unlock(&space_info->lock);
1649 * The normal path here is an unused block group is passed here,
1650 * then trimming is handled in the transaction commit path.
1651 * Async discard interposes before this to do the trimming
1652 * before coming down the unused block group path as trimming
1653 * will no longer be done later in the transaction commit path.
1655 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1659 * DISCARD can flip during remount. On zoned filesystems, we
1660 * need to reset sequential-required zones.
1662 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1663 btrfs_is_zoned(fs_info);
1665 /* Implicit trim during transaction commit. */
1667 btrfs_freeze_block_group(block_group);
1670 * Btrfs_remove_chunk will abort the transaction if things go
1673 ret = btrfs_remove_chunk(trans, block_group->start);
1677 btrfs_unfreeze_block_group(block_group);
1682 * If we're not mounted with -odiscard, we can just forget
1683 * about this block group. Otherwise we'll need to wait
1684 * until transaction commit to do the actual discard.
1687 spin_lock(&fs_info->unused_bgs_lock);
1689 * A concurrent scrub might have added us to the list
1690 * fs_info->unused_bgs, so use a list_move operation
1691 * to add the block group to the deleted_bgs list.
1693 list_move(&block_group->bg_list,
1694 &trans->transaction->deleted_bgs);
1695 spin_unlock(&fs_info->unused_bgs_lock);
1696 btrfs_get_block_group(block_group);
1699 btrfs_end_transaction(trans);
1701 btrfs_put_block_group(block_group);
1702 spin_lock(&fs_info->unused_bgs_lock);
1704 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1705 spin_unlock(&fs_info->unused_bgs_lock);
1706 mutex_unlock(&fs_info->reclaim_bgs_lock);
1710 btrfs_end_transaction(trans);
1711 spin_lock(&fs_info->unused_bgs_lock);
1712 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1713 spin_unlock(&fs_info->unused_bgs_lock);
1714 mutex_unlock(&fs_info->reclaim_bgs_lock);
1715 btrfs_put_block_group(block_group);
1716 btrfs_discard_punt_unused_bgs_list(fs_info);
1719 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1721 struct btrfs_fs_info *fs_info = bg->fs_info;
1723 spin_lock(&fs_info->unused_bgs_lock);
1724 if (list_empty(&bg->bg_list)) {
1725 btrfs_get_block_group(bg);
1726 trace_btrfs_add_unused_block_group(bg);
1727 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1728 } else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1729 /* Pull out the block group from the reclaim_bgs list. */
1730 trace_btrfs_add_unused_block_group(bg);
1731 list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1733 spin_unlock(&fs_info->unused_bgs_lock);
1737 * We want block groups with a low number of used bytes to be in the beginning
1738 * of the list, so they will get reclaimed first.
1740 static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1741 const struct list_head *b)
1743 const struct btrfs_block_group *bg1, *bg2;
1745 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1746 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1748 return bg1->used > bg2->used;
1751 static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1753 if (btrfs_is_zoned(fs_info))
1754 return btrfs_zoned_should_reclaim(fs_info);
1758 static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1760 const struct btrfs_space_info *space_info = bg->space_info;
1761 const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1762 const u64 new_val = bg->used;
1763 const u64 old_val = new_val + bytes_freed;
1766 if (reclaim_thresh == 0)
1769 thresh = mult_perc(bg->length, reclaim_thresh);
1772 * If we were below the threshold before don't reclaim, we are likely a
1773 * brand new block group and we don't want to relocate new block groups.
1775 if (old_val < thresh)
1777 if (new_val >= thresh)
1782 void btrfs_reclaim_bgs_work(struct work_struct *work)
1784 struct btrfs_fs_info *fs_info =
1785 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1786 struct btrfs_block_group *bg;
1787 struct btrfs_space_info *space_info;
1789 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1792 if (btrfs_fs_closing(fs_info))
1795 if (!btrfs_should_reclaim(fs_info))
1798 sb_start_write(fs_info->sb);
1800 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1801 sb_end_write(fs_info->sb);
1806 * Long running balances can keep us blocked here for eternity, so
1807 * simply skip reclaim if we're unable to get the mutex.
1809 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1810 btrfs_exclop_finish(fs_info);
1811 sb_end_write(fs_info->sb);
1815 spin_lock(&fs_info->unused_bgs_lock);
1817 * Sort happens under lock because we can't simply splice it and sort.
1818 * The block groups might still be in use and reachable via bg_list,
1819 * and their presence in the reclaim_bgs list must be preserved.
1821 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1822 while (!list_empty(&fs_info->reclaim_bgs)) {
1826 bg = list_first_entry(&fs_info->reclaim_bgs,
1827 struct btrfs_block_group,
1829 list_del_init(&bg->bg_list);
1831 space_info = bg->space_info;
1832 spin_unlock(&fs_info->unused_bgs_lock);
1834 /* Don't race with allocators so take the groups_sem */
1835 down_write(&space_info->groups_sem);
1837 spin_lock(&bg->lock);
1838 if (bg->reserved || bg->pinned || bg->ro) {
1840 * We want to bail if we made new allocations or have
1841 * outstanding allocations in this block group. We do
1842 * the ro check in case balance is currently acting on
1845 spin_unlock(&bg->lock);
1846 up_write(&space_info->groups_sem);
1849 if (bg->used == 0) {
1851 * It is possible that we trigger relocation on a block
1852 * group as its extents are deleted and it first goes
1853 * below the threshold, then shortly after goes empty.
1855 * In this case, relocating it does delete it, but has
1856 * some overhead in relocation specific metadata, looking
1857 * for the non-existent extents and running some extra
1858 * transactions, which we can avoid by using one of the
1859 * other mechanisms for dealing with empty block groups.
1861 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1862 btrfs_mark_bg_unused(bg);
1863 spin_unlock(&bg->lock);
1864 up_write(&space_info->groups_sem);
1869 * The block group might no longer meet the reclaim condition by
1870 * the time we get around to reclaiming it, so to avoid
1871 * reclaiming overly full block_groups, skip reclaiming them.
1873 * Since the decision making process also depends on the amount
1874 * being freed, pass in a fake giant value to skip that extra
1875 * check, which is more meaningful when adding to the list in
1878 if (!should_reclaim_block_group(bg, bg->length)) {
1879 spin_unlock(&bg->lock);
1880 up_write(&space_info->groups_sem);
1883 spin_unlock(&bg->lock);
1886 * Get out fast, in case we're read-only or unmounting the
1887 * filesystem. It is OK to drop block groups from the list even
1888 * for the read-only case. As we did sb_start_write(),
1889 * "mount -o remount,ro" won't happen and read-only filesystem
1890 * means it is forced read-only due to a fatal error. So, it
1891 * never gets back to read-write to let us reclaim again.
1893 if (btrfs_need_cleaner_sleep(fs_info)) {
1894 up_write(&space_info->groups_sem);
1899 * Cache the zone_unusable value before turning the block group
1900 * to read only. As soon as the blog group is read only it's
1901 * zone_unusable value gets moved to the block group's read-only
1902 * bytes and isn't available for calculations anymore.
1904 zone_unusable = bg->zone_unusable;
1905 ret = inc_block_group_ro(bg, 0);
1906 up_write(&space_info->groups_sem);
1911 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1913 div64_u64(bg->used * 100, bg->length),
1914 div64_u64(zone_unusable * 100, bg->length));
1915 trace_btrfs_reclaim_block_group(bg);
1916 ret = btrfs_relocate_chunk(fs_info, bg->start);
1918 btrfs_dec_block_group_ro(bg);
1919 btrfs_err(fs_info, "error relocating chunk %llu",
1925 btrfs_mark_bg_to_reclaim(bg);
1926 btrfs_put_block_group(bg);
1928 mutex_unlock(&fs_info->reclaim_bgs_lock);
1930 * Reclaiming all the block groups in the list can take really
1931 * long. Prioritize cleaning up unused block groups.
1933 btrfs_delete_unused_bgs(fs_info);
1935 * If we are interrupted by a balance, we can just bail out. The
1936 * cleaner thread restart again if necessary.
1938 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1940 spin_lock(&fs_info->unused_bgs_lock);
1942 spin_unlock(&fs_info->unused_bgs_lock);
1943 mutex_unlock(&fs_info->reclaim_bgs_lock);
1945 btrfs_exclop_finish(fs_info);
1946 sb_end_write(fs_info->sb);
1949 void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1951 spin_lock(&fs_info->unused_bgs_lock);
1952 if (!list_empty(&fs_info->reclaim_bgs))
1953 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1954 spin_unlock(&fs_info->unused_bgs_lock);
1957 void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1959 struct btrfs_fs_info *fs_info = bg->fs_info;
1961 spin_lock(&fs_info->unused_bgs_lock);
1962 if (list_empty(&bg->bg_list)) {
1963 btrfs_get_block_group(bg);
1964 trace_btrfs_add_reclaim_block_group(bg);
1965 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1967 spin_unlock(&fs_info->unused_bgs_lock);
1970 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1971 struct btrfs_path *path)
1973 struct btrfs_chunk_map *map;
1974 struct btrfs_block_group_item bg;
1975 struct extent_buffer *leaf;
1980 slot = path->slots[0];
1981 leaf = path->nodes[0];
1983 map = btrfs_find_chunk_map(fs_info, key->objectid, key->offset);
1986 "logical %llu len %llu found bg but no related chunk",
1987 key->objectid, key->offset);
1991 if (map->start != key->objectid || map->chunk_len != key->offset) {
1993 "block group %llu len %llu mismatch with chunk %llu len %llu",
1994 key->objectid, key->offset, map->start, map->chunk_len);
1999 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
2001 flags = btrfs_stack_block_group_flags(&bg) &
2002 BTRFS_BLOCK_GROUP_TYPE_MASK;
2004 if (flags != (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2006 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
2007 key->objectid, key->offset, flags,
2008 (BTRFS_BLOCK_GROUP_TYPE_MASK & map->type));
2013 btrfs_free_chunk_map(map);
2017 static int find_first_block_group(struct btrfs_fs_info *fs_info,
2018 struct btrfs_path *path,
2019 struct btrfs_key *key)
2021 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2023 struct btrfs_key found_key;
2025 btrfs_for_each_slot(root, key, &found_key, path, ret) {
2026 if (found_key.objectid >= key->objectid &&
2027 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2028 return read_bg_from_eb(fs_info, &found_key, path);
2034 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2036 u64 extra_flags = chunk_to_extended(flags) &
2037 BTRFS_EXTENDED_PROFILE_MASK;
2039 write_seqlock(&fs_info->profiles_lock);
2040 if (flags & BTRFS_BLOCK_GROUP_DATA)
2041 fs_info->avail_data_alloc_bits |= extra_flags;
2042 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2043 fs_info->avail_metadata_alloc_bits |= extra_flags;
2044 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2045 fs_info->avail_system_alloc_bits |= extra_flags;
2046 write_sequnlock(&fs_info->profiles_lock);
2050 * Map a physical disk address to a list of logical addresses.
2052 * @fs_info: the filesystem
2053 * @chunk_start: logical address of block group
2054 * @physical: physical address to map to logical addresses
2055 * @logical: return array of logical addresses which map to @physical
2056 * @naddrs: length of @logical
2057 * @stripe_len: size of IO stripe for the given block group
2059 * Maps a particular @physical disk address to a list of @logical addresses.
2060 * Used primarily to exclude those portions of a block group that contain super
2063 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2064 u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2066 struct btrfs_chunk_map *map;
2069 u64 data_stripe_length;
2074 map = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2078 data_stripe_length = map->stripe_size;
2079 io_stripe_size = BTRFS_STRIPE_LEN;
2080 chunk_start = map->start;
2082 /* For RAID5/6 adjust to a full IO stripe length */
2083 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2084 io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2086 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2092 for (i = 0; i < map->num_stripes; i++) {
2093 bool already_inserted = false;
2098 if (!in_range(physical, map->stripes[i].physical,
2099 data_stripe_length))
2102 stripe_nr = (physical - map->stripes[i].physical) >>
2103 BTRFS_STRIPE_LEN_SHIFT;
2104 offset = (physical - map->stripes[i].physical) &
2105 BTRFS_STRIPE_LEN_MASK;
2107 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2108 BTRFS_BLOCK_GROUP_RAID10))
2109 stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2112 * The remaining case would be for RAID56, multiply by
2113 * nr_data_stripes(). Alternatively, just use rmap_len below
2114 * instead of map->stripe_len
2116 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2118 /* Ensure we don't add duplicate addresses */
2119 for (j = 0; j < nr; j++) {
2120 if (buf[j] == bytenr) {
2121 already_inserted = true;
2126 if (!already_inserted)
2132 *stripe_len = io_stripe_size;
2134 btrfs_free_chunk_map(map);
2138 static int exclude_super_stripes(struct btrfs_block_group *cache)
2140 struct btrfs_fs_info *fs_info = cache->fs_info;
2141 const bool zoned = btrfs_is_zoned(fs_info);
2147 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2148 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2149 cache->bytes_super += stripe_len;
2150 ret = set_extent_bit(&fs_info->excluded_extents, cache->start,
2151 cache->start + stripe_len - 1,
2152 EXTENT_UPTODATE, NULL);
2157 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2158 bytenr = btrfs_sb_offset(i);
2159 ret = btrfs_rmap_block(fs_info, cache->start,
2160 bytenr, &logical, &nr, &stripe_len);
2164 /* Shouldn't have super stripes in sequential zones */
2168 "zoned: block group %llu must not contain super block",
2174 u64 len = min_t(u64, stripe_len,
2175 cache->start + cache->length - logical[nr]);
2177 cache->bytes_super += len;
2178 ret = set_extent_bit(&fs_info->excluded_extents, logical[nr],
2179 logical[nr] + len - 1,
2180 EXTENT_UPTODATE, NULL);
2192 static struct btrfs_block_group *btrfs_create_block_group_cache(
2193 struct btrfs_fs_info *fs_info, u64 start)
2195 struct btrfs_block_group *cache;
2197 cache = kzalloc(sizeof(*cache), GFP_NOFS);
2201 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2203 if (!cache->free_space_ctl) {
2208 cache->start = start;
2210 cache->fs_info = fs_info;
2211 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2213 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2215 refcount_set(&cache->refs, 1);
2216 spin_lock_init(&cache->lock);
2217 init_rwsem(&cache->data_rwsem);
2218 INIT_LIST_HEAD(&cache->list);
2219 INIT_LIST_HEAD(&cache->cluster_list);
2220 INIT_LIST_HEAD(&cache->bg_list);
2221 INIT_LIST_HEAD(&cache->ro_list);
2222 INIT_LIST_HEAD(&cache->discard_list);
2223 INIT_LIST_HEAD(&cache->dirty_list);
2224 INIT_LIST_HEAD(&cache->io_list);
2225 INIT_LIST_HEAD(&cache->active_bg_list);
2226 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2227 atomic_set(&cache->frozen, 0);
2228 mutex_init(&cache->free_space_lock);
2234 * Iterate all chunks and verify that each of them has the corresponding block
2237 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2243 struct btrfs_chunk_map *map;
2244 struct btrfs_block_group *bg;
2247 * btrfs_find_chunk_map() will return the first chunk map
2248 * intersecting the range, so setting @length to 1 is enough to
2249 * get the first chunk.
2251 map = btrfs_find_chunk_map(fs_info, start, 1);
2255 bg = btrfs_lookup_block_group(fs_info, map->start);
2258 "chunk start=%llu len=%llu doesn't have corresponding block group",
2259 map->start, map->chunk_len);
2261 btrfs_free_chunk_map(map);
2264 if (bg->start != map->start || bg->length != map->chunk_len ||
2265 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2266 (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2268 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2269 map->start, map->chunk_len,
2270 map->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2271 bg->start, bg->length,
2272 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2274 btrfs_free_chunk_map(map);
2275 btrfs_put_block_group(bg);
2278 start = map->start + map->chunk_len;
2279 btrfs_free_chunk_map(map);
2280 btrfs_put_block_group(bg);
2285 static int read_one_block_group(struct btrfs_fs_info *info,
2286 struct btrfs_block_group_item *bgi,
2287 const struct btrfs_key *key,
2290 struct btrfs_block_group *cache;
2291 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2294 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2296 cache = btrfs_create_block_group_cache(info, key->objectid);
2300 cache->length = key->offset;
2301 cache->used = btrfs_stack_block_group_used(bgi);
2302 cache->commit_used = cache->used;
2303 cache->flags = btrfs_stack_block_group_flags(bgi);
2304 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2306 set_free_space_tree_thresholds(cache);
2310 * When we mount with old space cache, we need to
2311 * set BTRFS_DC_CLEAR and set dirty flag.
2313 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2314 * truncate the old free space cache inode and
2316 * b) Setting 'dirty flag' makes sure that we flush
2317 * the new space cache info onto disk.
2319 if (btrfs_test_opt(info, SPACE_CACHE))
2320 cache->disk_cache_state = BTRFS_DC_CLEAR;
2322 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2323 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2325 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2331 ret = btrfs_load_block_group_zone_info(cache, false);
2333 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2339 * We need to exclude the super stripes now so that the space info has
2340 * super bytes accounted for, otherwise we'll think we have more space
2341 * than we actually do.
2343 ret = exclude_super_stripes(cache);
2345 /* We may have excluded something, so call this just in case. */
2346 btrfs_free_excluded_extents(cache);
2351 * For zoned filesystem, space after the allocation offset is the only
2352 * free space for a block group. So, we don't need any caching work.
2353 * btrfs_calc_zone_unusable() will set the amount of free space and
2354 * zone_unusable space.
2356 * For regular filesystem, check for two cases, either we are full, and
2357 * therefore don't need to bother with the caching work since we won't
2358 * find any space, or we are empty, and we can just add all the space
2359 * in and be done with it. This saves us _a_lot_ of time, particularly
2362 if (btrfs_is_zoned(info)) {
2363 btrfs_calc_zone_unusable(cache);
2364 /* Should not have any excluded extents. Just in case, though. */
2365 btrfs_free_excluded_extents(cache);
2366 } else if (cache->length == cache->used) {
2367 cache->cached = BTRFS_CACHE_FINISHED;
2368 btrfs_free_excluded_extents(cache);
2369 } else if (cache->used == 0) {
2370 cache->cached = BTRFS_CACHE_FINISHED;
2371 ret = btrfs_add_new_free_space(cache, cache->start,
2372 cache->start + cache->length, NULL);
2373 btrfs_free_excluded_extents(cache);
2378 ret = btrfs_add_block_group_cache(info, cache);
2380 btrfs_remove_free_space_cache(cache);
2383 trace_btrfs_add_block_group(info, cache, 0);
2384 btrfs_add_bg_to_space_info(info, cache);
2386 set_avail_alloc_bits(info, cache->flags);
2387 if (btrfs_chunk_writeable(info, cache->start)) {
2388 if (cache->used == 0) {
2389 ASSERT(list_empty(&cache->bg_list));
2390 if (btrfs_test_opt(info, DISCARD_ASYNC))
2391 btrfs_discard_queue_work(&info->discard_ctl, cache);
2393 btrfs_mark_bg_unused(cache);
2396 inc_block_group_ro(cache, 1);
2401 btrfs_put_block_group(cache);
2405 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2407 struct rb_node *node;
2410 for (node = rb_first_cached(&fs_info->mapping_tree); node; node = rb_next(node)) {
2411 struct btrfs_chunk_map *map;
2412 struct btrfs_block_group *bg;
2414 map = rb_entry(node, struct btrfs_chunk_map, rb_node);
2415 bg = btrfs_create_block_group_cache(fs_info, map->start);
2421 /* Fill dummy cache as FULL */
2422 bg->length = map->chunk_len;
2423 bg->flags = map->type;
2424 bg->cached = BTRFS_CACHE_FINISHED;
2425 bg->used = map->chunk_len;
2426 bg->flags = map->type;
2427 ret = btrfs_add_block_group_cache(fs_info, bg);
2429 * We may have some valid block group cache added already, in
2430 * that case we skip to the next one.
2432 if (ret == -EEXIST) {
2434 btrfs_put_block_group(bg);
2439 btrfs_remove_free_space_cache(bg);
2440 btrfs_put_block_group(bg);
2444 btrfs_add_bg_to_space_info(fs_info, bg);
2446 set_avail_alloc_bits(fs_info, bg->flags);
2449 btrfs_init_global_block_rsv(fs_info);
2453 int btrfs_read_block_groups(struct btrfs_fs_info *info)
2455 struct btrfs_root *root = btrfs_block_group_root(info);
2456 struct btrfs_path *path;
2458 struct btrfs_block_group *cache;
2459 struct btrfs_space_info *space_info;
2460 struct btrfs_key key;
2465 * Either no extent root (with ibadroots rescue option) or we have
2466 * unsupported RO options. The fs can never be mounted read-write, so no
2467 * need to waste time searching block group items.
2469 * This also allows new extent tree related changes to be RO compat,
2470 * no need for a full incompat flag.
2472 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2473 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2474 return fill_dummy_bgs(info);
2478 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2479 path = btrfs_alloc_path();
2483 cache_gen = btrfs_super_cache_generation(info->super_copy);
2484 if (btrfs_test_opt(info, SPACE_CACHE) &&
2485 btrfs_super_generation(info->super_copy) != cache_gen)
2487 if (btrfs_test_opt(info, CLEAR_CACHE))
2491 struct btrfs_block_group_item bgi;
2492 struct extent_buffer *leaf;
2495 ret = find_first_block_group(info, path, &key);
2501 leaf = path->nodes[0];
2502 slot = path->slots[0];
2504 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2507 btrfs_item_key_to_cpu(leaf, &key, slot);
2508 btrfs_release_path(path);
2509 ret = read_one_block_group(info, &bgi, &key, need_clear);
2512 key.objectid += key.offset;
2515 btrfs_release_path(path);
2517 list_for_each_entry(space_info, &info->space_info, list) {
2520 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2521 if (list_empty(&space_info->block_groups[i]))
2523 cache = list_first_entry(&space_info->block_groups[i],
2524 struct btrfs_block_group,
2526 btrfs_sysfs_add_block_group_type(cache);
2529 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2530 (BTRFS_BLOCK_GROUP_RAID10 |
2531 BTRFS_BLOCK_GROUP_RAID1_MASK |
2532 BTRFS_BLOCK_GROUP_RAID56_MASK |
2533 BTRFS_BLOCK_GROUP_DUP)))
2536 * Avoid allocating from un-mirrored block group if there are
2537 * mirrored block groups.
2539 list_for_each_entry(cache,
2540 &space_info->block_groups[BTRFS_RAID_RAID0],
2542 inc_block_group_ro(cache, 1);
2543 list_for_each_entry(cache,
2544 &space_info->block_groups[BTRFS_RAID_SINGLE],
2546 inc_block_group_ro(cache, 1);
2549 btrfs_init_global_block_rsv(info);
2550 ret = check_chunk_block_group_mappings(info);
2552 btrfs_free_path(path);
2554 * We've hit some error while reading the extent tree, and have
2555 * rescue=ibadroots mount option.
2556 * Try to fill the tree using dummy block groups so that the user can
2557 * continue to mount and grab their data.
2559 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2560 ret = fill_dummy_bgs(info);
2565 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2568 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2571 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2572 struct btrfs_block_group *block_group)
2574 struct btrfs_fs_info *fs_info = trans->fs_info;
2575 struct btrfs_block_group_item bgi;
2576 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2577 struct btrfs_key key;
2578 u64 old_commit_used;
2581 spin_lock(&block_group->lock);
2582 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2583 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2584 block_group->global_root_id);
2585 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2586 old_commit_used = block_group->commit_used;
2587 block_group->commit_used = block_group->used;
2588 key.objectid = block_group->start;
2589 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2590 key.offset = block_group->length;
2591 spin_unlock(&block_group->lock);
2593 ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2595 spin_lock(&block_group->lock);
2596 block_group->commit_used = old_commit_used;
2597 spin_unlock(&block_group->lock);
2603 static int insert_dev_extent(struct btrfs_trans_handle *trans,
2604 struct btrfs_device *device, u64 chunk_offset,
2605 u64 start, u64 num_bytes)
2607 struct btrfs_fs_info *fs_info = device->fs_info;
2608 struct btrfs_root *root = fs_info->dev_root;
2609 struct btrfs_path *path;
2610 struct btrfs_dev_extent *extent;
2611 struct extent_buffer *leaf;
2612 struct btrfs_key key;
2615 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2616 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2617 path = btrfs_alloc_path();
2621 key.objectid = device->devid;
2622 key.type = BTRFS_DEV_EXTENT_KEY;
2624 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2628 leaf = path->nodes[0];
2629 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2630 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2631 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2632 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2633 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2635 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2636 btrfs_mark_buffer_dirty(trans, leaf);
2638 btrfs_free_path(path);
2643 * This function belongs to phase 2.
2645 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2648 static int insert_dev_extents(struct btrfs_trans_handle *trans,
2649 u64 chunk_offset, u64 chunk_size)
2651 struct btrfs_fs_info *fs_info = trans->fs_info;
2652 struct btrfs_device *device;
2653 struct btrfs_chunk_map *map;
2658 map = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2660 return PTR_ERR(map);
2663 * Take the device list mutex to prevent races with the final phase of
2664 * a device replace operation that replaces the device object associated
2665 * with the map's stripes, because the device object's id can change
2666 * at any time during that final phase of the device replace operation
2667 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2668 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2669 * resulting in persisting a device extent item with such ID.
2671 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2672 for (i = 0; i < map->num_stripes; i++) {
2673 device = map->stripes[i].dev;
2674 dev_offset = map->stripes[i].physical;
2676 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2681 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2683 btrfs_free_chunk_map(map);
2688 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2691 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2694 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2696 struct btrfs_fs_info *fs_info = trans->fs_info;
2697 struct btrfs_block_group *block_group;
2700 while (!list_empty(&trans->new_bgs)) {
2703 block_group = list_first_entry(&trans->new_bgs,
2704 struct btrfs_block_group,
2709 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2711 ret = insert_block_group_item(trans, block_group);
2713 btrfs_abort_transaction(trans, ret);
2714 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2715 &block_group->runtime_flags)) {
2716 mutex_lock(&fs_info->chunk_mutex);
2717 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2718 mutex_unlock(&fs_info->chunk_mutex);
2720 btrfs_abort_transaction(trans, ret);
2722 ret = insert_dev_extents(trans, block_group->start,
2723 block_group->length);
2725 btrfs_abort_transaction(trans, ret);
2726 add_block_group_free_space(trans, block_group);
2729 * If we restriped during balance, we may have added a new raid
2730 * type, so now add the sysfs entries when it is safe to do so.
2731 * We don't have to worry about locking here as it's handled in
2732 * btrfs_sysfs_add_block_group_type.
2734 if (block_group->space_info->block_group_kobjs[index] == NULL)
2735 btrfs_sysfs_add_block_group_type(block_group);
2737 /* Already aborted the transaction if it failed. */
2739 btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2740 list_del_init(&block_group->bg_list);
2741 clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2744 * If the block group is still unused, add it to the list of
2745 * unused block groups. The block group may have been created in
2746 * order to satisfy a space reservation, in which case the
2747 * extent allocation only happens later. But often we don't
2748 * actually need to allocate space that we previously reserved,
2749 * so the block group may become unused for a long time. For
2750 * example for metadata we generally reserve space for a worst
2751 * possible scenario, but then don't end up allocating all that
2752 * space or none at all (due to no need to COW, extent buffers
2753 * were already COWed in the current transaction and still
2754 * unwritten, tree heights lower than the maximum possible
2755 * height, etc). For data we generally reserve the axact amount
2756 * of space we are going to allocate later, the exception is
2757 * when using compression, as we must reserve space based on the
2758 * uncompressed data size, because the compression is only done
2759 * when writeback triggered and we don't know how much space we
2760 * are actually going to need, so we reserve the uncompressed
2761 * size because the data may be uncompressible in the worst case.
2766 spin_lock(&block_group->lock);
2767 used = btrfs_is_block_group_used(block_group);
2768 spin_unlock(&block_group->lock);
2771 btrfs_mark_bg_unused(block_group);
2774 btrfs_trans_release_chunk_metadata(trans);
2778 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2779 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2781 static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2786 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2787 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2789 /* If we have a smaller fs index based on 128MiB. */
2790 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2793 offset = div64_u64(offset, div);
2794 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2798 struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2800 u64 chunk_offset, u64 size)
2802 struct btrfs_fs_info *fs_info = trans->fs_info;
2803 struct btrfs_block_group *cache;
2806 btrfs_set_log_full_commit(trans);
2808 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2810 return ERR_PTR(-ENOMEM);
2813 * Mark it as new before adding it to the rbtree of block groups or any
2814 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2815 * before the new flag is set.
2817 set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2819 cache->length = size;
2820 set_free_space_tree_thresholds(cache);
2821 cache->flags = type;
2822 cache->cached = BTRFS_CACHE_FINISHED;
2823 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2825 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2826 set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2828 ret = btrfs_load_block_group_zone_info(cache, true);
2830 btrfs_put_block_group(cache);
2831 return ERR_PTR(ret);
2834 ret = exclude_super_stripes(cache);
2836 /* We may have excluded something, so call this just in case */
2837 btrfs_free_excluded_extents(cache);
2838 btrfs_put_block_group(cache);
2839 return ERR_PTR(ret);
2842 ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2843 btrfs_free_excluded_extents(cache);
2845 btrfs_put_block_group(cache);
2846 return ERR_PTR(ret);
2850 * Ensure the corresponding space_info object is created and
2851 * assigned to our block group. We want our bg to be added to the rbtree
2852 * with its ->space_info set.
2854 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2855 ASSERT(cache->space_info);
2857 ret = btrfs_add_block_group_cache(fs_info, cache);
2859 btrfs_remove_free_space_cache(cache);
2860 btrfs_put_block_group(cache);
2861 return ERR_PTR(ret);
2865 * Now that our block group has its ->space_info set and is inserted in
2866 * the rbtree, update the space info's counters.
2868 trace_btrfs_add_block_group(fs_info, cache, 1);
2869 btrfs_add_bg_to_space_info(fs_info, cache);
2870 btrfs_update_global_block_rsv(fs_info);
2872 #ifdef CONFIG_BTRFS_DEBUG
2873 if (btrfs_should_fragment_free_space(cache)) {
2874 cache->space_info->bytes_used += size >> 1;
2875 fragment_free_space(cache);
2879 list_add_tail(&cache->bg_list, &trans->new_bgs);
2880 btrfs_inc_delayed_refs_rsv_bg_inserts(fs_info);
2882 set_avail_alloc_bits(fs_info, type);
2887 * Mark one block group RO, can be called several times for the same block
2890 * @cache: the destination block group
2891 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2892 * ensure we still have some free space after marking this
2895 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2896 bool do_chunk_alloc)
2898 struct btrfs_fs_info *fs_info = cache->fs_info;
2899 struct btrfs_trans_handle *trans;
2900 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2903 bool dirty_bg_running;
2906 * This can only happen when we are doing read-only scrub on read-only
2908 * In that case we should not start a new transaction on read-only fs.
2909 * Thus here we skip all chunk allocations.
2911 if (sb_rdonly(fs_info->sb)) {
2912 mutex_lock(&fs_info->ro_block_group_mutex);
2913 ret = inc_block_group_ro(cache, 0);
2914 mutex_unlock(&fs_info->ro_block_group_mutex);
2919 trans = btrfs_join_transaction(root);
2921 return PTR_ERR(trans);
2923 dirty_bg_running = false;
2926 * We're not allowed to set block groups readonly after the dirty
2927 * block group cache has started writing. If it already started,
2928 * back off and let this transaction commit.
2930 mutex_lock(&fs_info->ro_block_group_mutex);
2931 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2932 u64 transid = trans->transid;
2934 mutex_unlock(&fs_info->ro_block_group_mutex);
2935 btrfs_end_transaction(trans);
2937 ret = btrfs_wait_for_commit(fs_info, transid);
2940 dirty_bg_running = true;
2942 } while (dirty_bg_running);
2944 if (do_chunk_alloc) {
2946 * If we are changing raid levels, try to allocate a
2947 * corresponding block group with the new raid level.
2949 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2950 if (alloc_flags != cache->flags) {
2951 ret = btrfs_chunk_alloc(trans, alloc_flags,
2954 * ENOSPC is allowed here, we may have enough space
2955 * already allocated at the new raid level to carry on
2964 ret = inc_block_group_ro(cache, 0);
2967 if (ret == -ETXTBSY)
2971 * Skip chunk allocation if the bg is SYSTEM, this is to avoid system
2972 * chunk allocation storm to exhaust the system chunk array. Otherwise
2973 * we still want to try our best to mark the block group read-only.
2975 if (!do_chunk_alloc && ret == -ENOSPC &&
2976 (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
2979 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2980 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2984 * We have allocated a new chunk. We also need to activate that chunk to
2985 * grant metadata tickets for zoned filesystem.
2987 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2991 ret = inc_block_group_ro(cache, 0);
2992 if (ret == -ETXTBSY)
2995 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2996 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2997 mutex_lock(&fs_info->chunk_mutex);
2998 check_system_chunk(trans, alloc_flags);
2999 mutex_unlock(&fs_info->chunk_mutex);
3002 mutex_unlock(&fs_info->ro_block_group_mutex);
3004 btrfs_end_transaction(trans);
3008 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
3010 struct btrfs_space_info *sinfo = cache->space_info;
3015 spin_lock(&sinfo->lock);
3016 spin_lock(&cache->lock);
3018 if (btrfs_is_zoned(cache->fs_info)) {
3019 /* Migrate zone_unusable bytes back */
3020 cache->zone_unusable =
3021 (cache->alloc_offset - cache->used) +
3022 (cache->length - cache->zone_capacity);
3023 sinfo->bytes_zone_unusable += cache->zone_unusable;
3024 sinfo->bytes_readonly -= cache->zone_unusable;
3026 num_bytes = cache->length - cache->reserved -
3027 cache->pinned - cache->bytes_super -
3028 cache->zone_unusable - cache->used;
3029 sinfo->bytes_readonly -= num_bytes;
3030 list_del_init(&cache->ro_list);
3032 spin_unlock(&cache->lock);
3033 spin_unlock(&sinfo->lock);
3036 static int update_block_group_item(struct btrfs_trans_handle *trans,
3037 struct btrfs_path *path,
3038 struct btrfs_block_group *cache)
3040 struct btrfs_fs_info *fs_info = trans->fs_info;
3042 struct btrfs_root *root = btrfs_block_group_root(fs_info);
3044 struct extent_buffer *leaf;
3045 struct btrfs_block_group_item bgi;
3046 struct btrfs_key key;
3047 u64 old_commit_used;
3051 * Block group items update can be triggered out of commit transaction
3052 * critical section, thus we need a consistent view of used bytes.
3053 * We cannot use cache->used directly outside of the spin lock, as it
3056 spin_lock(&cache->lock);
3057 old_commit_used = cache->commit_used;
3059 /* No change in used bytes, can safely skip it. */
3060 if (cache->commit_used == used) {
3061 spin_unlock(&cache->lock);
3064 cache->commit_used = used;
3065 spin_unlock(&cache->lock);
3067 key.objectid = cache->start;
3068 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3069 key.offset = cache->length;
3071 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3078 leaf = path->nodes[0];
3079 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3080 btrfs_set_stack_block_group_used(&bgi, used);
3081 btrfs_set_stack_block_group_chunk_objectid(&bgi,
3082 cache->global_root_id);
3083 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3084 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3085 btrfs_mark_buffer_dirty(trans, leaf);
3087 btrfs_release_path(path);
3089 * We didn't update the block group item, need to revert commit_used
3090 * unless the block group item didn't exist yet - this is to prevent a
3091 * race with a concurrent insertion of the block group item, with
3092 * insert_block_group_item(), that happened just after we attempted to
3093 * update. In that case we would reset commit_used to 0 just after the
3094 * insertion set it to a value greater than 0 - if the block group later
3095 * becomes with 0 used bytes, we would incorrectly skip its update.
3097 if (ret < 0 && ret != -ENOENT) {
3098 spin_lock(&cache->lock);
3099 cache->commit_used = old_commit_used;
3100 spin_unlock(&cache->lock);
3106 static int cache_save_setup(struct btrfs_block_group *block_group,
3107 struct btrfs_trans_handle *trans,
3108 struct btrfs_path *path)
3110 struct btrfs_fs_info *fs_info = block_group->fs_info;
3111 struct inode *inode = NULL;
3112 struct extent_changeset *data_reserved = NULL;
3114 int dcs = BTRFS_DC_ERROR;
3119 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3123 * If this block group is smaller than 100 megs don't bother caching the
3126 if (block_group->length < (100 * SZ_1M)) {
3127 spin_lock(&block_group->lock);
3128 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3129 spin_unlock(&block_group->lock);
3133 if (TRANS_ABORTED(trans))
3136 inode = lookup_free_space_inode(block_group, path);
3137 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3138 ret = PTR_ERR(inode);
3139 btrfs_release_path(path);
3143 if (IS_ERR(inode)) {
3147 if (block_group->ro)
3150 ret = create_free_space_inode(trans, block_group, path);
3157 * We want to set the generation to 0, that way if anything goes wrong
3158 * from here on out we know not to trust this cache when we load up next
3161 BTRFS_I(inode)->generation = 0;
3162 ret = btrfs_update_inode(trans, BTRFS_I(inode));
3165 * So theoretically we could recover from this, simply set the
3166 * super cache generation to 0 so we know to invalidate the
3167 * cache, but then we'd have to keep track of the block groups
3168 * that fail this way so we know we _have_ to reset this cache
3169 * before the next commit or risk reading stale cache. So to
3170 * limit our exposure to horrible edge cases lets just abort the
3171 * transaction, this only happens in really bad situations
3174 btrfs_abort_transaction(trans, ret);
3179 /* We've already setup this transaction, go ahead and exit */
3180 if (block_group->cache_generation == trans->transid &&
3181 i_size_read(inode)) {
3182 dcs = BTRFS_DC_SETUP;
3186 if (i_size_read(inode) > 0) {
3187 ret = btrfs_check_trunc_cache_free_space(fs_info,
3188 &fs_info->global_block_rsv);
3192 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3197 spin_lock(&block_group->lock);
3198 if (block_group->cached != BTRFS_CACHE_FINISHED ||
3199 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3201 * don't bother trying to write stuff out _if_
3202 * a) we're not cached,
3203 * b) we're with nospace_cache mount option,
3204 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3206 dcs = BTRFS_DC_WRITTEN;
3207 spin_unlock(&block_group->lock);
3210 spin_unlock(&block_group->lock);
3213 * We hit an ENOSPC when setting up the cache in this transaction, just
3214 * skip doing the setup, we've already cleared the cache so we're safe.
3216 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3222 * Try to preallocate enough space based on how big the block group is.
3223 * Keep in mind this has to include any pinned space which could end up
3224 * taking up quite a bit since it's not folded into the other space
3227 cache_size = div_u64(block_group->length, SZ_256M);
3232 cache_size *= fs_info->sectorsize;
3234 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3239 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3240 cache_size, cache_size,
3243 * Our cache requires contiguous chunks so that we don't modify a bunch
3244 * of metadata or split extents when writing the cache out, which means
3245 * we can enospc if we are heavily fragmented in addition to just normal
3246 * out of space conditions. So if we hit this just skip setting up any
3247 * other block groups for this transaction, maybe we'll unpin enough
3248 * space the next time around.
3251 dcs = BTRFS_DC_SETUP;
3252 else if (ret == -ENOSPC)
3253 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3258 btrfs_release_path(path);
3260 spin_lock(&block_group->lock);
3261 if (!ret && dcs == BTRFS_DC_SETUP)
3262 block_group->cache_generation = trans->transid;
3263 block_group->disk_cache_state = dcs;
3264 spin_unlock(&block_group->lock);
3266 extent_changeset_free(data_reserved);
3270 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3272 struct btrfs_fs_info *fs_info = trans->fs_info;
3273 struct btrfs_block_group *cache, *tmp;
3274 struct btrfs_transaction *cur_trans = trans->transaction;
3275 struct btrfs_path *path;
3277 if (list_empty(&cur_trans->dirty_bgs) ||
3278 !btrfs_test_opt(fs_info, SPACE_CACHE))
3281 path = btrfs_alloc_path();
3285 /* Could add new block groups, use _safe just in case */
3286 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3288 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3289 cache_save_setup(cache, trans, path);
3292 btrfs_free_path(path);
3297 * Transaction commit does final block group cache writeback during a critical
3298 * section where nothing is allowed to change the FS. This is required in
3299 * order for the cache to actually match the block group, but can introduce a
3300 * lot of latency into the commit.
3302 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3303 * There's a chance we'll have to redo some of it if the block group changes
3304 * again during the commit, but it greatly reduces the commit latency by
3305 * getting rid of the easy block groups while we're still allowing others to
3308 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3310 struct btrfs_fs_info *fs_info = trans->fs_info;
3311 struct btrfs_block_group *cache;
3312 struct btrfs_transaction *cur_trans = trans->transaction;
3315 struct btrfs_path *path = NULL;
3317 struct list_head *io = &cur_trans->io_bgs;
3320 spin_lock(&cur_trans->dirty_bgs_lock);
3321 if (list_empty(&cur_trans->dirty_bgs)) {
3322 spin_unlock(&cur_trans->dirty_bgs_lock);
3325 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3326 spin_unlock(&cur_trans->dirty_bgs_lock);
3329 /* Make sure all the block groups on our dirty list actually exist */
3330 btrfs_create_pending_block_groups(trans);
3333 path = btrfs_alloc_path();
3341 * cache_write_mutex is here only to save us from balance or automatic
3342 * removal of empty block groups deleting this block group while we are
3343 * writing out the cache
3345 mutex_lock(&trans->transaction->cache_write_mutex);
3346 while (!list_empty(&dirty)) {
3347 bool drop_reserve = true;
3349 cache = list_first_entry(&dirty, struct btrfs_block_group,
3352 * This can happen if something re-dirties a block group that
3353 * is already under IO. Just wait for it to finish and then do
3356 if (!list_empty(&cache->io_list)) {
3357 list_del_init(&cache->io_list);
3358 btrfs_wait_cache_io(trans, cache, path);
3359 btrfs_put_block_group(cache);
3364 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3365 * it should update the cache_state. Don't delete until after
3368 * Since we're not running in the commit critical section
3369 * we need the dirty_bgs_lock to protect from update_block_group
3371 spin_lock(&cur_trans->dirty_bgs_lock);
3372 list_del_init(&cache->dirty_list);
3373 spin_unlock(&cur_trans->dirty_bgs_lock);
3377 cache_save_setup(cache, trans, path);
3379 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3380 cache->io_ctl.inode = NULL;
3381 ret = btrfs_write_out_cache(trans, cache, path);
3382 if (ret == 0 && cache->io_ctl.inode) {
3386 * The cache_write_mutex is protecting the
3387 * io_list, also refer to the definition of
3388 * btrfs_transaction::io_bgs for more details
3390 list_add_tail(&cache->io_list, io);
3393 * If we failed to write the cache, the
3394 * generation will be bad and life goes on
3400 ret = update_block_group_item(trans, path, cache);
3402 * Our block group might still be attached to the list
3403 * of new block groups in the transaction handle of some
3404 * other task (struct btrfs_trans_handle->new_bgs). This
3405 * means its block group item isn't yet in the extent
3406 * tree. If this happens ignore the error, as we will
3407 * try again later in the critical section of the
3408 * transaction commit.
3410 if (ret == -ENOENT) {
3412 spin_lock(&cur_trans->dirty_bgs_lock);
3413 if (list_empty(&cache->dirty_list)) {
3414 list_add_tail(&cache->dirty_list,
3415 &cur_trans->dirty_bgs);
3416 btrfs_get_block_group(cache);
3417 drop_reserve = false;
3419 spin_unlock(&cur_trans->dirty_bgs_lock);
3421 btrfs_abort_transaction(trans, ret);
3425 /* If it's not on the io list, we need to put the block group */
3427 btrfs_put_block_group(cache);
3429 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3431 * Avoid blocking other tasks for too long. It might even save
3432 * us from writing caches for block groups that are going to be
3435 mutex_unlock(&trans->transaction->cache_write_mutex);
3438 mutex_lock(&trans->transaction->cache_write_mutex);
3440 mutex_unlock(&trans->transaction->cache_write_mutex);
3443 * Go through delayed refs for all the stuff we've just kicked off
3444 * and then loop back (just once)
3447 ret = btrfs_run_delayed_refs(trans, 0);
3448 if (!ret && loops == 0) {
3450 spin_lock(&cur_trans->dirty_bgs_lock);
3451 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3453 * dirty_bgs_lock protects us from concurrent block group
3454 * deletes too (not just cache_write_mutex).
3456 if (!list_empty(&dirty)) {
3457 spin_unlock(&cur_trans->dirty_bgs_lock);
3460 spin_unlock(&cur_trans->dirty_bgs_lock);
3464 spin_lock(&cur_trans->dirty_bgs_lock);
3465 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3466 spin_unlock(&cur_trans->dirty_bgs_lock);
3467 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3470 btrfs_free_path(path);
3474 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3476 struct btrfs_fs_info *fs_info = trans->fs_info;
3477 struct btrfs_block_group *cache;
3478 struct btrfs_transaction *cur_trans = trans->transaction;
3481 struct btrfs_path *path;
3482 struct list_head *io = &cur_trans->io_bgs;
3484 path = btrfs_alloc_path();
3489 * Even though we are in the critical section of the transaction commit,
3490 * we can still have concurrent tasks adding elements to this
3491 * transaction's list of dirty block groups. These tasks correspond to
3492 * endio free space workers started when writeback finishes for a
3493 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3494 * allocate new block groups as a result of COWing nodes of the root
3495 * tree when updating the free space inode. The writeback for the space
3496 * caches is triggered by an earlier call to
3497 * btrfs_start_dirty_block_groups() and iterations of the following
3499 * Also we want to do the cache_save_setup first and then run the
3500 * delayed refs to make sure we have the best chance at doing this all
3503 spin_lock(&cur_trans->dirty_bgs_lock);
3504 while (!list_empty(&cur_trans->dirty_bgs)) {
3505 cache = list_first_entry(&cur_trans->dirty_bgs,
3506 struct btrfs_block_group,
3510 * This can happen if cache_save_setup re-dirties a block group
3511 * that is already under IO. Just wait for it to finish and
3512 * then do it all again
3514 if (!list_empty(&cache->io_list)) {
3515 spin_unlock(&cur_trans->dirty_bgs_lock);
3516 list_del_init(&cache->io_list);
3517 btrfs_wait_cache_io(trans, cache, path);
3518 btrfs_put_block_group(cache);
3519 spin_lock(&cur_trans->dirty_bgs_lock);
3523 * Don't remove from the dirty list until after we've waited on
3526 list_del_init(&cache->dirty_list);
3527 spin_unlock(&cur_trans->dirty_bgs_lock);
3530 cache_save_setup(cache, trans, path);
3533 ret = btrfs_run_delayed_refs(trans, U64_MAX);
3535 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3536 cache->io_ctl.inode = NULL;
3537 ret = btrfs_write_out_cache(trans, cache, path);
3538 if (ret == 0 && cache->io_ctl.inode) {
3540 list_add_tail(&cache->io_list, io);
3543 * If we failed to write the cache, the
3544 * generation will be bad and life goes on
3550 ret = update_block_group_item(trans, path, cache);
3552 * One of the free space endio workers might have
3553 * created a new block group while updating a free space
3554 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3555 * and hasn't released its transaction handle yet, in
3556 * which case the new block group is still attached to
3557 * its transaction handle and its creation has not
3558 * finished yet (no block group item in the extent tree
3559 * yet, etc). If this is the case, wait for all free
3560 * space endio workers to finish and retry. This is a
3561 * very rare case so no need for a more efficient and
3564 if (ret == -ENOENT) {
3565 wait_event(cur_trans->writer_wait,
3566 atomic_read(&cur_trans->num_writers) == 1);
3567 ret = update_block_group_item(trans, path, cache);
3570 btrfs_abort_transaction(trans, ret);
3573 /* If its not on the io list, we need to put the block group */
3575 btrfs_put_block_group(cache);
3576 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3577 spin_lock(&cur_trans->dirty_bgs_lock);
3579 spin_unlock(&cur_trans->dirty_bgs_lock);
3582 * Refer to the definition of io_bgs member for details why it's safe
3583 * to use it without any locking
3585 while (!list_empty(io)) {
3586 cache = list_first_entry(io, struct btrfs_block_group,
3588 list_del_init(&cache->io_list);
3589 btrfs_wait_cache_io(trans, cache, path);
3590 btrfs_put_block_group(cache);
3593 btrfs_free_path(path);
3597 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3598 u64 bytenr, u64 num_bytes, bool alloc)
3600 struct btrfs_fs_info *info = trans->fs_info;
3601 struct btrfs_space_info *space_info;
3602 struct btrfs_block_group *cache;
3604 bool reclaim = false;
3605 bool bg_already_dirty = true;
3608 /* Block accounting for super block */
3609 spin_lock(&info->delalloc_root_lock);
3610 old_val = btrfs_super_bytes_used(info->super_copy);
3612 old_val += num_bytes;
3614 old_val -= num_bytes;
3615 btrfs_set_super_bytes_used(info->super_copy, old_val);
3616 spin_unlock(&info->delalloc_root_lock);
3618 cache = btrfs_lookup_block_group(info, bytenr);
3622 /* An extent can not span multiple block groups. */
3623 ASSERT(bytenr + num_bytes <= cache->start + cache->length);
3625 space_info = cache->space_info;
3626 factor = btrfs_bg_type_to_factor(cache->flags);
3629 * If this block group has free space cache written out, we need to make
3630 * sure to load it if we are removing space. This is because we need
3631 * the unpinning stage to actually add the space back to the block group,
3632 * otherwise we will leak space.
3634 if (!alloc && !btrfs_block_group_done(cache))
3635 btrfs_cache_block_group(cache, true);
3637 spin_lock(&space_info->lock);
3638 spin_lock(&cache->lock);
3640 if (btrfs_test_opt(info, SPACE_CACHE) &&
3641 cache->disk_cache_state < BTRFS_DC_CLEAR)
3642 cache->disk_cache_state = BTRFS_DC_CLEAR;
3644 old_val = cache->used;
3646 old_val += num_bytes;
3647 cache->used = old_val;
3648 cache->reserved -= num_bytes;
3649 space_info->bytes_reserved -= num_bytes;
3650 space_info->bytes_used += num_bytes;
3651 space_info->disk_used += num_bytes * factor;
3652 spin_unlock(&cache->lock);
3653 spin_unlock(&space_info->lock);
3655 old_val -= num_bytes;
3656 cache->used = old_val;
3657 cache->pinned += num_bytes;
3658 btrfs_space_info_update_bytes_pinned(info, space_info, num_bytes);
3659 space_info->bytes_used -= num_bytes;
3660 space_info->disk_used -= num_bytes * factor;
3662 reclaim = should_reclaim_block_group(cache, num_bytes);
3664 spin_unlock(&cache->lock);
3665 spin_unlock(&space_info->lock);
3667 set_extent_bit(&trans->transaction->pinned_extents, bytenr,
3668 bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
3671 spin_lock(&trans->transaction->dirty_bgs_lock);
3672 if (list_empty(&cache->dirty_list)) {
3673 list_add_tail(&cache->dirty_list, &trans->transaction->dirty_bgs);
3674 bg_already_dirty = false;
3675 btrfs_get_block_group(cache);
3677 spin_unlock(&trans->transaction->dirty_bgs_lock);
3680 * No longer have used bytes in this block group, queue it for deletion.
3681 * We do this after adding the block group to the dirty list to avoid
3682 * races between cleaner kthread and space cache writeout.
3684 if (!alloc && old_val == 0) {
3685 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3686 btrfs_mark_bg_unused(cache);
3687 } else if (!alloc && reclaim) {
3688 btrfs_mark_bg_to_reclaim(cache);
3691 btrfs_put_block_group(cache);
3693 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3694 if (!bg_already_dirty)
3695 btrfs_inc_delayed_refs_rsv_bg_updates(info);
3701 * Update the block_group and space info counters.
3703 * @cache: The cache we are manipulating
3704 * @ram_bytes: The number of bytes of file content, and will be same to
3705 * @num_bytes except for the compress path.
3706 * @num_bytes: The number of bytes in question
3707 * @delalloc: The blocks are allocated for the delalloc write
3709 * This is called by the allocator when it reserves space. If this is a
3710 * reservation and the block group has become read only we cannot make the
3711 * reservation and return -EAGAIN, otherwise this function always succeeds.
3713 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3714 u64 ram_bytes, u64 num_bytes, int delalloc,
3715 bool force_wrong_size_class)
3717 struct btrfs_space_info *space_info = cache->space_info;
3718 enum btrfs_block_group_size_class size_class;
3721 spin_lock(&space_info->lock);
3722 spin_lock(&cache->lock);
3728 if (btrfs_block_group_should_use_size_class(cache)) {
3729 size_class = btrfs_calc_block_group_size_class(num_bytes);
3730 ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3734 cache->reserved += num_bytes;
3735 space_info->bytes_reserved += num_bytes;
3736 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3737 space_info->flags, num_bytes, 1);
3738 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3739 space_info, -ram_bytes);
3741 cache->delalloc_bytes += num_bytes;
3744 * Compression can use less space than we reserved, so wake tickets if
3747 if (num_bytes < ram_bytes)
3748 btrfs_try_granting_tickets(cache->fs_info, space_info);
3750 spin_unlock(&cache->lock);
3751 spin_unlock(&space_info->lock);
3756 * Update the block_group and space info counters.
3758 * @cache: The cache we are manipulating
3759 * @num_bytes: The number of bytes in question
3760 * @delalloc: The blocks are allocated for the delalloc write
3762 * This is called by somebody who is freeing space that was never actually used
3763 * on disk. For example if you reserve some space for a new leaf in transaction
3764 * A and before transaction A commits you free that leaf, you call this with
3765 * reserve set to 0 in order to clear the reservation.
3767 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3768 u64 num_bytes, int delalloc)
3770 struct btrfs_space_info *space_info = cache->space_info;
3772 spin_lock(&space_info->lock);
3773 spin_lock(&cache->lock);
3775 space_info->bytes_readonly += num_bytes;
3776 cache->reserved -= num_bytes;
3777 space_info->bytes_reserved -= num_bytes;
3778 space_info->max_extent_size = 0;
3781 cache->delalloc_bytes -= num_bytes;
3782 spin_unlock(&cache->lock);
3784 btrfs_try_granting_tickets(cache->fs_info, space_info);
3785 spin_unlock(&space_info->lock);
3788 static void force_metadata_allocation(struct btrfs_fs_info *info)
3790 struct list_head *head = &info->space_info;
3791 struct btrfs_space_info *found;
3793 list_for_each_entry(found, head, list) {
3794 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3795 found->force_alloc = CHUNK_ALLOC_FORCE;
3799 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3800 struct btrfs_space_info *sinfo, int force)
3802 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3805 if (force == CHUNK_ALLOC_FORCE)
3809 * in limited mode, we want to have some free space up to
3810 * about 1% of the FS size.
3812 if (force == CHUNK_ALLOC_LIMITED) {
3813 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3814 thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3816 if (sinfo->total_bytes - bytes_used < thresh)
3820 if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3825 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3827 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3829 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3832 static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3834 struct btrfs_block_group *bg;
3838 * Check if we have enough space in the system space info because we
3839 * will need to update device items in the chunk btree and insert a new
3840 * chunk item in the chunk btree as well. This will allocate a new
3841 * system block group if needed.
3843 check_system_chunk(trans, flags);
3845 bg = btrfs_create_chunk(trans, flags);
3851 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3853 * Normally we are not expected to fail with -ENOSPC here, since we have
3854 * previously reserved space in the system space_info and allocated one
3855 * new system chunk if necessary. However there are three exceptions:
3857 * 1) We may have enough free space in the system space_info but all the
3858 * existing system block groups have a profile which can not be used
3859 * for extent allocation.
3861 * This happens when mounting in degraded mode. For example we have a
3862 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3863 * using the other device in degraded mode. If we then allocate a chunk,
3864 * we may have enough free space in the existing system space_info, but
3865 * none of the block groups can be used for extent allocation since they
3866 * have a RAID1 profile, and because we are in degraded mode with a
3867 * single device, we are forced to allocate a new system chunk with a
3868 * SINGLE profile. Making check_system_chunk() iterate over all system
3869 * block groups and check if they have a usable profile and enough space
3870 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3871 * try again after forcing allocation of a new system chunk. Like this
3872 * we avoid paying the cost of that search in normal circumstances, when
3873 * we were not mounted in degraded mode;
3875 * 2) We had enough free space info the system space_info, and one suitable
3876 * block group to allocate from when we called check_system_chunk()
3877 * above. However right after we called it, the only system block group
3878 * with enough free space got turned into RO mode by a running scrub,
3879 * and in this case we have to allocate a new one and retry. We only
3880 * need do this allocate and retry once, since we have a transaction
3881 * handle and scrub uses the commit root to search for block groups;
3883 * 3) We had one system block group with enough free space when we called
3884 * check_system_chunk(), but after that, right before we tried to
3885 * allocate the last extent buffer we needed, a discard operation came
3886 * in and it temporarily removed the last free space entry from the
3887 * block group (discard removes a free space entry, discards it, and
3888 * then adds back the entry to the block group cache).
3890 if (ret == -ENOSPC) {
3891 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3892 struct btrfs_block_group *sys_bg;
3894 sys_bg = btrfs_create_chunk(trans, sys_flags);
3895 if (IS_ERR(sys_bg)) {
3896 ret = PTR_ERR(sys_bg);
3897 btrfs_abort_transaction(trans, ret);
3901 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3903 btrfs_abort_transaction(trans, ret);
3907 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3909 btrfs_abort_transaction(trans, ret);
3913 btrfs_abort_transaction(trans, ret);
3917 btrfs_trans_release_chunk_metadata(trans);
3920 return ERR_PTR(ret);
3922 btrfs_get_block_group(bg);
3927 * Chunk allocation is done in 2 phases:
3929 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3930 * the chunk, the chunk mapping, create its block group and add the items
3931 * that belong in the chunk btree to it - more specifically, we need to
3932 * update device items in the chunk btree and add a new chunk item to it.
3934 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3935 * group item to the extent btree and the device extent items to the devices
3938 * This is done to prevent deadlocks. For example when COWing a node from the
3939 * extent btree we are holding a write lock on the node's parent and if we
3940 * trigger chunk allocation and attempted to insert the new block group item
3941 * in the extent btree right way, we could deadlock because the path for the
3942 * insertion can include that parent node. At first glance it seems impossible
3943 * to trigger chunk allocation after starting a transaction since tasks should
3944 * reserve enough transaction units (metadata space), however while that is true
3945 * most of the time, chunk allocation may still be triggered for several reasons:
3947 * 1) When reserving metadata, we check if there is enough free space in the
3948 * metadata space_info and therefore don't trigger allocation of a new chunk.
3949 * However later when the task actually tries to COW an extent buffer from
3950 * the extent btree or from the device btree for example, it is forced to
3951 * allocate a new block group (chunk) because the only one that had enough
3952 * free space was just turned to RO mode by a running scrub for example (or
3953 * device replace, block group reclaim thread, etc), so we can not use it
3954 * for allocating an extent and end up being forced to allocate a new one;
3956 * 2) Because we only check that the metadata space_info has enough free bytes,
3957 * we end up not allocating a new metadata chunk in that case. However if
3958 * the filesystem was mounted in degraded mode, none of the existing block
3959 * groups might be suitable for extent allocation due to their incompatible
3960 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3961 * use a RAID1 profile, in degraded mode using a single device). In this case
3962 * when the task attempts to COW some extent buffer of the extent btree for
3963 * example, it will trigger allocation of a new metadata block group with a
3964 * suitable profile (SINGLE profile in the example of the degraded mount of
3965 * the RAID1 filesystem);
3967 * 3) The task has reserved enough transaction units / metadata space, but when
3968 * it attempts to COW an extent buffer from the extent or device btree for
3969 * example, it does not find any free extent in any metadata block group,
3970 * therefore forced to try to allocate a new metadata block group.
3971 * This is because some other task allocated all available extents in the
3972 * meanwhile - this typically happens with tasks that don't reserve space
3973 * properly, either intentionally or as a bug. One example where this is
3974 * done intentionally is fsync, as it does not reserve any transaction units
3975 * and ends up allocating a variable number of metadata extents for log
3976 * tree extent buffers;
3978 * 4) The task has reserved enough transaction units / metadata space, but right
3979 * before it tries to allocate the last extent buffer it needs, a discard
3980 * operation comes in and, temporarily, removes the last free space entry from
3981 * the only metadata block group that had free space (discard starts by
3982 * removing a free space entry from a block group, then does the discard
3983 * operation and, once it's done, it adds back the free space entry to the
3986 * We also need this 2 phases setup when adding a device to a filesystem with
3987 * a seed device - we must create new metadata and system chunks without adding
3988 * any of the block group items to the chunk, extent and device btrees. If we
3989 * did not do it this way, we would get ENOSPC when attempting to update those
3990 * btrees, since all the chunks from the seed device are read-only.
3992 * Phase 1 does the updates and insertions to the chunk btree because if we had
3993 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3994 * parallel, we risk having too many system chunks allocated by many tasks if
3995 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3996 * extreme case this leads to exhaustion of the system chunk array in the
3997 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3998 * and with RAID filesystems (so we have more device items in the chunk btree).
3999 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
4000 * the system chunk array due to concurrent allocations") provides more details.
4002 * Allocation of system chunks does not happen through this function. A task that
4003 * needs to update the chunk btree (the only btree that uses system chunks), must
4004 * preallocate chunk space by calling either check_system_chunk() or
4005 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
4006 * metadata chunk or when removing a chunk, while the later is used before doing
4007 * a modification to the chunk btree - use cases for the later are adding,
4008 * removing and resizing a device as well as relocation of a system chunk.
4009 * See the comment below for more details.
4011 * The reservation of system space, done through check_system_chunk(), as well
4012 * as all the updates and insertions into the chunk btree must be done while
4013 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4014 * an extent buffer from the chunks btree we never trigger allocation of a new
4015 * system chunk, which would result in a deadlock (trying to lock twice an
4016 * extent buffer of the chunk btree, first time before triggering the chunk
4017 * allocation and the second time during chunk allocation while attempting to
4018 * update the chunks btree). The system chunk array is also updated while holding
4019 * that mutex. The same logic applies to removing chunks - we must reserve system
4020 * space, update the chunk btree and the system chunk array in the superblock
4021 * while holding fs_info->chunk_mutex.
4023 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4025 * If @force is CHUNK_ALLOC_FORCE:
4026 * - return 1 if it successfully allocates a chunk,
4027 * - return errors including -ENOSPC otherwise.
4028 * If @force is NOT CHUNK_ALLOC_FORCE:
4029 * - return 0 if it doesn't need to allocate a new chunk,
4030 * - return 1 if it successfully allocates a chunk,
4031 * - return errors including -ENOSPC otherwise.
4033 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4034 enum btrfs_chunk_alloc_enum force)
4036 struct btrfs_fs_info *fs_info = trans->fs_info;
4037 struct btrfs_space_info *space_info;
4038 struct btrfs_block_group *ret_bg;
4039 bool wait_for_alloc = false;
4040 bool should_alloc = false;
4041 bool from_extent_allocation = false;
4044 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4045 from_extent_allocation = true;
4046 force = CHUNK_ALLOC_FORCE;
4049 /* Don't re-enter if we're already allocating a chunk */
4050 if (trans->allocating_chunk)
4053 * Allocation of system chunks can not happen through this path, as we
4054 * could end up in a deadlock if we are allocating a data or metadata
4055 * chunk and there is another task modifying the chunk btree.
4057 * This is because while we are holding the chunk mutex, we will attempt
4058 * to add the new chunk item to the chunk btree or update an existing
4059 * device item in the chunk btree, while the other task that is modifying
4060 * the chunk btree is attempting to COW an extent buffer while holding a
4061 * lock on it and on its parent - if the COW operation triggers a system
4062 * chunk allocation, then we can deadlock because we are holding the
4063 * chunk mutex and we may need to access that extent buffer or its parent
4064 * in order to add the chunk item or update a device item.
4066 * Tasks that want to modify the chunk tree should reserve system space
4067 * before updating the chunk btree, by calling either
4068 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4069 * It's possible that after a task reserves the space, it still ends up
4070 * here - this happens in the cases described above at do_chunk_alloc().
4071 * The task will have to either retry or fail.
4073 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4076 space_info = btrfs_find_space_info(fs_info, flags);
4080 spin_lock(&space_info->lock);
4081 if (force < space_info->force_alloc)
4082 force = space_info->force_alloc;
4083 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4084 if (space_info->full) {
4085 /* No more free physical space */
4090 spin_unlock(&space_info->lock);
4092 } else if (!should_alloc) {
4093 spin_unlock(&space_info->lock);
4095 } else if (space_info->chunk_alloc) {
4097 * Someone is already allocating, so we need to block
4098 * until this someone is finished and then loop to
4099 * recheck if we should continue with our allocation
4102 wait_for_alloc = true;
4103 force = CHUNK_ALLOC_NO_FORCE;
4104 spin_unlock(&space_info->lock);
4105 mutex_lock(&fs_info->chunk_mutex);
4106 mutex_unlock(&fs_info->chunk_mutex);
4108 /* Proceed with allocation */
4109 space_info->chunk_alloc = 1;
4110 wait_for_alloc = false;
4111 spin_unlock(&space_info->lock);
4115 } while (wait_for_alloc);
4117 mutex_lock(&fs_info->chunk_mutex);
4118 trans->allocating_chunk = true;
4121 * If we have mixed data/metadata chunks we want to make sure we keep
4122 * allocating mixed chunks instead of individual chunks.
4124 if (btrfs_mixed_space_info(space_info))
4125 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4128 * if we're doing a data chunk, go ahead and make sure that
4129 * we keep a reasonable number of metadata chunks allocated in the
4132 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4133 fs_info->data_chunk_allocations++;
4134 if (!(fs_info->data_chunk_allocations %
4135 fs_info->metadata_ratio))
4136 force_metadata_allocation(fs_info);
4139 ret_bg = do_chunk_alloc(trans, flags);
4140 trans->allocating_chunk = false;
4142 if (IS_ERR(ret_bg)) {
4143 ret = PTR_ERR(ret_bg);
4144 } else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4146 * New block group is likely to be used soon. Try to activate
4147 * it now. Failure is OK for now.
4149 btrfs_zone_activate(ret_bg);
4153 btrfs_put_block_group(ret_bg);
4155 spin_lock(&space_info->lock);
4158 space_info->full = 1;
4163 space_info->max_extent_size = 0;
4166 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4168 space_info->chunk_alloc = 0;
4169 spin_unlock(&space_info->lock);
4170 mutex_unlock(&fs_info->chunk_mutex);
4175 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4179 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4181 num_dev = fs_info->fs_devices->rw_devices;
4186 static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4190 struct btrfs_fs_info *fs_info = trans->fs_info;
4191 struct btrfs_space_info *info;
4196 * Needed because we can end up allocating a system chunk and for an
4197 * atomic and race free space reservation in the chunk block reserve.
4199 lockdep_assert_held(&fs_info->chunk_mutex);
4201 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4202 spin_lock(&info->lock);
4203 left = info->total_bytes - btrfs_space_info_used(info, true);
4204 spin_unlock(&info->lock);
4206 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4207 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4209 btrfs_dump_space_info(fs_info, info, 0, 0);
4213 u64 flags = btrfs_system_alloc_profile(fs_info);
4214 struct btrfs_block_group *bg;
4217 * Ignore failure to create system chunk. We might end up not
4218 * needing it, as we might not need to COW all nodes/leafs from
4219 * the paths we visit in the chunk tree (they were already COWed
4220 * or created in the current transaction for example).
4222 bg = btrfs_create_chunk(trans, flags);
4227 * We have a new chunk. We also need to activate it for
4230 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
4235 * If we fail to add the chunk item here, we end up
4236 * trying again at phase 2 of chunk allocation, at
4237 * btrfs_create_pending_block_groups(). So ignore
4238 * any error here. An ENOSPC here could happen, due to
4239 * the cases described at do_chunk_alloc() - the system
4240 * block group we just created was just turned into RO
4241 * mode by a scrub for example, or a running discard
4242 * temporarily removed its free space entries, etc.
4244 btrfs_chunk_alloc_add_chunk_item(trans, bg);
4249 ret = btrfs_block_rsv_add(fs_info,
4250 &fs_info->chunk_block_rsv,
4251 bytes, BTRFS_RESERVE_NO_FLUSH);
4253 trans->chunk_bytes_reserved += bytes;
4258 * Reserve space in the system space for allocating or removing a chunk.
4259 * The caller must be holding fs_info->chunk_mutex.
4261 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4263 struct btrfs_fs_info *fs_info = trans->fs_info;
4264 const u64 num_devs = get_profile_num_devs(fs_info, type);
4267 /* num_devs device items to update and 1 chunk item to add or remove. */
4268 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4269 btrfs_calc_insert_metadata_size(fs_info, 1);
4271 reserve_chunk_space(trans, bytes, type);
4275 * Reserve space in the system space, if needed, for doing a modification to the
4278 * @trans: A transaction handle.
4279 * @is_item_insertion: Indicate if the modification is for inserting a new item
4280 * in the chunk btree or if it's for the deletion or update
4281 * of an existing item.
4283 * This is used in a context where we need to update the chunk btree outside
4284 * block group allocation and removal, to avoid a deadlock with a concurrent
4285 * task that is allocating a metadata or data block group and therefore needs to
4286 * update the chunk btree while holding the chunk mutex. After the update to the
4287 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4290 void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4291 bool is_item_insertion)
4293 struct btrfs_fs_info *fs_info = trans->fs_info;
4296 if (is_item_insertion)
4297 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4299 bytes = btrfs_calc_metadata_size(fs_info, 1);
4301 mutex_lock(&fs_info->chunk_mutex);
4302 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4303 mutex_unlock(&fs_info->chunk_mutex);
4306 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4308 struct btrfs_block_group *block_group;
4310 block_group = btrfs_lookup_first_block_group(info, 0);
4311 while (block_group) {
4312 btrfs_wait_block_group_cache_done(block_group);
4313 spin_lock(&block_group->lock);
4314 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4315 &block_group->runtime_flags)) {
4316 struct inode *inode = block_group->inode;
4318 block_group->inode = NULL;
4319 spin_unlock(&block_group->lock);
4321 ASSERT(block_group->io_ctl.inode == NULL);
4324 spin_unlock(&block_group->lock);
4326 block_group = btrfs_next_block_group(block_group);
4331 * Must be called only after stopping all workers, since we could have block
4332 * group caching kthreads running, and therefore they could race with us if we
4333 * freed the block groups before stopping them.
4335 int btrfs_free_block_groups(struct btrfs_fs_info *info)
4337 struct btrfs_block_group *block_group;
4338 struct btrfs_space_info *space_info;
4339 struct btrfs_caching_control *caching_ctl;
4342 if (btrfs_is_zoned(info)) {
4343 if (info->active_meta_bg) {
4344 btrfs_put_block_group(info->active_meta_bg);
4345 info->active_meta_bg = NULL;
4347 if (info->active_system_bg) {
4348 btrfs_put_block_group(info->active_system_bg);
4349 info->active_system_bg = NULL;
4353 write_lock(&info->block_group_cache_lock);
4354 while (!list_empty(&info->caching_block_groups)) {
4355 caching_ctl = list_entry(info->caching_block_groups.next,
4356 struct btrfs_caching_control, list);
4357 list_del(&caching_ctl->list);
4358 btrfs_put_caching_control(caching_ctl);
4360 write_unlock(&info->block_group_cache_lock);
4362 spin_lock(&info->unused_bgs_lock);
4363 while (!list_empty(&info->unused_bgs)) {
4364 block_group = list_first_entry(&info->unused_bgs,
4365 struct btrfs_block_group,
4367 list_del_init(&block_group->bg_list);
4368 btrfs_put_block_group(block_group);
4371 while (!list_empty(&info->reclaim_bgs)) {
4372 block_group = list_first_entry(&info->reclaim_bgs,
4373 struct btrfs_block_group,
4375 list_del_init(&block_group->bg_list);
4376 btrfs_put_block_group(block_group);
4378 spin_unlock(&info->unused_bgs_lock);
4380 spin_lock(&info->zone_active_bgs_lock);
4381 while (!list_empty(&info->zone_active_bgs)) {
4382 block_group = list_first_entry(&info->zone_active_bgs,
4383 struct btrfs_block_group,
4385 list_del_init(&block_group->active_bg_list);
4386 btrfs_put_block_group(block_group);
4388 spin_unlock(&info->zone_active_bgs_lock);
4390 write_lock(&info->block_group_cache_lock);
4391 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4392 block_group = rb_entry(n, struct btrfs_block_group,
4394 rb_erase_cached(&block_group->cache_node,
4395 &info->block_group_cache_tree);
4396 RB_CLEAR_NODE(&block_group->cache_node);
4397 write_unlock(&info->block_group_cache_lock);
4399 down_write(&block_group->space_info->groups_sem);
4400 list_del(&block_group->list);
4401 up_write(&block_group->space_info->groups_sem);
4404 * We haven't cached this block group, which means we could
4405 * possibly have excluded extents on this block group.
4407 if (block_group->cached == BTRFS_CACHE_NO ||
4408 block_group->cached == BTRFS_CACHE_ERROR)
4409 btrfs_free_excluded_extents(block_group);
4411 btrfs_remove_free_space_cache(block_group);
4412 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4413 ASSERT(list_empty(&block_group->dirty_list));
4414 ASSERT(list_empty(&block_group->io_list));
4415 ASSERT(list_empty(&block_group->bg_list));
4416 ASSERT(refcount_read(&block_group->refs) == 1);
4417 ASSERT(block_group->swap_extents == 0);
4418 btrfs_put_block_group(block_group);
4420 write_lock(&info->block_group_cache_lock);
4422 write_unlock(&info->block_group_cache_lock);
4424 btrfs_release_global_block_rsv(info);
4426 while (!list_empty(&info->space_info)) {
4427 space_info = list_entry(info->space_info.next,
4428 struct btrfs_space_info,
4432 * Do not hide this behind enospc_debug, this is actually
4433 * important and indicates a real bug if this happens.
4435 if (WARN_ON(space_info->bytes_pinned > 0 ||
4436 space_info->bytes_may_use > 0))
4437 btrfs_dump_space_info(info, space_info, 0, 0);
4440 * If there was a failure to cleanup a log tree, very likely due
4441 * to an IO failure on a writeback attempt of one or more of its
4442 * extent buffers, we could not do proper (and cheap) unaccounting
4443 * of their reserved space, so don't warn on bytes_reserved > 0 in
4446 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4447 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4448 if (WARN_ON(space_info->bytes_reserved > 0))
4449 btrfs_dump_space_info(info, space_info, 0, 0);
4452 WARN_ON(space_info->reclaim_size > 0);
4453 list_del(&space_info->list);
4454 btrfs_sysfs_remove_space_info(space_info);
4459 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4461 atomic_inc(&cache->frozen);
4464 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4466 struct btrfs_fs_info *fs_info = block_group->fs_info;
4469 spin_lock(&block_group->lock);
4470 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4471 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4472 spin_unlock(&block_group->lock);
4475 struct btrfs_chunk_map *map;
4477 map = btrfs_find_chunk_map(fs_info, block_group->start, 1);
4478 /* Logic error, can't happen. */
4481 btrfs_remove_chunk_map(fs_info, map);
4483 /* Once for our lookup reference. */
4484 btrfs_free_chunk_map(map);
4487 * We may have left one free space entry and other possible
4488 * tasks trimming this block group have left 1 entry each one.
4491 btrfs_remove_free_space_cache(block_group);
4495 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4499 spin_lock(&bg->lock);
4504 spin_unlock(&bg->lock);
4509 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4511 spin_lock(&bg->lock);
4513 ASSERT(bg->swap_extents >= amount);
4514 bg->swap_extents -= amount;
4515 spin_unlock(&bg->lock);
4518 enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4520 if (size <= SZ_128K)
4521 return BTRFS_BG_SZ_SMALL;
4523 return BTRFS_BG_SZ_MEDIUM;
4524 return BTRFS_BG_SZ_LARGE;
4528 * Handle a block group allocating an extent in a size class
4530 * @bg: The block group we allocated in.
4531 * @size_class: The size class of the allocation.
4532 * @force_wrong_size_class: Whether we are desperate enough to allow
4533 * mismatched size classes.
4535 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4536 * case of a race that leads to the wrong size class without
4537 * force_wrong_size_class set.
4539 * find_free_extent will skip block groups with a mismatched size class until
4540 * it really needs to avoid ENOSPC. In that case it will set
4541 * force_wrong_size_class. However, if a block group is newly allocated and
4542 * doesn't yet have a size class, then it is possible for two allocations of
4543 * different sizes to race and both try to use it. The loser is caught here and
4546 int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4547 enum btrfs_block_group_size_class size_class,
4548 bool force_wrong_size_class)
4550 ASSERT(size_class != BTRFS_BG_SZ_NONE);
4552 /* The new allocation is in the right size class, do nothing */
4553 if (bg->size_class == size_class)
4556 * The new allocation is in a mismatched size class.
4557 * This means one of two things:
4559 * 1. Two tasks in find_free_extent for different size_classes raced
4560 * and hit the same empty block_group. Make the loser try again.
4561 * 2. A call to find_free_extent got desperate enough to set
4562 * 'force_wrong_slab'. Don't change the size_class, but allow the
4565 if (bg->size_class != BTRFS_BG_SZ_NONE) {
4566 if (force_wrong_size_class)
4571 * The happy new block group case: the new allocation is the first
4572 * one in the block_group so we set size_class.
4574 bg->size_class = size_class;
4579 bool btrfs_block_group_should_use_size_class(struct btrfs_block_group *bg)
4581 if (btrfs_is_zoned(bg->fs_info))
4583 if (!btrfs_is_block_group_data_only(bg))