1 // SPDX-License-Identifier: GPL-2.0
5 #include "block-group.h"
6 #include "space-info.h"
8 #include "free-space-cache.h"
9 #include "free-space-tree.h"
12 #include "transaction.h"
13 #include "ref-verify.h"
16 #include "delalloc-space.h"
19 * Return target flags in extended format or 0 if restripe for this chunk_type
22 * Should be called with balance_lock held
24 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
26 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
32 if (flags & BTRFS_BLOCK_GROUP_DATA &&
33 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
34 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
35 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
36 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
37 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
38 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
39 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
40 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
47 * @flags: available profiles in extended format (see ctree.h)
49 * Return reduced profile in chunk format. If profile changing is in progress
50 * (either running or paused) picks the target profile (if it's already
51 * available), otherwise falls back to plain reducing.
53 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
55 u64 num_devices = fs_info->fs_devices->rw_devices;
61 * See if restripe for this chunk_type is in progress, if so try to
62 * reduce to the target profile
64 spin_lock(&fs_info->balance_lock);
65 target = get_restripe_target(fs_info, flags);
67 /* Pick target profile only if it's already available */
68 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
69 spin_unlock(&fs_info->balance_lock);
70 return extended_to_chunk(target);
73 spin_unlock(&fs_info->balance_lock);
75 /* First, mask out the RAID levels which aren't possible */
76 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
77 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
78 allowed |= btrfs_raid_array[raid_type].bg_flag;
82 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
83 allowed = BTRFS_BLOCK_GROUP_RAID6;
84 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
85 allowed = BTRFS_BLOCK_GROUP_RAID5;
86 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
87 allowed = BTRFS_BLOCK_GROUP_RAID10;
88 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
89 allowed = BTRFS_BLOCK_GROUP_RAID1;
90 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
91 allowed = BTRFS_BLOCK_GROUP_RAID0;
93 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
95 return extended_to_chunk(flags | allowed);
98 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
105 seq = read_seqbegin(&fs_info->profiles_lock);
107 if (flags & BTRFS_BLOCK_GROUP_DATA)
108 flags |= fs_info->avail_data_alloc_bits;
109 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
110 flags |= fs_info->avail_system_alloc_bits;
111 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
112 flags |= fs_info->avail_metadata_alloc_bits;
113 } while (read_seqretry(&fs_info->profiles_lock, seq));
115 return btrfs_reduce_alloc_profile(fs_info, flags);
118 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
120 return get_alloc_profile(fs_info, orig_flags);
123 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
125 atomic_inc(&cache->count);
128 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
130 if (atomic_dec_and_test(&cache->count)) {
131 WARN_ON(cache->pinned > 0);
132 WARN_ON(cache->reserved > 0);
135 * If not empty, someone is still holding mutex of
136 * full_stripe_lock, which can only be released by caller.
137 * And it will definitely cause use-after-free when caller
138 * tries to release full stripe lock.
140 * No better way to resolve, but only to warn.
142 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
143 kfree(cache->free_space_ctl);
149 * This adds the block group to the fs_info rb tree for the block group cache
151 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
152 struct btrfs_block_group_cache *block_group)
155 struct rb_node *parent = NULL;
156 struct btrfs_block_group_cache *cache;
158 spin_lock(&info->block_group_cache_lock);
159 p = &info->block_group_cache_tree.rb_node;
163 cache = rb_entry(parent, struct btrfs_block_group_cache,
165 if (block_group->key.objectid < cache->key.objectid) {
167 } else if (block_group->key.objectid > cache->key.objectid) {
170 spin_unlock(&info->block_group_cache_lock);
175 rb_link_node(&block_group->cache_node, parent, p);
176 rb_insert_color(&block_group->cache_node,
177 &info->block_group_cache_tree);
179 if (info->first_logical_byte > block_group->key.objectid)
180 info->first_logical_byte = block_group->key.objectid;
182 spin_unlock(&info->block_group_cache_lock);
188 * This will return the block group at or after bytenr if contains is 0, else
189 * it will return the block group that contains the bytenr
191 static struct btrfs_block_group_cache *block_group_cache_tree_search(
192 struct btrfs_fs_info *info, u64 bytenr, int contains)
194 struct btrfs_block_group_cache *cache, *ret = NULL;
198 spin_lock(&info->block_group_cache_lock);
199 n = info->block_group_cache_tree.rb_node;
202 cache = rb_entry(n, struct btrfs_block_group_cache,
204 end = cache->key.objectid + cache->key.offset - 1;
205 start = cache->key.objectid;
207 if (bytenr < start) {
208 if (!contains && (!ret || start < ret->key.objectid))
211 } else if (bytenr > start) {
212 if (contains && bytenr <= end) {
223 btrfs_get_block_group(ret);
224 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
225 info->first_logical_byte = ret->key.objectid;
227 spin_unlock(&info->block_group_cache_lock);
233 * Return the block group that starts at or after bytenr
235 struct btrfs_block_group_cache *btrfs_lookup_first_block_group(
236 struct btrfs_fs_info *info, u64 bytenr)
238 return block_group_cache_tree_search(info, bytenr, 0);
242 * Return the block group that contains the given bytenr
244 struct btrfs_block_group_cache *btrfs_lookup_block_group(
245 struct btrfs_fs_info *info, u64 bytenr)
247 return block_group_cache_tree_search(info, bytenr, 1);
250 struct btrfs_block_group_cache *btrfs_next_block_group(
251 struct btrfs_block_group_cache *cache)
253 struct btrfs_fs_info *fs_info = cache->fs_info;
254 struct rb_node *node;
256 spin_lock(&fs_info->block_group_cache_lock);
258 /* If our block group was removed, we need a full search. */
259 if (RB_EMPTY_NODE(&cache->cache_node)) {
260 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
262 spin_unlock(&fs_info->block_group_cache_lock);
263 btrfs_put_block_group(cache);
264 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
266 node = rb_next(&cache->cache_node);
267 btrfs_put_block_group(cache);
269 cache = rb_entry(node, struct btrfs_block_group_cache,
271 btrfs_get_block_group(cache);
274 spin_unlock(&fs_info->block_group_cache_lock);
278 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
280 struct btrfs_block_group_cache *bg;
283 bg = btrfs_lookup_block_group(fs_info, bytenr);
287 spin_lock(&bg->lock);
291 atomic_inc(&bg->nocow_writers);
292 spin_unlock(&bg->lock);
294 /* No put on block group, done by btrfs_dec_nocow_writers */
296 btrfs_put_block_group(bg);
301 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
303 struct btrfs_block_group_cache *bg;
305 bg = btrfs_lookup_block_group(fs_info, bytenr);
307 if (atomic_dec_and_test(&bg->nocow_writers))
308 wake_up_var(&bg->nocow_writers);
310 * Once for our lookup and once for the lookup done by a previous call
311 * to btrfs_inc_nocow_writers()
313 btrfs_put_block_group(bg);
314 btrfs_put_block_group(bg);
317 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
319 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
322 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
325 struct btrfs_block_group_cache *bg;
327 bg = btrfs_lookup_block_group(fs_info, start);
329 if (atomic_dec_and_test(&bg->reservations))
330 wake_up_var(&bg->reservations);
331 btrfs_put_block_group(bg);
334 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
336 struct btrfs_space_info *space_info = bg->space_info;
340 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
344 * Our block group is read only but before we set it to read only,
345 * some task might have had allocated an extent from it already, but it
346 * has not yet created a respective ordered extent (and added it to a
347 * root's list of ordered extents).
348 * Therefore wait for any task currently allocating extents, since the
349 * block group's reservations counter is incremented while a read lock
350 * on the groups' semaphore is held and decremented after releasing
351 * the read access on that semaphore and creating the ordered extent.
353 down_write(&space_info->groups_sem);
354 up_write(&space_info->groups_sem);
356 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
359 struct btrfs_caching_control *btrfs_get_caching_control(
360 struct btrfs_block_group_cache *cache)
362 struct btrfs_caching_control *ctl;
364 spin_lock(&cache->lock);
365 if (!cache->caching_ctl) {
366 spin_unlock(&cache->lock);
370 ctl = cache->caching_ctl;
371 refcount_inc(&ctl->count);
372 spin_unlock(&cache->lock);
376 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
378 if (refcount_dec_and_test(&ctl->count))
383 * When we wait for progress in the block group caching, its because our
384 * allocation attempt failed at least once. So, we must sleep and let some
385 * progress happen before we try again.
387 * This function will sleep at least once waiting for new free space to show
388 * up, and then it will check the block group free space numbers for our min
389 * num_bytes. Another option is to have it go ahead and look in the rbtree for
390 * a free extent of a given size, but this is a good start.
392 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
393 * any of the information in this block group.
395 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
398 struct btrfs_caching_control *caching_ctl;
400 caching_ctl = btrfs_get_caching_control(cache);
404 wait_event(caching_ctl->wait, btrfs_block_group_cache_done(cache) ||
405 (cache->free_space_ctl->free_space >= num_bytes));
407 btrfs_put_caching_control(caching_ctl);
410 int btrfs_wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
412 struct btrfs_caching_control *caching_ctl;
415 caching_ctl = btrfs_get_caching_control(cache);
417 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
419 wait_event(caching_ctl->wait, btrfs_block_group_cache_done(cache));
420 if (cache->cached == BTRFS_CACHE_ERROR)
422 btrfs_put_caching_control(caching_ctl);
426 #ifdef CONFIG_BTRFS_DEBUG
427 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
429 struct btrfs_fs_info *fs_info = block_group->fs_info;
430 u64 start = block_group->key.objectid;
431 u64 len = block_group->key.offset;
432 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
433 fs_info->nodesize : fs_info->sectorsize;
434 u64 step = chunk << 1;
436 while (len > chunk) {
437 btrfs_remove_free_space(block_group, start, chunk);
448 * This is only called by btrfs_cache_block_group, since we could have freed
449 * extents we need to check the pinned_extents for any extents that can't be
450 * used yet since their free space will be released as soon as the transaction
453 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
456 struct btrfs_fs_info *info = block_group->fs_info;
457 u64 extent_start, extent_end, size, total_added = 0;
460 while (start < end) {
461 ret = find_first_extent_bit(info->pinned_extents, start,
462 &extent_start, &extent_end,
463 EXTENT_DIRTY | EXTENT_UPTODATE,
468 if (extent_start <= start) {
469 start = extent_end + 1;
470 } else if (extent_start > start && extent_start < end) {
471 size = extent_start - start;
473 ret = btrfs_add_free_space(block_group, start,
475 BUG_ON(ret); /* -ENOMEM or logic error */
476 start = extent_end + 1;
485 ret = btrfs_add_free_space(block_group, start, size);
486 BUG_ON(ret); /* -ENOMEM or logic error */
492 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
494 struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
495 struct btrfs_fs_info *fs_info = block_group->fs_info;
496 struct btrfs_root *extent_root = fs_info->extent_root;
497 struct btrfs_path *path;
498 struct extent_buffer *leaf;
499 struct btrfs_key key;
506 path = btrfs_alloc_path();
510 last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
512 #ifdef CONFIG_BTRFS_DEBUG
514 * If we're fragmenting we don't want to make anybody think we can
515 * allocate from this block group until we've had a chance to fragment
518 if (btrfs_should_fragment_free_space(block_group))
522 * We don't want to deadlock with somebody trying to allocate a new
523 * extent for the extent root while also trying to search the extent
524 * root to add free space. So we skip locking and search the commit
525 * root, since its read-only
527 path->skip_locking = 1;
528 path->search_commit_root = 1;
529 path->reada = READA_FORWARD;
533 key.type = BTRFS_EXTENT_ITEM_KEY;
536 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
540 leaf = path->nodes[0];
541 nritems = btrfs_header_nritems(leaf);
544 if (btrfs_fs_closing(fs_info) > 1) {
549 if (path->slots[0] < nritems) {
550 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
552 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
556 if (need_resched() ||
557 rwsem_is_contended(&fs_info->commit_root_sem)) {
559 caching_ctl->progress = last;
560 btrfs_release_path(path);
561 up_read(&fs_info->commit_root_sem);
562 mutex_unlock(&caching_ctl->mutex);
564 mutex_lock(&caching_ctl->mutex);
565 down_read(&fs_info->commit_root_sem);
569 ret = btrfs_next_leaf(extent_root, path);
574 leaf = path->nodes[0];
575 nritems = btrfs_header_nritems(leaf);
579 if (key.objectid < last) {
582 key.type = BTRFS_EXTENT_ITEM_KEY;
585 caching_ctl->progress = last;
586 btrfs_release_path(path);
590 if (key.objectid < block_group->key.objectid) {
595 if (key.objectid >= block_group->key.objectid +
596 block_group->key.offset)
599 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
600 key.type == BTRFS_METADATA_ITEM_KEY) {
601 total_found += add_new_free_space(block_group, last,
603 if (key.type == BTRFS_METADATA_ITEM_KEY)
604 last = key.objectid +
607 last = key.objectid + key.offset;
609 if (total_found > CACHING_CTL_WAKE_UP) {
612 wake_up(&caching_ctl->wait);
619 total_found += add_new_free_space(block_group, last,
620 block_group->key.objectid +
621 block_group->key.offset);
622 caching_ctl->progress = (u64)-1;
625 btrfs_free_path(path);
629 static noinline void caching_thread(struct btrfs_work *work)
631 struct btrfs_block_group_cache *block_group;
632 struct btrfs_fs_info *fs_info;
633 struct btrfs_caching_control *caching_ctl;
636 caching_ctl = container_of(work, struct btrfs_caching_control, work);
637 block_group = caching_ctl->block_group;
638 fs_info = block_group->fs_info;
640 mutex_lock(&caching_ctl->mutex);
641 down_read(&fs_info->commit_root_sem);
643 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
644 ret = load_free_space_tree(caching_ctl);
646 ret = load_extent_tree_free(caching_ctl);
648 spin_lock(&block_group->lock);
649 block_group->caching_ctl = NULL;
650 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
651 spin_unlock(&block_group->lock);
653 #ifdef CONFIG_BTRFS_DEBUG
654 if (btrfs_should_fragment_free_space(block_group)) {
657 spin_lock(&block_group->space_info->lock);
658 spin_lock(&block_group->lock);
659 bytes_used = block_group->key.offset -
660 btrfs_block_group_used(&block_group->item);
661 block_group->space_info->bytes_used += bytes_used >> 1;
662 spin_unlock(&block_group->lock);
663 spin_unlock(&block_group->space_info->lock);
664 fragment_free_space(block_group);
668 caching_ctl->progress = (u64)-1;
670 up_read(&fs_info->commit_root_sem);
671 btrfs_free_excluded_extents(block_group);
672 mutex_unlock(&caching_ctl->mutex);
674 wake_up(&caching_ctl->wait);
676 btrfs_put_caching_control(caching_ctl);
677 btrfs_put_block_group(block_group);
680 int btrfs_cache_block_group(struct btrfs_block_group_cache *cache,
684 struct btrfs_fs_info *fs_info = cache->fs_info;
685 struct btrfs_caching_control *caching_ctl;
688 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
692 INIT_LIST_HEAD(&caching_ctl->list);
693 mutex_init(&caching_ctl->mutex);
694 init_waitqueue_head(&caching_ctl->wait);
695 caching_ctl->block_group = cache;
696 caching_ctl->progress = cache->key.objectid;
697 refcount_set(&caching_ctl->count, 1);
698 btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
699 caching_thread, NULL, NULL);
701 spin_lock(&cache->lock);
703 * This should be a rare occasion, but this could happen I think in the
704 * case where one thread starts to load the space cache info, and then
705 * some other thread starts a transaction commit which tries to do an
706 * allocation while the other thread is still loading the space cache
707 * info. The previous loop should have kept us from choosing this block
708 * group, but if we've moved to the state where we will wait on caching
709 * block groups we need to first check if we're doing a fast load here,
710 * so we can wait for it to finish, otherwise we could end up allocating
711 * from a block group who's cache gets evicted for one reason or
714 while (cache->cached == BTRFS_CACHE_FAST) {
715 struct btrfs_caching_control *ctl;
717 ctl = cache->caching_ctl;
718 refcount_inc(&ctl->count);
719 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
720 spin_unlock(&cache->lock);
724 finish_wait(&ctl->wait, &wait);
725 btrfs_put_caching_control(ctl);
726 spin_lock(&cache->lock);
729 if (cache->cached != BTRFS_CACHE_NO) {
730 spin_unlock(&cache->lock);
734 WARN_ON(cache->caching_ctl);
735 cache->caching_ctl = caching_ctl;
736 cache->cached = BTRFS_CACHE_FAST;
737 spin_unlock(&cache->lock);
739 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
740 mutex_lock(&caching_ctl->mutex);
741 ret = load_free_space_cache(cache);
743 spin_lock(&cache->lock);
745 cache->caching_ctl = NULL;
746 cache->cached = BTRFS_CACHE_FINISHED;
747 cache->last_byte_to_unpin = (u64)-1;
748 caching_ctl->progress = (u64)-1;
750 if (load_cache_only) {
751 cache->caching_ctl = NULL;
752 cache->cached = BTRFS_CACHE_NO;
754 cache->cached = BTRFS_CACHE_STARTED;
755 cache->has_caching_ctl = 1;
758 spin_unlock(&cache->lock);
759 #ifdef CONFIG_BTRFS_DEBUG
761 btrfs_should_fragment_free_space(cache)) {
764 spin_lock(&cache->space_info->lock);
765 spin_lock(&cache->lock);
766 bytes_used = cache->key.offset -
767 btrfs_block_group_used(&cache->item);
768 cache->space_info->bytes_used += bytes_used >> 1;
769 spin_unlock(&cache->lock);
770 spin_unlock(&cache->space_info->lock);
771 fragment_free_space(cache);
774 mutex_unlock(&caching_ctl->mutex);
776 wake_up(&caching_ctl->wait);
778 btrfs_put_caching_control(caching_ctl);
779 btrfs_free_excluded_extents(cache);
784 * We're either using the free space tree or no caching at all.
785 * Set cached to the appropriate value and wakeup any waiters.
787 spin_lock(&cache->lock);
788 if (load_cache_only) {
789 cache->caching_ctl = NULL;
790 cache->cached = BTRFS_CACHE_NO;
792 cache->cached = BTRFS_CACHE_STARTED;
793 cache->has_caching_ctl = 1;
795 spin_unlock(&cache->lock);
796 wake_up(&caching_ctl->wait);
799 if (load_cache_only) {
800 btrfs_put_caching_control(caching_ctl);
804 down_write(&fs_info->commit_root_sem);
805 refcount_inc(&caching_ctl->count);
806 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
807 up_write(&fs_info->commit_root_sem);
809 btrfs_get_block_group(cache);
811 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
816 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
818 u64 extra_flags = chunk_to_extended(flags) &
819 BTRFS_EXTENDED_PROFILE_MASK;
821 write_seqlock(&fs_info->profiles_lock);
822 if (flags & BTRFS_BLOCK_GROUP_DATA)
823 fs_info->avail_data_alloc_bits &= ~extra_flags;
824 if (flags & BTRFS_BLOCK_GROUP_METADATA)
825 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
826 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
827 fs_info->avail_system_alloc_bits &= ~extra_flags;
828 write_sequnlock(&fs_info->profiles_lock);
832 * Clear incompat bits for the following feature(s):
834 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
835 * in the whole filesystem
837 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
839 if (flags & BTRFS_BLOCK_GROUP_RAID56_MASK) {
840 struct list_head *head = &fs_info->space_info;
841 struct btrfs_space_info *sinfo;
843 list_for_each_entry_rcu(sinfo, head, list) {
846 down_read(&sinfo->groups_sem);
847 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
849 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
851 up_read(&sinfo->groups_sem);
856 btrfs_clear_fs_incompat(fs_info, RAID56);
860 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
861 u64 group_start, struct extent_map *em)
863 struct btrfs_fs_info *fs_info = trans->fs_info;
864 struct btrfs_root *root = fs_info->extent_root;
865 struct btrfs_path *path;
866 struct btrfs_block_group_cache *block_group;
867 struct btrfs_free_cluster *cluster;
868 struct btrfs_root *tree_root = fs_info->tree_root;
869 struct btrfs_key key;
871 struct kobject *kobj = NULL;
875 struct btrfs_caching_control *caching_ctl = NULL;
877 bool remove_rsv = false;
879 block_group = btrfs_lookup_block_group(fs_info, group_start);
880 BUG_ON(!block_group);
881 BUG_ON(!block_group->ro);
883 trace_btrfs_remove_block_group(block_group);
885 * Free the reserved super bytes from this block group before
888 btrfs_free_excluded_extents(block_group);
889 btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
890 block_group->key.offset);
892 memcpy(&key, &block_group->key, sizeof(key));
893 index = btrfs_bg_flags_to_raid_index(block_group->flags);
894 factor = btrfs_bg_type_to_factor(block_group->flags);
896 /* make sure this block group isn't part of an allocation cluster */
897 cluster = &fs_info->data_alloc_cluster;
898 spin_lock(&cluster->refill_lock);
899 btrfs_return_cluster_to_free_space(block_group, cluster);
900 spin_unlock(&cluster->refill_lock);
903 * make sure this block group isn't part of a metadata
906 cluster = &fs_info->meta_alloc_cluster;
907 spin_lock(&cluster->refill_lock);
908 btrfs_return_cluster_to_free_space(block_group, cluster);
909 spin_unlock(&cluster->refill_lock);
911 path = btrfs_alloc_path();
918 * get the inode first so any iput calls done for the io_list
919 * aren't the final iput (no unlinks allowed now)
921 inode = lookup_free_space_inode(block_group, path);
923 mutex_lock(&trans->transaction->cache_write_mutex);
925 * Make sure our free space cache IO is done before removing the
928 spin_lock(&trans->transaction->dirty_bgs_lock);
929 if (!list_empty(&block_group->io_list)) {
930 list_del_init(&block_group->io_list);
932 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
934 spin_unlock(&trans->transaction->dirty_bgs_lock);
935 btrfs_wait_cache_io(trans, block_group, path);
936 btrfs_put_block_group(block_group);
937 spin_lock(&trans->transaction->dirty_bgs_lock);
940 if (!list_empty(&block_group->dirty_list)) {
941 list_del_init(&block_group->dirty_list);
943 btrfs_put_block_group(block_group);
945 spin_unlock(&trans->transaction->dirty_bgs_lock);
946 mutex_unlock(&trans->transaction->cache_write_mutex);
948 if (!IS_ERR(inode)) {
949 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
951 btrfs_add_delayed_iput(inode);
955 /* One for the block groups ref */
956 spin_lock(&block_group->lock);
957 if (block_group->iref) {
958 block_group->iref = 0;
959 block_group->inode = NULL;
960 spin_unlock(&block_group->lock);
963 spin_unlock(&block_group->lock);
965 /* One for our lookup ref */
966 btrfs_add_delayed_iput(inode);
969 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
970 key.offset = block_group->key.objectid;
973 ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
977 btrfs_release_path(path);
979 ret = btrfs_del_item(trans, tree_root, path);
982 btrfs_release_path(path);
985 spin_lock(&fs_info->block_group_cache_lock);
986 rb_erase(&block_group->cache_node,
987 &fs_info->block_group_cache_tree);
988 RB_CLEAR_NODE(&block_group->cache_node);
990 if (fs_info->first_logical_byte == block_group->key.objectid)
991 fs_info->first_logical_byte = (u64)-1;
992 spin_unlock(&fs_info->block_group_cache_lock);
994 down_write(&block_group->space_info->groups_sem);
996 * we must use list_del_init so people can check to see if they
997 * are still on the list after taking the semaphore
999 list_del_init(&block_group->list);
1000 if (list_empty(&block_group->space_info->block_groups[index])) {
1001 kobj = block_group->space_info->block_group_kobjs[index];
1002 block_group->space_info->block_group_kobjs[index] = NULL;
1003 clear_avail_alloc_bits(fs_info, block_group->flags);
1005 up_write(&block_group->space_info->groups_sem);
1006 clear_incompat_bg_bits(fs_info, block_group->flags);
1012 if (block_group->has_caching_ctl)
1013 caching_ctl = btrfs_get_caching_control(block_group);
1014 if (block_group->cached == BTRFS_CACHE_STARTED)
1015 btrfs_wait_block_group_cache_done(block_group);
1016 if (block_group->has_caching_ctl) {
1017 down_write(&fs_info->commit_root_sem);
1019 struct btrfs_caching_control *ctl;
1021 list_for_each_entry(ctl,
1022 &fs_info->caching_block_groups, list)
1023 if (ctl->block_group == block_group) {
1025 refcount_inc(&caching_ctl->count);
1030 list_del_init(&caching_ctl->list);
1031 up_write(&fs_info->commit_root_sem);
1033 /* Once for the caching bgs list and once for us. */
1034 btrfs_put_caching_control(caching_ctl);
1035 btrfs_put_caching_control(caching_ctl);
1039 spin_lock(&trans->transaction->dirty_bgs_lock);
1040 WARN_ON(!list_empty(&block_group->dirty_list));
1041 WARN_ON(!list_empty(&block_group->io_list));
1042 spin_unlock(&trans->transaction->dirty_bgs_lock);
1044 btrfs_remove_free_space_cache(block_group);
1046 spin_lock(&block_group->space_info->lock);
1047 list_del_init(&block_group->ro_list);
1049 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1050 WARN_ON(block_group->space_info->total_bytes
1051 < block_group->key.offset);
1052 WARN_ON(block_group->space_info->bytes_readonly
1053 < block_group->key.offset);
1054 WARN_ON(block_group->space_info->disk_total
1055 < block_group->key.offset * factor);
1057 block_group->space_info->total_bytes -= block_group->key.offset;
1058 block_group->space_info->bytes_readonly -= block_group->key.offset;
1059 block_group->space_info->disk_total -= block_group->key.offset * factor;
1061 spin_unlock(&block_group->space_info->lock);
1063 memcpy(&key, &block_group->key, sizeof(key));
1065 mutex_lock(&fs_info->chunk_mutex);
1066 spin_lock(&block_group->lock);
1067 block_group->removed = 1;
1069 * At this point trimming can't start on this block group, because we
1070 * removed the block group from the tree fs_info->block_group_cache_tree
1071 * so no one can't find it anymore and even if someone already got this
1072 * block group before we removed it from the rbtree, they have already
1073 * incremented block_group->trimming - if they didn't, they won't find
1074 * any free space entries because we already removed them all when we
1075 * called btrfs_remove_free_space_cache().
1077 * And we must not remove the extent map from the fs_info->mapping_tree
1078 * to prevent the same logical address range and physical device space
1079 * ranges from being reused for a new block group. This is because our
1080 * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1081 * completely transactionless, so while it is trimming a range the
1082 * currently running transaction might finish and a new one start,
1083 * allowing for new block groups to be created that can reuse the same
1084 * physical device locations unless we take this special care.
1086 * There may also be an implicit trim operation if the file system
1087 * is mounted with -odiscard. The same protections must remain
1088 * in place until the extents have been discarded completely when
1089 * the transaction commit has completed.
1091 remove_em = (atomic_read(&block_group->trimming) == 0);
1092 spin_unlock(&block_group->lock);
1094 mutex_unlock(&fs_info->chunk_mutex);
1096 ret = remove_block_group_free_space(trans, block_group);
1100 btrfs_put_block_group(block_group);
1101 btrfs_put_block_group(block_group);
1103 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1109 ret = btrfs_del_item(trans, root, path);
1114 struct extent_map_tree *em_tree;
1116 em_tree = &fs_info->mapping_tree;
1117 write_lock(&em_tree->lock);
1118 remove_extent_mapping(em_tree, em);
1119 write_unlock(&em_tree->lock);
1120 /* once for the tree */
1121 free_extent_map(em);
1125 btrfs_delayed_refs_rsv_release(fs_info, 1);
1126 btrfs_free_path(path);
1130 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1131 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1133 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1134 struct extent_map *em;
1135 struct map_lookup *map;
1136 unsigned int num_items;
1138 read_lock(&em_tree->lock);
1139 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1140 read_unlock(&em_tree->lock);
1141 ASSERT(em && em->start == chunk_offset);
1144 * We need to reserve 3 + N units from the metadata space info in order
1145 * to remove a block group (done at btrfs_remove_chunk() and at
1146 * btrfs_remove_block_group()), which are used for:
1148 * 1 unit for adding the free space inode's orphan (located in the tree
1150 * 1 unit for deleting the block group item (located in the extent
1152 * 1 unit for deleting the free space item (located in tree of tree
1154 * N units for deleting N device extent items corresponding to each
1155 * stripe (located in the device tree).
1157 * In order to remove a block group we also need to reserve units in the
1158 * system space info in order to update the chunk tree (update one or
1159 * more device items and remove one chunk item), but this is done at
1160 * btrfs_remove_chunk() through a call to check_system_chunk().
1162 map = em->map_lookup;
1163 num_items = 3 + map->num_stripes;
1164 free_extent_map(em);
1166 return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
1171 * Mark block group @cache read-only, so later write won't happen to block
1174 * If @force is not set, this function will only mark the block group readonly
1175 * if we have enough free space (1M) in other metadata/system block groups.
1176 * If @force is not set, this function will mark the block group readonly
1177 * without checking free space.
1179 * NOTE: This function doesn't care if other block groups can contain all the
1180 * data in this block group. That check should be done by relocation routine,
1181 * not this function.
1183 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
1185 struct btrfs_space_info *sinfo = cache->space_info;
1188 u64 min_allocable_bytes;
1192 * We need some metadata space and system metadata space for
1193 * allocating chunks in some corner cases until we force to set
1194 * it to be readonly.
1197 (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
1199 min_allocable_bytes = SZ_1M;
1201 min_allocable_bytes = 0;
1203 spin_lock(&sinfo->lock);
1204 spin_lock(&cache->lock);
1212 num_bytes = cache->key.offset - cache->reserved - cache->pinned -
1213 cache->bytes_super - btrfs_block_group_used(&cache->item);
1214 sinfo_used = btrfs_space_info_used(sinfo, true);
1217 * sinfo_used + num_bytes should always <= sinfo->total_bytes.
1219 * Here we make sure if we mark this bg RO, we still have enough
1220 * free space as buffer (if min_allocable_bytes is not 0).
1222 if (sinfo_used + num_bytes + min_allocable_bytes <=
1223 sinfo->total_bytes) {
1224 sinfo->bytes_readonly += num_bytes;
1226 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1230 spin_unlock(&cache->lock);
1231 spin_unlock(&sinfo->lock);
1232 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1233 btrfs_info(cache->fs_info,
1234 "unable to make block group %llu ro",
1235 cache->key.objectid);
1236 btrfs_info(cache->fs_info,
1237 "sinfo_used=%llu bg_num_bytes=%llu min_allocable=%llu",
1238 sinfo_used, num_bytes, min_allocable_bytes);
1239 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1245 * Process the unused_bgs list and remove any that don't have any allocated
1246 * space inside of them.
1248 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1250 struct btrfs_block_group_cache *block_group;
1251 struct btrfs_space_info *space_info;
1252 struct btrfs_trans_handle *trans;
1255 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1258 spin_lock(&fs_info->unused_bgs_lock);
1259 while (!list_empty(&fs_info->unused_bgs)) {
1263 block_group = list_first_entry(&fs_info->unused_bgs,
1264 struct btrfs_block_group_cache,
1266 list_del_init(&block_group->bg_list);
1268 space_info = block_group->space_info;
1270 if (ret || btrfs_mixed_space_info(space_info)) {
1271 btrfs_put_block_group(block_group);
1274 spin_unlock(&fs_info->unused_bgs_lock);
1276 mutex_lock(&fs_info->delete_unused_bgs_mutex);
1278 /* Don't want to race with allocators so take the groups_sem */
1279 down_write(&space_info->groups_sem);
1280 spin_lock(&block_group->lock);
1281 if (block_group->reserved || block_group->pinned ||
1282 btrfs_block_group_used(&block_group->item) ||
1284 list_is_singular(&block_group->list)) {
1286 * We want to bail if we made new allocations or have
1287 * outstanding allocations in this block group. We do
1288 * the ro check in case balance is currently acting on
1291 trace_btrfs_skip_unused_block_group(block_group);
1292 spin_unlock(&block_group->lock);
1293 up_write(&space_info->groups_sem);
1296 spin_unlock(&block_group->lock);
1298 /* We don't want to force the issue, only flip if it's ok. */
1299 ret = inc_block_group_ro(block_group, 0);
1300 up_write(&space_info->groups_sem);
1307 * Want to do this before we do anything else so we can recover
1308 * properly if we fail to join the transaction.
1310 trans = btrfs_start_trans_remove_block_group(fs_info,
1311 block_group->key.objectid);
1312 if (IS_ERR(trans)) {
1313 btrfs_dec_block_group_ro(block_group);
1314 ret = PTR_ERR(trans);
1319 * We could have pending pinned extents for this block group,
1320 * just delete them, we don't care about them anymore.
1322 start = block_group->key.objectid;
1323 end = start + block_group->key.offset - 1;
1325 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1326 * btrfs_finish_extent_commit(). If we are at transaction N,
1327 * another task might be running finish_extent_commit() for the
1328 * previous transaction N - 1, and have seen a range belonging
1329 * to the block group in freed_extents[] before we were able to
1330 * clear the whole block group range from freed_extents[]. This
1331 * means that task can lookup for the block group after we
1332 * unpinned it from freed_extents[] and removed it, leading to
1333 * a BUG_ON() at btrfs_unpin_extent_range().
1335 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1336 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
1339 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1340 btrfs_dec_block_group_ro(block_group);
1343 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
1346 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1347 btrfs_dec_block_group_ro(block_group);
1350 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1352 /* Reset pinned so btrfs_put_block_group doesn't complain */
1353 spin_lock(&space_info->lock);
1354 spin_lock(&block_group->lock);
1356 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1357 -block_group->pinned);
1358 space_info->bytes_readonly += block_group->pinned;
1359 percpu_counter_add_batch(&space_info->total_bytes_pinned,
1360 -block_group->pinned,
1361 BTRFS_TOTAL_BYTES_PINNED_BATCH);
1362 block_group->pinned = 0;
1364 spin_unlock(&block_group->lock);
1365 spin_unlock(&space_info->lock);
1367 /* DISCARD can flip during remount */
1368 trimming = btrfs_test_opt(fs_info, DISCARD);
1370 /* Implicit trim during transaction commit. */
1372 btrfs_get_block_group_trimming(block_group);
1375 * Btrfs_remove_chunk will abort the transaction if things go
1378 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
1382 btrfs_put_block_group_trimming(block_group);
1387 * If we're not mounted with -odiscard, we can just forget
1388 * about this block group. Otherwise we'll need to wait
1389 * until transaction commit to do the actual discard.
1392 spin_lock(&fs_info->unused_bgs_lock);
1394 * A concurrent scrub might have added us to the list
1395 * fs_info->unused_bgs, so use a list_move operation
1396 * to add the block group to the deleted_bgs list.
1398 list_move(&block_group->bg_list,
1399 &trans->transaction->deleted_bgs);
1400 spin_unlock(&fs_info->unused_bgs_lock);
1401 btrfs_get_block_group(block_group);
1404 btrfs_end_transaction(trans);
1406 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1407 btrfs_put_block_group(block_group);
1408 spin_lock(&fs_info->unused_bgs_lock);
1410 spin_unlock(&fs_info->unused_bgs_lock);
1413 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
1415 struct btrfs_fs_info *fs_info = bg->fs_info;
1417 spin_lock(&fs_info->unused_bgs_lock);
1418 if (list_empty(&bg->bg_list)) {
1419 btrfs_get_block_group(bg);
1420 trace_btrfs_add_unused_block_group(bg);
1421 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1423 spin_unlock(&fs_info->unused_bgs_lock);
1426 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1427 struct btrfs_path *path,
1428 struct btrfs_key *key)
1430 struct btrfs_root *root = fs_info->extent_root;
1432 struct btrfs_key found_key;
1433 struct extent_buffer *leaf;
1434 struct btrfs_block_group_item bg;
1438 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1443 slot = path->slots[0];
1444 leaf = path->nodes[0];
1445 if (slot >= btrfs_header_nritems(leaf)) {
1446 ret = btrfs_next_leaf(root, path);
1453 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1455 if (found_key.objectid >= key->objectid &&
1456 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1457 struct extent_map_tree *em_tree;
1458 struct extent_map *em;
1460 em_tree = &root->fs_info->mapping_tree;
1461 read_lock(&em_tree->lock);
1462 em = lookup_extent_mapping(em_tree, found_key.objectid,
1464 read_unlock(&em_tree->lock);
1467 "logical %llu len %llu found bg but no related chunk",
1468 found_key.objectid, found_key.offset);
1470 } else if (em->start != found_key.objectid ||
1471 em->len != found_key.offset) {
1473 "block group %llu len %llu mismatch with chunk %llu len %llu",
1474 found_key.objectid, found_key.offset,
1475 em->start, em->len);
1478 read_extent_buffer(leaf, &bg,
1479 btrfs_item_ptr_offset(leaf, slot),
1481 flags = btrfs_block_group_flags(&bg) &
1482 BTRFS_BLOCK_GROUP_TYPE_MASK;
1484 if (flags != (em->map_lookup->type &
1485 BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1487 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1489 found_key.offset, flags,
1490 (BTRFS_BLOCK_GROUP_TYPE_MASK &
1491 em->map_lookup->type));
1497 free_extent_map(em);
1506 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1508 u64 extra_flags = chunk_to_extended(flags) &
1509 BTRFS_EXTENDED_PROFILE_MASK;
1511 write_seqlock(&fs_info->profiles_lock);
1512 if (flags & BTRFS_BLOCK_GROUP_DATA)
1513 fs_info->avail_data_alloc_bits |= extra_flags;
1514 if (flags & BTRFS_BLOCK_GROUP_METADATA)
1515 fs_info->avail_metadata_alloc_bits |= extra_flags;
1516 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1517 fs_info->avail_system_alloc_bits |= extra_flags;
1518 write_sequnlock(&fs_info->profiles_lock);
1521 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
1523 struct btrfs_fs_info *fs_info = cache->fs_info;
1529 if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
1530 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
1531 cache->bytes_super += stripe_len;
1532 ret = btrfs_add_excluded_extent(fs_info, cache->key.objectid,
1538 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1539 bytenr = btrfs_sb_offset(i);
1540 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
1541 bytenr, &logical, &nr, &stripe_len);
1548 if (logical[nr] > cache->key.objectid +
1552 if (logical[nr] + stripe_len <= cache->key.objectid)
1555 start = logical[nr];
1556 if (start < cache->key.objectid) {
1557 start = cache->key.objectid;
1558 len = (logical[nr] + stripe_len) - start;
1560 len = min_t(u64, stripe_len,
1561 cache->key.objectid +
1562 cache->key.offset - start);
1565 cache->bytes_super += len;
1566 ret = btrfs_add_excluded_extent(fs_info, start, len);
1578 static void link_block_group(struct btrfs_block_group_cache *cache)
1580 struct btrfs_space_info *space_info = cache->space_info;
1581 int index = btrfs_bg_flags_to_raid_index(cache->flags);
1584 down_write(&space_info->groups_sem);
1585 if (list_empty(&space_info->block_groups[index]))
1587 list_add_tail(&cache->list, &space_info->block_groups[index]);
1588 up_write(&space_info->groups_sem);
1591 btrfs_sysfs_add_block_group_type(cache);
1594 static struct btrfs_block_group_cache *btrfs_create_block_group_cache(
1595 struct btrfs_fs_info *fs_info, u64 start, u64 size)
1597 struct btrfs_block_group_cache *cache;
1599 cache = kzalloc(sizeof(*cache), GFP_NOFS);
1603 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1605 if (!cache->free_space_ctl) {
1610 cache->key.objectid = start;
1611 cache->key.offset = size;
1612 cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1614 cache->fs_info = fs_info;
1615 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1616 set_free_space_tree_thresholds(cache);
1618 atomic_set(&cache->count, 1);
1619 spin_lock_init(&cache->lock);
1620 init_rwsem(&cache->data_rwsem);
1621 INIT_LIST_HEAD(&cache->list);
1622 INIT_LIST_HEAD(&cache->cluster_list);
1623 INIT_LIST_HEAD(&cache->bg_list);
1624 INIT_LIST_HEAD(&cache->ro_list);
1625 INIT_LIST_HEAD(&cache->dirty_list);
1626 INIT_LIST_HEAD(&cache->io_list);
1627 btrfs_init_free_space_ctl(cache);
1628 atomic_set(&cache->trimming, 0);
1629 mutex_init(&cache->free_space_lock);
1630 btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
1636 * Iterate all chunks and verify that each of them has the corresponding block
1639 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
1641 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
1642 struct extent_map *em;
1643 struct btrfs_block_group_cache *bg;
1648 read_lock(&map_tree->lock);
1650 * lookup_extent_mapping will return the first extent map
1651 * intersecting the range, so setting @len to 1 is enough to
1652 * get the first chunk.
1654 em = lookup_extent_mapping(map_tree, start, 1);
1655 read_unlock(&map_tree->lock);
1659 bg = btrfs_lookup_block_group(fs_info, em->start);
1662 "chunk start=%llu len=%llu doesn't have corresponding block group",
1663 em->start, em->len);
1665 free_extent_map(em);
1668 if (bg->key.objectid != em->start ||
1669 bg->key.offset != em->len ||
1670 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
1671 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1673 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1675 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
1676 bg->key.objectid, bg->key.offset,
1677 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
1679 free_extent_map(em);
1680 btrfs_put_block_group(bg);
1683 start = em->start + em->len;
1684 free_extent_map(em);
1685 btrfs_put_block_group(bg);
1690 int btrfs_read_block_groups(struct btrfs_fs_info *info)
1692 struct btrfs_path *path;
1694 struct btrfs_block_group_cache *cache;
1695 struct btrfs_space_info *space_info;
1696 struct btrfs_key key;
1697 struct btrfs_key found_key;
1698 struct extent_buffer *leaf;
1704 feature = btrfs_super_incompat_flags(info->super_copy);
1705 mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
1709 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1710 path = btrfs_alloc_path();
1713 path->reada = READA_FORWARD;
1715 cache_gen = btrfs_super_cache_generation(info->super_copy);
1716 if (btrfs_test_opt(info, SPACE_CACHE) &&
1717 btrfs_super_generation(info->super_copy) != cache_gen)
1719 if (btrfs_test_opt(info, CLEAR_CACHE))
1723 ret = find_first_block_group(info, path, &key);
1729 leaf = path->nodes[0];
1730 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1732 cache = btrfs_create_block_group_cache(info, found_key.objectid,
1741 * When we mount with old space cache, we need to
1742 * set BTRFS_DC_CLEAR and set dirty flag.
1744 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
1745 * truncate the old free space cache inode and
1747 * b) Setting 'dirty flag' makes sure that we flush
1748 * the new space cache info onto disk.
1750 if (btrfs_test_opt(info, SPACE_CACHE))
1751 cache->disk_cache_state = BTRFS_DC_CLEAR;
1754 read_extent_buffer(leaf, &cache->item,
1755 btrfs_item_ptr_offset(leaf, path->slots[0]),
1756 sizeof(cache->item));
1757 cache->flags = btrfs_block_group_flags(&cache->item);
1759 ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
1760 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
1762 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
1763 cache->key.objectid);
1764 btrfs_put_block_group(cache);
1769 key.objectid = found_key.objectid + found_key.offset;
1770 btrfs_release_path(path);
1773 * We need to exclude the super stripes now so that the space
1774 * info has super bytes accounted for, otherwise we'll think
1775 * we have more space than we actually do.
1777 ret = exclude_super_stripes(cache);
1780 * We may have excluded something, so call this just in
1783 btrfs_free_excluded_extents(cache);
1784 btrfs_put_block_group(cache);
1789 * Check for two cases, either we are full, and therefore
1790 * don't need to bother with the caching work since we won't
1791 * find any space, or we are empty, and we can just add all
1792 * the space in and be done with it. This saves us _a_lot_ of
1793 * time, particularly in the full case.
1795 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
1796 cache->last_byte_to_unpin = (u64)-1;
1797 cache->cached = BTRFS_CACHE_FINISHED;
1798 btrfs_free_excluded_extents(cache);
1799 } else if (btrfs_block_group_used(&cache->item) == 0) {
1800 cache->last_byte_to_unpin = (u64)-1;
1801 cache->cached = BTRFS_CACHE_FINISHED;
1802 add_new_free_space(cache, found_key.objectid,
1803 found_key.objectid +
1805 btrfs_free_excluded_extents(cache);
1808 ret = btrfs_add_block_group_cache(info, cache);
1810 btrfs_remove_free_space_cache(cache);
1811 btrfs_put_block_group(cache);
1815 trace_btrfs_add_block_group(info, cache, 0);
1816 btrfs_update_space_info(info, cache->flags, found_key.offset,
1817 btrfs_block_group_used(&cache->item),
1818 cache->bytes_super, &space_info);
1820 cache->space_info = space_info;
1822 link_block_group(cache);
1824 set_avail_alloc_bits(info, cache->flags);
1825 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
1826 inc_block_group_ro(cache, 1);
1827 } else if (btrfs_block_group_used(&cache->item) == 0) {
1828 ASSERT(list_empty(&cache->bg_list));
1829 btrfs_mark_bg_unused(cache);
1833 list_for_each_entry_rcu(space_info, &info->space_info, list) {
1834 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
1835 (BTRFS_BLOCK_GROUP_RAID10 |
1836 BTRFS_BLOCK_GROUP_RAID1_MASK |
1837 BTRFS_BLOCK_GROUP_RAID56_MASK |
1838 BTRFS_BLOCK_GROUP_DUP)))
1841 * Avoid allocating from un-mirrored block group if there are
1842 * mirrored block groups.
1844 list_for_each_entry(cache,
1845 &space_info->block_groups[BTRFS_RAID_RAID0],
1847 inc_block_group_ro(cache, 1);
1848 list_for_each_entry(cache,
1849 &space_info->block_groups[BTRFS_RAID_SINGLE],
1851 inc_block_group_ro(cache, 1);
1854 btrfs_init_global_block_rsv(info);
1855 ret = check_chunk_block_group_mappings(info);
1857 btrfs_free_path(path);
1861 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
1863 struct btrfs_fs_info *fs_info = trans->fs_info;
1864 struct btrfs_block_group_cache *block_group;
1865 struct btrfs_root *extent_root = fs_info->extent_root;
1866 struct btrfs_block_group_item item;
1867 struct btrfs_key key;
1870 if (!trans->can_flush_pending_bgs)
1873 while (!list_empty(&trans->new_bgs)) {
1874 block_group = list_first_entry(&trans->new_bgs,
1875 struct btrfs_block_group_cache,
1880 spin_lock(&block_group->lock);
1881 memcpy(&item, &block_group->item, sizeof(item));
1882 memcpy(&key, &block_group->key, sizeof(key));
1883 spin_unlock(&block_group->lock);
1885 ret = btrfs_insert_item(trans, extent_root, &key, &item,
1888 btrfs_abort_transaction(trans, ret);
1889 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
1891 btrfs_abort_transaction(trans, ret);
1892 add_block_group_free_space(trans, block_group);
1893 /* Already aborted the transaction if it failed. */
1895 btrfs_delayed_refs_rsv_release(fs_info, 1);
1896 list_del_init(&block_group->bg_list);
1898 btrfs_trans_release_chunk_metadata(trans);
1901 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
1902 u64 type, u64 chunk_offset, u64 size)
1904 struct btrfs_fs_info *fs_info = trans->fs_info;
1905 struct btrfs_block_group_cache *cache;
1908 btrfs_set_log_full_commit(trans);
1910 cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
1914 btrfs_set_block_group_used(&cache->item, bytes_used);
1915 btrfs_set_block_group_chunk_objectid(&cache->item,
1916 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1917 btrfs_set_block_group_flags(&cache->item, type);
1919 cache->flags = type;
1920 cache->last_byte_to_unpin = (u64)-1;
1921 cache->cached = BTRFS_CACHE_FINISHED;
1922 cache->needs_free_space = 1;
1923 ret = exclude_super_stripes(cache);
1925 /* We may have excluded something, so call this just in case */
1926 btrfs_free_excluded_extents(cache);
1927 btrfs_put_block_group(cache);
1931 add_new_free_space(cache, chunk_offset, chunk_offset + size);
1933 btrfs_free_excluded_extents(cache);
1935 #ifdef CONFIG_BTRFS_DEBUG
1936 if (btrfs_should_fragment_free_space(cache)) {
1937 u64 new_bytes_used = size - bytes_used;
1939 bytes_used += new_bytes_used >> 1;
1940 fragment_free_space(cache);
1944 * Ensure the corresponding space_info object is created and
1945 * assigned to our block group. We want our bg to be added to the rbtree
1946 * with its ->space_info set.
1948 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
1949 ASSERT(cache->space_info);
1951 ret = btrfs_add_block_group_cache(fs_info, cache);
1953 btrfs_remove_free_space_cache(cache);
1954 btrfs_put_block_group(cache);
1959 * Now that our block group has its ->space_info set and is inserted in
1960 * the rbtree, update the space info's counters.
1962 trace_btrfs_add_block_group(fs_info, cache, 1);
1963 btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
1964 cache->bytes_super, &cache->space_info);
1965 btrfs_update_global_block_rsv(fs_info);
1967 link_block_group(cache);
1969 list_add_tail(&cache->bg_list, &trans->new_bgs);
1970 trans->delayed_ref_updates++;
1971 btrfs_update_delayed_refs_rsv(trans);
1973 set_avail_alloc_bits(fs_info, type);
1977 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
1983 * if restripe for this chunk_type is on pick target profile and
1984 * return, otherwise do the usual balance
1986 stripped = get_restripe_target(fs_info, flags);
1988 return extended_to_chunk(stripped);
1990 num_devices = fs_info->fs_devices->rw_devices;
1992 stripped = BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID56_MASK |
1993 BTRFS_BLOCK_GROUP_RAID1_MASK | BTRFS_BLOCK_GROUP_RAID10;
1995 if (num_devices == 1) {
1996 stripped |= BTRFS_BLOCK_GROUP_DUP;
1997 stripped = flags & ~stripped;
1999 /* turn raid0 into single device chunks */
2000 if (flags & BTRFS_BLOCK_GROUP_RAID0)
2003 /* turn mirroring into duplication */
2004 if (flags & (BTRFS_BLOCK_GROUP_RAID1_MASK |
2005 BTRFS_BLOCK_GROUP_RAID10))
2006 return stripped | BTRFS_BLOCK_GROUP_DUP;
2008 /* they already had raid on here, just return */
2009 if (flags & stripped)
2012 stripped |= BTRFS_BLOCK_GROUP_DUP;
2013 stripped = flags & ~stripped;
2015 /* switch duplicated blocks with raid1 */
2016 if (flags & BTRFS_BLOCK_GROUP_DUP)
2017 return stripped | BTRFS_BLOCK_GROUP_RAID1;
2019 /* this is drive concat, leave it alone */
2025 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
2028 struct btrfs_fs_info *fs_info = cache->fs_info;
2029 struct btrfs_trans_handle *trans;
2034 trans = btrfs_join_transaction(fs_info->extent_root);
2036 return PTR_ERR(trans);
2039 * we're not allowed to set block groups readonly after the dirty
2040 * block groups cache has started writing. If it already started,
2041 * back off and let this transaction commit
2043 mutex_lock(&fs_info->ro_block_group_mutex);
2044 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2045 u64 transid = trans->transid;
2047 mutex_unlock(&fs_info->ro_block_group_mutex);
2048 btrfs_end_transaction(trans);
2050 ret = btrfs_wait_for_commit(fs_info, transid);
2057 * if we are changing raid levels, try to allocate a corresponding
2058 * block group with the new raid level.
2060 alloc_flags = update_block_group_flags(fs_info, cache->flags);
2061 if (alloc_flags != cache->flags) {
2062 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2064 * ENOSPC is allowed here, we may have enough space
2065 * already allocated at the new raid level to
2074 ret = inc_block_group_ro(cache, 0);
2077 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2078 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2081 ret = inc_block_group_ro(cache, 0);
2083 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2084 alloc_flags = update_block_group_flags(fs_info, cache->flags);
2085 mutex_lock(&fs_info->chunk_mutex);
2086 check_system_chunk(trans, alloc_flags);
2087 mutex_unlock(&fs_info->chunk_mutex);
2089 mutex_unlock(&fs_info->ro_block_group_mutex);
2091 btrfs_end_transaction(trans);
2095 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
2097 struct btrfs_space_info *sinfo = cache->space_info;
2102 spin_lock(&sinfo->lock);
2103 spin_lock(&cache->lock);
2105 num_bytes = cache->key.offset - cache->reserved -
2106 cache->pinned - cache->bytes_super -
2107 btrfs_block_group_used(&cache->item);
2108 sinfo->bytes_readonly -= num_bytes;
2109 list_del_init(&cache->ro_list);
2111 spin_unlock(&cache->lock);
2112 spin_unlock(&sinfo->lock);
2115 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2116 struct btrfs_path *path,
2117 struct btrfs_block_group_cache *cache)
2119 struct btrfs_fs_info *fs_info = trans->fs_info;
2121 struct btrfs_root *extent_root = fs_info->extent_root;
2123 struct extent_buffer *leaf;
2125 ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2132 leaf = path->nodes[0];
2133 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2134 write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2135 btrfs_mark_buffer_dirty(leaf);
2137 btrfs_release_path(path);
2142 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2143 struct btrfs_trans_handle *trans,
2144 struct btrfs_path *path)
2146 struct btrfs_fs_info *fs_info = block_group->fs_info;
2147 struct btrfs_root *root = fs_info->tree_root;
2148 struct inode *inode = NULL;
2149 struct extent_changeset *data_reserved = NULL;
2151 int dcs = BTRFS_DC_ERROR;
2157 * If this block group is smaller than 100 megs don't bother caching the
2160 if (block_group->key.offset < (100 * SZ_1M)) {
2161 spin_lock(&block_group->lock);
2162 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2163 spin_unlock(&block_group->lock);
2170 inode = lookup_free_space_inode(block_group, path);
2171 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2172 ret = PTR_ERR(inode);
2173 btrfs_release_path(path);
2177 if (IS_ERR(inode)) {
2181 if (block_group->ro)
2184 ret = create_free_space_inode(trans, block_group, path);
2191 * We want to set the generation to 0, that way if anything goes wrong
2192 * from here on out we know not to trust this cache when we load up next
2195 BTRFS_I(inode)->generation = 0;
2196 ret = btrfs_update_inode(trans, root, inode);
2199 * So theoretically we could recover from this, simply set the
2200 * super cache generation to 0 so we know to invalidate the
2201 * cache, but then we'd have to keep track of the block groups
2202 * that fail this way so we know we _have_ to reset this cache
2203 * before the next commit or risk reading stale cache. So to
2204 * limit our exposure to horrible edge cases lets just abort the
2205 * transaction, this only happens in really bad situations
2208 btrfs_abort_transaction(trans, ret);
2213 /* We've already setup this transaction, go ahead and exit */
2214 if (block_group->cache_generation == trans->transid &&
2215 i_size_read(inode)) {
2216 dcs = BTRFS_DC_SETUP;
2220 if (i_size_read(inode) > 0) {
2221 ret = btrfs_check_trunc_cache_free_space(fs_info,
2222 &fs_info->global_block_rsv);
2226 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2231 spin_lock(&block_group->lock);
2232 if (block_group->cached != BTRFS_CACHE_FINISHED ||
2233 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2235 * don't bother trying to write stuff out _if_
2236 * a) we're not cached,
2237 * b) we're with nospace_cache mount option,
2238 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2240 dcs = BTRFS_DC_WRITTEN;
2241 spin_unlock(&block_group->lock);
2244 spin_unlock(&block_group->lock);
2247 * We hit an ENOSPC when setting up the cache in this transaction, just
2248 * skip doing the setup, we've already cleared the cache so we're safe.
2250 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2256 * Try to preallocate enough space based on how big the block group is.
2257 * Keep in mind this has to include any pinned space which could end up
2258 * taking up quite a bit since it's not folded into the other space
2261 num_pages = div_u64(block_group->key.offset, SZ_256M);
2266 num_pages *= PAGE_SIZE;
2268 ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
2272 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2273 num_pages, num_pages,
2276 * Our cache requires contiguous chunks so that we don't modify a bunch
2277 * of metadata or split extents when writing the cache out, which means
2278 * we can enospc if we are heavily fragmented in addition to just normal
2279 * out of space conditions. So if we hit this just skip setting up any
2280 * other block groups for this transaction, maybe we'll unpin enough
2281 * space the next time around.
2284 dcs = BTRFS_DC_SETUP;
2285 else if (ret == -ENOSPC)
2286 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2291 btrfs_release_path(path);
2293 spin_lock(&block_group->lock);
2294 if (!ret && dcs == BTRFS_DC_SETUP)
2295 block_group->cache_generation = trans->transid;
2296 block_group->disk_cache_state = dcs;
2297 spin_unlock(&block_group->lock);
2299 extent_changeset_free(data_reserved);
2303 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2305 struct btrfs_fs_info *fs_info = trans->fs_info;
2306 struct btrfs_block_group_cache *cache, *tmp;
2307 struct btrfs_transaction *cur_trans = trans->transaction;
2308 struct btrfs_path *path;
2310 if (list_empty(&cur_trans->dirty_bgs) ||
2311 !btrfs_test_opt(fs_info, SPACE_CACHE))
2314 path = btrfs_alloc_path();
2318 /* Could add new block groups, use _safe just in case */
2319 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
2321 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2322 cache_save_setup(cache, trans, path);
2325 btrfs_free_path(path);
2330 * Transaction commit does final block group cache writeback during a critical
2331 * section where nothing is allowed to change the FS. This is required in
2332 * order for the cache to actually match the block group, but can introduce a
2333 * lot of latency into the commit.
2335 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2336 * There's a chance we'll have to redo some of it if the block group changes
2337 * again during the commit, but it greatly reduces the commit latency by
2338 * getting rid of the easy block groups while we're still allowing others to
2341 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
2343 struct btrfs_fs_info *fs_info = trans->fs_info;
2344 struct btrfs_block_group_cache *cache;
2345 struct btrfs_transaction *cur_trans = trans->transaction;
2348 struct btrfs_path *path = NULL;
2350 struct list_head *io = &cur_trans->io_bgs;
2351 int num_started = 0;
2354 spin_lock(&cur_trans->dirty_bgs_lock);
2355 if (list_empty(&cur_trans->dirty_bgs)) {
2356 spin_unlock(&cur_trans->dirty_bgs_lock);
2359 list_splice_init(&cur_trans->dirty_bgs, &dirty);
2360 spin_unlock(&cur_trans->dirty_bgs_lock);
2363 /* Make sure all the block groups on our dirty list actually exist */
2364 btrfs_create_pending_block_groups(trans);
2367 path = btrfs_alloc_path();
2373 * cache_write_mutex is here only to save us from balance or automatic
2374 * removal of empty block groups deleting this block group while we are
2375 * writing out the cache
2377 mutex_lock(&trans->transaction->cache_write_mutex);
2378 while (!list_empty(&dirty)) {
2379 bool drop_reserve = true;
2381 cache = list_first_entry(&dirty,
2382 struct btrfs_block_group_cache,
2385 * This can happen if something re-dirties a block group that
2386 * is already under IO. Just wait for it to finish and then do
2389 if (!list_empty(&cache->io_list)) {
2390 list_del_init(&cache->io_list);
2391 btrfs_wait_cache_io(trans, cache, path);
2392 btrfs_put_block_group(cache);
2397 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2398 * it should update the cache_state. Don't delete until after
2401 * Since we're not running in the commit critical section
2402 * we need the dirty_bgs_lock to protect from update_block_group
2404 spin_lock(&cur_trans->dirty_bgs_lock);
2405 list_del_init(&cache->dirty_list);
2406 spin_unlock(&cur_trans->dirty_bgs_lock);
2410 cache_save_setup(cache, trans, path);
2412 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
2413 cache->io_ctl.inode = NULL;
2414 ret = btrfs_write_out_cache(trans, cache, path);
2415 if (ret == 0 && cache->io_ctl.inode) {
2420 * The cache_write_mutex is protecting the
2421 * io_list, also refer to the definition of
2422 * btrfs_transaction::io_bgs for more details
2424 list_add_tail(&cache->io_list, io);
2427 * If we failed to write the cache, the
2428 * generation will be bad and life goes on
2434 ret = write_one_cache_group(trans, path, cache);
2436 * Our block group might still be attached to the list
2437 * of new block groups in the transaction handle of some
2438 * other task (struct btrfs_trans_handle->new_bgs). This
2439 * means its block group item isn't yet in the extent
2440 * tree. If this happens ignore the error, as we will
2441 * try again later in the critical section of the
2442 * transaction commit.
2444 if (ret == -ENOENT) {
2446 spin_lock(&cur_trans->dirty_bgs_lock);
2447 if (list_empty(&cache->dirty_list)) {
2448 list_add_tail(&cache->dirty_list,
2449 &cur_trans->dirty_bgs);
2450 btrfs_get_block_group(cache);
2451 drop_reserve = false;
2453 spin_unlock(&cur_trans->dirty_bgs_lock);
2455 btrfs_abort_transaction(trans, ret);
2459 /* If it's not on the io list, we need to put the block group */
2461 btrfs_put_block_group(cache);
2463 btrfs_delayed_refs_rsv_release(fs_info, 1);
2469 * Avoid blocking other tasks for too long. It might even save
2470 * us from writing caches for block groups that are going to be
2473 mutex_unlock(&trans->transaction->cache_write_mutex);
2474 mutex_lock(&trans->transaction->cache_write_mutex);
2476 mutex_unlock(&trans->transaction->cache_write_mutex);
2479 * Go through delayed refs for all the stuff we've just kicked off
2480 * and then loop back (just once)
2482 ret = btrfs_run_delayed_refs(trans, 0);
2483 if (!ret && loops == 0) {
2485 spin_lock(&cur_trans->dirty_bgs_lock);
2486 list_splice_init(&cur_trans->dirty_bgs, &dirty);
2488 * dirty_bgs_lock protects us from concurrent block group
2489 * deletes too (not just cache_write_mutex).
2491 if (!list_empty(&dirty)) {
2492 spin_unlock(&cur_trans->dirty_bgs_lock);
2495 spin_unlock(&cur_trans->dirty_bgs_lock);
2496 } else if (ret < 0) {
2497 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
2500 btrfs_free_path(path);
2504 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
2506 struct btrfs_fs_info *fs_info = trans->fs_info;
2507 struct btrfs_block_group_cache *cache;
2508 struct btrfs_transaction *cur_trans = trans->transaction;
2511 struct btrfs_path *path;
2512 struct list_head *io = &cur_trans->io_bgs;
2513 int num_started = 0;
2515 path = btrfs_alloc_path();
2520 * Even though we are in the critical section of the transaction commit,
2521 * we can still have concurrent tasks adding elements to this
2522 * transaction's list of dirty block groups. These tasks correspond to
2523 * endio free space workers started when writeback finishes for a
2524 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
2525 * allocate new block groups as a result of COWing nodes of the root
2526 * tree when updating the free space inode. The writeback for the space
2527 * caches is triggered by an earlier call to
2528 * btrfs_start_dirty_block_groups() and iterations of the following
2530 * Also we want to do the cache_save_setup first and then run the
2531 * delayed refs to make sure we have the best chance at doing this all
2534 spin_lock(&cur_trans->dirty_bgs_lock);
2535 while (!list_empty(&cur_trans->dirty_bgs)) {
2536 cache = list_first_entry(&cur_trans->dirty_bgs,
2537 struct btrfs_block_group_cache,
2541 * This can happen if cache_save_setup re-dirties a block group
2542 * that is already under IO. Just wait for it to finish and
2543 * then do it all again
2545 if (!list_empty(&cache->io_list)) {
2546 spin_unlock(&cur_trans->dirty_bgs_lock);
2547 list_del_init(&cache->io_list);
2548 btrfs_wait_cache_io(trans, cache, path);
2549 btrfs_put_block_group(cache);
2550 spin_lock(&cur_trans->dirty_bgs_lock);
2554 * Don't remove from the dirty list until after we've waited on
2557 list_del_init(&cache->dirty_list);
2558 spin_unlock(&cur_trans->dirty_bgs_lock);
2561 cache_save_setup(cache, trans, path);
2564 ret = btrfs_run_delayed_refs(trans,
2565 (unsigned long) -1);
2567 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
2568 cache->io_ctl.inode = NULL;
2569 ret = btrfs_write_out_cache(trans, cache, path);
2570 if (ret == 0 && cache->io_ctl.inode) {
2573 list_add_tail(&cache->io_list, io);
2576 * If we failed to write the cache, the
2577 * generation will be bad and life goes on
2583 ret = write_one_cache_group(trans, path, cache);
2585 * One of the free space endio workers might have
2586 * created a new block group while updating a free space
2587 * cache's inode (at inode.c:btrfs_finish_ordered_io())
2588 * and hasn't released its transaction handle yet, in
2589 * which case the new block group is still attached to
2590 * its transaction handle and its creation has not
2591 * finished yet (no block group item in the extent tree
2592 * yet, etc). If this is the case, wait for all free
2593 * space endio workers to finish and retry. This is a
2594 * a very rare case so no need for a more efficient and
2597 if (ret == -ENOENT) {
2598 wait_event(cur_trans->writer_wait,
2599 atomic_read(&cur_trans->num_writers) == 1);
2600 ret = write_one_cache_group(trans, path, cache);
2603 btrfs_abort_transaction(trans, ret);
2606 /* If its not on the io list, we need to put the block group */
2608 btrfs_put_block_group(cache);
2609 btrfs_delayed_refs_rsv_release(fs_info, 1);
2610 spin_lock(&cur_trans->dirty_bgs_lock);
2612 spin_unlock(&cur_trans->dirty_bgs_lock);
2615 * Refer to the definition of io_bgs member for details why it's safe
2616 * to use it without any locking
2618 while (!list_empty(io)) {
2619 cache = list_first_entry(io, struct btrfs_block_group_cache,
2621 list_del_init(&cache->io_list);
2622 btrfs_wait_cache_io(trans, cache, path);
2623 btrfs_put_block_group(cache);
2626 btrfs_free_path(path);
2630 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
2631 u64 bytenr, u64 num_bytes, int alloc)
2633 struct btrfs_fs_info *info = trans->fs_info;
2634 struct btrfs_block_group_cache *cache = NULL;
2635 u64 total = num_bytes;
2641 /* Block accounting for super block */
2642 spin_lock(&info->delalloc_root_lock);
2643 old_val = btrfs_super_bytes_used(info->super_copy);
2645 old_val += num_bytes;
2647 old_val -= num_bytes;
2648 btrfs_set_super_bytes_used(info->super_copy, old_val);
2649 spin_unlock(&info->delalloc_root_lock);
2652 cache = btrfs_lookup_block_group(info, bytenr);
2657 factor = btrfs_bg_type_to_factor(cache->flags);
2660 * If this block group has free space cache written out, we
2661 * need to make sure to load it if we are removing space. This
2662 * is because we need the unpinning stage to actually add the
2663 * space back to the block group, otherwise we will leak space.
2665 if (!alloc && cache->cached == BTRFS_CACHE_NO)
2666 btrfs_cache_block_group(cache, 1);
2668 byte_in_group = bytenr - cache->key.objectid;
2669 WARN_ON(byte_in_group > cache->key.offset);
2671 spin_lock(&cache->space_info->lock);
2672 spin_lock(&cache->lock);
2674 if (btrfs_test_opt(info, SPACE_CACHE) &&
2675 cache->disk_cache_state < BTRFS_DC_CLEAR)
2676 cache->disk_cache_state = BTRFS_DC_CLEAR;
2678 old_val = btrfs_block_group_used(&cache->item);
2679 num_bytes = min(total, cache->key.offset - byte_in_group);
2681 old_val += num_bytes;
2682 btrfs_set_block_group_used(&cache->item, old_val);
2683 cache->reserved -= num_bytes;
2684 cache->space_info->bytes_reserved -= num_bytes;
2685 cache->space_info->bytes_used += num_bytes;
2686 cache->space_info->disk_used += num_bytes * factor;
2687 spin_unlock(&cache->lock);
2688 spin_unlock(&cache->space_info->lock);
2690 old_val -= num_bytes;
2691 btrfs_set_block_group_used(&cache->item, old_val);
2692 cache->pinned += num_bytes;
2693 btrfs_space_info_update_bytes_pinned(info,
2694 cache->space_info, num_bytes);
2695 cache->space_info->bytes_used -= num_bytes;
2696 cache->space_info->disk_used -= num_bytes * factor;
2697 spin_unlock(&cache->lock);
2698 spin_unlock(&cache->space_info->lock);
2700 percpu_counter_add_batch(
2701 &cache->space_info->total_bytes_pinned,
2703 BTRFS_TOTAL_BYTES_PINNED_BATCH);
2704 set_extent_dirty(info->pinned_extents,
2705 bytenr, bytenr + num_bytes - 1,
2706 GFP_NOFS | __GFP_NOFAIL);
2709 spin_lock(&trans->transaction->dirty_bgs_lock);
2710 if (list_empty(&cache->dirty_list)) {
2711 list_add_tail(&cache->dirty_list,
2712 &trans->transaction->dirty_bgs);
2713 trans->delayed_ref_updates++;
2714 btrfs_get_block_group(cache);
2716 spin_unlock(&trans->transaction->dirty_bgs_lock);
2719 * No longer have used bytes in this block group, queue it for
2720 * deletion. We do this after adding the block group to the
2721 * dirty list to avoid races between cleaner kthread and space
2724 if (!alloc && old_val == 0)
2725 btrfs_mark_bg_unused(cache);
2727 btrfs_put_block_group(cache);
2729 bytenr += num_bytes;
2732 /* Modified block groups are accounted for in the delayed_refs_rsv. */
2733 btrfs_update_delayed_refs_rsv(trans);
2738 * btrfs_add_reserved_bytes - update the block_group and space info counters
2739 * @cache: The cache we are manipulating
2740 * @ram_bytes: The number of bytes of file content, and will be same to
2741 * @num_bytes except for the compress path.
2742 * @num_bytes: The number of bytes in question
2743 * @delalloc: The blocks are allocated for the delalloc write
2745 * This is called by the allocator when it reserves space. If this is a
2746 * reservation and the block group has become read only we cannot make the
2747 * reservation and return -EAGAIN, otherwise this function always succeeds.
2749 int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
2750 u64 ram_bytes, u64 num_bytes, int delalloc)
2752 struct btrfs_space_info *space_info = cache->space_info;
2755 spin_lock(&space_info->lock);
2756 spin_lock(&cache->lock);
2760 cache->reserved += num_bytes;
2761 space_info->bytes_reserved += num_bytes;
2762 trace_btrfs_space_reservation(cache->fs_info, "space_info",
2763 space_info->flags, num_bytes, 1);
2764 btrfs_space_info_update_bytes_may_use(cache->fs_info,
2765 space_info, -ram_bytes);
2767 cache->delalloc_bytes += num_bytes;
2769 spin_unlock(&cache->lock);
2770 spin_unlock(&space_info->lock);
2775 * btrfs_free_reserved_bytes - update the block_group and space info counters
2776 * @cache: The cache we are manipulating
2777 * @num_bytes: The number of bytes in question
2778 * @delalloc: The blocks are allocated for the delalloc write
2780 * This is called by somebody who is freeing space that was never actually used
2781 * on disk. For example if you reserve some space for a new leaf in transaction
2782 * A and before transaction A commits you free that leaf, you call this with
2783 * reserve set to 0 in order to clear the reservation.
2785 void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
2786 u64 num_bytes, int delalloc)
2788 struct btrfs_space_info *space_info = cache->space_info;
2790 spin_lock(&space_info->lock);
2791 spin_lock(&cache->lock);
2793 space_info->bytes_readonly += num_bytes;
2794 cache->reserved -= num_bytes;
2795 space_info->bytes_reserved -= num_bytes;
2796 space_info->max_extent_size = 0;
2799 cache->delalloc_bytes -= num_bytes;
2800 spin_unlock(&cache->lock);
2801 spin_unlock(&space_info->lock);
2804 static void force_metadata_allocation(struct btrfs_fs_info *info)
2806 struct list_head *head = &info->space_info;
2807 struct btrfs_space_info *found;
2810 list_for_each_entry_rcu(found, head, list) {
2811 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
2812 found->force_alloc = CHUNK_ALLOC_FORCE;
2817 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
2818 struct btrfs_space_info *sinfo, int force)
2820 u64 bytes_used = btrfs_space_info_used(sinfo, false);
2823 if (force == CHUNK_ALLOC_FORCE)
2827 * in limited mode, we want to have some free space up to
2828 * about 1% of the FS size.
2830 if (force == CHUNK_ALLOC_LIMITED) {
2831 thresh = btrfs_super_total_bytes(fs_info->super_copy);
2832 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
2834 if (sinfo->total_bytes - bytes_used < thresh)
2838 if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
2843 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
2845 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
2847 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2851 * If force is CHUNK_ALLOC_FORCE:
2852 * - return 1 if it successfully allocates a chunk,
2853 * - return errors including -ENOSPC otherwise.
2854 * If force is NOT CHUNK_ALLOC_FORCE:
2855 * - return 0 if it doesn't need to allocate a new chunk,
2856 * - return 1 if it successfully allocates a chunk,
2857 * - return errors including -ENOSPC otherwise.
2859 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
2860 enum btrfs_chunk_alloc_enum force)
2862 struct btrfs_fs_info *fs_info = trans->fs_info;
2863 struct btrfs_space_info *space_info;
2864 bool wait_for_alloc = false;
2865 bool should_alloc = false;
2868 /* Don't re-enter if we're already allocating a chunk */
2869 if (trans->allocating_chunk)
2872 space_info = btrfs_find_space_info(fs_info, flags);
2876 spin_lock(&space_info->lock);
2877 if (force < space_info->force_alloc)
2878 force = space_info->force_alloc;
2879 should_alloc = should_alloc_chunk(fs_info, space_info, force);
2880 if (space_info->full) {
2881 /* No more free physical space */
2886 spin_unlock(&space_info->lock);
2888 } else if (!should_alloc) {
2889 spin_unlock(&space_info->lock);
2891 } else if (space_info->chunk_alloc) {
2893 * Someone is already allocating, so we need to block
2894 * until this someone is finished and then loop to
2895 * recheck if we should continue with our allocation
2898 wait_for_alloc = true;
2899 spin_unlock(&space_info->lock);
2900 mutex_lock(&fs_info->chunk_mutex);
2901 mutex_unlock(&fs_info->chunk_mutex);
2903 /* Proceed with allocation */
2904 space_info->chunk_alloc = 1;
2905 wait_for_alloc = false;
2906 spin_unlock(&space_info->lock);
2910 } while (wait_for_alloc);
2912 mutex_lock(&fs_info->chunk_mutex);
2913 trans->allocating_chunk = true;
2916 * If we have mixed data/metadata chunks we want to make sure we keep
2917 * allocating mixed chunks instead of individual chunks.
2919 if (btrfs_mixed_space_info(space_info))
2920 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
2923 * if we're doing a data chunk, go ahead and make sure that
2924 * we keep a reasonable number of metadata chunks allocated in the
2927 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
2928 fs_info->data_chunk_allocations++;
2929 if (!(fs_info->data_chunk_allocations %
2930 fs_info->metadata_ratio))
2931 force_metadata_allocation(fs_info);
2935 * Check if we have enough space in SYSTEM chunk because we may need
2936 * to update devices.
2938 check_system_chunk(trans, flags);
2940 ret = btrfs_alloc_chunk(trans, flags);
2941 trans->allocating_chunk = false;
2943 spin_lock(&space_info->lock);
2946 space_info->full = 1;
2951 space_info->max_extent_size = 0;
2954 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
2956 space_info->chunk_alloc = 0;
2957 spin_unlock(&space_info->lock);
2958 mutex_unlock(&fs_info->chunk_mutex);
2960 * When we allocate a new chunk we reserve space in the chunk block
2961 * reserve to make sure we can COW nodes/leafs in the chunk tree or
2962 * add new nodes/leafs to it if we end up needing to do it when
2963 * inserting the chunk item and updating device items as part of the
2964 * second phase of chunk allocation, performed by
2965 * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
2966 * large number of new block groups to create in our transaction
2967 * handle's new_bgs list to avoid exhausting the chunk block reserve
2968 * in extreme cases - like having a single transaction create many new
2969 * block groups when starting to write out the free space caches of all
2970 * the block groups that were made dirty during the lifetime of the
2973 if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
2974 btrfs_create_pending_block_groups(trans);
2979 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
2983 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
2985 num_dev = fs_info->fs_devices->rw_devices;
2991 * If @is_allocation is true, reserve space in the system space info necessary
2992 * for allocating a chunk, otherwise if it's false, reserve space necessary for
2995 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
2997 struct btrfs_fs_info *fs_info = trans->fs_info;
2998 struct btrfs_space_info *info;
3005 * Needed because we can end up allocating a system chunk and for an
3006 * atomic and race free space reservation in the chunk block reserve.
3008 lockdep_assert_held(&fs_info->chunk_mutex);
3010 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3011 spin_lock(&info->lock);
3012 left = info->total_bytes - btrfs_space_info_used(info, true);
3013 spin_unlock(&info->lock);
3015 num_devs = get_profile_num_devs(fs_info, type);
3017 /* num_devs device items to update and 1 chunk item to add or remove */
3018 thresh = btrfs_calc_metadata_size(fs_info, num_devs) +
3019 btrfs_calc_insert_metadata_size(fs_info, 1);
3021 if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3022 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3023 left, thresh, type);
3024 btrfs_dump_space_info(fs_info, info, 0, 0);
3027 if (left < thresh) {
3028 u64 flags = btrfs_system_alloc_profile(fs_info);
3031 * Ignore failure to create system chunk. We might end up not
3032 * needing it, as we might not need to COW all nodes/leafs from
3033 * the paths we visit in the chunk tree (they were already COWed
3034 * or created in the current transaction for example).
3036 ret = btrfs_alloc_chunk(trans, flags);
3040 ret = btrfs_block_rsv_add(fs_info->chunk_root,
3041 &fs_info->chunk_block_rsv,
3042 thresh, BTRFS_RESERVE_NO_FLUSH);
3044 trans->chunk_bytes_reserved += thresh;
3048 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
3050 struct btrfs_block_group_cache *block_group;
3054 struct inode *inode;
3056 block_group = btrfs_lookup_first_block_group(info, last);
3057 while (block_group) {
3058 btrfs_wait_block_group_cache_done(block_group);
3059 spin_lock(&block_group->lock);
3060 if (block_group->iref)
3062 spin_unlock(&block_group->lock);
3063 block_group = btrfs_next_block_group(block_group);
3072 inode = block_group->inode;
3073 block_group->iref = 0;
3074 block_group->inode = NULL;
3075 spin_unlock(&block_group->lock);
3076 ASSERT(block_group->io_ctl.inode == NULL);
3078 last = block_group->key.objectid + block_group->key.offset;
3079 btrfs_put_block_group(block_group);
3084 * Must be called only after stopping all workers, since we could have block
3085 * group caching kthreads running, and therefore they could race with us if we
3086 * freed the block groups before stopping them.
3088 int btrfs_free_block_groups(struct btrfs_fs_info *info)
3090 struct btrfs_block_group_cache *block_group;
3091 struct btrfs_space_info *space_info;
3092 struct btrfs_caching_control *caching_ctl;
3095 down_write(&info->commit_root_sem);
3096 while (!list_empty(&info->caching_block_groups)) {
3097 caching_ctl = list_entry(info->caching_block_groups.next,
3098 struct btrfs_caching_control, list);
3099 list_del(&caching_ctl->list);
3100 btrfs_put_caching_control(caching_ctl);
3102 up_write(&info->commit_root_sem);
3104 spin_lock(&info->unused_bgs_lock);
3105 while (!list_empty(&info->unused_bgs)) {
3106 block_group = list_first_entry(&info->unused_bgs,
3107 struct btrfs_block_group_cache,
3109 list_del_init(&block_group->bg_list);
3110 btrfs_put_block_group(block_group);
3112 spin_unlock(&info->unused_bgs_lock);
3114 spin_lock(&info->block_group_cache_lock);
3115 while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
3116 block_group = rb_entry(n, struct btrfs_block_group_cache,
3118 rb_erase(&block_group->cache_node,
3119 &info->block_group_cache_tree);
3120 RB_CLEAR_NODE(&block_group->cache_node);
3121 spin_unlock(&info->block_group_cache_lock);
3123 down_write(&block_group->space_info->groups_sem);
3124 list_del(&block_group->list);
3125 up_write(&block_group->space_info->groups_sem);
3128 * We haven't cached this block group, which means we could
3129 * possibly have excluded extents on this block group.
3131 if (block_group->cached == BTRFS_CACHE_NO ||
3132 block_group->cached == BTRFS_CACHE_ERROR)
3133 btrfs_free_excluded_extents(block_group);
3135 btrfs_remove_free_space_cache(block_group);
3136 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
3137 ASSERT(list_empty(&block_group->dirty_list));
3138 ASSERT(list_empty(&block_group->io_list));
3139 ASSERT(list_empty(&block_group->bg_list));
3140 ASSERT(atomic_read(&block_group->count) == 1);
3141 btrfs_put_block_group(block_group);
3143 spin_lock(&info->block_group_cache_lock);
3145 spin_unlock(&info->block_group_cache_lock);
3148 * Now that all the block groups are freed, go through and free all the
3149 * space_info structs. This is only called during the final stages of
3150 * unmount, and so we know nobody is using them. We call
3151 * synchronize_rcu() once before we start, just to be on the safe side.
3155 btrfs_release_global_block_rsv(info);
3157 while (!list_empty(&info->space_info)) {
3158 space_info = list_entry(info->space_info.next,
3159 struct btrfs_space_info,
3163 * Do not hide this behind enospc_debug, this is actually
3164 * important and indicates a real bug if this happens.
3166 if (WARN_ON(space_info->bytes_pinned > 0 ||
3167 space_info->bytes_reserved > 0 ||
3168 space_info->bytes_may_use > 0))
3169 btrfs_dump_space_info(info, space_info, 0, 0);
3170 list_del(&space_info->list);
3171 btrfs_sysfs_remove_space_info(space_info);