1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
6 #include <linux/sched.h>
7 #include <linux/sched/signal.h>
8 #include <linux/pagemap.h>
9 #include <linux/writeback.h>
10 #include <linux/blkdev.h>
11 #include <linux/sort.h>
12 #include <linux/rcupdate.h>
13 #include <linux/kthread.h>
14 #include <linux/slab.h>
15 #include <linux/ratelimit.h>
16 #include <linux/percpu_counter.h>
17 #include <linux/lockdep.h>
18 #include <linux/crc32c.h>
21 #include "print-tree.h"
25 #include "free-space-cache.h"
26 #include "free-space-tree.h"
30 #include "ref-verify.h"
32 #undef SCRAMBLE_DELAYED_REFS
35 * control flags for do_chunk_alloc's force field
36 * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
37 * if we really need one.
39 * CHUNK_ALLOC_LIMITED means to only try and allocate one
40 * if we have very few chunks already allocated. This is
41 * used as part of the clustering code to help make sure
42 * we have a good pool of storage to cluster in, without
43 * filling the FS with empty chunks
45 * CHUNK_ALLOC_FORCE means it must try to allocate one
49 CHUNK_ALLOC_NO_FORCE = 0,
50 CHUNK_ALLOC_LIMITED = 1,
51 CHUNK_ALLOC_FORCE = 2,
54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
55 struct btrfs_delayed_ref_node *node, u64 parent,
56 u64 root_objectid, u64 owner_objectid,
57 u64 owner_offset, int refs_to_drop,
58 struct btrfs_delayed_extent_op *extra_op);
59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
60 struct extent_buffer *leaf,
61 struct btrfs_extent_item *ei);
62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
63 u64 parent, u64 root_objectid,
64 u64 flags, u64 owner, u64 offset,
65 struct btrfs_key *ins, int ref_mod);
66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
67 struct btrfs_delayed_ref_node *node,
68 struct btrfs_delayed_extent_op *extent_op);
69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
71 static int find_next_key(struct btrfs_path *path, int level,
72 struct btrfs_key *key);
73 static void dump_space_info(struct btrfs_fs_info *fs_info,
74 struct btrfs_space_info *info, u64 bytes,
75 int dump_block_groups);
76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
79 struct btrfs_space_info *space_info,
81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
82 struct btrfs_space_info *space_info,
86 block_group_cache_done(struct btrfs_block_group_cache *cache)
89 return cache->cached == BTRFS_CACHE_FINISHED ||
90 cache->cached == BTRFS_CACHE_ERROR;
93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
95 return (cache->flags & bits) == bits;
98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
100 atomic_inc(&cache->count);
103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
105 if (atomic_dec_and_test(&cache->count)) {
106 WARN_ON(cache->pinned > 0);
107 WARN_ON(cache->reserved > 0);
110 * If not empty, someone is still holding mutex of
111 * full_stripe_lock, which can only be released by caller.
112 * And it will definitely cause use-after-free when caller
113 * tries to release full stripe lock.
115 * No better way to resolve, but only to warn.
117 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
118 kfree(cache->free_space_ctl);
124 * this adds the block group to the fs_info rb tree for the block group
127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
128 struct btrfs_block_group_cache *block_group)
131 struct rb_node *parent = NULL;
132 struct btrfs_block_group_cache *cache;
134 spin_lock(&info->block_group_cache_lock);
135 p = &info->block_group_cache_tree.rb_node;
139 cache = rb_entry(parent, struct btrfs_block_group_cache,
141 if (block_group->key.objectid < cache->key.objectid) {
143 } else if (block_group->key.objectid > cache->key.objectid) {
146 spin_unlock(&info->block_group_cache_lock);
151 rb_link_node(&block_group->cache_node, parent, p);
152 rb_insert_color(&block_group->cache_node,
153 &info->block_group_cache_tree);
155 if (info->first_logical_byte > block_group->key.objectid)
156 info->first_logical_byte = block_group->key.objectid;
158 spin_unlock(&info->block_group_cache_lock);
164 * This will return the block group at or after bytenr if contains is 0, else
165 * it will return the block group that contains the bytenr
167 static struct btrfs_block_group_cache *
168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
171 struct btrfs_block_group_cache *cache, *ret = NULL;
175 spin_lock(&info->block_group_cache_lock);
176 n = info->block_group_cache_tree.rb_node;
179 cache = rb_entry(n, struct btrfs_block_group_cache,
181 end = cache->key.objectid + cache->key.offset - 1;
182 start = cache->key.objectid;
184 if (bytenr < start) {
185 if (!contains && (!ret || start < ret->key.objectid))
188 } else if (bytenr > start) {
189 if (contains && bytenr <= end) {
200 btrfs_get_block_group(ret);
201 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
202 info->first_logical_byte = ret->key.objectid;
204 spin_unlock(&info->block_group_cache_lock);
209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
210 u64 start, u64 num_bytes)
212 u64 end = start + num_bytes - 1;
213 set_extent_bits(&fs_info->freed_extents[0],
214 start, end, EXTENT_UPTODATE);
215 set_extent_bits(&fs_info->freed_extents[1],
216 start, end, EXTENT_UPTODATE);
220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
222 struct btrfs_fs_info *fs_info = cache->fs_info;
225 start = cache->key.objectid;
226 end = start + cache->key.offset - 1;
228 clear_extent_bits(&fs_info->freed_extents[0],
229 start, end, EXTENT_UPTODATE);
230 clear_extent_bits(&fs_info->freed_extents[1],
231 start, end, EXTENT_UPTODATE);
234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
236 struct btrfs_fs_info *fs_info = cache->fs_info;
242 if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
243 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
244 cache->bytes_super += stripe_len;
245 ret = add_excluded_extent(fs_info, cache->key.objectid,
251 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252 bytenr = btrfs_sb_offset(i);
253 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
254 bytenr, &logical, &nr, &stripe_len);
261 if (logical[nr] > cache->key.objectid +
265 if (logical[nr] + stripe_len <= cache->key.objectid)
269 if (start < cache->key.objectid) {
270 start = cache->key.objectid;
271 len = (logical[nr] + stripe_len) - start;
273 len = min_t(u64, stripe_len,
274 cache->key.objectid +
275 cache->key.offset - start);
278 cache->bytes_super += len;
279 ret = add_excluded_extent(fs_info, start, len);
291 static struct btrfs_caching_control *
292 get_caching_control(struct btrfs_block_group_cache *cache)
294 struct btrfs_caching_control *ctl;
296 spin_lock(&cache->lock);
297 if (!cache->caching_ctl) {
298 spin_unlock(&cache->lock);
302 ctl = cache->caching_ctl;
303 refcount_inc(&ctl->count);
304 spin_unlock(&cache->lock);
308 static void put_caching_control(struct btrfs_caching_control *ctl)
310 if (refcount_dec_and_test(&ctl->count))
314 #ifdef CONFIG_BTRFS_DEBUG
315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
317 struct btrfs_fs_info *fs_info = block_group->fs_info;
318 u64 start = block_group->key.objectid;
319 u64 len = block_group->key.offset;
320 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
321 fs_info->nodesize : fs_info->sectorsize;
322 u64 step = chunk << 1;
324 while (len > chunk) {
325 btrfs_remove_free_space(block_group, start, chunk);
336 * this is only called by cache_block_group, since we could have freed extents
337 * we need to check the pinned_extents for any extents that can't be used yet
338 * since their free space will be released as soon as the transaction commits.
340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
343 struct btrfs_fs_info *info = block_group->fs_info;
344 u64 extent_start, extent_end, size, total_added = 0;
347 while (start < end) {
348 ret = find_first_extent_bit(info->pinned_extents, start,
349 &extent_start, &extent_end,
350 EXTENT_DIRTY | EXTENT_UPTODATE,
355 if (extent_start <= start) {
356 start = extent_end + 1;
357 } else if (extent_start > start && extent_start < end) {
358 size = extent_start - start;
360 ret = btrfs_add_free_space(block_group, start,
362 BUG_ON(ret); /* -ENOMEM or logic error */
363 start = extent_end + 1;
372 ret = btrfs_add_free_space(block_group, start, size);
373 BUG_ON(ret); /* -ENOMEM or logic error */
379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
381 struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
382 struct btrfs_fs_info *fs_info = block_group->fs_info;
383 struct btrfs_root *extent_root = fs_info->extent_root;
384 struct btrfs_path *path;
385 struct extent_buffer *leaf;
386 struct btrfs_key key;
393 path = btrfs_alloc_path();
397 last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
399 #ifdef CONFIG_BTRFS_DEBUG
401 * If we're fragmenting we don't want to make anybody think we can
402 * allocate from this block group until we've had a chance to fragment
405 if (btrfs_should_fragment_free_space(block_group))
409 * We don't want to deadlock with somebody trying to allocate a new
410 * extent for the extent root while also trying to search the extent
411 * root to add free space. So we skip locking and search the commit
412 * root, since its read-only
414 path->skip_locking = 1;
415 path->search_commit_root = 1;
416 path->reada = READA_FORWARD;
420 key.type = BTRFS_EXTENT_ITEM_KEY;
423 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
427 leaf = path->nodes[0];
428 nritems = btrfs_header_nritems(leaf);
431 if (btrfs_fs_closing(fs_info) > 1) {
436 if (path->slots[0] < nritems) {
437 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
439 ret = find_next_key(path, 0, &key);
443 if (need_resched() ||
444 rwsem_is_contended(&fs_info->commit_root_sem)) {
446 caching_ctl->progress = last;
447 btrfs_release_path(path);
448 up_read(&fs_info->commit_root_sem);
449 mutex_unlock(&caching_ctl->mutex);
451 mutex_lock(&caching_ctl->mutex);
452 down_read(&fs_info->commit_root_sem);
456 ret = btrfs_next_leaf(extent_root, path);
461 leaf = path->nodes[0];
462 nritems = btrfs_header_nritems(leaf);
466 if (key.objectid < last) {
469 key.type = BTRFS_EXTENT_ITEM_KEY;
472 caching_ctl->progress = last;
473 btrfs_release_path(path);
477 if (key.objectid < block_group->key.objectid) {
482 if (key.objectid >= block_group->key.objectid +
483 block_group->key.offset)
486 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
487 key.type == BTRFS_METADATA_ITEM_KEY) {
488 total_found += add_new_free_space(block_group, last,
490 if (key.type == BTRFS_METADATA_ITEM_KEY)
491 last = key.objectid +
494 last = key.objectid + key.offset;
496 if (total_found > CACHING_CTL_WAKE_UP) {
499 wake_up(&caching_ctl->wait);
506 total_found += add_new_free_space(block_group, last,
507 block_group->key.objectid +
508 block_group->key.offset);
509 caching_ctl->progress = (u64)-1;
512 btrfs_free_path(path);
516 static noinline void caching_thread(struct btrfs_work *work)
518 struct btrfs_block_group_cache *block_group;
519 struct btrfs_fs_info *fs_info;
520 struct btrfs_caching_control *caching_ctl;
523 caching_ctl = container_of(work, struct btrfs_caching_control, work);
524 block_group = caching_ctl->block_group;
525 fs_info = block_group->fs_info;
527 mutex_lock(&caching_ctl->mutex);
528 down_read(&fs_info->commit_root_sem);
530 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
531 ret = load_free_space_tree(caching_ctl);
533 ret = load_extent_tree_free(caching_ctl);
535 spin_lock(&block_group->lock);
536 block_group->caching_ctl = NULL;
537 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
538 spin_unlock(&block_group->lock);
540 #ifdef CONFIG_BTRFS_DEBUG
541 if (btrfs_should_fragment_free_space(block_group)) {
544 spin_lock(&block_group->space_info->lock);
545 spin_lock(&block_group->lock);
546 bytes_used = block_group->key.offset -
547 btrfs_block_group_used(&block_group->item);
548 block_group->space_info->bytes_used += bytes_used >> 1;
549 spin_unlock(&block_group->lock);
550 spin_unlock(&block_group->space_info->lock);
551 fragment_free_space(block_group);
555 caching_ctl->progress = (u64)-1;
557 up_read(&fs_info->commit_root_sem);
558 free_excluded_extents(block_group);
559 mutex_unlock(&caching_ctl->mutex);
561 wake_up(&caching_ctl->wait);
563 put_caching_control(caching_ctl);
564 btrfs_put_block_group(block_group);
567 static int cache_block_group(struct btrfs_block_group_cache *cache,
571 struct btrfs_fs_info *fs_info = cache->fs_info;
572 struct btrfs_caching_control *caching_ctl;
575 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
579 INIT_LIST_HEAD(&caching_ctl->list);
580 mutex_init(&caching_ctl->mutex);
581 init_waitqueue_head(&caching_ctl->wait);
582 caching_ctl->block_group = cache;
583 caching_ctl->progress = cache->key.objectid;
584 refcount_set(&caching_ctl->count, 1);
585 btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
586 caching_thread, NULL, NULL);
588 spin_lock(&cache->lock);
590 * This should be a rare occasion, but this could happen I think in the
591 * case where one thread starts to load the space cache info, and then
592 * some other thread starts a transaction commit which tries to do an
593 * allocation while the other thread is still loading the space cache
594 * info. The previous loop should have kept us from choosing this block
595 * group, but if we've moved to the state where we will wait on caching
596 * block groups we need to first check if we're doing a fast load here,
597 * so we can wait for it to finish, otherwise we could end up allocating
598 * from a block group who's cache gets evicted for one reason or
601 while (cache->cached == BTRFS_CACHE_FAST) {
602 struct btrfs_caching_control *ctl;
604 ctl = cache->caching_ctl;
605 refcount_inc(&ctl->count);
606 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
607 spin_unlock(&cache->lock);
611 finish_wait(&ctl->wait, &wait);
612 put_caching_control(ctl);
613 spin_lock(&cache->lock);
616 if (cache->cached != BTRFS_CACHE_NO) {
617 spin_unlock(&cache->lock);
621 WARN_ON(cache->caching_ctl);
622 cache->caching_ctl = caching_ctl;
623 cache->cached = BTRFS_CACHE_FAST;
624 spin_unlock(&cache->lock);
626 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
627 mutex_lock(&caching_ctl->mutex);
628 ret = load_free_space_cache(fs_info, cache);
630 spin_lock(&cache->lock);
632 cache->caching_ctl = NULL;
633 cache->cached = BTRFS_CACHE_FINISHED;
634 cache->last_byte_to_unpin = (u64)-1;
635 caching_ctl->progress = (u64)-1;
637 if (load_cache_only) {
638 cache->caching_ctl = NULL;
639 cache->cached = BTRFS_CACHE_NO;
641 cache->cached = BTRFS_CACHE_STARTED;
642 cache->has_caching_ctl = 1;
645 spin_unlock(&cache->lock);
646 #ifdef CONFIG_BTRFS_DEBUG
648 btrfs_should_fragment_free_space(cache)) {
651 spin_lock(&cache->space_info->lock);
652 spin_lock(&cache->lock);
653 bytes_used = cache->key.offset -
654 btrfs_block_group_used(&cache->item);
655 cache->space_info->bytes_used += bytes_used >> 1;
656 spin_unlock(&cache->lock);
657 spin_unlock(&cache->space_info->lock);
658 fragment_free_space(cache);
661 mutex_unlock(&caching_ctl->mutex);
663 wake_up(&caching_ctl->wait);
665 put_caching_control(caching_ctl);
666 free_excluded_extents(cache);
671 * We're either using the free space tree or no caching at all.
672 * Set cached to the appropriate value and wakeup any waiters.
674 spin_lock(&cache->lock);
675 if (load_cache_only) {
676 cache->caching_ctl = NULL;
677 cache->cached = BTRFS_CACHE_NO;
679 cache->cached = BTRFS_CACHE_STARTED;
680 cache->has_caching_ctl = 1;
682 spin_unlock(&cache->lock);
683 wake_up(&caching_ctl->wait);
686 if (load_cache_only) {
687 put_caching_control(caching_ctl);
691 down_write(&fs_info->commit_root_sem);
692 refcount_inc(&caching_ctl->count);
693 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
694 up_write(&fs_info->commit_root_sem);
696 btrfs_get_block_group(cache);
698 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
704 * return the block group that starts at or after bytenr
706 static struct btrfs_block_group_cache *
707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
709 return block_group_cache_tree_search(info, bytenr, 0);
713 * return the block group that contains the given bytenr
715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
716 struct btrfs_fs_info *info,
719 return block_group_cache_tree_search(info, bytenr, 1);
722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
725 struct list_head *head = &info->space_info;
726 struct btrfs_space_info *found;
728 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
731 list_for_each_entry_rcu(found, head, list) {
732 if (found->flags & flags) {
741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
742 bool metadata, u64 root_objectid)
744 struct btrfs_space_info *space_info;
748 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
749 flags = BTRFS_BLOCK_GROUP_SYSTEM;
751 flags = BTRFS_BLOCK_GROUP_METADATA;
753 flags = BTRFS_BLOCK_GROUP_DATA;
756 space_info = __find_space_info(fs_info, flags);
758 percpu_counter_add_batch(&space_info->total_bytes_pinned, num_bytes,
759 BTRFS_TOTAL_BYTES_PINNED_BATCH);
763 * after adding space to the filesystem, we need to clear the full flags
764 * on all the space infos.
766 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
768 struct list_head *head = &info->space_info;
769 struct btrfs_space_info *found;
772 list_for_each_entry_rcu(found, head, list)
777 /* simple helper to search for an existing data extent at a given offset */
778 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
781 struct btrfs_key key;
782 struct btrfs_path *path;
784 path = btrfs_alloc_path();
788 key.objectid = start;
790 key.type = BTRFS_EXTENT_ITEM_KEY;
791 ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
792 btrfs_free_path(path);
797 * helper function to lookup reference count and flags of a tree block.
799 * the head node for delayed ref is used to store the sum of all the
800 * reference count modifications queued up in the rbtree. the head
801 * node may also store the extent flags to set. This way you can check
802 * to see what the reference count and extent flags would be if all of
803 * the delayed refs are not processed.
805 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
806 struct btrfs_fs_info *fs_info, u64 bytenr,
807 u64 offset, int metadata, u64 *refs, u64 *flags)
809 struct btrfs_delayed_ref_head *head;
810 struct btrfs_delayed_ref_root *delayed_refs;
811 struct btrfs_path *path;
812 struct btrfs_extent_item *ei;
813 struct extent_buffer *leaf;
814 struct btrfs_key key;
821 * If we don't have skinny metadata, don't bother doing anything
824 if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
825 offset = fs_info->nodesize;
829 path = btrfs_alloc_path();
834 path->skip_locking = 1;
835 path->search_commit_root = 1;
839 key.objectid = bytenr;
842 key.type = BTRFS_METADATA_ITEM_KEY;
844 key.type = BTRFS_EXTENT_ITEM_KEY;
846 ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
850 if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
851 if (path->slots[0]) {
853 btrfs_item_key_to_cpu(path->nodes[0], &key,
855 if (key.objectid == bytenr &&
856 key.type == BTRFS_EXTENT_ITEM_KEY &&
857 key.offset == fs_info->nodesize)
863 leaf = path->nodes[0];
864 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
865 if (item_size >= sizeof(*ei)) {
866 ei = btrfs_item_ptr(leaf, path->slots[0],
867 struct btrfs_extent_item);
868 num_refs = btrfs_extent_refs(leaf, ei);
869 extent_flags = btrfs_extent_flags(leaf, ei);
872 btrfs_print_v0_err(fs_info);
874 btrfs_abort_transaction(trans, ret);
876 btrfs_handle_fs_error(fs_info, ret, NULL);
881 BUG_ON(num_refs == 0);
891 delayed_refs = &trans->transaction->delayed_refs;
892 spin_lock(&delayed_refs->lock);
893 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
895 if (!mutex_trylock(&head->mutex)) {
896 refcount_inc(&head->refs);
897 spin_unlock(&delayed_refs->lock);
899 btrfs_release_path(path);
902 * Mutex was contended, block until it's released and try
905 mutex_lock(&head->mutex);
906 mutex_unlock(&head->mutex);
907 btrfs_put_delayed_ref_head(head);
910 spin_lock(&head->lock);
911 if (head->extent_op && head->extent_op->update_flags)
912 extent_flags |= head->extent_op->flags_to_set;
914 BUG_ON(num_refs == 0);
916 num_refs += head->ref_mod;
917 spin_unlock(&head->lock);
918 mutex_unlock(&head->mutex);
920 spin_unlock(&delayed_refs->lock);
922 WARN_ON(num_refs == 0);
926 *flags = extent_flags;
928 btrfs_free_path(path);
933 * Back reference rules. Back refs have three main goals:
935 * 1) differentiate between all holders of references to an extent so that
936 * when a reference is dropped we can make sure it was a valid reference
937 * before freeing the extent.
939 * 2) Provide enough information to quickly find the holders of an extent
940 * if we notice a given block is corrupted or bad.
942 * 3) Make it easy to migrate blocks for FS shrinking or storage pool
943 * maintenance. This is actually the same as #2, but with a slightly
944 * different use case.
946 * There are two kinds of back refs. The implicit back refs is optimized
947 * for pointers in non-shared tree blocks. For a given pointer in a block,
948 * back refs of this kind provide information about the block's owner tree
949 * and the pointer's key. These information allow us to find the block by
950 * b-tree searching. The full back refs is for pointers in tree blocks not
951 * referenced by their owner trees. The location of tree block is recorded
952 * in the back refs. Actually the full back refs is generic, and can be
953 * used in all cases the implicit back refs is used. The major shortcoming
954 * of the full back refs is its overhead. Every time a tree block gets
955 * COWed, we have to update back refs entry for all pointers in it.
957 * For a newly allocated tree block, we use implicit back refs for
958 * pointers in it. This means most tree related operations only involve
959 * implicit back refs. For a tree block created in old transaction, the
960 * only way to drop a reference to it is COW it. So we can detect the
961 * event that tree block loses its owner tree's reference and do the
962 * back refs conversion.
964 * When a tree block is COWed through a tree, there are four cases:
966 * The reference count of the block is one and the tree is the block's
967 * owner tree. Nothing to do in this case.
969 * The reference count of the block is one and the tree is not the
970 * block's owner tree. In this case, full back refs is used for pointers
971 * in the block. Remove these full back refs, add implicit back refs for
972 * every pointers in the new block.
974 * The reference count of the block is greater than one and the tree is
975 * the block's owner tree. In this case, implicit back refs is used for
976 * pointers in the block. Add full back refs for every pointers in the
977 * block, increase lower level extents' reference counts. The original
978 * implicit back refs are entailed to the new block.
980 * The reference count of the block is greater than one and the tree is
981 * not the block's owner tree. Add implicit back refs for every pointer in
982 * the new block, increase lower level extents' reference count.
984 * Back Reference Key composing:
986 * The key objectid corresponds to the first byte in the extent,
987 * The key type is used to differentiate between types of back refs.
988 * There are different meanings of the key offset for different types
991 * File extents can be referenced by:
993 * - multiple snapshots, subvolumes, or different generations in one subvol
994 * - different files inside a single subvolume
995 * - different offsets inside a file (bookend extents in file.c)
997 * The extent ref structure for the implicit back refs has fields for:
999 * - Objectid of the subvolume root
1000 * - objectid of the file holding the reference
1001 * - original offset in the file
1002 * - how many bookend extents
1004 * The key offset for the implicit back refs is hash of the first
1007 * The extent ref structure for the full back refs has field for:
1009 * - number of pointers in the tree leaf
1011 * The key offset for the implicit back refs is the first byte of
1014 * When a file extent is allocated, The implicit back refs is used.
1015 * the fields are filled in:
1017 * (root_key.objectid, inode objectid, offset in file, 1)
1019 * When a file extent is removed file truncation, we find the
1020 * corresponding implicit back refs and check the following fields:
1022 * (btrfs_header_owner(leaf), inode objectid, offset in file)
1024 * Btree extents can be referenced by:
1026 * - Different subvolumes
1028 * Both the implicit back refs and the full back refs for tree blocks
1029 * only consist of key. The key offset for the implicit back refs is
1030 * objectid of block's owner tree. The key offset for the full back refs
1031 * is the first byte of parent block.
1033 * When implicit back refs is used, information about the lowest key and
1034 * level of the tree block are required. These information are stored in
1035 * tree block info structure.
1039 * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1040 * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1041 * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1043 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1044 struct btrfs_extent_inline_ref *iref,
1045 enum btrfs_inline_ref_type is_data)
1047 int type = btrfs_extent_inline_ref_type(eb, iref);
1048 u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1050 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1051 type == BTRFS_SHARED_BLOCK_REF_KEY ||
1052 type == BTRFS_SHARED_DATA_REF_KEY ||
1053 type == BTRFS_EXTENT_DATA_REF_KEY) {
1054 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1055 if (type == BTRFS_TREE_BLOCK_REF_KEY)
1057 if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1058 ASSERT(eb->fs_info);
1060 * Every shared one has parent tree
1061 * block, which must be aligned to
1065 IS_ALIGNED(offset, eb->fs_info->nodesize))
1068 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1069 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1071 if (type == BTRFS_SHARED_DATA_REF_KEY) {
1072 ASSERT(eb->fs_info);
1074 * Every shared one has parent tree
1075 * block, which must be aligned to
1079 IS_ALIGNED(offset, eb->fs_info->nodesize))
1083 ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1088 btrfs_print_leaf((struct extent_buffer *)eb);
1089 btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1093 return BTRFS_REF_TYPE_INVALID;
1096 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1098 u32 high_crc = ~(u32)0;
1099 u32 low_crc = ~(u32)0;
1102 lenum = cpu_to_le64(root_objectid);
1103 high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1104 lenum = cpu_to_le64(owner);
1105 low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1106 lenum = cpu_to_le64(offset);
1107 low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1109 return ((u64)high_crc << 31) ^ (u64)low_crc;
1112 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1113 struct btrfs_extent_data_ref *ref)
1115 return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1116 btrfs_extent_data_ref_objectid(leaf, ref),
1117 btrfs_extent_data_ref_offset(leaf, ref));
1120 static int match_extent_data_ref(struct extent_buffer *leaf,
1121 struct btrfs_extent_data_ref *ref,
1122 u64 root_objectid, u64 owner, u64 offset)
1124 if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1125 btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1126 btrfs_extent_data_ref_offset(leaf, ref) != offset)
1131 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1132 struct btrfs_path *path,
1133 u64 bytenr, u64 parent,
1135 u64 owner, u64 offset)
1137 struct btrfs_root *root = trans->fs_info->extent_root;
1138 struct btrfs_key key;
1139 struct btrfs_extent_data_ref *ref;
1140 struct extent_buffer *leaf;
1146 key.objectid = bytenr;
1148 key.type = BTRFS_SHARED_DATA_REF_KEY;
1149 key.offset = parent;
1151 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1152 key.offset = hash_extent_data_ref(root_objectid,
1157 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1169 leaf = path->nodes[0];
1170 nritems = btrfs_header_nritems(leaf);
1172 if (path->slots[0] >= nritems) {
1173 ret = btrfs_next_leaf(root, path);
1179 leaf = path->nodes[0];
1180 nritems = btrfs_header_nritems(leaf);
1184 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1185 if (key.objectid != bytenr ||
1186 key.type != BTRFS_EXTENT_DATA_REF_KEY)
1189 ref = btrfs_item_ptr(leaf, path->slots[0],
1190 struct btrfs_extent_data_ref);
1192 if (match_extent_data_ref(leaf, ref, root_objectid,
1195 btrfs_release_path(path);
1207 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1208 struct btrfs_path *path,
1209 u64 bytenr, u64 parent,
1210 u64 root_objectid, u64 owner,
1211 u64 offset, int refs_to_add)
1213 struct btrfs_root *root = trans->fs_info->extent_root;
1214 struct btrfs_key key;
1215 struct extent_buffer *leaf;
1220 key.objectid = bytenr;
1222 key.type = BTRFS_SHARED_DATA_REF_KEY;
1223 key.offset = parent;
1224 size = sizeof(struct btrfs_shared_data_ref);
1226 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1227 key.offset = hash_extent_data_ref(root_objectid,
1229 size = sizeof(struct btrfs_extent_data_ref);
1232 ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1233 if (ret && ret != -EEXIST)
1236 leaf = path->nodes[0];
1238 struct btrfs_shared_data_ref *ref;
1239 ref = btrfs_item_ptr(leaf, path->slots[0],
1240 struct btrfs_shared_data_ref);
1242 btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1244 num_refs = btrfs_shared_data_ref_count(leaf, ref);
1245 num_refs += refs_to_add;
1246 btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1249 struct btrfs_extent_data_ref *ref;
1250 while (ret == -EEXIST) {
1251 ref = btrfs_item_ptr(leaf, path->slots[0],
1252 struct btrfs_extent_data_ref);
1253 if (match_extent_data_ref(leaf, ref, root_objectid,
1256 btrfs_release_path(path);
1258 ret = btrfs_insert_empty_item(trans, root, path, &key,
1260 if (ret && ret != -EEXIST)
1263 leaf = path->nodes[0];
1265 ref = btrfs_item_ptr(leaf, path->slots[0],
1266 struct btrfs_extent_data_ref);
1268 btrfs_set_extent_data_ref_root(leaf, ref,
1270 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1271 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1272 btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1274 num_refs = btrfs_extent_data_ref_count(leaf, ref);
1275 num_refs += refs_to_add;
1276 btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1279 btrfs_mark_buffer_dirty(leaf);
1282 btrfs_release_path(path);
1286 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1287 struct btrfs_path *path,
1288 int refs_to_drop, int *last_ref)
1290 struct btrfs_key key;
1291 struct btrfs_extent_data_ref *ref1 = NULL;
1292 struct btrfs_shared_data_ref *ref2 = NULL;
1293 struct extent_buffer *leaf;
1297 leaf = path->nodes[0];
1298 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1300 if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1301 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1302 struct btrfs_extent_data_ref);
1303 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1304 } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1305 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1306 struct btrfs_shared_data_ref);
1307 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1308 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
1309 btrfs_print_v0_err(trans->fs_info);
1310 btrfs_abort_transaction(trans, -EINVAL);
1316 BUG_ON(num_refs < refs_to_drop);
1317 num_refs -= refs_to_drop;
1319 if (num_refs == 0) {
1320 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1323 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1324 btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1325 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1326 btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1327 btrfs_mark_buffer_dirty(leaf);
1332 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1333 struct btrfs_extent_inline_ref *iref)
1335 struct btrfs_key key;
1336 struct extent_buffer *leaf;
1337 struct btrfs_extent_data_ref *ref1;
1338 struct btrfs_shared_data_ref *ref2;
1342 leaf = path->nodes[0];
1343 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1345 BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY);
1348 * If type is invalid, we should have bailed out earlier than
1351 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1352 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1353 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1354 ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1355 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1357 ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1358 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1360 } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1361 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1362 struct btrfs_extent_data_ref);
1363 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1364 } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1365 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1366 struct btrfs_shared_data_ref);
1367 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1374 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1375 struct btrfs_path *path,
1376 u64 bytenr, u64 parent,
1379 struct btrfs_root *root = trans->fs_info->extent_root;
1380 struct btrfs_key key;
1383 key.objectid = bytenr;
1385 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1386 key.offset = parent;
1388 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1389 key.offset = root_objectid;
1392 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1398 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1399 struct btrfs_path *path,
1400 u64 bytenr, u64 parent,
1403 struct btrfs_key key;
1406 key.objectid = bytenr;
1408 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1409 key.offset = parent;
1411 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1412 key.offset = root_objectid;
1415 ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1417 btrfs_release_path(path);
1421 static inline int extent_ref_type(u64 parent, u64 owner)
1424 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1426 type = BTRFS_SHARED_BLOCK_REF_KEY;
1428 type = BTRFS_TREE_BLOCK_REF_KEY;
1431 type = BTRFS_SHARED_DATA_REF_KEY;
1433 type = BTRFS_EXTENT_DATA_REF_KEY;
1438 static int find_next_key(struct btrfs_path *path, int level,
1439 struct btrfs_key *key)
1442 for (; level < BTRFS_MAX_LEVEL; level++) {
1443 if (!path->nodes[level])
1445 if (path->slots[level] + 1 >=
1446 btrfs_header_nritems(path->nodes[level]))
1449 btrfs_item_key_to_cpu(path->nodes[level], key,
1450 path->slots[level] + 1);
1452 btrfs_node_key_to_cpu(path->nodes[level], key,
1453 path->slots[level] + 1);
1460 * look for inline back ref. if back ref is found, *ref_ret is set
1461 * to the address of inline back ref, and 0 is returned.
1463 * if back ref isn't found, *ref_ret is set to the address where it
1464 * should be inserted, and -ENOENT is returned.
1466 * if insert is true and there are too many inline back refs, the path
1467 * points to the extent item, and -EAGAIN is returned.
1469 * NOTE: inline back refs are ordered in the same way that back ref
1470 * items in the tree are ordered.
1472 static noinline_for_stack
1473 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1474 struct btrfs_path *path,
1475 struct btrfs_extent_inline_ref **ref_ret,
1476 u64 bytenr, u64 num_bytes,
1477 u64 parent, u64 root_objectid,
1478 u64 owner, u64 offset, int insert)
1480 struct btrfs_fs_info *fs_info = trans->fs_info;
1481 struct btrfs_root *root = fs_info->extent_root;
1482 struct btrfs_key key;
1483 struct extent_buffer *leaf;
1484 struct btrfs_extent_item *ei;
1485 struct btrfs_extent_inline_ref *iref;
1495 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1498 key.objectid = bytenr;
1499 key.type = BTRFS_EXTENT_ITEM_KEY;
1500 key.offset = num_bytes;
1502 want = extent_ref_type(parent, owner);
1504 extra_size = btrfs_extent_inline_ref_size(want);
1505 path->keep_locks = 1;
1510 * Owner is our level, so we can just add one to get the level for the
1511 * block we are interested in.
1513 if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1514 key.type = BTRFS_METADATA_ITEM_KEY;
1519 ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1526 * We may be a newly converted file system which still has the old fat
1527 * extent entries for metadata, so try and see if we have one of those.
1529 if (ret > 0 && skinny_metadata) {
1530 skinny_metadata = false;
1531 if (path->slots[0]) {
1533 btrfs_item_key_to_cpu(path->nodes[0], &key,
1535 if (key.objectid == bytenr &&
1536 key.type == BTRFS_EXTENT_ITEM_KEY &&
1537 key.offset == num_bytes)
1541 key.objectid = bytenr;
1542 key.type = BTRFS_EXTENT_ITEM_KEY;
1543 key.offset = num_bytes;
1544 btrfs_release_path(path);
1549 if (ret && !insert) {
1552 } else if (WARN_ON(ret)) {
1557 leaf = path->nodes[0];
1558 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1559 if (unlikely(item_size < sizeof(*ei))) {
1561 btrfs_print_v0_err(fs_info);
1562 btrfs_abort_transaction(trans, err);
1566 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1567 flags = btrfs_extent_flags(leaf, ei);
1569 ptr = (unsigned long)(ei + 1);
1570 end = (unsigned long)ei + item_size;
1572 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1573 ptr += sizeof(struct btrfs_tree_block_info);
1577 if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1578 needed = BTRFS_REF_TYPE_DATA;
1580 needed = BTRFS_REF_TYPE_BLOCK;
1588 iref = (struct btrfs_extent_inline_ref *)ptr;
1589 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1590 if (type == BTRFS_REF_TYPE_INVALID) {
1598 ptr += btrfs_extent_inline_ref_size(type);
1602 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1603 struct btrfs_extent_data_ref *dref;
1604 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1605 if (match_extent_data_ref(leaf, dref, root_objectid,
1610 if (hash_extent_data_ref_item(leaf, dref) <
1611 hash_extent_data_ref(root_objectid, owner, offset))
1615 ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1617 if (parent == ref_offset) {
1621 if (ref_offset < parent)
1624 if (root_objectid == ref_offset) {
1628 if (ref_offset < root_objectid)
1632 ptr += btrfs_extent_inline_ref_size(type);
1634 if (err == -ENOENT && insert) {
1635 if (item_size + extra_size >=
1636 BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1641 * To add new inline back ref, we have to make sure
1642 * there is no corresponding back ref item.
1643 * For simplicity, we just do not add new inline back
1644 * ref if there is any kind of item for this block
1646 if (find_next_key(path, 0, &key) == 0 &&
1647 key.objectid == bytenr &&
1648 key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1653 *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1656 path->keep_locks = 0;
1657 btrfs_unlock_up_safe(path, 1);
1663 * helper to add new inline back ref
1665 static noinline_for_stack
1666 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1667 struct btrfs_path *path,
1668 struct btrfs_extent_inline_ref *iref,
1669 u64 parent, u64 root_objectid,
1670 u64 owner, u64 offset, int refs_to_add,
1671 struct btrfs_delayed_extent_op *extent_op)
1673 struct extent_buffer *leaf;
1674 struct btrfs_extent_item *ei;
1677 unsigned long item_offset;
1682 leaf = path->nodes[0];
1683 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1684 item_offset = (unsigned long)iref - (unsigned long)ei;
1686 type = extent_ref_type(parent, owner);
1687 size = btrfs_extent_inline_ref_size(type);
1689 btrfs_extend_item(fs_info, path, size);
1691 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1692 refs = btrfs_extent_refs(leaf, ei);
1693 refs += refs_to_add;
1694 btrfs_set_extent_refs(leaf, ei, refs);
1696 __run_delayed_extent_op(extent_op, leaf, ei);
1698 ptr = (unsigned long)ei + item_offset;
1699 end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1700 if (ptr < end - size)
1701 memmove_extent_buffer(leaf, ptr + size, ptr,
1704 iref = (struct btrfs_extent_inline_ref *)ptr;
1705 btrfs_set_extent_inline_ref_type(leaf, iref, type);
1706 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1707 struct btrfs_extent_data_ref *dref;
1708 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1709 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1710 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1711 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1712 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1713 } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1714 struct btrfs_shared_data_ref *sref;
1715 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1716 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1717 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1718 } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1719 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1721 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1723 btrfs_mark_buffer_dirty(leaf);
1726 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1727 struct btrfs_path *path,
1728 struct btrfs_extent_inline_ref **ref_ret,
1729 u64 bytenr, u64 num_bytes, u64 parent,
1730 u64 root_objectid, u64 owner, u64 offset)
1734 ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1735 num_bytes, parent, root_objectid,
1740 btrfs_release_path(path);
1743 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1744 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1747 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1748 root_objectid, owner, offset);
1754 * helper to update/remove inline back ref
1756 static noinline_for_stack
1757 void update_inline_extent_backref(struct btrfs_path *path,
1758 struct btrfs_extent_inline_ref *iref,
1760 struct btrfs_delayed_extent_op *extent_op,
1763 struct extent_buffer *leaf = path->nodes[0];
1764 struct btrfs_fs_info *fs_info = leaf->fs_info;
1765 struct btrfs_extent_item *ei;
1766 struct btrfs_extent_data_ref *dref = NULL;
1767 struct btrfs_shared_data_ref *sref = NULL;
1775 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1776 refs = btrfs_extent_refs(leaf, ei);
1777 WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1778 refs += refs_to_mod;
1779 btrfs_set_extent_refs(leaf, ei, refs);
1781 __run_delayed_extent_op(extent_op, leaf, ei);
1784 * If type is invalid, we should have bailed out after
1785 * lookup_inline_extent_backref().
1787 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1788 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1790 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1791 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1792 refs = btrfs_extent_data_ref_count(leaf, dref);
1793 } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1794 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1795 refs = btrfs_shared_data_ref_count(leaf, sref);
1798 BUG_ON(refs_to_mod != -1);
1801 BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1802 refs += refs_to_mod;
1805 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1806 btrfs_set_extent_data_ref_count(leaf, dref, refs);
1808 btrfs_set_shared_data_ref_count(leaf, sref, refs);
1811 size = btrfs_extent_inline_ref_size(type);
1812 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1813 ptr = (unsigned long)iref;
1814 end = (unsigned long)ei + item_size;
1815 if (ptr + size < end)
1816 memmove_extent_buffer(leaf, ptr, ptr + size,
1819 btrfs_truncate_item(fs_info, path, item_size, 1);
1821 btrfs_mark_buffer_dirty(leaf);
1824 static noinline_for_stack
1825 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1826 struct btrfs_path *path,
1827 u64 bytenr, u64 num_bytes, u64 parent,
1828 u64 root_objectid, u64 owner,
1829 u64 offset, int refs_to_add,
1830 struct btrfs_delayed_extent_op *extent_op)
1832 struct btrfs_extent_inline_ref *iref;
1835 ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1836 num_bytes, parent, root_objectid,
1839 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1840 update_inline_extent_backref(path, iref, refs_to_add,
1842 } else if (ret == -ENOENT) {
1843 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1844 root_objectid, owner, offset,
1845 refs_to_add, extent_op);
1851 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1852 struct btrfs_path *path,
1853 u64 bytenr, u64 parent, u64 root_objectid,
1854 u64 owner, u64 offset, int refs_to_add)
1857 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1858 BUG_ON(refs_to_add != 1);
1859 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1862 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1863 root_objectid, owner, offset,
1869 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1870 struct btrfs_path *path,
1871 struct btrfs_extent_inline_ref *iref,
1872 int refs_to_drop, int is_data, int *last_ref)
1876 BUG_ON(!is_data && refs_to_drop != 1);
1878 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
1880 } else if (is_data) {
1881 ret = remove_extent_data_ref(trans, path, refs_to_drop,
1885 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1890 #define in_range(b, first, len) ((b) >= (first) && (b) < (first) + (len))
1891 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1892 u64 *discarded_bytes)
1895 u64 bytes_left, end;
1896 u64 aligned_start = ALIGN(start, 1 << 9);
1898 if (WARN_ON(start != aligned_start)) {
1899 len -= aligned_start - start;
1900 len = round_down(len, 1 << 9);
1901 start = aligned_start;
1904 *discarded_bytes = 0;
1912 /* Skip any superblocks on this device. */
1913 for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1914 u64 sb_start = btrfs_sb_offset(j);
1915 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1916 u64 size = sb_start - start;
1918 if (!in_range(sb_start, start, bytes_left) &&
1919 !in_range(sb_end, start, bytes_left) &&
1920 !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1924 * Superblock spans beginning of range. Adjust start and
1927 if (sb_start <= start) {
1928 start += sb_end - start;
1933 bytes_left = end - start;
1938 ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1941 *discarded_bytes += size;
1942 else if (ret != -EOPNOTSUPP)
1951 bytes_left = end - start;
1955 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
1958 *discarded_bytes += bytes_left;
1963 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
1964 u64 num_bytes, u64 *actual_bytes)
1967 u64 discarded_bytes = 0;
1968 struct btrfs_bio *bbio = NULL;
1972 * Avoid races with device replace and make sure our bbio has devices
1973 * associated to its stripes that don't go away while we are discarding.
1975 btrfs_bio_counter_inc_blocked(fs_info);
1976 /* Tell the block device(s) that the sectors can be discarded */
1977 ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
1979 /* Error condition is -ENOMEM */
1981 struct btrfs_bio_stripe *stripe = bbio->stripes;
1985 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1987 struct request_queue *req_q;
1989 if (!stripe->dev->bdev) {
1990 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
1993 req_q = bdev_get_queue(stripe->dev->bdev);
1994 if (!blk_queue_discard(req_q))
1997 ret = btrfs_issue_discard(stripe->dev->bdev,
2002 discarded_bytes += bytes;
2003 else if (ret != -EOPNOTSUPP)
2004 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2007 * Just in case we get back EOPNOTSUPP for some reason,
2008 * just ignore the return value so we don't screw up
2009 * people calling discard_extent.
2013 btrfs_put_bbio(bbio);
2015 btrfs_bio_counter_dec(fs_info);
2018 *actual_bytes = discarded_bytes;
2021 if (ret == -EOPNOTSUPP)
2026 /* Can return -ENOMEM */
2027 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2028 struct btrfs_root *root,
2029 u64 bytenr, u64 num_bytes, u64 parent,
2030 u64 root_objectid, u64 owner, u64 offset)
2032 struct btrfs_fs_info *fs_info = root->fs_info;
2033 int old_ref_mod, new_ref_mod;
2036 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2037 root_objectid == BTRFS_TREE_LOG_OBJECTID);
2039 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2040 owner, offset, BTRFS_ADD_DELAYED_REF);
2042 if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2043 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2045 root_objectid, (int)owner,
2046 BTRFS_ADD_DELAYED_REF, NULL,
2047 &old_ref_mod, &new_ref_mod);
2049 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2051 root_objectid, owner, offset,
2052 0, BTRFS_ADD_DELAYED_REF,
2053 &old_ref_mod, &new_ref_mod);
2056 if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2057 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2059 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2066 * __btrfs_inc_extent_ref - insert backreference for a given extent
2068 * @trans: Handle of transaction
2070 * @node: The delayed ref node used to get the bytenr/length for
2071 * extent whose references are incremented.
2073 * @parent: If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2074 * BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2075 * bytenr of the parent block. Since new extents are always
2076 * created with indirect references, this will only be the case
2077 * when relocating a shared extent. In that case, root_objectid
2078 * will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2081 * @root_objectid: The id of the root where this modification has originated,
2082 * this can be either one of the well-known metadata trees or
2083 * the subvolume id which references this extent.
2085 * @owner: For data extents it is the inode number of the owning file.
2086 * For metadata extents this parameter holds the level in the
2087 * tree of the extent.
2089 * @offset: For metadata extents the offset is ignored and is currently
2090 * always passed as 0. For data extents it is the fileoffset
2091 * this extent belongs to.
2093 * @refs_to_add Number of references to add
2095 * @extent_op Pointer to a structure, holding information necessary when
2096 * updating a tree block's flags
2099 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2100 struct btrfs_delayed_ref_node *node,
2101 u64 parent, u64 root_objectid,
2102 u64 owner, u64 offset, int refs_to_add,
2103 struct btrfs_delayed_extent_op *extent_op)
2105 struct btrfs_path *path;
2106 struct extent_buffer *leaf;
2107 struct btrfs_extent_item *item;
2108 struct btrfs_key key;
2109 u64 bytenr = node->bytenr;
2110 u64 num_bytes = node->num_bytes;
2114 path = btrfs_alloc_path();
2118 path->reada = READA_FORWARD;
2119 path->leave_spinning = 1;
2120 /* this will setup the path even if it fails to insert the back ref */
2121 ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2122 parent, root_objectid, owner,
2123 offset, refs_to_add, extent_op);
2124 if ((ret < 0 && ret != -EAGAIN) || !ret)
2128 * Ok we had -EAGAIN which means we didn't have space to insert and
2129 * inline extent ref, so just update the reference count and add a
2132 leaf = path->nodes[0];
2133 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2134 item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2135 refs = btrfs_extent_refs(leaf, item);
2136 btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2138 __run_delayed_extent_op(extent_op, leaf, item);
2140 btrfs_mark_buffer_dirty(leaf);
2141 btrfs_release_path(path);
2143 path->reada = READA_FORWARD;
2144 path->leave_spinning = 1;
2145 /* now insert the actual backref */
2146 ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2147 owner, offset, refs_to_add);
2149 btrfs_abort_transaction(trans, ret);
2151 btrfs_free_path(path);
2155 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2156 struct btrfs_delayed_ref_node *node,
2157 struct btrfs_delayed_extent_op *extent_op,
2158 int insert_reserved)
2161 struct btrfs_delayed_data_ref *ref;
2162 struct btrfs_key ins;
2167 ins.objectid = node->bytenr;
2168 ins.offset = node->num_bytes;
2169 ins.type = BTRFS_EXTENT_ITEM_KEY;
2171 ref = btrfs_delayed_node_to_data_ref(node);
2172 trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2174 if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2175 parent = ref->parent;
2176 ref_root = ref->root;
2178 if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2180 flags |= extent_op->flags_to_set;
2181 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2182 flags, ref->objectid,
2185 } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2186 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2187 ref->objectid, ref->offset,
2188 node->ref_mod, extent_op);
2189 } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2190 ret = __btrfs_free_extent(trans, node, parent,
2191 ref_root, ref->objectid,
2192 ref->offset, node->ref_mod,
2200 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2201 struct extent_buffer *leaf,
2202 struct btrfs_extent_item *ei)
2204 u64 flags = btrfs_extent_flags(leaf, ei);
2205 if (extent_op->update_flags) {
2206 flags |= extent_op->flags_to_set;
2207 btrfs_set_extent_flags(leaf, ei, flags);
2210 if (extent_op->update_key) {
2211 struct btrfs_tree_block_info *bi;
2212 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2213 bi = (struct btrfs_tree_block_info *)(ei + 1);
2214 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2218 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2219 struct btrfs_delayed_ref_head *head,
2220 struct btrfs_delayed_extent_op *extent_op)
2222 struct btrfs_fs_info *fs_info = trans->fs_info;
2223 struct btrfs_key key;
2224 struct btrfs_path *path;
2225 struct btrfs_extent_item *ei;
2226 struct extent_buffer *leaf;
2230 int metadata = !extent_op->is_data;
2235 if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2238 path = btrfs_alloc_path();
2242 key.objectid = head->bytenr;
2245 key.type = BTRFS_METADATA_ITEM_KEY;
2246 key.offset = extent_op->level;
2248 key.type = BTRFS_EXTENT_ITEM_KEY;
2249 key.offset = head->num_bytes;
2253 path->reada = READA_FORWARD;
2254 path->leave_spinning = 1;
2255 ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2262 if (path->slots[0] > 0) {
2264 btrfs_item_key_to_cpu(path->nodes[0], &key,
2266 if (key.objectid == head->bytenr &&
2267 key.type == BTRFS_EXTENT_ITEM_KEY &&
2268 key.offset == head->num_bytes)
2272 btrfs_release_path(path);
2275 key.objectid = head->bytenr;
2276 key.offset = head->num_bytes;
2277 key.type = BTRFS_EXTENT_ITEM_KEY;
2286 leaf = path->nodes[0];
2287 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2289 if (unlikely(item_size < sizeof(*ei))) {
2291 btrfs_print_v0_err(fs_info);
2292 btrfs_abort_transaction(trans, err);
2296 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2297 __run_delayed_extent_op(extent_op, leaf, ei);
2299 btrfs_mark_buffer_dirty(leaf);
2301 btrfs_free_path(path);
2305 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2306 struct btrfs_delayed_ref_node *node,
2307 struct btrfs_delayed_extent_op *extent_op,
2308 int insert_reserved)
2311 struct btrfs_delayed_tree_ref *ref;
2315 ref = btrfs_delayed_node_to_tree_ref(node);
2316 trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2318 if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2319 parent = ref->parent;
2320 ref_root = ref->root;
2322 if (node->ref_mod != 1) {
2323 btrfs_err(trans->fs_info,
2324 "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2325 node->bytenr, node->ref_mod, node->action, ref_root,
2329 if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2330 BUG_ON(!extent_op || !extent_op->update_flags);
2331 ret = alloc_reserved_tree_block(trans, node, extent_op);
2332 } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2333 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2334 ref->level, 0, 1, extent_op);
2335 } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2336 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2337 ref->level, 0, 1, extent_op);
2344 /* helper function to actually process a single delayed ref entry */
2345 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2346 struct btrfs_delayed_ref_node *node,
2347 struct btrfs_delayed_extent_op *extent_op,
2348 int insert_reserved)
2352 if (trans->aborted) {
2353 if (insert_reserved)
2354 btrfs_pin_extent(trans->fs_info, node->bytenr,
2355 node->num_bytes, 1);
2359 if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2360 node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2361 ret = run_delayed_tree_ref(trans, node, extent_op,
2363 else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2364 node->type == BTRFS_SHARED_DATA_REF_KEY)
2365 ret = run_delayed_data_ref(trans, node, extent_op,
2372 static inline struct btrfs_delayed_ref_node *
2373 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2375 struct btrfs_delayed_ref_node *ref;
2377 if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
2381 * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2382 * This is to prevent a ref count from going down to zero, which deletes
2383 * the extent item from the extent tree, when there still are references
2384 * to add, which would fail because they would not find the extent item.
2386 if (!list_empty(&head->ref_add_list))
2387 return list_first_entry(&head->ref_add_list,
2388 struct btrfs_delayed_ref_node, add_list);
2390 ref = rb_entry(rb_first_cached(&head->ref_tree),
2391 struct btrfs_delayed_ref_node, ref_node);
2392 ASSERT(list_empty(&ref->add_list));
2396 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2397 struct btrfs_delayed_ref_head *head)
2399 spin_lock(&delayed_refs->lock);
2400 head->processing = 0;
2401 delayed_refs->num_heads_ready++;
2402 spin_unlock(&delayed_refs->lock);
2403 btrfs_delayed_ref_unlock(head);
2406 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2407 struct btrfs_delayed_ref_head *head)
2409 struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2414 head->extent_op = NULL;
2415 if (head->must_insert_reserved) {
2416 btrfs_free_delayed_extent_op(extent_op);
2419 spin_unlock(&head->lock);
2420 ret = run_delayed_extent_op(trans, head, extent_op);
2421 btrfs_free_delayed_extent_op(extent_op);
2422 return ret ? ret : 1;
2425 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2426 struct btrfs_delayed_ref_head *head)
2429 struct btrfs_fs_info *fs_info = trans->fs_info;
2430 struct btrfs_delayed_ref_root *delayed_refs;
2433 delayed_refs = &trans->transaction->delayed_refs;
2435 ret = cleanup_extent_op(trans, head);
2437 unselect_delayed_ref_head(delayed_refs, head);
2438 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2445 * Need to drop our head ref lock and re-acquire the delayed ref lock
2446 * and then re-check to make sure nobody got added.
2448 spin_unlock(&head->lock);
2449 spin_lock(&delayed_refs->lock);
2450 spin_lock(&head->lock);
2451 if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root) || head->extent_op) {
2452 spin_unlock(&head->lock);
2453 spin_unlock(&delayed_refs->lock);
2456 delayed_refs->num_heads--;
2457 rb_erase_cached(&head->href_node, &delayed_refs->href_root);
2458 RB_CLEAR_NODE(&head->href_node);
2459 spin_unlock(&head->lock);
2460 spin_unlock(&delayed_refs->lock);
2461 atomic_dec(&delayed_refs->num_entries);
2463 trace_run_delayed_ref_head(fs_info, head, 0);
2465 if (head->total_ref_mod < 0) {
2466 struct btrfs_space_info *space_info;
2470 flags = BTRFS_BLOCK_GROUP_DATA;
2471 else if (head->is_system)
2472 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2474 flags = BTRFS_BLOCK_GROUP_METADATA;
2475 space_info = __find_space_info(fs_info, flags);
2477 percpu_counter_add_batch(&space_info->total_bytes_pinned,
2479 BTRFS_TOTAL_BYTES_PINNED_BATCH);
2481 if (head->is_data) {
2482 spin_lock(&delayed_refs->lock);
2483 delayed_refs->pending_csums -= head->num_bytes;
2484 spin_unlock(&delayed_refs->lock);
2488 if (head->must_insert_reserved) {
2489 btrfs_pin_extent(fs_info, head->bytenr,
2490 head->num_bytes, 1);
2491 if (head->is_data) {
2492 ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2497 /* Also free its reserved qgroup space */
2498 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2499 head->qgroup_reserved);
2500 btrfs_delayed_ref_unlock(head);
2501 btrfs_put_delayed_ref_head(head);
2505 static struct btrfs_delayed_ref_head *btrfs_obtain_ref_head(
2506 struct btrfs_trans_handle *trans)
2508 struct btrfs_delayed_ref_root *delayed_refs =
2509 &trans->transaction->delayed_refs;
2510 struct btrfs_delayed_ref_head *head = NULL;
2513 spin_lock(&delayed_refs->lock);
2514 head = btrfs_select_ref_head(trans);
2516 spin_unlock(&delayed_refs->lock);
2521 * Grab the lock that says we are going to process all the refs for
2524 ret = btrfs_delayed_ref_lock(trans, head);
2525 spin_unlock(&delayed_refs->lock);
2528 * We may have dropped the spin lock to get the head mutex lock, and
2529 * that might have given someone else time to free the head. If that's
2530 * true, it has been removed from our list and we can move on.
2533 head = ERR_PTR(-EAGAIN);
2538 static int btrfs_run_delayed_refs_for_head(struct btrfs_trans_handle *trans,
2539 struct btrfs_delayed_ref_head *locked_ref,
2540 unsigned long *run_refs)
2542 struct btrfs_fs_info *fs_info = trans->fs_info;
2543 struct btrfs_delayed_ref_root *delayed_refs;
2544 struct btrfs_delayed_extent_op *extent_op;
2545 struct btrfs_delayed_ref_node *ref;
2546 int must_insert_reserved = 0;
2549 delayed_refs = &trans->transaction->delayed_refs;
2551 while ((ref = select_delayed_ref(locked_ref))) {
2553 btrfs_check_delayed_seq(fs_info, ref->seq)) {
2554 spin_unlock(&locked_ref->lock);
2555 unselect_delayed_ref_head(delayed_refs, locked_ref);
2561 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2562 RB_CLEAR_NODE(&ref->ref_node);
2563 if (!list_empty(&ref->add_list))
2564 list_del(&ref->add_list);
2566 * When we play the delayed ref, also correct the ref_mod on
2569 switch (ref->action) {
2570 case BTRFS_ADD_DELAYED_REF:
2571 case BTRFS_ADD_DELAYED_EXTENT:
2572 locked_ref->ref_mod -= ref->ref_mod;
2574 case BTRFS_DROP_DELAYED_REF:
2575 locked_ref->ref_mod += ref->ref_mod;
2580 atomic_dec(&delayed_refs->num_entries);
2583 * Record the must_insert_reserved flag before we drop the
2586 must_insert_reserved = locked_ref->must_insert_reserved;
2587 locked_ref->must_insert_reserved = 0;
2589 extent_op = locked_ref->extent_op;
2590 locked_ref->extent_op = NULL;
2591 spin_unlock(&locked_ref->lock);
2593 ret = run_one_delayed_ref(trans, ref, extent_op,
2594 must_insert_reserved);
2596 btrfs_free_delayed_extent_op(extent_op);
2598 unselect_delayed_ref_head(delayed_refs, locked_ref);
2599 btrfs_put_delayed_ref(ref);
2600 btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2605 btrfs_put_delayed_ref(ref);
2608 spin_lock(&locked_ref->lock);
2609 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2616 * Returns 0 on success or if called with an already aborted transaction.
2617 * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2619 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2622 struct btrfs_fs_info *fs_info = trans->fs_info;
2623 struct btrfs_delayed_ref_root *delayed_refs;
2624 struct btrfs_delayed_ref_node *ref;
2625 struct btrfs_delayed_ref_head *locked_ref = NULL;
2626 struct btrfs_delayed_extent_op *extent_op;
2627 ktime_t start = ktime_get();
2629 unsigned long count = 0;
2630 unsigned long actual_count = 0;
2631 int must_insert_reserved = 0;
2633 delayed_refs = &trans->transaction->delayed_refs;
2639 locked_ref = btrfs_obtain_ref_head(trans);
2642 else if (PTR_ERR(locked_ref) == -EAGAIN) {
2650 * We need to try and merge add/drops of the same ref since we
2651 * can run into issues with relocate dropping the implicit ref
2652 * and then it being added back again before the drop can
2653 * finish. If we merged anything we need to re-loop so we can
2655 * Or we can get node references of the same type that weren't
2656 * merged when created due to bumps in the tree mod seq, and
2657 * we need to merge them to prevent adding an inline extent
2658 * backref before dropping it (triggering a BUG_ON at
2659 * insert_inline_extent_backref()).
2661 spin_lock(&locked_ref->lock);
2662 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2664 ref = select_delayed_ref(locked_ref);
2666 if (ref && ref->seq &&
2667 btrfs_check_delayed_seq(fs_info, ref->seq)) {
2668 spin_unlock(&locked_ref->lock);
2669 unselect_delayed_ref_head(delayed_refs, locked_ref);
2677 * We're done processing refs in this ref_head, clean everything
2678 * up and move on to the next ref_head.
2681 ret = cleanup_ref_head(trans, locked_ref);
2683 /* We dropped our lock, we need to loop. */
2696 rb_erase_cached(&ref->ref_node, &locked_ref->ref_tree);
2697 RB_CLEAR_NODE(&ref->ref_node);
2698 if (!list_empty(&ref->add_list))
2699 list_del(&ref->add_list);
2701 * When we play the delayed ref, also correct the ref_mod on
2704 switch (ref->action) {
2705 case BTRFS_ADD_DELAYED_REF:
2706 case BTRFS_ADD_DELAYED_EXTENT:
2707 locked_ref->ref_mod -= ref->ref_mod;
2709 case BTRFS_DROP_DELAYED_REF:
2710 locked_ref->ref_mod += ref->ref_mod;
2715 atomic_dec(&delayed_refs->num_entries);
2718 * Record the must-insert_reserved flag before we drop the spin
2721 must_insert_reserved = locked_ref->must_insert_reserved;
2722 locked_ref->must_insert_reserved = 0;
2724 extent_op = locked_ref->extent_op;
2725 locked_ref->extent_op = NULL;
2726 spin_unlock(&locked_ref->lock);
2728 ret = run_one_delayed_ref(trans, ref, extent_op,
2729 must_insert_reserved);
2731 btrfs_free_delayed_extent_op(extent_op);
2733 unselect_delayed_ref_head(delayed_refs, locked_ref);
2734 btrfs_put_delayed_ref(ref);
2735 btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2740 btrfs_put_delayed_ref(ref);
2746 * We don't want to include ref heads since we can have empty ref heads
2747 * and those will drastically skew our runtime down since we just do
2748 * accounting, no actual extent tree updates.
2750 if (actual_count > 0) {
2751 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2755 * We weigh the current average higher than our current runtime
2756 * to avoid large swings in the average.
2758 spin_lock(&delayed_refs->lock);
2759 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2760 fs_info->avg_delayed_ref_runtime = avg >> 2; /* div by 4 */
2761 spin_unlock(&delayed_refs->lock);
2766 #ifdef SCRAMBLE_DELAYED_REFS
2768 * Normally delayed refs get processed in ascending bytenr order. This
2769 * correlates in most cases to the order added. To expose dependencies on this
2770 * order, we start to process the tree in the middle instead of the beginning
2772 static u64 find_middle(struct rb_root *root)
2774 struct rb_node *n = root->rb_node;
2775 struct btrfs_delayed_ref_node *entry;
2778 u64 first = 0, last = 0;
2782 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2783 first = entry->bytenr;
2787 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2788 last = entry->bytenr;
2793 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2794 WARN_ON(!entry->in_tree);
2796 middle = entry->bytenr;
2809 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2813 num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2814 sizeof(struct btrfs_extent_inline_ref));
2815 if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2816 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2819 * We don't ever fill up leaves all the way so multiply by 2 just to be
2820 * closer to what we're really going to want to use.
2822 return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2826 * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2827 * would require to store the csums for that many bytes.
2829 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2832 u64 num_csums_per_leaf;
2835 csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2836 num_csums_per_leaf = div64_u64(csum_size,
2837 (u64)btrfs_super_csum_size(fs_info->super_copy));
2838 num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2839 num_csums += num_csums_per_leaf - 1;
2840 num_csums = div64_u64(num_csums, num_csums_per_leaf);
2844 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2845 struct btrfs_fs_info *fs_info)
2847 struct btrfs_block_rsv *global_rsv;
2848 u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2849 u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2850 unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2851 u64 num_bytes, num_dirty_bgs_bytes;
2854 num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2855 num_heads = heads_to_leaves(fs_info, num_heads);
2857 num_bytes += (num_heads - 1) * fs_info->nodesize;
2859 num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2861 num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2863 global_rsv = &fs_info->global_block_rsv;
2866 * If we can't allocate any more chunks lets make sure we have _lots_ of
2867 * wiggle room since running delayed refs can create more delayed refs.
2869 if (global_rsv->space_info->full) {
2870 num_dirty_bgs_bytes <<= 1;
2874 spin_lock(&global_rsv->lock);
2875 if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2877 spin_unlock(&global_rsv->lock);
2881 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2882 struct btrfs_fs_info *fs_info)
2885 atomic_read(&trans->transaction->delayed_refs.num_entries);
2890 avg_runtime = fs_info->avg_delayed_ref_runtime;
2891 val = num_entries * avg_runtime;
2892 if (val >= NSEC_PER_SEC)
2894 if (val >= NSEC_PER_SEC / 2)
2897 return btrfs_check_space_for_delayed_refs(trans, fs_info);
2900 struct async_delayed_refs {
2901 struct btrfs_root *root;
2906 struct completion wait;
2907 struct btrfs_work work;
2910 static inline struct async_delayed_refs *
2911 to_async_delayed_refs(struct btrfs_work *work)
2913 return container_of(work, struct async_delayed_refs, work);
2916 static void delayed_ref_async_start(struct btrfs_work *work)
2918 struct async_delayed_refs *async = to_async_delayed_refs(work);
2919 struct btrfs_trans_handle *trans;
2920 struct btrfs_fs_info *fs_info = async->root->fs_info;
2923 /* if the commit is already started, we don't need to wait here */
2924 if (btrfs_transaction_blocked(fs_info))
2927 trans = btrfs_join_transaction(async->root);
2928 if (IS_ERR(trans)) {
2929 async->error = PTR_ERR(trans);
2934 * trans->sync means that when we call end_transaction, we won't
2935 * wait on delayed refs
2939 /* Don't bother flushing if we got into a different transaction */
2940 if (trans->transid > async->transid)
2943 ret = btrfs_run_delayed_refs(trans, async->count);
2947 ret = btrfs_end_transaction(trans);
2948 if (ret && !async->error)
2952 complete(&async->wait);
2957 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2958 unsigned long count, u64 transid, int wait)
2960 struct async_delayed_refs *async;
2963 async = kmalloc(sizeof(*async), GFP_NOFS);
2967 async->root = fs_info->tree_root;
2968 async->count = count;
2970 async->transid = transid;
2975 init_completion(&async->wait);
2977 btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2978 delayed_ref_async_start, NULL, NULL);
2980 btrfs_queue_work(fs_info->extent_workers, &async->work);
2983 wait_for_completion(&async->wait);
2992 * this starts processing the delayed reference count updates and
2993 * extent insertions we have queued up so far. count can be
2994 * 0, which means to process everything in the tree at the start
2995 * of the run (but not newly added entries), or it can be some target
2996 * number you'd like to process.
2998 * Returns 0 on success or if called with an aborted transaction
2999 * Returns <0 on error and aborts the transaction
3001 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3002 unsigned long count)
3004 struct btrfs_fs_info *fs_info = trans->fs_info;
3005 struct rb_node *node;
3006 struct btrfs_delayed_ref_root *delayed_refs;
3007 struct btrfs_delayed_ref_head *head;
3009 int run_all = count == (unsigned long)-1;
3010 bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3012 /* We'll clean this up in btrfs_cleanup_transaction */
3016 if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3019 delayed_refs = &trans->transaction->delayed_refs;
3021 count = atomic_read(&delayed_refs->num_entries) * 2;
3024 #ifdef SCRAMBLE_DELAYED_REFS
3025 delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3027 trans->can_flush_pending_bgs = false;
3028 ret = __btrfs_run_delayed_refs(trans, count);
3030 btrfs_abort_transaction(trans, ret);
3035 if (!list_empty(&trans->new_bgs))
3036 btrfs_create_pending_block_groups(trans);
3038 spin_lock(&delayed_refs->lock);
3039 node = rb_first_cached(&delayed_refs->href_root);
3041 spin_unlock(&delayed_refs->lock);
3044 head = rb_entry(node, struct btrfs_delayed_ref_head,
3046 refcount_inc(&head->refs);
3047 spin_unlock(&delayed_refs->lock);
3049 /* Mutex was contended, block until it's released and retry. */
3050 mutex_lock(&head->mutex);
3051 mutex_unlock(&head->mutex);
3053 btrfs_put_delayed_ref_head(head);
3058 trans->can_flush_pending_bgs = can_flush_pending_bgs;
3062 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3063 struct btrfs_fs_info *fs_info,
3064 u64 bytenr, u64 num_bytes, u64 flags,
3065 int level, int is_data)
3067 struct btrfs_delayed_extent_op *extent_op;
3070 extent_op = btrfs_alloc_delayed_extent_op();
3074 extent_op->flags_to_set = flags;
3075 extent_op->update_flags = true;
3076 extent_op->update_key = false;
3077 extent_op->is_data = is_data ? true : false;
3078 extent_op->level = level;
3080 ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3081 num_bytes, extent_op);
3083 btrfs_free_delayed_extent_op(extent_op);
3087 static noinline int check_delayed_ref(struct btrfs_root *root,
3088 struct btrfs_path *path,
3089 u64 objectid, u64 offset, u64 bytenr)
3091 struct btrfs_delayed_ref_head *head;
3092 struct btrfs_delayed_ref_node *ref;
3093 struct btrfs_delayed_data_ref *data_ref;
3094 struct btrfs_delayed_ref_root *delayed_refs;
3095 struct btrfs_transaction *cur_trans;
3096 struct rb_node *node;
3099 spin_lock(&root->fs_info->trans_lock);
3100 cur_trans = root->fs_info->running_transaction;
3102 refcount_inc(&cur_trans->use_count);
3103 spin_unlock(&root->fs_info->trans_lock);
3107 delayed_refs = &cur_trans->delayed_refs;
3108 spin_lock(&delayed_refs->lock);
3109 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3111 spin_unlock(&delayed_refs->lock);
3112 btrfs_put_transaction(cur_trans);
3116 if (!mutex_trylock(&head->mutex)) {
3117 refcount_inc(&head->refs);
3118 spin_unlock(&delayed_refs->lock);
3120 btrfs_release_path(path);
3123 * Mutex was contended, block until it's released and let
3126 mutex_lock(&head->mutex);
3127 mutex_unlock(&head->mutex);
3128 btrfs_put_delayed_ref_head(head);
3129 btrfs_put_transaction(cur_trans);
3132 spin_unlock(&delayed_refs->lock);
3134 spin_lock(&head->lock);
3136 * XXX: We should replace this with a proper search function in the
3139 for (node = rb_first_cached(&head->ref_tree); node;
3140 node = rb_next(node)) {
3141 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3142 /* If it's a shared ref we know a cross reference exists */
3143 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3148 data_ref = btrfs_delayed_node_to_data_ref(ref);
3151 * If our ref doesn't match the one we're currently looking at
3152 * then we have a cross reference.
3154 if (data_ref->root != root->root_key.objectid ||
3155 data_ref->objectid != objectid ||
3156 data_ref->offset != offset) {
3161 spin_unlock(&head->lock);
3162 mutex_unlock(&head->mutex);
3163 btrfs_put_transaction(cur_trans);
3167 static noinline int check_committed_ref(struct btrfs_root *root,
3168 struct btrfs_path *path,
3169 u64 objectid, u64 offset, u64 bytenr)
3171 struct btrfs_fs_info *fs_info = root->fs_info;
3172 struct btrfs_root *extent_root = fs_info->extent_root;
3173 struct extent_buffer *leaf;
3174 struct btrfs_extent_data_ref *ref;
3175 struct btrfs_extent_inline_ref *iref;
3176 struct btrfs_extent_item *ei;
3177 struct btrfs_key key;
3182 key.objectid = bytenr;
3183 key.offset = (u64)-1;
3184 key.type = BTRFS_EXTENT_ITEM_KEY;
3186 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3189 BUG_ON(ret == 0); /* Corruption */
3192 if (path->slots[0] == 0)
3196 leaf = path->nodes[0];
3197 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3199 if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3203 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3204 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3206 if (item_size != sizeof(*ei) +
3207 btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3210 if (btrfs_extent_generation(leaf, ei) <=
3211 btrfs_root_last_snapshot(&root->root_item))
3214 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3216 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3217 if (type != BTRFS_EXTENT_DATA_REF_KEY)
3220 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3221 if (btrfs_extent_refs(leaf, ei) !=
3222 btrfs_extent_data_ref_count(leaf, ref) ||
3223 btrfs_extent_data_ref_root(leaf, ref) !=
3224 root->root_key.objectid ||
3225 btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3226 btrfs_extent_data_ref_offset(leaf, ref) != offset)
3234 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3237 struct btrfs_path *path;
3240 path = btrfs_alloc_path();
3245 ret = check_committed_ref(root, path, objectid,
3247 if (ret && ret != -ENOENT)
3250 ret = check_delayed_ref(root, path, objectid, offset, bytenr);
3251 } while (ret == -EAGAIN);
3254 btrfs_free_path(path);
3255 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3260 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3261 struct btrfs_root *root,
3262 struct extent_buffer *buf,
3263 int full_backref, int inc)
3265 struct btrfs_fs_info *fs_info = root->fs_info;
3271 struct btrfs_key key;
3272 struct btrfs_file_extent_item *fi;
3276 int (*process_func)(struct btrfs_trans_handle *,
3277 struct btrfs_root *,
3278 u64, u64, u64, u64, u64, u64);
3281 if (btrfs_is_testing(fs_info))
3284 ref_root = btrfs_header_owner(buf);
3285 nritems = btrfs_header_nritems(buf);
3286 level = btrfs_header_level(buf);
3288 if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3292 process_func = btrfs_inc_extent_ref;
3294 process_func = btrfs_free_extent;
3297 parent = buf->start;
3301 for (i = 0; i < nritems; i++) {
3303 btrfs_item_key_to_cpu(buf, &key, i);
3304 if (key.type != BTRFS_EXTENT_DATA_KEY)
3306 fi = btrfs_item_ptr(buf, i,
3307 struct btrfs_file_extent_item);
3308 if (btrfs_file_extent_type(buf, fi) ==
3309 BTRFS_FILE_EXTENT_INLINE)
3311 bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3315 num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3316 key.offset -= btrfs_file_extent_offset(buf, fi);
3317 ret = process_func(trans, root, bytenr, num_bytes,
3318 parent, ref_root, key.objectid,
3323 bytenr = btrfs_node_blockptr(buf, i);
3324 num_bytes = fs_info->nodesize;
3325 ret = process_func(trans, root, bytenr, num_bytes,
3326 parent, ref_root, level - 1, 0);
3336 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3337 struct extent_buffer *buf, int full_backref)
3339 return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3342 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3343 struct extent_buffer *buf, int full_backref)
3345 return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3348 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3349 struct btrfs_fs_info *fs_info,
3350 struct btrfs_path *path,
3351 struct btrfs_block_group_cache *cache)
3354 struct btrfs_root *extent_root = fs_info->extent_root;
3356 struct extent_buffer *leaf;
3358 ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3365 leaf = path->nodes[0];
3366 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3367 write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3368 btrfs_mark_buffer_dirty(leaf);
3370 btrfs_release_path(path);
3375 static struct btrfs_block_group_cache *
3376 next_block_group(struct btrfs_fs_info *fs_info,
3377 struct btrfs_block_group_cache *cache)
3379 struct rb_node *node;
3381 spin_lock(&fs_info->block_group_cache_lock);
3383 /* If our block group was removed, we need a full search. */
3384 if (RB_EMPTY_NODE(&cache->cache_node)) {
3385 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3387 spin_unlock(&fs_info->block_group_cache_lock);
3388 btrfs_put_block_group(cache);
3389 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3391 node = rb_next(&cache->cache_node);
3392 btrfs_put_block_group(cache);
3394 cache = rb_entry(node, struct btrfs_block_group_cache,
3396 btrfs_get_block_group(cache);
3399 spin_unlock(&fs_info->block_group_cache_lock);
3403 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3404 struct btrfs_trans_handle *trans,
3405 struct btrfs_path *path)
3407 struct btrfs_fs_info *fs_info = block_group->fs_info;
3408 struct btrfs_root *root = fs_info->tree_root;
3409 struct inode *inode = NULL;
3410 struct extent_changeset *data_reserved = NULL;
3412 int dcs = BTRFS_DC_ERROR;
3418 * If this block group is smaller than 100 megs don't bother caching the
3421 if (block_group->key.offset < (100 * SZ_1M)) {
3422 spin_lock(&block_group->lock);
3423 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3424 spin_unlock(&block_group->lock);
3431 inode = lookup_free_space_inode(fs_info, block_group, path);
3432 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3433 ret = PTR_ERR(inode);
3434 btrfs_release_path(path);
3438 if (IS_ERR(inode)) {
3442 if (block_group->ro)
3445 ret = create_free_space_inode(fs_info, trans, block_group,
3453 * We want to set the generation to 0, that way if anything goes wrong
3454 * from here on out we know not to trust this cache when we load up next
3457 BTRFS_I(inode)->generation = 0;
3458 ret = btrfs_update_inode(trans, root, inode);
3461 * So theoretically we could recover from this, simply set the
3462 * super cache generation to 0 so we know to invalidate the
3463 * cache, but then we'd have to keep track of the block groups
3464 * that fail this way so we know we _have_ to reset this cache
3465 * before the next commit or risk reading stale cache. So to
3466 * limit our exposure to horrible edge cases lets just abort the
3467 * transaction, this only happens in really bad situations
3470 btrfs_abort_transaction(trans, ret);
3475 /* We've already setup this transaction, go ahead and exit */
3476 if (block_group->cache_generation == trans->transid &&
3477 i_size_read(inode)) {
3478 dcs = BTRFS_DC_SETUP;
3482 if (i_size_read(inode) > 0) {
3483 ret = btrfs_check_trunc_cache_free_space(fs_info,
3484 &fs_info->global_block_rsv);
3488 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3493 spin_lock(&block_group->lock);
3494 if (block_group->cached != BTRFS_CACHE_FINISHED ||
3495 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3497 * don't bother trying to write stuff out _if_
3498 * a) we're not cached,
3499 * b) we're with nospace_cache mount option,
3500 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3502 dcs = BTRFS_DC_WRITTEN;
3503 spin_unlock(&block_group->lock);
3506 spin_unlock(&block_group->lock);
3509 * We hit an ENOSPC when setting up the cache in this transaction, just
3510 * skip doing the setup, we've already cleared the cache so we're safe.
3512 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3518 * Try to preallocate enough space based on how big the block group is.
3519 * Keep in mind this has to include any pinned space which could end up
3520 * taking up quite a bit since it's not folded into the other space
3523 num_pages = div_u64(block_group->key.offset, SZ_256M);
3528 num_pages *= PAGE_SIZE;
3530 ret = btrfs_check_data_free_space(inode, &data_reserved, 0, num_pages);
3534 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3535 num_pages, num_pages,
3538 * Our cache requires contiguous chunks so that we don't modify a bunch
3539 * of metadata or split extents when writing the cache out, which means
3540 * we can enospc if we are heavily fragmented in addition to just normal
3541 * out of space conditions. So if we hit this just skip setting up any
3542 * other block groups for this transaction, maybe we'll unpin enough
3543 * space the next time around.
3546 dcs = BTRFS_DC_SETUP;
3547 else if (ret == -ENOSPC)
3548 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3553 btrfs_release_path(path);
3555 spin_lock(&block_group->lock);
3556 if (!ret && dcs == BTRFS_DC_SETUP)
3557 block_group->cache_generation = trans->transid;
3558 block_group->disk_cache_state = dcs;
3559 spin_unlock(&block_group->lock);
3561 extent_changeset_free(data_reserved);
3565 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3566 struct btrfs_fs_info *fs_info)
3568 struct btrfs_block_group_cache *cache, *tmp;
3569 struct btrfs_transaction *cur_trans = trans->transaction;
3570 struct btrfs_path *path;
3572 if (list_empty(&cur_trans->dirty_bgs) ||
3573 !btrfs_test_opt(fs_info, SPACE_CACHE))
3576 path = btrfs_alloc_path();
3580 /* Could add new block groups, use _safe just in case */
3581 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3583 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3584 cache_save_setup(cache, trans, path);
3587 btrfs_free_path(path);
3592 * transaction commit does final block group cache writeback during a
3593 * critical section where nothing is allowed to change the FS. This is
3594 * required in order for the cache to actually match the block group,
3595 * but can introduce a lot of latency into the commit.
3597 * So, btrfs_start_dirty_block_groups is here to kick off block group
3598 * cache IO. There's a chance we'll have to redo some of it if the
3599 * block group changes again during the commit, but it greatly reduces
3600 * the commit latency by getting rid of the easy block groups while
3601 * we're still allowing others to join the commit.
3603 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3605 struct btrfs_fs_info *fs_info = trans->fs_info;
3606 struct btrfs_block_group_cache *cache;
3607 struct btrfs_transaction *cur_trans = trans->transaction;
3610 struct btrfs_path *path = NULL;
3612 struct list_head *io = &cur_trans->io_bgs;
3613 int num_started = 0;
3616 spin_lock(&cur_trans->dirty_bgs_lock);
3617 if (list_empty(&cur_trans->dirty_bgs)) {
3618 spin_unlock(&cur_trans->dirty_bgs_lock);
3621 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3622 spin_unlock(&cur_trans->dirty_bgs_lock);
3626 * make sure all the block groups on our dirty list actually
3629 btrfs_create_pending_block_groups(trans);
3632 path = btrfs_alloc_path();
3638 * cache_write_mutex is here only to save us from balance or automatic
3639 * removal of empty block groups deleting this block group while we are
3640 * writing out the cache
3642 mutex_lock(&trans->transaction->cache_write_mutex);
3643 while (!list_empty(&dirty)) {
3644 cache = list_first_entry(&dirty,
3645 struct btrfs_block_group_cache,
3648 * this can happen if something re-dirties a block
3649 * group that is already under IO. Just wait for it to
3650 * finish and then do it all again
3652 if (!list_empty(&cache->io_list)) {
3653 list_del_init(&cache->io_list);
3654 btrfs_wait_cache_io(trans, cache, path);
3655 btrfs_put_block_group(cache);
3660 * btrfs_wait_cache_io uses the cache->dirty_list to decide
3661 * if it should update the cache_state. Don't delete
3662 * until after we wait.
3664 * Since we're not running in the commit critical section
3665 * we need the dirty_bgs_lock to protect from update_block_group
3667 spin_lock(&cur_trans->dirty_bgs_lock);
3668 list_del_init(&cache->dirty_list);
3669 spin_unlock(&cur_trans->dirty_bgs_lock);
3673 cache_save_setup(cache, trans, path);
3675 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3676 cache->io_ctl.inode = NULL;
3677 ret = btrfs_write_out_cache(fs_info, trans,
3679 if (ret == 0 && cache->io_ctl.inode) {
3684 * The cache_write_mutex is protecting the
3685 * io_list, also refer to the definition of
3686 * btrfs_transaction::io_bgs for more details
3688 list_add_tail(&cache->io_list, io);
3691 * if we failed to write the cache, the
3692 * generation will be bad and life goes on
3698 ret = write_one_cache_group(trans, fs_info,
3701 * Our block group might still be attached to the list
3702 * of new block groups in the transaction handle of some
3703 * other task (struct btrfs_trans_handle->new_bgs). This
3704 * means its block group item isn't yet in the extent
3705 * tree. If this happens ignore the error, as we will
3706 * try again later in the critical section of the
3707 * transaction commit.
3709 if (ret == -ENOENT) {
3711 spin_lock(&cur_trans->dirty_bgs_lock);
3712 if (list_empty(&cache->dirty_list)) {
3713 list_add_tail(&cache->dirty_list,
3714 &cur_trans->dirty_bgs);
3715 btrfs_get_block_group(cache);
3717 spin_unlock(&cur_trans->dirty_bgs_lock);
3719 btrfs_abort_transaction(trans, ret);
3723 /* if its not on the io list, we need to put the block group */
3725 btrfs_put_block_group(cache);
3731 * Avoid blocking other tasks for too long. It might even save
3732 * us from writing caches for block groups that are going to be
3735 mutex_unlock(&trans->transaction->cache_write_mutex);
3736 mutex_lock(&trans->transaction->cache_write_mutex);
3738 mutex_unlock(&trans->transaction->cache_write_mutex);
3741 * go through delayed refs for all the stuff we've just kicked off
3742 * and then loop back (just once)
3744 ret = btrfs_run_delayed_refs(trans, 0);
3745 if (!ret && loops == 0) {
3747 spin_lock(&cur_trans->dirty_bgs_lock);
3748 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3750 * dirty_bgs_lock protects us from concurrent block group
3751 * deletes too (not just cache_write_mutex).
3753 if (!list_empty(&dirty)) {
3754 spin_unlock(&cur_trans->dirty_bgs_lock);
3757 spin_unlock(&cur_trans->dirty_bgs_lock);
3758 } else if (ret < 0) {
3759 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3762 btrfs_free_path(path);
3766 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3767 struct btrfs_fs_info *fs_info)
3769 struct btrfs_block_group_cache *cache;
3770 struct btrfs_transaction *cur_trans = trans->transaction;
3773 struct btrfs_path *path;
3774 struct list_head *io = &cur_trans->io_bgs;
3775 int num_started = 0;
3777 path = btrfs_alloc_path();
3782 * Even though we are in the critical section of the transaction commit,
3783 * we can still have concurrent tasks adding elements to this
3784 * transaction's list of dirty block groups. These tasks correspond to
3785 * endio free space workers started when writeback finishes for a
3786 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3787 * allocate new block groups as a result of COWing nodes of the root
3788 * tree when updating the free space inode. The writeback for the space
3789 * caches is triggered by an earlier call to
3790 * btrfs_start_dirty_block_groups() and iterations of the following
3792 * Also we want to do the cache_save_setup first and then run the
3793 * delayed refs to make sure we have the best chance at doing this all
3796 spin_lock(&cur_trans->dirty_bgs_lock);
3797 while (!list_empty(&cur_trans->dirty_bgs)) {
3798 cache = list_first_entry(&cur_trans->dirty_bgs,
3799 struct btrfs_block_group_cache,
3803 * this can happen if cache_save_setup re-dirties a block
3804 * group that is already under IO. Just wait for it to
3805 * finish and then do it all again
3807 if (!list_empty(&cache->io_list)) {
3808 spin_unlock(&cur_trans->dirty_bgs_lock);
3809 list_del_init(&cache->io_list);
3810 btrfs_wait_cache_io(trans, cache, path);
3811 btrfs_put_block_group(cache);
3812 spin_lock(&cur_trans->dirty_bgs_lock);
3816 * don't remove from the dirty list until after we've waited
3819 list_del_init(&cache->dirty_list);
3820 spin_unlock(&cur_trans->dirty_bgs_lock);
3823 cache_save_setup(cache, trans, path);
3826 ret = btrfs_run_delayed_refs(trans,
3827 (unsigned long) -1);
3829 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3830 cache->io_ctl.inode = NULL;
3831 ret = btrfs_write_out_cache(fs_info, trans,
3833 if (ret == 0 && cache->io_ctl.inode) {
3836 list_add_tail(&cache->io_list, io);
3839 * if we failed to write the cache, the
3840 * generation will be bad and life goes on
3846 ret = write_one_cache_group(trans, fs_info,
3849 * One of the free space endio workers might have
3850 * created a new block group while updating a free space
3851 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3852 * and hasn't released its transaction handle yet, in
3853 * which case the new block group is still attached to
3854 * its transaction handle and its creation has not
3855 * finished yet (no block group item in the extent tree
3856 * yet, etc). If this is the case, wait for all free
3857 * space endio workers to finish and retry. This is a
3858 * a very rare case so no need for a more efficient and
3861 if (ret == -ENOENT) {
3862 wait_event(cur_trans->writer_wait,
3863 atomic_read(&cur_trans->num_writers) == 1);
3864 ret = write_one_cache_group(trans, fs_info,
3868 btrfs_abort_transaction(trans, ret);
3871 /* if its not on the io list, we need to put the block group */
3873 btrfs_put_block_group(cache);
3874 spin_lock(&cur_trans->dirty_bgs_lock);
3876 spin_unlock(&cur_trans->dirty_bgs_lock);
3879 * Refer to the definition of io_bgs member for details why it's safe
3880 * to use it without any locking
3882 while (!list_empty(io)) {
3883 cache = list_first_entry(io, struct btrfs_block_group_cache,
3885 list_del_init(&cache->io_list);
3886 btrfs_wait_cache_io(trans, cache, path);
3887 btrfs_put_block_group(cache);
3890 btrfs_free_path(path);
3894 int btrfs_extent_readonly(struct btrfs_fs_info *fs_info, u64 bytenr)
3896 struct btrfs_block_group_cache *block_group;
3899 block_group = btrfs_lookup_block_group(fs_info, bytenr);
3900 if (!block_group || block_group->ro)
3903 btrfs_put_block_group(block_group);
3907 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3909 struct btrfs_block_group_cache *bg;
3912 bg = btrfs_lookup_block_group(fs_info, bytenr);
3916 spin_lock(&bg->lock);
3920 atomic_inc(&bg->nocow_writers);
3921 spin_unlock(&bg->lock);
3923 /* no put on block group, done by btrfs_dec_nocow_writers */
3925 btrfs_put_block_group(bg);
3931 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3933 struct btrfs_block_group_cache *bg;
3935 bg = btrfs_lookup_block_group(fs_info, bytenr);
3937 if (atomic_dec_and_test(&bg->nocow_writers))
3938 wake_up_var(&bg->nocow_writers);
3940 * Once for our lookup and once for the lookup done by a previous call
3941 * to btrfs_inc_nocow_writers()
3943 btrfs_put_block_group(bg);
3944 btrfs_put_block_group(bg);
3947 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3949 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
3952 static const char *alloc_name(u64 flags)
3955 case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3957 case BTRFS_BLOCK_GROUP_METADATA:
3959 case BTRFS_BLOCK_GROUP_DATA:
3961 case BTRFS_BLOCK_GROUP_SYSTEM:
3965 return "invalid-combination";
3969 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
3972 struct btrfs_space_info *space_info;
3976 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
3980 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
3987 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3988 INIT_LIST_HEAD(&space_info->block_groups[i]);
3989 init_rwsem(&space_info->groups_sem);
3990 spin_lock_init(&space_info->lock);
3991 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3992 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3993 init_waitqueue_head(&space_info->wait);
3994 INIT_LIST_HEAD(&space_info->ro_bgs);
3995 INIT_LIST_HEAD(&space_info->tickets);
3996 INIT_LIST_HEAD(&space_info->priority_tickets);
3998 ret = kobject_init_and_add(&space_info->kobj, &space_info_ktype,
3999 info->space_info_kobj, "%s",
4000 alloc_name(space_info->flags));
4002 percpu_counter_destroy(&space_info->total_bytes_pinned);
4007 list_add_rcu(&space_info->list, &info->space_info);
4008 if (flags & BTRFS_BLOCK_GROUP_DATA)
4009 info->data_sinfo = space_info;
4014 static void update_space_info(struct btrfs_fs_info *info, u64 flags,
4015 u64 total_bytes, u64 bytes_used,
4017 struct btrfs_space_info **space_info)
4019 struct btrfs_space_info *found;
4022 factor = btrfs_bg_type_to_factor(flags);
4024 found = __find_space_info(info, flags);
4026 spin_lock(&found->lock);
4027 found->total_bytes += total_bytes;
4028 found->disk_total += total_bytes * factor;
4029 found->bytes_used += bytes_used;
4030 found->disk_used += bytes_used * factor;
4031 found->bytes_readonly += bytes_readonly;
4032 if (total_bytes > 0)
4034 space_info_add_new_bytes(info, found, total_bytes -
4035 bytes_used - bytes_readonly);
4036 spin_unlock(&found->lock);
4037 *space_info = found;
4040 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4042 u64 extra_flags = chunk_to_extended(flags) &
4043 BTRFS_EXTENDED_PROFILE_MASK;
4045 write_seqlock(&fs_info->profiles_lock);
4046 if (flags & BTRFS_BLOCK_GROUP_DATA)
4047 fs_info->avail_data_alloc_bits |= extra_flags;
4048 if (flags & BTRFS_BLOCK_GROUP_METADATA)
4049 fs_info->avail_metadata_alloc_bits |= extra_flags;
4050 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4051 fs_info->avail_system_alloc_bits |= extra_flags;
4052 write_sequnlock(&fs_info->profiles_lock);
4056 * returns target flags in extended format or 0 if restripe for this
4057 * chunk_type is not in progress
4059 * should be called with balance_lock held
4061 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4063 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4069 if (flags & BTRFS_BLOCK_GROUP_DATA &&
4070 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4071 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4072 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4073 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4074 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4075 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4076 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4077 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4084 * @flags: available profiles in extended format (see ctree.h)
4086 * Returns reduced profile in chunk format. If profile changing is in
4087 * progress (either running or paused) picks the target profile (if it's
4088 * already available), otherwise falls back to plain reducing.
4090 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
4092 u64 num_devices = fs_info->fs_devices->rw_devices;
4098 * see if restripe for this chunk_type is in progress, if so
4099 * try to reduce to the target profile
4101 spin_lock(&fs_info->balance_lock);
4102 target = get_restripe_target(fs_info, flags);
4104 /* pick target profile only if it's already available */
4105 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4106 spin_unlock(&fs_info->balance_lock);
4107 return extended_to_chunk(target);
4110 spin_unlock(&fs_info->balance_lock);
4112 /* First, mask out the RAID levels which aren't possible */
4113 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4114 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4115 allowed |= btrfs_raid_array[raid_type].bg_flag;
4119 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4120 allowed = BTRFS_BLOCK_GROUP_RAID6;
4121 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4122 allowed = BTRFS_BLOCK_GROUP_RAID5;
4123 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4124 allowed = BTRFS_BLOCK_GROUP_RAID10;
4125 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4126 allowed = BTRFS_BLOCK_GROUP_RAID1;
4127 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4128 allowed = BTRFS_BLOCK_GROUP_RAID0;
4130 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4132 return extended_to_chunk(flags | allowed);
4135 static u64 get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
4142 seq = read_seqbegin(&fs_info->profiles_lock);
4144 if (flags & BTRFS_BLOCK_GROUP_DATA)
4145 flags |= fs_info->avail_data_alloc_bits;
4146 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4147 flags |= fs_info->avail_system_alloc_bits;
4148 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4149 flags |= fs_info->avail_metadata_alloc_bits;
4150 } while (read_seqretry(&fs_info->profiles_lock, seq));
4152 return btrfs_reduce_alloc_profile(fs_info, flags);
4155 static u64 get_alloc_profile_by_root(struct btrfs_root *root, int data)
4157 struct btrfs_fs_info *fs_info = root->fs_info;
4162 flags = BTRFS_BLOCK_GROUP_DATA;
4163 else if (root == fs_info->chunk_root)
4164 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4166 flags = BTRFS_BLOCK_GROUP_METADATA;
4168 ret = get_alloc_profile(fs_info, flags);
4172 u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
4174 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
4177 u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
4179 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4182 u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
4184 return get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4187 static u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
4188 bool may_use_included)
4191 return s_info->bytes_used + s_info->bytes_reserved +
4192 s_info->bytes_pinned + s_info->bytes_readonly +
4193 (may_use_included ? s_info->bytes_may_use : 0);
4196 int btrfs_alloc_data_chunk_ondemand(struct btrfs_inode *inode, u64 bytes)
4198 struct btrfs_root *root = inode->root;
4199 struct btrfs_fs_info *fs_info = root->fs_info;
4200 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
4203 int need_commit = 2;
4204 int have_pinned_space;
4206 /* make sure bytes are sectorsize aligned */
4207 bytes = ALIGN(bytes, fs_info->sectorsize);
4209 if (btrfs_is_free_space_inode(inode)) {
4211 ASSERT(current->journal_info);
4215 /* make sure we have enough space to handle the data first */
4216 spin_lock(&data_sinfo->lock);
4217 used = btrfs_space_info_used(data_sinfo, true);
4219 if (used + bytes > data_sinfo->total_bytes) {
4220 struct btrfs_trans_handle *trans;
4223 * if we don't have enough free bytes in this space then we need
4224 * to alloc a new chunk.
4226 if (!data_sinfo->full) {
4229 data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4230 spin_unlock(&data_sinfo->lock);
4232 alloc_target = btrfs_data_alloc_profile(fs_info);
4234 * It is ugly that we don't call nolock join
4235 * transaction for the free space inode case here.
4236 * But it is safe because we only do the data space
4237 * reservation for the free space cache in the
4238 * transaction context, the common join transaction
4239 * just increase the counter of the current transaction
4240 * handler, doesn't try to acquire the trans_lock of
4243 trans = btrfs_join_transaction(root);
4245 return PTR_ERR(trans);
4247 ret = do_chunk_alloc(trans, alloc_target,
4248 CHUNK_ALLOC_NO_FORCE);
4249 btrfs_end_transaction(trans);
4254 have_pinned_space = 1;
4263 * If we don't have enough pinned space to deal with this
4264 * allocation, and no removed chunk in current transaction,
4265 * don't bother committing the transaction.
4267 have_pinned_space = __percpu_counter_compare(
4268 &data_sinfo->total_bytes_pinned,
4269 used + bytes - data_sinfo->total_bytes,
4270 BTRFS_TOTAL_BYTES_PINNED_BATCH);
4271 spin_unlock(&data_sinfo->lock);
4273 /* commit the current transaction and try again */
4278 if (need_commit > 0) {
4279 btrfs_start_delalloc_roots(fs_info, -1);
4280 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0,
4284 trans = btrfs_join_transaction(root);
4286 return PTR_ERR(trans);
4287 if (have_pinned_space >= 0 ||
4288 test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4289 &trans->transaction->flags) ||
4291 ret = btrfs_commit_transaction(trans);
4295 * The cleaner kthread might still be doing iput
4296 * operations. Wait for it to finish so that
4297 * more space is released.
4299 mutex_lock(&fs_info->cleaner_delayed_iput_mutex);
4300 mutex_unlock(&fs_info->cleaner_delayed_iput_mutex);
4303 btrfs_end_transaction(trans);
4307 trace_btrfs_space_reservation(fs_info,
4308 "space_info:enospc",
4309 data_sinfo->flags, bytes, 1);
4312 data_sinfo->bytes_may_use += bytes;
4313 trace_btrfs_space_reservation(fs_info, "space_info",
4314 data_sinfo->flags, bytes, 1);
4315 spin_unlock(&data_sinfo->lock);
4320 int btrfs_check_data_free_space(struct inode *inode,
4321 struct extent_changeset **reserved, u64 start, u64 len)
4323 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4326 /* align the range */
4327 len = round_up(start + len, fs_info->sectorsize) -
4328 round_down(start, fs_info->sectorsize);
4329 start = round_down(start, fs_info->sectorsize);
4331 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode), len);
4335 /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4336 ret = btrfs_qgroup_reserve_data(inode, reserved, start, len);
4338 btrfs_free_reserved_data_space_noquota(inode, start, len);
4345 * Called if we need to clear a data reservation for this inode
4346 * Normally in a error case.
4348 * This one will *NOT* use accurate qgroup reserved space API, just for case
4349 * which we can't sleep and is sure it won't affect qgroup reserved space.
4350 * Like clear_bit_hook().
4352 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4355 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4356 struct btrfs_space_info *data_sinfo;
4358 /* Make sure the range is aligned to sectorsize */
4359 len = round_up(start + len, fs_info->sectorsize) -
4360 round_down(start, fs_info->sectorsize);
4361 start = round_down(start, fs_info->sectorsize);
4363 data_sinfo = fs_info->data_sinfo;
4364 spin_lock(&data_sinfo->lock);
4365 if (WARN_ON(data_sinfo->bytes_may_use < len))
4366 data_sinfo->bytes_may_use = 0;
4368 data_sinfo->bytes_may_use -= len;
4369 trace_btrfs_space_reservation(fs_info, "space_info",
4370 data_sinfo->flags, len, 0);
4371 spin_unlock(&data_sinfo->lock);
4375 * Called if we need to clear a data reservation for this inode
4376 * Normally in a error case.
4378 * This one will handle the per-inode data rsv map for accurate reserved
4381 void btrfs_free_reserved_data_space(struct inode *inode,
4382 struct extent_changeset *reserved, u64 start, u64 len)
4384 struct btrfs_root *root = BTRFS_I(inode)->root;
4386 /* Make sure the range is aligned to sectorsize */
4387 len = round_up(start + len, root->fs_info->sectorsize) -
4388 round_down(start, root->fs_info->sectorsize);
4389 start = round_down(start, root->fs_info->sectorsize);
4391 btrfs_free_reserved_data_space_noquota(inode, start, len);
4392 btrfs_qgroup_free_data(inode, reserved, start, len);
4395 static void force_metadata_allocation(struct btrfs_fs_info *info)
4397 struct list_head *head = &info->space_info;
4398 struct btrfs_space_info *found;
4401 list_for_each_entry_rcu(found, head, list) {
4402 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4403 found->force_alloc = CHUNK_ALLOC_FORCE;
4408 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4410 return (global->size << 1);
4413 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
4414 struct btrfs_space_info *sinfo, int force)
4416 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4417 u64 bytes_used = btrfs_space_info_used(sinfo, false);
4420 if (force == CHUNK_ALLOC_FORCE)
4424 * We need to take into account the global rsv because for all intents
4425 * and purposes it's used space. Don't worry about locking the
4426 * global_rsv, it doesn't change except when the transaction commits.
4428 if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4429 bytes_used += calc_global_rsv_need_space(global_rsv);
4432 * in limited mode, we want to have some free space up to
4433 * about 1% of the FS size.
4435 if (force == CHUNK_ALLOC_LIMITED) {
4436 thresh = btrfs_super_total_bytes(fs_info->super_copy);
4437 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4439 if (sinfo->total_bytes - bytes_used < thresh)
4443 if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
4448 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4452 if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4453 BTRFS_BLOCK_GROUP_RAID0 |
4454 BTRFS_BLOCK_GROUP_RAID5 |
4455 BTRFS_BLOCK_GROUP_RAID6))
4456 num_dev = fs_info->fs_devices->rw_devices;
4457 else if (type & BTRFS_BLOCK_GROUP_RAID1)
4460 num_dev = 1; /* DUP or single */
4466 * If @is_allocation is true, reserve space in the system space info necessary
4467 * for allocating a chunk, otherwise if it's false, reserve space necessary for
4470 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4472 struct btrfs_fs_info *fs_info = trans->fs_info;
4473 struct btrfs_space_info *info;
4480 * Needed because we can end up allocating a system chunk and for an
4481 * atomic and race free space reservation in the chunk block reserve.
4483 lockdep_assert_held(&fs_info->chunk_mutex);
4485 info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4486 spin_lock(&info->lock);
4487 left = info->total_bytes - btrfs_space_info_used(info, true);
4488 spin_unlock(&info->lock);
4490 num_devs = get_profile_num_devs(fs_info, type);
4492 /* num_devs device items to update and 1 chunk item to add or remove */
4493 thresh = btrfs_calc_trunc_metadata_size(fs_info, num_devs) +
4494 btrfs_calc_trans_metadata_size(fs_info, 1);
4496 if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4497 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4498 left, thresh, type);
4499 dump_space_info(fs_info, info, 0, 0);
4502 if (left < thresh) {
4503 u64 flags = btrfs_system_alloc_profile(fs_info);
4506 * Ignore failure to create system chunk. We might end up not
4507 * needing it, as we might not need to COW all nodes/leafs from
4508 * the paths we visit in the chunk tree (they were already COWed
4509 * or created in the current transaction for example).
4511 ret = btrfs_alloc_chunk(trans, flags);
4515 ret = btrfs_block_rsv_add(fs_info->chunk_root,
4516 &fs_info->chunk_block_rsv,
4517 thresh, BTRFS_RESERVE_NO_FLUSH);
4519 trans->chunk_bytes_reserved += thresh;
4524 * If force is CHUNK_ALLOC_FORCE:
4525 * - return 1 if it successfully allocates a chunk,
4526 * - return errors including -ENOSPC otherwise.
4527 * If force is NOT CHUNK_ALLOC_FORCE:
4528 * - return 0 if it doesn't need to allocate a new chunk,
4529 * - return 1 if it successfully allocates a chunk,
4530 * - return errors including -ENOSPC otherwise.
4532 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4535 struct btrfs_fs_info *fs_info = trans->fs_info;
4536 struct btrfs_space_info *space_info;
4537 bool wait_for_alloc = false;
4538 bool should_alloc = false;
4541 /* Don't re-enter if we're already allocating a chunk */
4542 if (trans->allocating_chunk)
4545 space_info = __find_space_info(fs_info, flags);
4549 spin_lock(&space_info->lock);
4550 if (force < space_info->force_alloc)
4551 force = space_info->force_alloc;
4552 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4553 if (space_info->full) {
4554 /* No more free physical space */
4559 spin_unlock(&space_info->lock);
4561 } else if (!should_alloc) {
4562 spin_unlock(&space_info->lock);
4564 } else if (space_info->chunk_alloc) {
4566 * Someone is already allocating, so we need to block
4567 * until this someone is finished and then loop to
4568 * recheck if we should continue with our allocation
4571 wait_for_alloc = true;
4572 spin_unlock(&space_info->lock);
4573 mutex_lock(&fs_info->chunk_mutex);
4574 mutex_unlock(&fs_info->chunk_mutex);
4576 /* Proceed with allocation */
4577 space_info->chunk_alloc = 1;
4578 wait_for_alloc = false;
4579 spin_unlock(&space_info->lock);
4583 } while (wait_for_alloc);
4585 mutex_lock(&fs_info->chunk_mutex);
4586 trans->allocating_chunk = true;
4589 * If we have mixed data/metadata chunks we want to make sure we keep
4590 * allocating mixed chunks instead of individual chunks.
4592 if (btrfs_mixed_space_info(space_info))
4593 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4596 * if we're doing a data chunk, go ahead and make sure that
4597 * we keep a reasonable number of metadata chunks allocated in the
4600 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4601 fs_info->data_chunk_allocations++;
4602 if (!(fs_info->data_chunk_allocations %
4603 fs_info->metadata_ratio))
4604 force_metadata_allocation(fs_info);
4608 * Check if we have enough space in SYSTEM chunk because we may need
4609 * to update devices.
4611 check_system_chunk(trans, flags);
4613 ret = btrfs_alloc_chunk(trans, flags);
4614 trans->allocating_chunk = false;
4616 spin_lock(&space_info->lock);
4619 space_info->full = 1;
4626 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4628 space_info->chunk_alloc = 0;
4629 spin_unlock(&space_info->lock);
4630 mutex_unlock(&fs_info->chunk_mutex);
4632 * When we allocate a new chunk we reserve space in the chunk block
4633 * reserve to make sure we can COW nodes/leafs in the chunk tree or
4634 * add new nodes/leafs to it if we end up needing to do it when
4635 * inserting the chunk item and updating device items as part of the
4636 * second phase of chunk allocation, performed by
4637 * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4638 * large number of new block groups to create in our transaction
4639 * handle's new_bgs list to avoid exhausting the chunk block reserve
4640 * in extreme cases - like having a single transaction create many new
4641 * block groups when starting to write out the free space caches of all
4642 * the block groups that were made dirty during the lifetime of the
4645 if (trans->can_flush_pending_bgs &&
4646 trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4647 btrfs_create_pending_block_groups(trans);
4648 btrfs_trans_release_chunk_metadata(trans);
4653 static int can_overcommit(struct btrfs_fs_info *fs_info,
4654 struct btrfs_space_info *space_info, u64 bytes,
4655 enum btrfs_reserve_flush_enum flush,
4658 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4665 /* Don't overcommit when in mixed mode. */
4666 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4670 profile = btrfs_system_alloc_profile(fs_info);
4672 profile = btrfs_metadata_alloc_profile(fs_info);
4674 used = btrfs_space_info_used(space_info, false);
4677 * We only want to allow over committing if we have lots of actual space
4678 * free, but if we don't have enough space to handle the global reserve
4679 * space then we could end up having a real enospc problem when trying
4680 * to allocate a chunk or some other such important allocation.
4682 spin_lock(&global_rsv->lock);
4683 space_size = calc_global_rsv_need_space(global_rsv);
4684 spin_unlock(&global_rsv->lock);
4685 if (used + space_size >= space_info->total_bytes)
4688 used += space_info->bytes_may_use;
4690 avail = atomic64_read(&fs_info->free_chunk_space);
4693 * If we have dup, raid1 or raid10 then only half of the free
4694 * space is actually useable. For raid56, the space info used
4695 * doesn't include the parity drive, so we don't have to
4698 factor = btrfs_bg_type_to_factor(profile);
4699 avail = div_u64(avail, factor);
4702 * If we aren't flushing all things, let us overcommit up to
4703 * 1/2th of the space. If we can flush, don't let us overcommit
4704 * too much, let it overcommit up to 1/8 of the space.
4706 if (flush == BTRFS_RESERVE_FLUSH_ALL)
4711 if (used + bytes < space_info->total_bytes + avail)
4716 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
4717 unsigned long nr_pages, int nr_items)
4719 struct super_block *sb = fs_info->sb;
4721 if (down_read_trylock(&sb->s_umount)) {
4722 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4723 up_read(&sb->s_umount);
4726 * We needn't worry the filesystem going from r/w to r/o though
4727 * we don't acquire ->s_umount mutex, because the filesystem
4728 * should guarantee the delalloc inodes list be empty after
4729 * the filesystem is readonly(all dirty pages are written to
4732 btrfs_start_delalloc_roots(fs_info, nr_items);
4733 if (!current->journal_info)
4734 btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
4738 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
4744 bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
4745 nr = div64_u64(to_reclaim, bytes);
4751 #define EXTENT_SIZE_PER_ITEM SZ_256K
4754 * shrink metadata reservation for delalloc
4756 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
4757 u64 orig, bool wait_ordered)
4759 struct btrfs_space_info *space_info;
4760 struct btrfs_trans_handle *trans;
4765 unsigned long nr_pages;
4768 /* Calc the number of the pages we need flush for space reservation */
4769 items = calc_reclaim_items_nr(fs_info, to_reclaim);
4770 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4772 trans = (struct btrfs_trans_handle *)current->journal_info;
4773 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4775 delalloc_bytes = percpu_counter_sum_positive(
4776 &fs_info->delalloc_bytes);
4777 if (delalloc_bytes == 0) {
4781 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4786 while (delalloc_bytes && loops < 3) {
4787 max_reclaim = min(delalloc_bytes, to_reclaim);
4788 nr_pages = max_reclaim >> PAGE_SHIFT;
4789 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
4791 * We need to wait for the async pages to actually start before
4794 max_reclaim = atomic_read(&fs_info->async_delalloc_pages);
4798 if (max_reclaim <= nr_pages)
4801 max_reclaim -= nr_pages;
4803 wait_event(fs_info->async_submit_wait,
4804 atomic_read(&fs_info->async_delalloc_pages) <=
4807 spin_lock(&space_info->lock);
4808 if (list_empty(&space_info->tickets) &&
4809 list_empty(&space_info->priority_tickets)) {
4810 spin_unlock(&space_info->lock);
4813 spin_unlock(&space_info->lock);
4816 if (wait_ordered && !trans) {
4817 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
4819 time_left = schedule_timeout_killable(1);
4823 delalloc_bytes = percpu_counter_sum_positive(
4824 &fs_info->delalloc_bytes);
4828 struct reserve_ticket {
4831 struct list_head list;
4832 wait_queue_head_t wait;
4836 * maybe_commit_transaction - possibly commit the transaction if its ok to
4837 * @root - the root we're allocating for
4838 * @bytes - the number of bytes we want to reserve
4839 * @force - force the commit
4841 * This will check to make sure that committing the transaction will actually
4842 * get us somewhere and then commit the transaction if it does. Otherwise it
4843 * will return -ENOSPC.
4845 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
4846 struct btrfs_space_info *space_info)
4848 struct reserve_ticket *ticket = NULL;
4849 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
4850 struct btrfs_trans_handle *trans;
4853 trans = (struct btrfs_trans_handle *)current->journal_info;
4857 spin_lock(&space_info->lock);
4858 if (!list_empty(&space_info->priority_tickets))
4859 ticket = list_first_entry(&space_info->priority_tickets,
4860 struct reserve_ticket, list);
4861 else if (!list_empty(&space_info->tickets))
4862 ticket = list_first_entry(&space_info->tickets,
4863 struct reserve_ticket, list);
4864 bytes = (ticket) ? ticket->bytes : 0;
4865 spin_unlock(&space_info->lock);
4870 /* See if there is enough pinned space to make this reservation */
4871 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4873 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
4877 * See if there is some space in the delayed insertion reservation for
4880 if (space_info != delayed_rsv->space_info)
4883 spin_lock(&delayed_rsv->lock);
4884 if (delayed_rsv->size > bytes)
4887 bytes -= delayed_rsv->size;
4888 spin_unlock(&delayed_rsv->lock);
4890 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
4892 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) {
4897 trans = btrfs_join_transaction(fs_info->extent_root);
4901 return btrfs_commit_transaction(trans);
4905 * Try to flush some data based on policy set by @state. This is only advisory
4906 * and may fail for various reasons. The caller is supposed to examine the
4907 * state of @space_info to detect the outcome.
4909 static void flush_space(struct btrfs_fs_info *fs_info,
4910 struct btrfs_space_info *space_info, u64 num_bytes,
4913 struct btrfs_root *root = fs_info->extent_root;
4914 struct btrfs_trans_handle *trans;
4919 case FLUSH_DELAYED_ITEMS_NR:
4920 case FLUSH_DELAYED_ITEMS:
4921 if (state == FLUSH_DELAYED_ITEMS_NR)
4922 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
4926 trans = btrfs_join_transaction(root);
4927 if (IS_ERR(trans)) {
4928 ret = PTR_ERR(trans);
4931 ret = btrfs_run_delayed_items_nr(trans, nr);
4932 btrfs_end_transaction(trans);
4934 case FLUSH_DELALLOC:
4935 case FLUSH_DELALLOC_WAIT:
4936 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
4937 state == FLUSH_DELALLOC_WAIT);
4940 trans = btrfs_join_transaction(root);
4941 if (IS_ERR(trans)) {
4942 ret = PTR_ERR(trans);
4945 ret = do_chunk_alloc(trans,
4946 btrfs_metadata_alloc_profile(fs_info),
4947 CHUNK_ALLOC_NO_FORCE);
4948 btrfs_end_transaction(trans);
4949 if (ret > 0 || ret == -ENOSPC)
4953 ret = may_commit_transaction(fs_info, space_info);
4960 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
4966 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
4967 struct btrfs_space_info *space_info,
4970 struct reserve_ticket *ticket;
4975 list_for_each_entry(ticket, &space_info->tickets, list)
4976 to_reclaim += ticket->bytes;
4977 list_for_each_entry(ticket, &space_info->priority_tickets, list)
4978 to_reclaim += ticket->bytes;
4982 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4983 if (can_overcommit(fs_info, space_info, to_reclaim,
4984 BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4987 used = btrfs_space_info_used(space_info, true);
4989 if (can_overcommit(fs_info, space_info, SZ_1M,
4990 BTRFS_RESERVE_FLUSH_ALL, system_chunk))
4991 expected = div_factor_fine(space_info->total_bytes, 95);
4993 expected = div_factor_fine(space_info->total_bytes, 90);
4995 if (used > expected)
4996 to_reclaim = used - expected;
4999 to_reclaim = min(to_reclaim, space_info->bytes_may_use +
5000 space_info->bytes_reserved);
5004 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
5005 struct btrfs_space_info *space_info,
5006 u64 used, bool system_chunk)
5008 u64 thresh = div_factor_fine(space_info->total_bytes, 98);
5010 /* If we're just plain full then async reclaim just slows us down. */
5011 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
5014 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5018 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
5019 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
5022 static void wake_all_tickets(struct list_head *head)
5024 struct reserve_ticket *ticket;
5026 while (!list_empty(head)) {
5027 ticket = list_first_entry(head, struct reserve_ticket, list);
5028 list_del_init(&ticket->list);
5029 ticket->error = -ENOSPC;
5030 wake_up(&ticket->wait);
5035 * This is for normal flushers, we can wait all goddamned day if we want to. We
5036 * will loop and continuously try to flush as long as we are making progress.
5037 * We count progress as clearing off tickets each time we have to loop.
5039 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
5041 struct btrfs_fs_info *fs_info;
5042 struct btrfs_space_info *space_info;
5045 int commit_cycles = 0;
5046 u64 last_tickets_id;
5048 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
5049 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5051 spin_lock(&space_info->lock);
5052 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5055 space_info->flush = 0;
5056 spin_unlock(&space_info->lock);
5059 last_tickets_id = space_info->tickets_id;
5060 spin_unlock(&space_info->lock);
5062 flush_state = FLUSH_DELAYED_ITEMS_NR;
5064 flush_space(fs_info, space_info, to_reclaim, flush_state);
5065 spin_lock(&space_info->lock);
5066 if (list_empty(&space_info->tickets)) {
5067 space_info->flush = 0;
5068 spin_unlock(&space_info->lock);
5071 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
5074 if (last_tickets_id == space_info->tickets_id) {
5077 last_tickets_id = space_info->tickets_id;
5078 flush_state = FLUSH_DELAYED_ITEMS_NR;
5083 if (flush_state > COMMIT_TRANS) {
5085 if (commit_cycles > 2) {
5086 wake_all_tickets(&space_info->tickets);
5087 space_info->flush = 0;
5089 flush_state = FLUSH_DELAYED_ITEMS_NR;
5092 spin_unlock(&space_info->lock);
5093 } while (flush_state <= COMMIT_TRANS);
5096 void btrfs_init_async_reclaim_work(struct work_struct *work)
5098 INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5101 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5102 struct btrfs_space_info *space_info,
5103 struct reserve_ticket *ticket)
5106 int flush_state = FLUSH_DELAYED_ITEMS_NR;
5108 spin_lock(&space_info->lock);
5109 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
5112 spin_unlock(&space_info->lock);
5115 spin_unlock(&space_info->lock);
5118 flush_space(fs_info, space_info, to_reclaim, flush_state);
5120 spin_lock(&space_info->lock);
5121 if (ticket->bytes == 0) {
5122 spin_unlock(&space_info->lock);
5125 spin_unlock(&space_info->lock);
5128 * Priority flushers can't wait on delalloc without
5131 if (flush_state == FLUSH_DELALLOC ||
5132 flush_state == FLUSH_DELALLOC_WAIT)
5133 flush_state = ALLOC_CHUNK;
5134 } while (flush_state < COMMIT_TRANS);
5137 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5138 struct btrfs_space_info *space_info,
5139 struct reserve_ticket *ticket, u64 orig_bytes)
5145 spin_lock(&space_info->lock);
5146 while (ticket->bytes > 0 && ticket->error == 0) {
5147 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5152 spin_unlock(&space_info->lock);
5156 finish_wait(&ticket->wait, &wait);
5157 spin_lock(&space_info->lock);
5160 ret = ticket->error;
5161 if (!list_empty(&ticket->list))
5162 list_del_init(&ticket->list);
5163 if (ticket->bytes && ticket->bytes < orig_bytes) {
5164 u64 num_bytes = orig_bytes - ticket->bytes;
5165 space_info->bytes_may_use -= num_bytes;
5166 trace_btrfs_space_reservation(fs_info, "space_info",
5167 space_info->flags, num_bytes, 0);
5169 spin_unlock(&space_info->lock);
5175 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5176 * @root - the root we're allocating for
5177 * @space_info - the space info we want to allocate from
5178 * @orig_bytes - the number of bytes we want
5179 * @flush - whether or not we can flush to make our reservation
5181 * This will reserve orig_bytes number of bytes from the space info associated
5182 * with the block_rsv. If there is not enough space it will make an attempt to
5183 * flush out space to make room. It will do this by flushing delalloc if
5184 * possible or committing the transaction. If flush is 0 then no attempts to
5185 * regain reservations will be made and this will fail if there is not enough
5188 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
5189 struct btrfs_space_info *space_info,
5191 enum btrfs_reserve_flush_enum flush,
5194 struct reserve_ticket ticket;
5199 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5201 spin_lock(&space_info->lock);
5203 used = btrfs_space_info_used(space_info, true);
5206 * If we have enough space then hooray, make our reservation and carry
5207 * on. If not see if we can overcommit, and if we can, hooray carry on.
5208 * If not things get more complicated.
5210 if (used + orig_bytes <= space_info->total_bytes) {
5211 space_info->bytes_may_use += orig_bytes;
5212 trace_btrfs_space_reservation(fs_info, "space_info",
5213 space_info->flags, orig_bytes, 1);
5215 } else if (can_overcommit(fs_info, space_info, orig_bytes, flush,
5217 space_info->bytes_may_use += orig_bytes;
5218 trace_btrfs_space_reservation(fs_info, "space_info",
5219 space_info->flags, orig_bytes, 1);
5224 * If we couldn't make a reservation then setup our reservation ticket
5225 * and kick the async worker if it's not already running.
5227 * If we are a priority flusher then we just need to add our ticket to
5228 * the list and we will do our own flushing further down.
5230 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5231 ticket.bytes = orig_bytes;
5233 init_waitqueue_head(&ticket.wait);
5234 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5235 list_add_tail(&ticket.list, &space_info->tickets);
5236 if (!space_info->flush) {
5237 space_info->flush = 1;
5238 trace_btrfs_trigger_flush(fs_info,
5242 queue_work(system_unbound_wq,
5243 &fs_info->async_reclaim_work);
5246 list_add_tail(&ticket.list,
5247 &space_info->priority_tickets);
5249 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5252 * We will do the space reservation dance during log replay,
5253 * which means we won't have fs_info->fs_root set, so don't do
5254 * the async reclaim as we will panic.
5256 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
5257 need_do_async_reclaim(fs_info, space_info,
5258 used, system_chunk) &&
5259 !work_busy(&fs_info->async_reclaim_work)) {
5260 trace_btrfs_trigger_flush(fs_info, space_info->flags,
5261 orig_bytes, flush, "preempt");
5262 queue_work(system_unbound_wq,
5263 &fs_info->async_reclaim_work);
5266 spin_unlock(&space_info->lock);
5267 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5270 if (flush == BTRFS_RESERVE_FLUSH_ALL)
5271 return wait_reserve_ticket(fs_info, space_info, &ticket,
5275 priority_reclaim_metadata_space(fs_info, space_info, &ticket);
5276 spin_lock(&space_info->lock);
5278 if (ticket.bytes < orig_bytes) {
5279 u64 num_bytes = orig_bytes - ticket.bytes;
5280 space_info->bytes_may_use -= num_bytes;
5281 trace_btrfs_space_reservation(fs_info, "space_info",
5286 list_del_init(&ticket.list);
5289 spin_unlock(&space_info->lock);
5290 ASSERT(list_empty(&ticket.list));
5295 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5296 * @root - the root we're allocating for
5297 * @block_rsv - the block_rsv we're allocating for
5298 * @orig_bytes - the number of bytes we want
5299 * @flush - whether or not we can flush to make our reservation
5301 * This will reserve orgi_bytes number of bytes from the space info associated
5302 * with the block_rsv. If there is not enough space it will make an attempt to
5303 * flush out space to make room. It will do this by flushing delalloc if
5304 * possible or committing the transaction. If flush is 0 then no attempts to
5305 * regain reservations will be made and this will fail if there is not enough
5308 static int reserve_metadata_bytes(struct btrfs_root *root,
5309 struct btrfs_block_rsv *block_rsv,
5311 enum btrfs_reserve_flush_enum flush)
5313 struct btrfs_fs_info *fs_info = root->fs_info;
5314 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5316 bool system_chunk = (root == fs_info->chunk_root);
5318 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
5319 orig_bytes, flush, system_chunk);
5320 if (ret == -ENOSPC &&
5321 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5322 if (block_rsv != global_rsv &&
5323 !block_rsv_use_bytes(global_rsv, orig_bytes))
5326 if (ret == -ENOSPC) {
5327 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
5328 block_rsv->space_info->flags,
5331 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
5332 dump_space_info(fs_info, block_rsv->space_info,
5338 static struct btrfs_block_rsv *get_block_rsv(
5339 const struct btrfs_trans_handle *trans,
5340 const struct btrfs_root *root)
5342 struct btrfs_fs_info *fs_info = root->fs_info;
5343 struct btrfs_block_rsv *block_rsv = NULL;
5345 if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5346 (root == fs_info->csum_root && trans->adding_csums) ||
5347 (root == fs_info->uuid_root))
5348 block_rsv = trans->block_rsv;
5351 block_rsv = root->block_rsv;
5354 block_rsv = &fs_info->empty_block_rsv;
5359 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5363 spin_lock(&block_rsv->lock);
5364 if (block_rsv->reserved >= num_bytes) {
5365 block_rsv->reserved -= num_bytes;
5366 if (block_rsv->reserved < block_rsv->size)
5367 block_rsv->full = 0;
5370 spin_unlock(&block_rsv->lock);
5374 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5375 u64 num_bytes, bool update_size)
5377 spin_lock(&block_rsv->lock);
5378 block_rsv->reserved += num_bytes;
5380 block_rsv->size += num_bytes;
5381 else if (block_rsv->reserved >= block_rsv->size)
5382 block_rsv->full = 1;
5383 spin_unlock(&block_rsv->lock);
5386 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5387 struct btrfs_block_rsv *dest, u64 num_bytes,
5390 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5393 if (global_rsv->space_info != dest->space_info)
5396 spin_lock(&global_rsv->lock);
5397 min_bytes = div_factor(global_rsv->size, min_factor);
5398 if (global_rsv->reserved < min_bytes + num_bytes) {
5399 spin_unlock(&global_rsv->lock);
5402 global_rsv->reserved -= num_bytes;
5403 if (global_rsv->reserved < global_rsv->size)
5404 global_rsv->full = 0;
5405 spin_unlock(&global_rsv->lock);
5407 block_rsv_add_bytes(dest, num_bytes, true);
5412 * This is for space we already have accounted in space_info->bytes_may_use, so
5413 * basically when we're returning space from block_rsv's.
5415 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5416 struct btrfs_space_info *space_info,
5419 struct reserve_ticket *ticket;
5420 struct list_head *head;
5422 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5423 bool check_overcommit = false;
5425 spin_lock(&space_info->lock);
5426 head = &space_info->priority_tickets;
5429 * If we are over our limit then we need to check and see if we can
5430 * overcommit, and if we can't then we just need to free up our space
5431 * and not satisfy any requests.
5433 used = btrfs_space_info_used(space_info, true);
5434 if (used - num_bytes >= space_info->total_bytes)
5435 check_overcommit = true;
5437 while (!list_empty(head) && num_bytes) {
5438 ticket = list_first_entry(head, struct reserve_ticket,
5441 * We use 0 bytes because this space is already reserved, so
5442 * adding the ticket space would be a double count.
5444 if (check_overcommit &&
5445 !can_overcommit(fs_info, space_info, 0, flush, false))
5447 if (num_bytes >= ticket->bytes) {
5448 list_del_init(&ticket->list);
5449 num_bytes -= ticket->bytes;
5451 space_info->tickets_id++;
5452 wake_up(&ticket->wait);
5454 ticket->bytes -= num_bytes;
5459 if (num_bytes && head == &space_info->priority_tickets) {
5460 head = &space_info->tickets;
5461 flush = BTRFS_RESERVE_FLUSH_ALL;
5464 space_info->bytes_may_use -= num_bytes;
5465 trace_btrfs_space_reservation(fs_info, "space_info",
5466 space_info->flags, num_bytes, 0);
5467 spin_unlock(&space_info->lock);
5471 * This is for newly allocated space that isn't accounted in
5472 * space_info->bytes_may_use yet. So if we allocate a chunk or unpin an extent
5473 * we use this helper.
5475 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5476 struct btrfs_space_info *space_info,
5479 struct reserve_ticket *ticket;
5480 struct list_head *head = &space_info->priority_tickets;
5483 while (!list_empty(head) && num_bytes) {
5484 ticket = list_first_entry(head, struct reserve_ticket,
5486 if (num_bytes >= ticket->bytes) {
5487 trace_btrfs_space_reservation(fs_info, "space_info",
5490 list_del_init(&ticket->list);
5491 num_bytes -= ticket->bytes;
5492 space_info->bytes_may_use += ticket->bytes;
5494 space_info->tickets_id++;
5495 wake_up(&ticket->wait);
5497 trace_btrfs_space_reservation(fs_info, "space_info",
5500 space_info->bytes_may_use += num_bytes;
5501 ticket->bytes -= num_bytes;
5506 if (num_bytes && head == &space_info->priority_tickets) {
5507 head = &space_info->tickets;
5512 static u64 block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5513 struct btrfs_block_rsv *block_rsv,
5514 struct btrfs_block_rsv *dest, u64 num_bytes,
5515 u64 *qgroup_to_release_ret)
5517 struct btrfs_space_info *space_info = block_rsv->space_info;
5518 u64 qgroup_to_release = 0;
5521 spin_lock(&block_rsv->lock);
5522 if (num_bytes == (u64)-1) {
5523 num_bytes = block_rsv->size;
5524 qgroup_to_release = block_rsv->qgroup_rsv_size;
5526 block_rsv->size -= num_bytes;
5527 if (block_rsv->reserved >= block_rsv->size) {
5528 num_bytes = block_rsv->reserved - block_rsv->size;
5529 block_rsv->reserved = block_rsv->size;
5530 block_rsv->full = 1;
5534 if (block_rsv->qgroup_rsv_reserved >= block_rsv->qgroup_rsv_size) {
5535 qgroup_to_release = block_rsv->qgroup_rsv_reserved -
5536 block_rsv->qgroup_rsv_size;
5537 block_rsv->qgroup_rsv_reserved = block_rsv->qgroup_rsv_size;
5539 qgroup_to_release = 0;
5541 spin_unlock(&block_rsv->lock);
5544 if (num_bytes > 0) {
5546 spin_lock(&dest->lock);
5550 bytes_to_add = dest->size - dest->reserved;
5551 bytes_to_add = min(num_bytes, bytes_to_add);
5552 dest->reserved += bytes_to_add;
5553 if (dest->reserved >= dest->size)
5555 num_bytes -= bytes_to_add;
5557 spin_unlock(&dest->lock);
5560 space_info_add_old_bytes(fs_info, space_info,
5563 if (qgroup_to_release_ret)
5564 *qgroup_to_release_ret = qgroup_to_release;
5568 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5569 struct btrfs_block_rsv *dst, u64 num_bytes,
5574 ret = block_rsv_use_bytes(src, num_bytes);
5578 block_rsv_add_bytes(dst, num_bytes, update_size);
5582 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5584 memset(rsv, 0, sizeof(*rsv));
5585 spin_lock_init(&rsv->lock);
5589 void btrfs_init_metadata_block_rsv(struct btrfs_fs_info *fs_info,
5590 struct btrfs_block_rsv *rsv,
5591 unsigned short type)
5593 btrfs_init_block_rsv(rsv, type);
5594 rsv->space_info = __find_space_info(fs_info,
5595 BTRFS_BLOCK_GROUP_METADATA);
5598 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_fs_info *fs_info,
5599 unsigned short type)
5601 struct btrfs_block_rsv *block_rsv;
5603 block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5607 btrfs_init_metadata_block_rsv(fs_info, block_rsv, type);
5611 void btrfs_free_block_rsv(struct btrfs_fs_info *fs_info,
5612 struct btrfs_block_rsv *rsv)
5616 btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5620 int btrfs_block_rsv_add(struct btrfs_root *root,
5621 struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5622 enum btrfs_reserve_flush_enum flush)
5629 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5631 block_rsv_add_bytes(block_rsv, num_bytes, true);
5636 int btrfs_block_rsv_check(struct btrfs_block_rsv *block_rsv, int min_factor)
5644 spin_lock(&block_rsv->lock);
5645 num_bytes = div_factor(block_rsv->size, min_factor);
5646 if (block_rsv->reserved >= num_bytes)
5648 spin_unlock(&block_rsv->lock);
5653 int btrfs_block_rsv_refill(struct btrfs_root *root,
5654 struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5655 enum btrfs_reserve_flush_enum flush)
5663 spin_lock(&block_rsv->lock);
5664 num_bytes = min_reserved;
5665 if (block_rsv->reserved >= num_bytes)
5668 num_bytes -= block_rsv->reserved;
5669 spin_unlock(&block_rsv->lock);
5674 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5676 block_rsv_add_bytes(block_rsv, num_bytes, false);
5684 * btrfs_inode_rsv_refill - refill the inode block rsv.
5685 * @inode - the inode we are refilling.
5686 * @flush - the flusing restriction.
5688 * Essentially the same as btrfs_block_rsv_refill, except it uses the
5689 * block_rsv->size as the minimum size. We'll either refill the missing amount
5690 * or return if we already have enough space. This will also handle the resreve
5691 * tracepoint for the reserved amount.
5693 static int btrfs_inode_rsv_refill(struct btrfs_inode *inode,
5694 enum btrfs_reserve_flush_enum flush)
5696 struct btrfs_root *root = inode->root;
5697 struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5699 u64 qgroup_num_bytes = 0;
5702 spin_lock(&block_rsv->lock);
5703 if (block_rsv->reserved < block_rsv->size)
5704 num_bytes = block_rsv->size - block_rsv->reserved;
5705 if (block_rsv->qgroup_rsv_reserved < block_rsv->qgroup_rsv_size)
5706 qgroup_num_bytes = block_rsv->qgroup_rsv_size -
5707 block_rsv->qgroup_rsv_reserved;
5708 spin_unlock(&block_rsv->lock);
5713 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_num_bytes, true);
5716 ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5718 block_rsv_add_bytes(block_rsv, num_bytes, false);
5719 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5720 btrfs_ino(inode), num_bytes, 1);
5722 /* Don't forget to increase qgroup_rsv_reserved */
5723 spin_lock(&block_rsv->lock);
5724 block_rsv->qgroup_rsv_reserved += qgroup_num_bytes;
5725 spin_unlock(&block_rsv->lock);
5727 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5732 * btrfs_inode_rsv_release - release any excessive reservation.
5733 * @inode - the inode we need to release from.
5734 * @qgroup_free - free or convert qgroup meta.
5735 * Unlike normal operation, qgroup meta reservation needs to know if we are
5736 * freeing qgroup reservation or just converting it into per-trans. Normally
5737 * @qgroup_free is true for error handling, and false for normal release.
5739 * This is the same as btrfs_block_rsv_release, except that it handles the
5740 * tracepoint for the reservation.
5742 static void btrfs_inode_rsv_release(struct btrfs_inode *inode, bool qgroup_free)
5744 struct btrfs_fs_info *fs_info = inode->root->fs_info;
5745 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5746 struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5748 u64 qgroup_to_release = 0;
5751 * Since we statically set the block_rsv->size we just want to say we
5752 * are releasing 0 bytes, and then we'll just get the reservation over
5755 released = block_rsv_release_bytes(fs_info, block_rsv, global_rsv, 0,
5756 &qgroup_to_release);
5758 trace_btrfs_space_reservation(fs_info, "delalloc",
5759 btrfs_ino(inode), released, 0);
5761 btrfs_qgroup_free_meta_prealloc(inode->root, qgroup_to_release);
5763 btrfs_qgroup_convert_reserved_meta(inode->root,
5767 void btrfs_block_rsv_release(struct btrfs_fs_info *fs_info,
5768 struct btrfs_block_rsv *block_rsv,
5771 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5773 if (global_rsv == block_rsv ||
5774 block_rsv->space_info != global_rsv->space_info)
5776 block_rsv_release_bytes(fs_info, block_rsv, global_rsv, num_bytes, NULL);
5779 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5781 struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5782 struct btrfs_space_info *sinfo = block_rsv->space_info;
5786 * The global block rsv is based on the size of the extent tree, the
5787 * checksum tree and the root tree. If the fs is empty we want to set
5788 * it to a minimal amount for safety.
5790 num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5791 btrfs_root_used(&fs_info->csum_root->root_item) +
5792 btrfs_root_used(&fs_info->tree_root->root_item);
5793 num_bytes = max_t(u64, num_bytes, SZ_16M);
5795 spin_lock(&sinfo->lock);
5796 spin_lock(&block_rsv->lock);
5798 block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5800 if (block_rsv->reserved < block_rsv->size) {
5801 num_bytes = btrfs_space_info_used(sinfo, true);
5802 if (sinfo->total_bytes > num_bytes) {
5803 num_bytes = sinfo->total_bytes - num_bytes;
5804 num_bytes = min(num_bytes,
5805 block_rsv->size - block_rsv->reserved);
5806 block_rsv->reserved += num_bytes;
5807 sinfo->bytes_may_use += num_bytes;
5808 trace_btrfs_space_reservation(fs_info, "space_info",
5809 sinfo->flags, num_bytes,
5812 } else if (block_rsv->reserved > block_rsv->size) {
5813 num_bytes = block_rsv->reserved - block_rsv->size;
5814 sinfo->bytes_may_use -= num_bytes;
5815 trace_btrfs_space_reservation(fs_info, "space_info",
5816 sinfo->flags, num_bytes, 0);
5817 block_rsv->reserved = block_rsv->size;
5820 if (block_rsv->reserved == block_rsv->size)
5821 block_rsv->full = 1;
5823 block_rsv->full = 0;
5825 spin_unlock(&block_rsv->lock);
5826 spin_unlock(&sinfo->lock);
5829 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5831 struct btrfs_space_info *space_info;
5833 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5834 fs_info->chunk_block_rsv.space_info = space_info;
5836 space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5837 fs_info->global_block_rsv.space_info = space_info;
5838 fs_info->trans_block_rsv.space_info = space_info;
5839 fs_info->empty_block_rsv.space_info = space_info;
5840 fs_info->delayed_block_rsv.space_info = space_info;
5842 fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5843 fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5844 fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5845 fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5846 if (fs_info->quota_root)
5847 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5848 fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5850 update_global_block_rsv(fs_info);
5853 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5855 block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5857 WARN_ON(fs_info->trans_block_rsv.size > 0);
5858 WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5859 WARN_ON(fs_info->chunk_block_rsv.size > 0);
5860 WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5861 WARN_ON(fs_info->delayed_block_rsv.size > 0);
5862 WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5867 * To be called after all the new block groups attached to the transaction
5868 * handle have been created (btrfs_create_pending_block_groups()).
5870 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5872 struct btrfs_fs_info *fs_info = trans->fs_info;
5874 if (!trans->chunk_bytes_reserved)
5877 WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5879 block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5880 trans->chunk_bytes_reserved, NULL);
5881 trans->chunk_bytes_reserved = 0;
5885 * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5886 * root: the root of the parent directory
5887 * rsv: block reservation
5888 * items: the number of items that we need do reservation
5889 * use_global_rsv: allow fallback to the global block reservation
5891 * This function is used to reserve the space for snapshot/subvolume
5892 * creation and deletion. Those operations are different with the
5893 * common file/directory operations, they change two fs/file trees
5894 * and root tree, the number of items that the qgroup reserves is
5895 * different with the free space reservation. So we can not use
5896 * the space reservation mechanism in start_transaction().
5898 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5899 struct btrfs_block_rsv *rsv, int items,
5900 bool use_global_rsv)
5902 u64 qgroup_num_bytes = 0;
5905 struct btrfs_fs_info *fs_info = root->fs_info;
5906 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5908 if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
5909 /* One for parent inode, two for dir entries */
5910 qgroup_num_bytes = 3 * fs_info->nodesize;
5911 ret = btrfs_qgroup_reserve_meta_prealloc(root,
5912 qgroup_num_bytes, true);
5917 num_bytes = btrfs_calc_trans_metadata_size(fs_info, items);
5918 rsv->space_info = __find_space_info(fs_info,
5919 BTRFS_BLOCK_GROUP_METADATA);
5920 ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5921 BTRFS_RESERVE_FLUSH_ALL);
5923 if (ret == -ENOSPC && use_global_rsv)
5924 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, true);
5926 if (ret && qgroup_num_bytes)
5927 btrfs_qgroup_free_meta_prealloc(root, qgroup_num_bytes);
5932 void btrfs_subvolume_release_metadata(struct btrfs_fs_info *fs_info,
5933 struct btrfs_block_rsv *rsv)
5935 btrfs_block_rsv_release(fs_info, rsv, (u64)-1);
5938 static void btrfs_calculate_inode_block_rsv_size(struct btrfs_fs_info *fs_info,
5939 struct btrfs_inode *inode)
5941 struct btrfs_block_rsv *block_rsv = &inode->block_rsv;
5942 u64 reserve_size = 0;
5943 u64 qgroup_rsv_size = 0;
5945 unsigned outstanding_extents;
5947 lockdep_assert_held(&inode->lock);
5948 outstanding_extents = inode->outstanding_extents;
5949 if (outstanding_extents)
5950 reserve_size = btrfs_calc_trans_metadata_size(fs_info,
5951 outstanding_extents + 1);
5952 csum_leaves = btrfs_csum_bytes_to_leaves(fs_info,
5954 reserve_size += btrfs_calc_trans_metadata_size(fs_info,
5957 * For qgroup rsv, the calculation is very simple:
5958 * account one nodesize for each outstanding extent
5960 * This is overestimating in most cases.
5962 qgroup_rsv_size = outstanding_extents * fs_info->nodesize;
5964 spin_lock(&block_rsv->lock);
5965 block_rsv->size = reserve_size;
5966 block_rsv->qgroup_rsv_size = qgroup_rsv_size;
5967 spin_unlock(&block_rsv->lock);
5970 int btrfs_delalloc_reserve_metadata(struct btrfs_inode *inode, u64 num_bytes)
5972 struct btrfs_fs_info *fs_info = inode->root->fs_info;
5973 unsigned nr_extents;
5974 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5976 bool delalloc_lock = true;
5978 /* If we are a free space inode we need to not flush since we will be in
5979 * the middle of a transaction commit. We also don't need the delalloc
5980 * mutex since we won't race with anybody. We need this mostly to make
5981 * lockdep shut its filthy mouth.
5983 * If we have a transaction open (can happen if we call truncate_block
5984 * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5986 if (btrfs_is_free_space_inode(inode)) {
5987 flush = BTRFS_RESERVE_NO_FLUSH;
5988 delalloc_lock = false;
5990 if (current->journal_info)
5991 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5993 if (btrfs_transaction_in_commit(fs_info))
5994 schedule_timeout(1);
5998 mutex_lock(&inode->delalloc_mutex);
6000 num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6002 /* Add our new extents and calculate the new rsv size. */
6003 spin_lock(&inode->lock);
6004 nr_extents = count_max_extents(num_bytes);
6005 btrfs_mod_outstanding_extents(inode, nr_extents);
6006 inode->csum_bytes += num_bytes;
6007 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6008 spin_unlock(&inode->lock);
6010 ret = btrfs_inode_rsv_refill(inode, flush);
6015 mutex_unlock(&inode->delalloc_mutex);
6019 spin_lock(&inode->lock);
6020 nr_extents = count_max_extents(num_bytes);
6021 btrfs_mod_outstanding_extents(inode, -nr_extents);
6022 inode->csum_bytes -= num_bytes;
6023 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6024 spin_unlock(&inode->lock);
6026 btrfs_inode_rsv_release(inode, true);
6028 mutex_unlock(&inode->delalloc_mutex);
6033 * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6034 * @inode: the inode to release the reservation for.
6035 * @num_bytes: the number of bytes we are releasing.
6036 * @qgroup_free: free qgroup reservation or convert it to per-trans reservation
6038 * This will release the metadata reservation for an inode. This can be called
6039 * once we complete IO for a given set of bytes to release their metadata
6040 * reservations, or on error for the same reason.
6042 void btrfs_delalloc_release_metadata(struct btrfs_inode *inode, u64 num_bytes,
6045 struct btrfs_fs_info *fs_info = inode->root->fs_info;
6047 num_bytes = ALIGN(num_bytes, fs_info->sectorsize);
6048 spin_lock(&inode->lock);
6049 inode->csum_bytes -= num_bytes;
6050 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6051 spin_unlock(&inode->lock);
6053 if (btrfs_is_testing(fs_info))
6056 btrfs_inode_rsv_release(inode, qgroup_free);
6060 * btrfs_delalloc_release_extents - release our outstanding_extents
6061 * @inode: the inode to balance the reservation for.
6062 * @num_bytes: the number of bytes we originally reserved with
6063 * @qgroup_free: do we need to free qgroup meta reservation or convert them.
6065 * When we reserve space we increase outstanding_extents for the extents we may
6066 * add. Once we've set the range as delalloc or created our ordered extents we
6067 * have outstanding_extents to track the real usage, so we use this to free our
6068 * temporarily tracked outstanding_extents. This _must_ be used in conjunction
6069 * with btrfs_delalloc_reserve_metadata.
6071 void btrfs_delalloc_release_extents(struct btrfs_inode *inode, u64 num_bytes,
6074 struct btrfs_fs_info *fs_info = inode->root->fs_info;
6075 unsigned num_extents;
6077 spin_lock(&inode->lock);
6078 num_extents = count_max_extents(num_bytes);
6079 btrfs_mod_outstanding_extents(inode, -num_extents);
6080 btrfs_calculate_inode_block_rsv_size(fs_info, inode);
6081 spin_unlock(&inode->lock);
6083 if (btrfs_is_testing(fs_info))
6086 btrfs_inode_rsv_release(inode, qgroup_free);
6090 * btrfs_delalloc_reserve_space - reserve data and metadata space for
6092 * @inode: inode we're writing to
6093 * @start: start range we are writing to
6094 * @len: how long the range we are writing to
6095 * @reserved: mandatory parameter, record actually reserved qgroup ranges of
6096 * current reservation.
6098 * This will do the following things
6100 * o reserve space in data space info for num bytes
6101 * and reserve precious corresponding qgroup space
6102 * (Done in check_data_free_space)
6104 * o reserve space for metadata space, based on the number of outstanding
6105 * extents and how much csums will be needed
6106 * also reserve metadata space in a per root over-reserve method.
6107 * o add to the inodes->delalloc_bytes
6108 * o add it to the fs_info's delalloc inodes list.
6109 * (Above 3 all done in delalloc_reserve_metadata)
6111 * Return 0 for success
6112 * Return <0 for error(-ENOSPC or -EQUOT)
6114 int btrfs_delalloc_reserve_space(struct inode *inode,
6115 struct extent_changeset **reserved, u64 start, u64 len)
6119 ret = btrfs_check_data_free_space(inode, reserved, start, len);
6122 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode), len);
6124 btrfs_free_reserved_data_space(inode, *reserved, start, len);
6129 * btrfs_delalloc_release_space - release data and metadata space for delalloc
6130 * @inode: inode we're releasing space for
6131 * @start: start position of the space already reserved
6132 * @len: the len of the space already reserved
6133 * @release_bytes: the len of the space we consumed or didn't use
6135 * This function will release the metadata space that was not used and will
6136 * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6137 * list if there are no delalloc bytes left.
6138 * Also it will handle the qgroup reserved space.
6140 void btrfs_delalloc_release_space(struct inode *inode,
6141 struct extent_changeset *reserved,
6142 u64 start, u64 len, bool qgroup_free)
6144 btrfs_delalloc_release_metadata(BTRFS_I(inode), len, qgroup_free);
6145 btrfs_free_reserved_data_space(inode, reserved, start, len);
6148 static int update_block_group(struct btrfs_trans_handle *trans,
6149 struct btrfs_fs_info *info, u64 bytenr,
6150 u64 num_bytes, int alloc)
6152 struct btrfs_block_group_cache *cache = NULL;
6153 u64 total = num_bytes;
6158 /* block accounting for super block */
6159 spin_lock(&info->delalloc_root_lock);
6160 old_val = btrfs_super_bytes_used(info->super_copy);
6162 old_val += num_bytes;
6164 old_val -= num_bytes;
6165 btrfs_set_super_bytes_used(info->super_copy, old_val);
6166 spin_unlock(&info->delalloc_root_lock);
6169 cache = btrfs_lookup_block_group(info, bytenr);
6172 factor = btrfs_bg_type_to_factor(cache->flags);
6175 * If this block group has free space cache written out, we
6176 * need to make sure to load it if we are removing space. This
6177 * is because we need the unpinning stage to actually add the
6178 * space back to the block group, otherwise we will leak space.
6180 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6181 cache_block_group(cache, 1);
6183 byte_in_group = bytenr - cache->key.objectid;
6184 WARN_ON(byte_in_group > cache->key.offset);
6186 spin_lock(&cache->space_info->lock);
6187 spin_lock(&cache->lock);
6189 if (btrfs_test_opt(info, SPACE_CACHE) &&
6190 cache->disk_cache_state < BTRFS_DC_CLEAR)
6191 cache->disk_cache_state = BTRFS_DC_CLEAR;
6193 old_val = btrfs_block_group_used(&cache->item);
6194 num_bytes = min(total, cache->key.offset - byte_in_group);
6196 old_val += num_bytes;
6197 btrfs_set_block_group_used(&cache->item, old_val);
6198 cache->reserved -= num_bytes;
6199 cache->space_info->bytes_reserved -= num_bytes;
6200 cache->space_info->bytes_used += num_bytes;
6201 cache->space_info->disk_used += num_bytes * factor;
6202 spin_unlock(&cache->lock);
6203 spin_unlock(&cache->space_info->lock);
6205 old_val -= num_bytes;
6206 btrfs_set_block_group_used(&cache->item, old_val);
6207 cache->pinned += num_bytes;
6208 cache->space_info->bytes_pinned += num_bytes;
6209 cache->space_info->bytes_used -= num_bytes;
6210 cache->space_info->disk_used -= num_bytes * factor;
6211 spin_unlock(&cache->lock);
6212 spin_unlock(&cache->space_info->lock);
6214 trace_btrfs_space_reservation(info, "pinned",
6215 cache->space_info->flags,
6217 percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6219 BTRFS_TOTAL_BYTES_PINNED_BATCH);
6220 set_extent_dirty(info->pinned_extents,
6221 bytenr, bytenr + num_bytes - 1,
6222 GFP_NOFS | __GFP_NOFAIL);
6225 spin_lock(&trans->transaction->dirty_bgs_lock);
6226 if (list_empty(&cache->dirty_list)) {
6227 list_add_tail(&cache->dirty_list,
6228 &trans->transaction->dirty_bgs);
6229 trans->transaction->num_dirty_bgs++;
6230 btrfs_get_block_group(cache);
6232 spin_unlock(&trans->transaction->dirty_bgs_lock);
6235 * No longer have used bytes in this block group, queue it for
6236 * deletion. We do this after adding the block group to the
6237 * dirty list to avoid races between cleaner kthread and space
6240 if (!alloc && old_val == 0)
6241 btrfs_mark_bg_unused(cache);
6243 btrfs_put_block_group(cache);
6245 bytenr += num_bytes;
6250 static u64 first_logical_byte(struct btrfs_fs_info *fs_info, u64 search_start)
6252 struct btrfs_block_group_cache *cache;
6255 spin_lock(&fs_info->block_group_cache_lock);
6256 bytenr = fs_info->first_logical_byte;
6257 spin_unlock(&fs_info->block_group_cache_lock);
6259 if (bytenr < (u64)-1)
6262 cache = btrfs_lookup_first_block_group(fs_info, search_start);
6266 bytenr = cache->key.objectid;
6267 btrfs_put_block_group(cache);
6272 static int pin_down_extent(struct btrfs_fs_info *fs_info,
6273 struct btrfs_block_group_cache *cache,
6274 u64 bytenr, u64 num_bytes, int reserved)
6276 spin_lock(&cache->space_info->lock);
6277 spin_lock(&cache->lock);
6278 cache->pinned += num_bytes;
6279 cache->space_info->bytes_pinned += num_bytes;
6281 cache->reserved -= num_bytes;
6282 cache->space_info->bytes_reserved -= num_bytes;
6284 spin_unlock(&cache->lock);
6285 spin_unlock(&cache->space_info->lock);
6287 trace_btrfs_space_reservation(fs_info, "pinned",
6288 cache->space_info->flags, num_bytes, 1);
6289 percpu_counter_add_batch(&cache->space_info->total_bytes_pinned,
6290 num_bytes, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6291 set_extent_dirty(fs_info->pinned_extents, bytenr,
6292 bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6297 * this function must be called within transaction
6299 int btrfs_pin_extent(struct btrfs_fs_info *fs_info,
6300 u64 bytenr, u64 num_bytes, int reserved)
6302 struct btrfs_block_group_cache *cache;
6304 cache = btrfs_lookup_block_group(fs_info, bytenr);
6305 BUG_ON(!cache); /* Logic error */
6307 pin_down_extent(fs_info, cache, bytenr, num_bytes, reserved);
6309 btrfs_put_block_group(cache);
6314 * this function must be called within transaction
6316 int btrfs_pin_extent_for_log_replay(struct btrfs_fs_info *fs_info,
6317 u64 bytenr, u64 num_bytes)
6319 struct btrfs_block_group_cache *cache;
6322 cache = btrfs_lookup_block_group(fs_info, bytenr);
6327 * pull in the free space cache (if any) so that our pin
6328 * removes the free space from the cache. We have load_only set
6329 * to one because the slow code to read in the free extents does check
6330 * the pinned extents.
6332 cache_block_group(cache, 1);
6334 pin_down_extent(fs_info, cache, bytenr, num_bytes, 0);
6336 /* remove us from the free space cache (if we're there at all) */
6337 ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6338 btrfs_put_block_group(cache);
6342 static int __exclude_logged_extent(struct btrfs_fs_info *fs_info,
6343 u64 start, u64 num_bytes)
6346 struct btrfs_block_group_cache *block_group;
6347 struct btrfs_caching_control *caching_ctl;
6349 block_group = btrfs_lookup_block_group(fs_info, start);
6353 cache_block_group(block_group, 0);
6354 caching_ctl = get_caching_control(block_group);
6358 BUG_ON(!block_group_cache_done(block_group));
6359 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6361 mutex_lock(&caching_ctl->mutex);
6363 if (start >= caching_ctl->progress) {
6364 ret = add_excluded_extent(fs_info, start, num_bytes);
6365 } else if (start + num_bytes <= caching_ctl->progress) {
6366 ret = btrfs_remove_free_space(block_group,
6369 num_bytes = caching_ctl->progress - start;
6370 ret = btrfs_remove_free_space(block_group,
6375 num_bytes = (start + num_bytes) -
6376 caching_ctl->progress;
6377 start = caching_ctl->progress;
6378 ret = add_excluded_extent(fs_info, start, num_bytes);
6381 mutex_unlock(&caching_ctl->mutex);
6382 put_caching_control(caching_ctl);
6384 btrfs_put_block_group(block_group);
6388 int btrfs_exclude_logged_extents(struct btrfs_fs_info *fs_info,
6389 struct extent_buffer *eb)
6391 struct btrfs_file_extent_item *item;
6392 struct btrfs_key key;
6397 if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS))
6400 for (i = 0; i < btrfs_header_nritems(eb); i++) {
6401 btrfs_item_key_to_cpu(eb, &key, i);
6402 if (key.type != BTRFS_EXTENT_DATA_KEY)
6404 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6405 found_type = btrfs_file_extent_type(eb, item);
6406 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6408 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6410 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6411 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6412 ret = __exclude_logged_extent(fs_info, key.objectid, key.offset);
6421 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6423 atomic_inc(&bg->reservations);
6426 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6429 struct btrfs_block_group_cache *bg;
6431 bg = btrfs_lookup_block_group(fs_info, start);
6433 if (atomic_dec_and_test(&bg->reservations))
6434 wake_up_var(&bg->reservations);
6435 btrfs_put_block_group(bg);
6438 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6440 struct btrfs_space_info *space_info = bg->space_info;
6444 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6448 * Our block group is read only but before we set it to read only,
6449 * some task might have had allocated an extent from it already, but it
6450 * has not yet created a respective ordered extent (and added it to a
6451 * root's list of ordered extents).
6452 * Therefore wait for any task currently allocating extents, since the
6453 * block group's reservations counter is incremented while a read lock
6454 * on the groups' semaphore is held and decremented after releasing
6455 * the read access on that semaphore and creating the ordered extent.
6457 down_write(&space_info->groups_sem);
6458 up_write(&space_info->groups_sem);
6460 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
6464 * btrfs_add_reserved_bytes - update the block_group and space info counters
6465 * @cache: The cache we are manipulating
6466 * @ram_bytes: The number of bytes of file content, and will be same to
6467 * @num_bytes except for the compress path.
6468 * @num_bytes: The number of bytes in question
6469 * @delalloc: The blocks are allocated for the delalloc write
6471 * This is called by the allocator when it reserves space. If this is a
6472 * reservation and the block group has become read only we cannot make the
6473 * reservation and return -EAGAIN, otherwise this function always succeeds.
6475 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6476 u64 ram_bytes, u64 num_bytes, int delalloc)
6478 struct btrfs_space_info *space_info = cache->space_info;
6481 spin_lock(&space_info->lock);
6482 spin_lock(&cache->lock);
6486 cache->reserved += num_bytes;
6487 space_info->bytes_reserved += num_bytes;
6488 space_info->bytes_may_use -= ram_bytes;
6490 cache->delalloc_bytes += num_bytes;
6492 spin_unlock(&cache->lock);
6493 spin_unlock(&space_info->lock);
6498 * btrfs_free_reserved_bytes - update the block_group and space info counters
6499 * @cache: The cache we are manipulating
6500 * @num_bytes: The number of bytes in question
6501 * @delalloc: The blocks are allocated for the delalloc write
6503 * This is called by somebody who is freeing space that was never actually used
6504 * on disk. For example if you reserve some space for a new leaf in transaction
6505 * A and before transaction A commits you free that leaf, you call this with
6506 * reserve set to 0 in order to clear the reservation.
6509 static void btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6510 u64 num_bytes, int delalloc)
6512 struct btrfs_space_info *space_info = cache->space_info;
6514 spin_lock(&space_info->lock);
6515 spin_lock(&cache->lock);
6517 space_info->bytes_readonly += num_bytes;
6518 cache->reserved -= num_bytes;
6519 space_info->bytes_reserved -= num_bytes;
6522 cache->delalloc_bytes -= num_bytes;
6523 spin_unlock(&cache->lock);
6524 spin_unlock(&space_info->lock);
6526 void btrfs_prepare_extent_commit(struct btrfs_fs_info *fs_info)
6528 struct btrfs_caching_control *next;
6529 struct btrfs_caching_control *caching_ctl;
6530 struct btrfs_block_group_cache *cache;
6532 down_write(&fs_info->commit_root_sem);
6534 list_for_each_entry_safe(caching_ctl, next,
6535 &fs_info->caching_block_groups, list) {
6536 cache = caching_ctl->block_group;
6537 if (block_group_cache_done(cache)) {
6538 cache->last_byte_to_unpin = (u64)-1;
6539 list_del_init(&caching_ctl->list);
6540 put_caching_control(caching_ctl);
6542 cache->last_byte_to_unpin = caching_ctl->progress;
6546 if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6547 fs_info->pinned_extents = &fs_info->freed_extents[1];
6549 fs_info->pinned_extents = &fs_info->freed_extents[0];
6551 up_write(&fs_info->commit_root_sem);
6553 update_global_block_rsv(fs_info);
6557 * Returns the free cluster for the given space info and sets empty_cluster to
6558 * what it should be based on the mount options.
6560 static struct btrfs_free_cluster *
6561 fetch_cluster_info(struct btrfs_fs_info *fs_info,
6562 struct btrfs_space_info *space_info, u64 *empty_cluster)
6564 struct btrfs_free_cluster *ret = NULL;
6567 if (btrfs_mixed_space_info(space_info))
6570 if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6571 ret = &fs_info->meta_alloc_cluster;
6572 if (btrfs_test_opt(fs_info, SSD))
6573 *empty_cluster = SZ_2M;
6575 *empty_cluster = SZ_64K;
6576 } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
6577 btrfs_test_opt(fs_info, SSD_SPREAD)) {
6578 *empty_cluster = SZ_2M;
6579 ret = &fs_info->data_alloc_cluster;
6585 static int unpin_extent_range(struct btrfs_fs_info *fs_info,
6587 const bool return_free_space)
6589 struct btrfs_block_group_cache *cache = NULL;
6590 struct btrfs_space_info *space_info;
6591 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6592 struct btrfs_free_cluster *cluster = NULL;
6594 u64 total_unpinned = 0;
6595 u64 empty_cluster = 0;
6598 while (start <= end) {
6601 start >= cache->key.objectid + cache->key.offset) {
6603 btrfs_put_block_group(cache);
6605 cache = btrfs_lookup_block_group(fs_info, start);
6606 BUG_ON(!cache); /* Logic error */
6608 cluster = fetch_cluster_info(fs_info,
6611 empty_cluster <<= 1;
6614 len = cache->key.objectid + cache->key.offset - start;
6615 len = min(len, end + 1 - start);
6617 if (start < cache->last_byte_to_unpin) {
6618 len = min(len, cache->last_byte_to_unpin - start);
6619 if (return_free_space)
6620 btrfs_add_free_space(cache, start, len);
6624 total_unpinned += len;
6625 space_info = cache->space_info;
6628 * If this space cluster has been marked as fragmented and we've
6629 * unpinned enough in this block group to potentially allow a
6630 * cluster to be created inside of it go ahead and clear the
6633 if (cluster && cluster->fragmented &&
6634 total_unpinned > empty_cluster) {
6635 spin_lock(&cluster->lock);
6636 cluster->fragmented = 0;
6637 spin_unlock(&cluster->lock);
6640 spin_lock(&space_info->lock);
6641 spin_lock(&cache->lock);
6642 cache->pinned -= len;
6643 space_info->bytes_pinned -= len;
6645 trace_btrfs_space_reservation(fs_info, "pinned",
6646 space_info->flags, len, 0);
6647 space_info->max_extent_size = 0;
6648 percpu_counter_add_batch(&space_info->total_bytes_pinned,
6649 -len, BTRFS_TOTAL_BYTES_PINNED_BATCH);
6651 space_info->bytes_readonly += len;
6654 spin_unlock(&cache->lock);
6655 if (!readonly && return_free_space &&
6656 global_rsv->space_info == space_info) {
6659 spin_lock(&global_rsv->lock);
6660 if (!global_rsv->full) {
6661 to_add = min(len, global_rsv->size -
6662 global_rsv->reserved);
6663 global_rsv->reserved += to_add;
6664 space_info->bytes_may_use += to_add;
6665 if (global_rsv->reserved >= global_rsv->size)
6666 global_rsv->full = 1;
6667 trace_btrfs_space_reservation(fs_info,
6673 spin_unlock(&global_rsv->lock);
6674 /* Add to any tickets we may have */
6676 space_info_add_new_bytes(fs_info, space_info,
6679 spin_unlock(&space_info->lock);
6683 btrfs_put_block_group(cache);
6687 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
6689 struct btrfs_fs_info *fs_info = trans->fs_info;
6690 struct btrfs_block_group_cache *block_group, *tmp;
6691 struct list_head *deleted_bgs;
6692 struct extent_io_tree *unpin;
6697 if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6698 unpin = &fs_info->freed_extents[1];
6700 unpin = &fs_info->freed_extents[0];
6702 while (!trans->aborted) {
6703 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6704 ret = find_first_extent_bit(unpin, 0, &start, &end,
6705 EXTENT_DIRTY, NULL);
6707 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6711 if (btrfs_test_opt(fs_info, DISCARD))
6712 ret = btrfs_discard_extent(fs_info, start,
6713 end + 1 - start, NULL);
6715 clear_extent_dirty(unpin, start, end);
6716 unpin_extent_range(fs_info, start, end, true);
6717 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6722 * Transaction is finished. We don't need the lock anymore. We
6723 * do need to clean up the block groups in case of a transaction
6726 deleted_bgs = &trans->transaction->deleted_bgs;
6727 list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6731 if (!trans->aborted)
6732 ret = btrfs_discard_extent(fs_info,
6733 block_group->key.objectid,
6734 block_group->key.offset,
6737 list_del_init(&block_group->bg_list);
6738 btrfs_put_block_group_trimming(block_group);
6739 btrfs_put_block_group(block_group);
6742 const char *errstr = btrfs_decode_error(ret);
6744 "discard failed while removing blockgroup: errno=%d %s",
6752 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6753 struct btrfs_delayed_ref_node *node, u64 parent,
6754 u64 root_objectid, u64 owner_objectid,
6755 u64 owner_offset, int refs_to_drop,
6756 struct btrfs_delayed_extent_op *extent_op)
6758 struct btrfs_fs_info *info = trans->fs_info;
6759 struct btrfs_key key;
6760 struct btrfs_path *path;
6761 struct btrfs_root *extent_root = info->extent_root;
6762 struct extent_buffer *leaf;
6763 struct btrfs_extent_item *ei;
6764 struct btrfs_extent_inline_ref *iref;
6767 int extent_slot = 0;
6768 int found_extent = 0;
6772 u64 bytenr = node->bytenr;
6773 u64 num_bytes = node->num_bytes;
6775 bool skinny_metadata = btrfs_fs_incompat(info, SKINNY_METADATA);
6777 path = btrfs_alloc_path();
6781 path->reada = READA_FORWARD;
6782 path->leave_spinning = 1;
6784 is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6785 BUG_ON(!is_data && refs_to_drop != 1);
6788 skinny_metadata = false;
6790 ret = lookup_extent_backref(trans, path, &iref, bytenr, num_bytes,
6791 parent, root_objectid, owner_objectid,
6794 extent_slot = path->slots[0];
6795 while (extent_slot >= 0) {
6796 btrfs_item_key_to_cpu(path->nodes[0], &key,
6798 if (key.objectid != bytenr)
6800 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6801 key.offset == num_bytes) {
6805 if (key.type == BTRFS_METADATA_ITEM_KEY &&
6806 key.offset == owner_objectid) {
6810 if (path->slots[0] - extent_slot > 5)
6815 if (!found_extent) {
6817 ret = remove_extent_backref(trans, path, NULL,
6819 is_data, &last_ref);
6821 btrfs_abort_transaction(trans, ret);
6824 btrfs_release_path(path);
6825 path->leave_spinning = 1;
6827 key.objectid = bytenr;
6828 key.type = BTRFS_EXTENT_ITEM_KEY;
6829 key.offset = num_bytes;
6831 if (!is_data && skinny_metadata) {
6832 key.type = BTRFS_METADATA_ITEM_KEY;
6833 key.offset = owner_objectid;
6836 ret = btrfs_search_slot(trans, extent_root,
6838 if (ret > 0 && skinny_metadata && path->slots[0]) {
6840 * Couldn't find our skinny metadata item,
6841 * see if we have ye olde extent item.
6844 btrfs_item_key_to_cpu(path->nodes[0], &key,
6846 if (key.objectid == bytenr &&
6847 key.type == BTRFS_EXTENT_ITEM_KEY &&
6848 key.offset == num_bytes)
6852 if (ret > 0 && skinny_metadata) {
6853 skinny_metadata = false;
6854 key.objectid = bytenr;
6855 key.type = BTRFS_EXTENT_ITEM_KEY;
6856 key.offset = num_bytes;
6857 btrfs_release_path(path);
6858 ret = btrfs_search_slot(trans, extent_root,
6864 "umm, got %d back from search, was looking for %llu",
6867 btrfs_print_leaf(path->nodes[0]);
6870 btrfs_abort_transaction(trans, ret);
6873 extent_slot = path->slots[0];
6875 } else if (WARN_ON(ret == -ENOENT)) {
6876 btrfs_print_leaf(path->nodes[0]);
6878 "unable to find ref byte nr %llu parent %llu root %llu owner %llu offset %llu",
6879 bytenr, parent, root_objectid, owner_objectid,
6881 btrfs_abort_transaction(trans, ret);
6884 btrfs_abort_transaction(trans, ret);
6888 leaf = path->nodes[0];
6889 item_size = btrfs_item_size_nr(leaf, extent_slot);
6890 if (unlikely(item_size < sizeof(*ei))) {
6892 btrfs_print_v0_err(info);
6893 btrfs_abort_transaction(trans, ret);
6896 ei = btrfs_item_ptr(leaf, extent_slot,
6897 struct btrfs_extent_item);
6898 if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6899 key.type == BTRFS_EXTENT_ITEM_KEY) {
6900 struct btrfs_tree_block_info *bi;
6901 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6902 bi = (struct btrfs_tree_block_info *)(ei + 1);
6903 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6906 refs = btrfs_extent_refs(leaf, ei);
6907 if (refs < refs_to_drop) {
6909 "trying to drop %d refs but we only have %Lu for bytenr %Lu",
6910 refs_to_drop, refs, bytenr);
6912 btrfs_abort_transaction(trans, ret);
6915 refs -= refs_to_drop;
6919 __run_delayed_extent_op(extent_op, leaf, ei);
6921 * In the case of inline back ref, reference count will
6922 * be updated by remove_extent_backref
6925 BUG_ON(!found_extent);
6927 btrfs_set_extent_refs(leaf, ei, refs);
6928 btrfs_mark_buffer_dirty(leaf);
6931 ret = remove_extent_backref(trans, path, iref,
6932 refs_to_drop, is_data,
6935 btrfs_abort_transaction(trans, ret);
6941 BUG_ON(is_data && refs_to_drop !=
6942 extent_data_ref_count(path, iref));
6944 BUG_ON(path->slots[0] != extent_slot);
6946 BUG_ON(path->slots[0] != extent_slot + 1);
6947 path->slots[0] = extent_slot;
6953 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6956 btrfs_abort_transaction(trans, ret);
6959 btrfs_release_path(path);
6962 ret = btrfs_del_csums(trans, info, bytenr, num_bytes);
6964 btrfs_abort_transaction(trans, ret);
6969 ret = add_to_free_space_tree(trans, bytenr, num_bytes);
6971 btrfs_abort_transaction(trans, ret);
6975 ret = update_block_group(trans, info, bytenr, num_bytes, 0);
6977 btrfs_abort_transaction(trans, ret);
6981 btrfs_release_path(path);
6984 btrfs_free_path(path);
6989 * when we free an block, it is possible (and likely) that we free the last
6990 * delayed ref for that extent as well. This searches the delayed ref tree for
6991 * a given extent, and if there are no other delayed refs to be processed, it
6992 * removes it from the tree.
6994 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6997 struct btrfs_delayed_ref_head *head;
6998 struct btrfs_delayed_ref_root *delayed_refs;
7001 delayed_refs = &trans->transaction->delayed_refs;
7002 spin_lock(&delayed_refs->lock);
7003 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
7005 goto out_delayed_unlock;
7007 spin_lock(&head->lock);
7008 if (!RB_EMPTY_ROOT(&head->ref_tree.rb_root))
7011 if (head->extent_op) {
7012 if (!head->must_insert_reserved)
7014 btrfs_free_delayed_extent_op(head->extent_op);
7015 head->extent_op = NULL;
7019 * waiting for the lock here would deadlock. If someone else has it
7020 * locked they are already in the process of dropping it anyway
7022 if (!mutex_trylock(&head->mutex))
7026 * at this point we have a head with no other entries. Go
7027 * ahead and process it.
7029 rb_erase_cached(&head->href_node, &delayed_refs->href_root);
7030 RB_CLEAR_NODE(&head->href_node);
7031 atomic_dec(&delayed_refs->num_entries);
7034 * we don't take a ref on the node because we're removing it from the
7035 * tree, so we just steal the ref the tree was holding.
7037 delayed_refs->num_heads--;
7038 if (head->processing == 0)
7039 delayed_refs->num_heads_ready--;
7040 head->processing = 0;
7041 spin_unlock(&head->lock);
7042 spin_unlock(&delayed_refs->lock);
7044 BUG_ON(head->extent_op);
7045 if (head->must_insert_reserved)
7048 mutex_unlock(&head->mutex);
7049 btrfs_put_delayed_ref_head(head);
7052 spin_unlock(&head->lock);
7055 spin_unlock(&delayed_refs->lock);
7059 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7060 struct btrfs_root *root,
7061 struct extent_buffer *buf,
7062 u64 parent, int last_ref)
7064 struct btrfs_fs_info *fs_info = root->fs_info;
7068 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7069 int old_ref_mod, new_ref_mod;
7071 btrfs_ref_tree_mod(root, buf->start, buf->len, parent,
7072 root->root_key.objectid,
7073 btrfs_header_level(buf), 0,
7074 BTRFS_DROP_DELAYED_REF);
7075 ret = btrfs_add_delayed_tree_ref(trans, buf->start,
7077 root->root_key.objectid,
7078 btrfs_header_level(buf),
7079 BTRFS_DROP_DELAYED_REF, NULL,
7080 &old_ref_mod, &new_ref_mod);
7081 BUG_ON(ret); /* -ENOMEM */
7082 pin = old_ref_mod >= 0 && new_ref_mod < 0;
7085 if (last_ref && btrfs_header_generation(buf) == trans->transid) {
7086 struct btrfs_block_group_cache *cache;
7088 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7089 ret = check_ref_cleanup(trans, buf->start);
7095 cache = btrfs_lookup_block_group(fs_info, buf->start);
7097 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7098 pin_down_extent(fs_info, cache, buf->start,
7100 btrfs_put_block_group(cache);
7104 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7106 btrfs_add_free_space(cache, buf->start, buf->len);
7107 btrfs_free_reserved_bytes(cache, buf->len, 0);
7108 btrfs_put_block_group(cache);
7109 trace_btrfs_reserved_extent_free(fs_info, buf->start, buf->len);
7113 add_pinned_bytes(fs_info, buf->len, true,
7114 root->root_key.objectid);
7118 * Deleting the buffer, clear the corrupt flag since it doesn't
7121 clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7125 /* Can return -ENOMEM */
7126 int btrfs_free_extent(struct btrfs_trans_handle *trans,
7127 struct btrfs_root *root,
7128 u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7129 u64 owner, u64 offset)
7131 struct btrfs_fs_info *fs_info = root->fs_info;
7132 int old_ref_mod, new_ref_mod;
7135 if (btrfs_is_testing(fs_info))
7138 if (root_objectid != BTRFS_TREE_LOG_OBJECTID)
7139 btrfs_ref_tree_mod(root, bytenr, num_bytes, parent,
7140 root_objectid, owner, offset,
7141 BTRFS_DROP_DELAYED_REF);
7144 * tree log blocks never actually go into the extent allocation
7145 * tree, just update pinning info and exit early.
7147 if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7148 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7149 /* unlocks the pinned mutex */
7150 btrfs_pin_extent(fs_info, bytenr, num_bytes, 1);
7151 old_ref_mod = new_ref_mod = 0;
7153 } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7154 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
7156 root_objectid, (int)owner,
7157 BTRFS_DROP_DELAYED_REF, NULL,
7158 &old_ref_mod, &new_ref_mod);
7160 ret = btrfs_add_delayed_data_ref(trans, bytenr,
7162 root_objectid, owner, offset,
7163 0, BTRFS_DROP_DELAYED_REF,
7164 &old_ref_mod, &new_ref_mod);
7167 if (ret == 0 && old_ref_mod >= 0 && new_ref_mod < 0) {
7168 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
7170 add_pinned_bytes(fs_info, num_bytes, metadata, root_objectid);
7177 * when we wait for progress in the block group caching, its because
7178 * our allocation attempt failed at least once. So, we must sleep
7179 * and let some progress happen before we try again.
7181 * This function will sleep at least once waiting for new free space to
7182 * show up, and then it will check the block group free space numbers
7183 * for our min num_bytes. Another option is to have it go ahead
7184 * and look in the rbtree for a free extent of a given size, but this
7187 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7188 * any of the information in this block group.
7190 static noinline void
7191 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7194 struct btrfs_caching_control *caching_ctl;
7196 caching_ctl = get_caching_control(cache);
7200 wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7201 (cache->free_space_ctl->free_space >= num_bytes));
7203 put_caching_control(caching_ctl);
7207 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7209 struct btrfs_caching_control *caching_ctl;
7212 caching_ctl = get_caching_control(cache);
7214 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7216 wait_event(caching_ctl->wait, block_group_cache_done(cache));
7217 if (cache->cached == BTRFS_CACHE_ERROR)
7219 put_caching_control(caching_ctl);
7223 enum btrfs_loop_type {
7224 LOOP_CACHING_NOWAIT = 0,
7225 LOOP_CACHING_WAIT = 1,
7226 LOOP_ALLOC_CHUNK = 2,
7227 LOOP_NO_EMPTY_SIZE = 3,
7231 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7235 down_read(&cache->data_rwsem);
7239 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7242 btrfs_get_block_group(cache);
7244 down_read(&cache->data_rwsem);
7247 static struct btrfs_block_group_cache *
7248 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7249 struct btrfs_free_cluster *cluster,
7252 struct btrfs_block_group_cache *used_bg = NULL;
7254 spin_lock(&cluster->refill_lock);
7256 used_bg = cluster->block_group;
7260 if (used_bg == block_group)
7263 btrfs_get_block_group(used_bg);
7268 if (down_read_trylock(&used_bg->data_rwsem))
7271 spin_unlock(&cluster->refill_lock);
7273 /* We should only have one-level nested. */
7274 down_read_nested(&used_bg->data_rwsem, SINGLE_DEPTH_NESTING);
7276 spin_lock(&cluster->refill_lock);
7277 if (used_bg == cluster->block_group)
7280 up_read(&used_bg->data_rwsem);
7281 btrfs_put_block_group(used_bg);
7286 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7290 up_read(&cache->data_rwsem);
7291 btrfs_put_block_group(cache);
7295 * walks the btree of allocated extents and find a hole of a given size.
7296 * The key ins is changed to record the hole:
7297 * ins->objectid == start position
7298 * ins->flags = BTRFS_EXTENT_ITEM_KEY
7299 * ins->offset == the size of the hole.
7300 * Any available blocks before search_start are skipped.
7302 * If there is no suitable free space, we will record the max size of
7303 * the free space extent currently.
7305 static noinline int find_free_extent(struct btrfs_fs_info *fs_info,
7306 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7307 u64 hint_byte, struct btrfs_key *ins,
7308 u64 flags, int delalloc)
7311 struct btrfs_root *root = fs_info->extent_root;
7312 struct btrfs_free_cluster *last_ptr = NULL;
7313 struct btrfs_block_group_cache *block_group = NULL;
7314 u64 search_start = 0;
7315 u64 max_extent_size = 0;
7316 u64 empty_cluster = 0;
7317 struct btrfs_space_info *space_info;
7319 int index = btrfs_bg_flags_to_raid_index(flags);
7320 bool failed_cluster_refill = false;
7321 bool failed_alloc = false;
7322 bool use_cluster = true;
7323 bool have_caching_bg = false;
7324 bool orig_have_caching_bg = false;
7325 bool full_search = false;
7327 WARN_ON(num_bytes < fs_info->sectorsize);
7328 ins->type = BTRFS_EXTENT_ITEM_KEY;
7332 trace_find_free_extent(fs_info, num_bytes, empty_size, flags);
7334 space_info = __find_space_info(fs_info, flags);
7336 btrfs_err(fs_info, "No space info for %llu", flags);
7341 * If our free space is heavily fragmented we may not be able to make
7342 * big contiguous allocations, so instead of doing the expensive search
7343 * for free space, simply return ENOSPC with our max_extent_size so we
7344 * can go ahead and search for a more manageable chunk.
7346 * If our max_extent_size is large enough for our allocation simply
7347 * disable clustering since we will likely not be able to find enough
7348 * space to create a cluster and induce latency trying.
7350 if (unlikely(space_info->max_extent_size)) {
7351 spin_lock(&space_info->lock);
7352 if (space_info->max_extent_size &&
7353 num_bytes > space_info->max_extent_size) {
7354 ins->offset = space_info->max_extent_size;
7355 spin_unlock(&space_info->lock);
7357 } else if (space_info->max_extent_size) {
7358 use_cluster = false;
7360 spin_unlock(&space_info->lock);
7363 last_ptr = fetch_cluster_info(fs_info, space_info, &empty_cluster);
7365 spin_lock(&last_ptr->lock);
7366 if (last_ptr->block_group)
7367 hint_byte = last_ptr->window_start;
7368 if (last_ptr->fragmented) {
7370 * We still set window_start so we can keep track of the
7371 * last place we found an allocation to try and save
7374 hint_byte = last_ptr->window_start;
7375 use_cluster = false;
7377 spin_unlock(&last_ptr->lock);
7380 search_start = max(search_start, first_logical_byte(fs_info, 0));
7381 search_start = max(search_start, hint_byte);
7382 if (search_start == hint_byte) {
7383 block_group = btrfs_lookup_block_group(fs_info, search_start);
7385 * we don't want to use the block group if it doesn't match our
7386 * allocation bits, or if its not cached.
7388 * However if we are re-searching with an ideal block group
7389 * picked out then we don't care that the block group is cached.
7391 if (block_group && block_group_bits(block_group, flags) &&
7392 block_group->cached != BTRFS_CACHE_NO) {
7393 down_read(&space_info->groups_sem);
7394 if (list_empty(&block_group->list) ||
7397 * someone is removing this block group,
7398 * we can't jump into the have_block_group
7399 * target because our list pointers are not
7402 btrfs_put_block_group(block_group);
7403 up_read(&space_info->groups_sem);
7405 index = btrfs_bg_flags_to_raid_index(
7406 block_group->flags);
7407 btrfs_lock_block_group(block_group, delalloc);
7408 goto have_block_group;
7410 } else if (block_group) {
7411 btrfs_put_block_group(block_group);
7415 have_caching_bg = false;
7416 if (index == 0 || index == btrfs_bg_flags_to_raid_index(flags))
7418 down_read(&space_info->groups_sem);
7419 list_for_each_entry(block_group, &space_info->block_groups[index],
7424 /* If the block group is read-only, we can skip it entirely. */
7425 if (unlikely(block_group->ro))
7428 btrfs_grab_block_group(block_group, delalloc);
7429 search_start = block_group->key.objectid;
7432 * this can happen if we end up cycling through all the
7433 * raid types, but we want to make sure we only allocate
7434 * for the proper type.
7436 if (!block_group_bits(block_group, flags)) {
7437 u64 extra = BTRFS_BLOCK_GROUP_DUP |
7438 BTRFS_BLOCK_GROUP_RAID1 |
7439 BTRFS_BLOCK_GROUP_RAID5 |
7440 BTRFS_BLOCK_GROUP_RAID6 |
7441 BTRFS_BLOCK_GROUP_RAID10;
7444 * if they asked for extra copies and this block group
7445 * doesn't provide them, bail. This does allow us to
7446 * fill raid0 from raid1.
7448 if ((flags & extra) && !(block_group->flags & extra))
7453 cached = block_group_cache_done(block_group);
7454 if (unlikely(!cached)) {
7455 have_caching_bg = true;
7456 ret = cache_block_group(block_group, 0);
7461 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7465 * Ok we want to try and use the cluster allocator, so
7468 if (last_ptr && use_cluster) {
7469 struct btrfs_block_group_cache *used_block_group;
7470 unsigned long aligned_cluster;
7472 * the refill lock keeps out other
7473 * people trying to start a new cluster
7475 used_block_group = btrfs_lock_cluster(block_group,
7478 if (!used_block_group)
7479 goto refill_cluster;
7481 if (used_block_group != block_group &&
7482 (used_block_group->ro ||
7483 !block_group_bits(used_block_group, flags)))
7484 goto release_cluster;
7486 offset = btrfs_alloc_from_cluster(used_block_group,
7489 used_block_group->key.objectid,
7492 /* we have a block, we're done */
7493 spin_unlock(&last_ptr->refill_lock);
7494 trace_btrfs_reserve_extent_cluster(
7496 search_start, num_bytes);
7497 if (used_block_group != block_group) {
7498 btrfs_release_block_group(block_group,
7500 block_group = used_block_group;
7505 WARN_ON(last_ptr->block_group != used_block_group);
7507 /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7508 * set up a new clusters, so lets just skip it
7509 * and let the allocator find whatever block
7510 * it can find. If we reach this point, we
7511 * will have tried the cluster allocator
7512 * plenty of times and not have found
7513 * anything, so we are likely way too
7514 * fragmented for the clustering stuff to find
7517 * However, if the cluster is taken from the
7518 * current block group, release the cluster
7519 * first, so that we stand a better chance of
7520 * succeeding in the unclustered
7522 if (loop >= LOOP_NO_EMPTY_SIZE &&
7523 used_block_group != block_group) {
7524 spin_unlock(&last_ptr->refill_lock);
7525 btrfs_release_block_group(used_block_group,
7527 goto unclustered_alloc;
7531 * this cluster didn't work out, free it and
7534 btrfs_return_cluster_to_free_space(NULL, last_ptr);
7536 if (used_block_group != block_group)
7537 btrfs_release_block_group(used_block_group,
7540 if (loop >= LOOP_NO_EMPTY_SIZE) {
7541 spin_unlock(&last_ptr->refill_lock);
7542 goto unclustered_alloc;
7545 aligned_cluster = max_t(unsigned long,
7546 empty_cluster + empty_size,
7547 block_group->full_stripe_len);
7549 /* allocate a cluster in this block group */
7550 ret = btrfs_find_space_cluster(fs_info, block_group,
7551 last_ptr, search_start,
7556 * now pull our allocation out of this
7559 offset = btrfs_alloc_from_cluster(block_group,
7565 /* we found one, proceed */
7566 spin_unlock(&last_ptr->refill_lock);
7567 trace_btrfs_reserve_extent_cluster(
7568 block_group, search_start,
7572 } else if (!cached && loop > LOOP_CACHING_NOWAIT
7573 && !failed_cluster_refill) {
7574 spin_unlock(&last_ptr->refill_lock);
7576 failed_cluster_refill = true;
7577 wait_block_group_cache_progress(block_group,
7578 num_bytes + empty_cluster + empty_size);
7579 goto have_block_group;
7583 * at this point we either didn't find a cluster
7584 * or we weren't able to allocate a block from our
7585 * cluster. Free the cluster we've been trying
7586 * to use, and go to the next block group
7588 btrfs_return_cluster_to_free_space(NULL, last_ptr);
7589 spin_unlock(&last_ptr->refill_lock);
7595 * We are doing an unclustered alloc, set the fragmented flag so
7596 * we don't bother trying to setup a cluster again until we get
7599 if (unlikely(last_ptr)) {
7600 spin_lock(&last_ptr->lock);
7601 last_ptr->fragmented = 1;
7602 spin_unlock(&last_ptr->lock);
7605 struct btrfs_free_space_ctl *ctl =
7606 block_group->free_space_ctl;
7608 spin_lock(&ctl->tree_lock);
7609 if (ctl->free_space <
7610 num_bytes + empty_cluster + empty_size) {
7611 if (ctl->free_space > max_extent_size)
7612 max_extent_size = ctl->free_space;
7613 spin_unlock(&ctl->tree_lock);
7616 spin_unlock(&ctl->tree_lock);
7619 offset = btrfs_find_space_for_alloc(block_group, search_start,
7620 num_bytes, empty_size,
7623 * If we didn't find a chunk, and we haven't failed on this
7624 * block group before, and this block group is in the middle of
7625 * caching and we are ok with waiting, then go ahead and wait
7626 * for progress to be made, and set failed_alloc to true.
7628 * If failed_alloc is true then we've already waited on this
7629 * block group once and should move on to the next block group.
7631 if (!offset && !failed_alloc && !cached &&
7632 loop > LOOP_CACHING_NOWAIT) {
7633 wait_block_group_cache_progress(block_group,
7634 num_bytes + empty_size);
7635 failed_alloc = true;
7636 goto have_block_group;
7637 } else if (!offset) {
7641 search_start = round_up(offset, fs_info->stripesize);
7643 /* move on to the next group */
7644 if (search_start + num_bytes >
7645 block_group->key.objectid + block_group->key.offset) {
7646 btrfs_add_free_space(block_group, offset, num_bytes);
7650 if (offset < search_start)
7651 btrfs_add_free_space(block_group, offset,
7652 search_start - offset);
7654 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7655 num_bytes, delalloc);
7656 if (ret == -EAGAIN) {
7657 btrfs_add_free_space(block_group, offset, num_bytes);
7660 btrfs_inc_block_group_reservations(block_group);
7662 /* we are all good, lets return */
7663 ins->objectid = search_start;
7664 ins->offset = num_bytes;
7666 trace_btrfs_reserve_extent(block_group, search_start, num_bytes);
7667 btrfs_release_block_group(block_group, delalloc);
7670 failed_cluster_refill = false;
7671 failed_alloc = false;
7672 BUG_ON(btrfs_bg_flags_to_raid_index(block_group->flags) !=
7674 btrfs_release_block_group(block_group, delalloc);
7677 up_read(&space_info->groups_sem);
7679 if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7680 && !orig_have_caching_bg)
7681 orig_have_caching_bg = true;
7683 if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7686 if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7690 * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7691 * caching kthreads as we move along
7692 * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7693 * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7694 * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7697 if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7699 if (loop == LOOP_CACHING_NOWAIT) {
7701 * We want to skip the LOOP_CACHING_WAIT step if we
7702 * don't have any uncached bgs and we've already done a
7703 * full search through.
7705 if (orig_have_caching_bg || !full_search)
7706 loop = LOOP_CACHING_WAIT;
7708 loop = LOOP_ALLOC_CHUNK;
7713 if (loop == LOOP_ALLOC_CHUNK) {
7714 struct btrfs_trans_handle *trans;
7717 trans = current->journal_info;
7721 trans = btrfs_join_transaction(root);
7723 if (IS_ERR(trans)) {
7724 ret = PTR_ERR(trans);
7728 ret = do_chunk_alloc(trans, flags, CHUNK_ALLOC_FORCE);
7731 * If we can't allocate a new chunk we've already looped
7732 * through at least once, move on to the NO_EMPTY_SIZE
7736 loop = LOOP_NO_EMPTY_SIZE;
7739 * Do not bail out on ENOSPC since we
7740 * can do more things.
7742 if (ret < 0 && ret != -ENOSPC)
7743 btrfs_abort_transaction(trans, ret);
7747 btrfs_end_transaction(trans);
7752 if (loop == LOOP_NO_EMPTY_SIZE) {
7754 * Don't loop again if we already have no empty_size and
7757 if (empty_size == 0 &&
7758 empty_cluster == 0) {
7767 } else if (!ins->objectid) {
7769 } else if (ins->objectid) {
7770 if (!use_cluster && last_ptr) {
7771 spin_lock(&last_ptr->lock);
7772 last_ptr->window_start = ins->objectid;
7773 spin_unlock(&last_ptr->lock);
7778 if (ret == -ENOSPC) {
7779 spin_lock(&space_info->lock);
7780 space_info->max_extent_size = max_extent_size;
7781 spin_unlock(&space_info->lock);
7782 ins->offset = max_extent_size;
7787 static void dump_space_info(struct btrfs_fs_info *fs_info,
7788 struct btrfs_space_info *info, u64 bytes,
7789 int dump_block_groups)
7791 struct btrfs_block_group_cache *cache;
7794 spin_lock(&info->lock);
7795 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7797 info->total_bytes - btrfs_space_info_used(info, true),
7798 info->full ? "" : "not ");
7800 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7801 info->total_bytes, info->bytes_used, info->bytes_pinned,
7802 info->bytes_reserved, info->bytes_may_use,
7803 info->bytes_readonly);
7804 spin_unlock(&info->lock);
7806 if (!dump_block_groups)
7809 down_read(&info->groups_sem);
7811 list_for_each_entry(cache, &info->block_groups[index], list) {
7812 spin_lock(&cache->lock);
7814 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7815 cache->key.objectid, cache->key.offset,
7816 btrfs_block_group_used(&cache->item), cache->pinned,
7817 cache->reserved, cache->ro ? "[readonly]" : "");
7818 btrfs_dump_free_space(cache, bytes);
7819 spin_unlock(&cache->lock);
7821 if (++index < BTRFS_NR_RAID_TYPES)
7823 up_read(&info->groups_sem);
7827 * btrfs_reserve_extent - entry point to the extent allocator. Tries to find a
7828 * hole that is at least as big as @num_bytes.
7830 * @root - The root that will contain this extent
7832 * @ram_bytes - The amount of space in ram that @num_bytes take. This
7833 * is used for accounting purposes. This value differs
7834 * from @num_bytes only in the case of compressed extents.
7836 * @num_bytes - Number of bytes to allocate on-disk.
7838 * @min_alloc_size - Indicates the minimum amount of space that the
7839 * allocator should try to satisfy. In some cases
7840 * @num_bytes may be larger than what is required and if
7841 * the filesystem is fragmented then allocation fails.
7842 * However, the presence of @min_alloc_size gives a
7843 * chance to try and satisfy the smaller allocation.
7845 * @empty_size - A hint that you plan on doing more COW. This is the
7846 * size in bytes the allocator should try to find free
7847 * next to the block it returns. This is just a hint and
7848 * may be ignored by the allocator.
7850 * @hint_byte - Hint to the allocator to start searching above the byte
7851 * address passed. It might be ignored.
7853 * @ins - This key is modified to record the found hole. It will
7854 * have the following values:
7855 * ins->objectid == start position
7856 * ins->flags = BTRFS_EXTENT_ITEM_KEY
7857 * ins->offset == the size of the hole.
7859 * @is_data - Boolean flag indicating whether an extent is
7860 * allocated for data (true) or metadata (false)
7862 * @delalloc - Boolean flag indicating whether this allocation is for
7863 * delalloc or not. If 'true' data_rwsem of block groups
7864 * is going to be acquired.
7867 * Returns 0 when an allocation succeeded or < 0 when an error occurred. In
7868 * case -ENOSPC is returned then @ins->offset will contain the size of the
7869 * largest available hole the allocator managed to find.
7871 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7872 u64 num_bytes, u64 min_alloc_size,
7873 u64 empty_size, u64 hint_byte,
7874 struct btrfs_key *ins, int is_data, int delalloc)
7876 struct btrfs_fs_info *fs_info = root->fs_info;
7877 bool final_tried = num_bytes == min_alloc_size;
7881 flags = get_alloc_profile_by_root(root, is_data);
7883 WARN_ON(num_bytes < fs_info->sectorsize);
7884 ret = find_free_extent(fs_info, ram_bytes, num_bytes, empty_size,
7885 hint_byte, ins, flags, delalloc);
7886 if (!ret && !is_data) {
7887 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7888 } else if (ret == -ENOSPC) {
7889 if (!final_tried && ins->offset) {
7890 num_bytes = min(num_bytes >> 1, ins->offset);
7891 num_bytes = round_down(num_bytes,
7892 fs_info->sectorsize);
7893 num_bytes = max(num_bytes, min_alloc_size);
7894 ram_bytes = num_bytes;
7895 if (num_bytes == min_alloc_size)
7898 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7899 struct btrfs_space_info *sinfo;
7901 sinfo = __find_space_info(fs_info, flags);
7903 "allocation failed flags %llu, wanted %llu",
7906 dump_space_info(fs_info, sinfo, num_bytes, 1);
7913 static int __btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7915 int pin, int delalloc)
7917 struct btrfs_block_group_cache *cache;
7920 cache = btrfs_lookup_block_group(fs_info, start);
7922 btrfs_err(fs_info, "Unable to find block group for %llu",
7928 pin_down_extent(fs_info, cache, start, len, 1);
7930 if (btrfs_test_opt(fs_info, DISCARD))
7931 ret = btrfs_discard_extent(fs_info, start, len, NULL);
7932 btrfs_add_free_space(cache, start, len);
7933 btrfs_free_reserved_bytes(cache, len, delalloc);
7934 trace_btrfs_reserved_extent_free(fs_info, start, len);
7937 btrfs_put_block_group(cache);
7941 int btrfs_free_reserved_extent(struct btrfs_fs_info *fs_info,
7942 u64 start, u64 len, int delalloc)
7944 return __btrfs_free_reserved_extent(fs_info, start, len, 0, delalloc);
7947 int btrfs_free_and_pin_reserved_extent(struct btrfs_fs_info *fs_info,
7950 return __btrfs_free_reserved_extent(fs_info, start, len, 1, 0);
7953 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7954 u64 parent, u64 root_objectid,
7955 u64 flags, u64 owner, u64 offset,
7956 struct btrfs_key *ins, int ref_mod)
7958 struct btrfs_fs_info *fs_info = trans->fs_info;
7960 struct btrfs_extent_item *extent_item;
7961 struct btrfs_extent_inline_ref *iref;
7962 struct btrfs_path *path;
7963 struct extent_buffer *leaf;
7968 type = BTRFS_SHARED_DATA_REF_KEY;
7970 type = BTRFS_EXTENT_DATA_REF_KEY;
7972 size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7974 path = btrfs_alloc_path();
7978 path->leave_spinning = 1;
7979 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7982 btrfs_free_path(path);
7986 leaf = path->nodes[0];
7987 extent_item = btrfs_item_ptr(leaf, path->slots[0],
7988 struct btrfs_extent_item);
7989 btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7990 btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7991 btrfs_set_extent_flags(leaf, extent_item,
7992 flags | BTRFS_EXTENT_FLAG_DATA);
7994 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7995 btrfs_set_extent_inline_ref_type(leaf, iref, type);
7997 struct btrfs_shared_data_ref *ref;
7998 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7999 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8000 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8002 struct btrfs_extent_data_ref *ref;
8003 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8004 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8005 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8006 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8007 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8010 btrfs_mark_buffer_dirty(path->nodes[0]);
8011 btrfs_free_path(path);
8013 ret = remove_from_free_space_tree(trans, ins->objectid, ins->offset);
8017 ret = update_block_group(trans, fs_info, ins->objectid, ins->offset, 1);
8018 if (ret) { /* -ENOENT, logic error */
8019 btrfs_err(fs_info, "update block group failed for %llu %llu",
8020 ins->objectid, ins->offset);
8023 trace_btrfs_reserved_extent_alloc(fs_info, ins->objectid, ins->offset);
8027 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8028 struct btrfs_delayed_ref_node *node,
8029 struct btrfs_delayed_extent_op *extent_op)
8031 struct btrfs_fs_info *fs_info = trans->fs_info;
8033 struct btrfs_extent_item *extent_item;
8034 struct btrfs_key extent_key;
8035 struct btrfs_tree_block_info *block_info;
8036 struct btrfs_extent_inline_ref *iref;
8037 struct btrfs_path *path;
8038 struct extent_buffer *leaf;
8039 struct btrfs_delayed_tree_ref *ref;
8040 u32 size = sizeof(*extent_item) + sizeof(*iref);
8042 u64 flags = extent_op->flags_to_set;
8043 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8045 ref = btrfs_delayed_node_to_tree_ref(node);
8047 extent_key.objectid = node->bytenr;
8048 if (skinny_metadata) {
8049 extent_key.offset = ref->level;
8050 extent_key.type = BTRFS_METADATA_ITEM_KEY;
8051 num_bytes = fs_info->nodesize;
8053 extent_key.offset = node->num_bytes;
8054 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
8055 size += sizeof(*block_info);
8056 num_bytes = node->num_bytes;
8059 path = btrfs_alloc_path();
8061 btrfs_free_and_pin_reserved_extent(fs_info,
8062 extent_key.objectid,
8067 path->leave_spinning = 1;
8068 ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8071 btrfs_free_path(path);
8072 btrfs_free_and_pin_reserved_extent(fs_info,
8073 extent_key.objectid,
8078 leaf = path->nodes[0];
8079 extent_item = btrfs_item_ptr(leaf, path->slots[0],
8080 struct btrfs_extent_item);
8081 btrfs_set_extent_refs(leaf, extent_item, 1);
8082 btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8083 btrfs_set_extent_flags(leaf, extent_item,
8084 flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8086 if (skinny_metadata) {
8087 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8089 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8090 btrfs_set_tree_block_key(leaf, block_info, &extent_op->key);
8091 btrfs_set_tree_block_level(leaf, block_info, ref->level);
8092 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8095 if (node->type == BTRFS_SHARED_BLOCK_REF_KEY) {
8096 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8097 btrfs_set_extent_inline_ref_type(leaf, iref,
8098 BTRFS_SHARED_BLOCK_REF_KEY);
8099 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->parent);
8101 btrfs_set_extent_inline_ref_type(leaf, iref,
8102 BTRFS_TREE_BLOCK_REF_KEY);
8103 btrfs_set_extent_inline_ref_offset(leaf, iref, ref->root);
8106 btrfs_mark_buffer_dirty(leaf);
8107 btrfs_free_path(path);
8109 ret = remove_from_free_space_tree(trans, extent_key.objectid,
8114 ret = update_block_group(trans, fs_info, extent_key.objectid,
8115 fs_info->nodesize, 1);
8116 if (ret) { /* -ENOENT, logic error */
8117 btrfs_err(fs_info, "update block group failed for %llu %llu",
8118 extent_key.objectid, extent_key.offset);
8122 trace_btrfs_reserved_extent_alloc(fs_info, extent_key.objectid,
8127 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8128 struct btrfs_root *root, u64 owner,
8129 u64 offset, u64 ram_bytes,
8130 struct btrfs_key *ins)
8134 BUG_ON(root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
8136 btrfs_ref_tree_mod(root, ins->objectid, ins->offset, 0,
8137 root->root_key.objectid, owner, offset,
8138 BTRFS_ADD_DELAYED_EXTENT);
8140 ret = btrfs_add_delayed_data_ref(trans, ins->objectid,
8142 root->root_key.objectid, owner,
8144 BTRFS_ADD_DELAYED_EXTENT, NULL, NULL);
8149 * this is used by the tree logging recovery code. It records that
8150 * an extent has been allocated and makes sure to clear the free
8151 * space cache bits as well
8153 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8154 u64 root_objectid, u64 owner, u64 offset,
8155 struct btrfs_key *ins)
8157 struct btrfs_fs_info *fs_info = trans->fs_info;
8159 struct btrfs_block_group_cache *block_group;
8160 struct btrfs_space_info *space_info;
8163 * Mixed block groups will exclude before processing the log so we only
8164 * need to do the exclude dance if this fs isn't mixed.
8166 if (!btrfs_fs_incompat(fs_info, MIXED_GROUPS)) {
8167 ret = __exclude_logged_extent(fs_info, ins->objectid,
8173 block_group = btrfs_lookup_block_group(fs_info, ins->objectid);
8177 space_info = block_group->space_info;
8178 spin_lock(&space_info->lock);
8179 spin_lock(&block_group->lock);
8180 space_info->bytes_reserved += ins->offset;
8181 block_group->reserved += ins->offset;
8182 spin_unlock(&block_group->lock);
8183 spin_unlock(&space_info->lock);
8185 ret = alloc_reserved_file_extent(trans, 0, root_objectid, 0, owner,
8187 btrfs_put_block_group(block_group);
8191 static struct extent_buffer *
8192 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8193 u64 bytenr, int level, u64 owner)
8195 struct btrfs_fs_info *fs_info = root->fs_info;
8196 struct extent_buffer *buf;
8198 buf = btrfs_find_create_tree_block(fs_info, bytenr);
8203 * Extra safety check in case the extent tree is corrupted and extent
8204 * allocator chooses to use a tree block which is already used and
8207 if (buf->lock_owner == current->pid) {
8208 btrfs_err_rl(fs_info,
8209 "tree block %llu owner %llu already locked by pid=%d, extent tree corruption detected",
8210 buf->start, btrfs_header_owner(buf), current->pid);
8211 free_extent_buffer(buf);
8212 return ERR_PTR(-EUCLEAN);
8215 btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8216 btrfs_tree_lock(buf);
8217 clean_tree_block(fs_info, buf);
8218 clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8220 btrfs_set_lock_blocking(buf);
8221 set_extent_buffer_uptodate(buf);
8223 memzero_extent_buffer(buf, 0, sizeof(struct btrfs_header));
8224 btrfs_set_header_level(buf, level);
8225 btrfs_set_header_bytenr(buf, buf->start);
8226 btrfs_set_header_generation(buf, trans->transid);
8227 btrfs_set_header_backref_rev(buf, BTRFS_MIXED_BACKREF_REV);
8228 btrfs_set_header_owner(buf, owner);
8229 write_extent_buffer_fsid(buf, fs_info->fsid);
8230 write_extent_buffer_chunk_tree_uuid(buf, fs_info->chunk_tree_uuid);
8231 if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8232 buf->log_index = root->log_transid % 2;
8234 * we allow two log transactions at a time, use different
8235 * EXENT bit to differentiate dirty pages.
8237 if (buf->log_index == 0)
8238 set_extent_dirty(&root->dirty_log_pages, buf->start,
8239 buf->start + buf->len - 1, GFP_NOFS);
8241 set_extent_new(&root->dirty_log_pages, buf->start,
8242 buf->start + buf->len - 1);
8244 buf->log_index = -1;
8245 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8246 buf->start + buf->len - 1, GFP_NOFS);
8248 trans->dirty = true;
8249 /* this returns a buffer locked for blocking */
8253 static struct btrfs_block_rsv *
8254 use_block_rsv(struct btrfs_trans_handle *trans,
8255 struct btrfs_root *root, u32 blocksize)
8257 struct btrfs_fs_info *fs_info = root->fs_info;
8258 struct btrfs_block_rsv *block_rsv;
8259 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
8261 bool global_updated = false;
8263 block_rsv = get_block_rsv(trans, root);
8265 if (unlikely(block_rsv->size == 0))
8268 ret = block_rsv_use_bytes(block_rsv, blocksize);
8272 if (block_rsv->failfast)
8273 return ERR_PTR(ret);
8275 if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8276 global_updated = true;
8277 update_global_block_rsv(fs_info);
8281 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
8282 static DEFINE_RATELIMIT_STATE(_rs,
8283 DEFAULT_RATELIMIT_INTERVAL * 10,
8284 /*DEFAULT_RATELIMIT_BURST*/ 1);
8285 if (__ratelimit(&_rs))
8287 "BTRFS: block rsv returned %d\n", ret);
8290 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8291 BTRFS_RESERVE_NO_FLUSH);
8295 * If we couldn't reserve metadata bytes try and use some from
8296 * the global reserve if its space type is the same as the global
8299 if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8300 block_rsv->space_info == global_rsv->space_info) {
8301 ret = block_rsv_use_bytes(global_rsv, blocksize);
8305 return ERR_PTR(ret);
8308 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8309 struct btrfs_block_rsv *block_rsv, u32 blocksize)
8311 block_rsv_add_bytes(block_rsv, blocksize, false);
8312 block_rsv_release_bytes(fs_info, block_rsv, NULL, 0, NULL);
8316 * finds a free extent and does all the dirty work required for allocation
8317 * returns the tree buffer or an ERR_PTR on error.
8319 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8320 struct btrfs_root *root,
8321 u64 parent, u64 root_objectid,
8322 const struct btrfs_disk_key *key,
8323 int level, u64 hint,
8326 struct btrfs_fs_info *fs_info = root->fs_info;
8327 struct btrfs_key ins;
8328 struct btrfs_block_rsv *block_rsv;
8329 struct extent_buffer *buf;
8330 struct btrfs_delayed_extent_op *extent_op;
8333 u32 blocksize = fs_info->nodesize;
8334 bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
8336 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8337 if (btrfs_is_testing(fs_info)) {
8338 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8339 level, root_objectid);
8341 root->alloc_bytenr += blocksize;
8346 block_rsv = use_block_rsv(trans, root, blocksize);
8347 if (IS_ERR(block_rsv))
8348 return ERR_CAST(block_rsv);
8350 ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8351 empty_size, hint, &ins, 0, 0);
8355 buf = btrfs_init_new_buffer(trans, root, ins.objectid, level,
8359 goto out_free_reserved;
8362 if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8364 parent = ins.objectid;
8365 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8369 if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8370 extent_op = btrfs_alloc_delayed_extent_op();
8376 memcpy(&extent_op->key, key, sizeof(extent_op->key));
8378 memset(&extent_op->key, 0, sizeof(extent_op->key));
8379 extent_op->flags_to_set = flags;
8380 extent_op->update_key = skinny_metadata ? false : true;
8381 extent_op->update_flags = true;
8382 extent_op->is_data = false;
8383 extent_op->level = level;
8385 btrfs_ref_tree_mod(root, ins.objectid, ins.offset, parent,
8386 root_objectid, level, 0,
8387 BTRFS_ADD_DELAYED_EXTENT);
8388 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
8390 root_objectid, level,
8391 BTRFS_ADD_DELAYED_EXTENT,
8392 extent_op, NULL, NULL);
8394 goto out_free_delayed;
8399 btrfs_free_delayed_extent_op(extent_op);
8401 free_extent_buffer(buf);
8403 btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 0);
8405 unuse_block_rsv(fs_info, block_rsv, blocksize);
8406 return ERR_PTR(ret);
8409 struct walk_control {
8410 u64 refs[BTRFS_MAX_LEVEL];
8411 u64 flags[BTRFS_MAX_LEVEL];
8412 struct btrfs_key update_progress;
8422 #define DROP_REFERENCE 1
8423 #define UPDATE_BACKREF 2
8425 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8426 struct btrfs_root *root,
8427 struct walk_control *wc,
8428 struct btrfs_path *path)
8430 struct btrfs_fs_info *fs_info = root->fs_info;
8436 struct btrfs_key key;
8437 struct extent_buffer *eb;
8442 if (path->slots[wc->level] < wc->reada_slot) {
8443 wc->reada_count = wc->reada_count * 2 / 3;
8444 wc->reada_count = max(wc->reada_count, 2);
8446 wc->reada_count = wc->reada_count * 3 / 2;
8447 wc->reada_count = min_t(int, wc->reada_count,
8448 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
8451 eb = path->nodes[wc->level];
8452 nritems = btrfs_header_nritems(eb);
8454 for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8455 if (nread >= wc->reada_count)
8459 bytenr = btrfs_node_blockptr(eb, slot);
8460 generation = btrfs_node_ptr_generation(eb, slot);
8462 if (slot == path->slots[wc->level])
8465 if (wc->stage == UPDATE_BACKREF &&
8466 generation <= root->root_key.offset)
8469 /* We don't lock the tree block, it's OK to be racy here */
8470 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr,
8471 wc->level - 1, 1, &refs,
8473 /* We don't care about errors in readahead. */
8478 if (wc->stage == DROP_REFERENCE) {
8482 if (wc->level == 1 &&
8483 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8485 if (!wc->update_ref ||
8486 generation <= root->root_key.offset)
8488 btrfs_node_key_to_cpu(eb, &key, slot);
8489 ret = btrfs_comp_cpu_keys(&key,
8490 &wc->update_progress);
8494 if (wc->level == 1 &&
8495 (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8499 readahead_tree_block(fs_info, bytenr);
8502 wc->reada_slot = slot;
8506 * helper to process tree block while walking down the tree.
8508 * when wc->stage == UPDATE_BACKREF, this function updates
8509 * back refs for pointers in the block.
8511 * NOTE: return value 1 means we should stop walking down.
8513 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8514 struct btrfs_root *root,
8515 struct btrfs_path *path,
8516 struct walk_control *wc, int lookup_info)
8518 struct btrfs_fs_info *fs_info = root->fs_info;
8519 int level = wc->level;
8520 struct extent_buffer *eb = path->nodes[level];
8521 u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8524 if (wc->stage == UPDATE_BACKREF &&
8525 btrfs_header_owner(eb) != root->root_key.objectid)
8529 * when reference count of tree block is 1, it won't increase
8530 * again. once full backref flag is set, we never clear it.
8533 ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8534 (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8535 BUG_ON(!path->locks[level]);
8536 ret = btrfs_lookup_extent_info(trans, fs_info,
8537 eb->start, level, 1,
8540 BUG_ON(ret == -ENOMEM);
8543 BUG_ON(wc->refs[level] == 0);
8546 if (wc->stage == DROP_REFERENCE) {
8547 if (wc->refs[level] > 1)
8550 if (path->locks[level] && !wc->keep_locks) {
8551 btrfs_tree_unlock_rw(eb, path->locks[level]);
8552 path->locks[level] = 0;
8557 /* wc->stage == UPDATE_BACKREF */
8558 if (!(wc->flags[level] & flag)) {
8559 BUG_ON(!path->locks[level]);
8560 ret = btrfs_inc_ref(trans, root, eb, 1);
8561 BUG_ON(ret); /* -ENOMEM */
8562 ret = btrfs_dec_ref(trans, root, eb, 0);
8563 BUG_ON(ret); /* -ENOMEM */
8564 ret = btrfs_set_disk_extent_flags(trans, fs_info, eb->start,
8566 btrfs_header_level(eb), 0);
8567 BUG_ON(ret); /* -ENOMEM */
8568 wc->flags[level] |= flag;
8572 * the block is shared by multiple trees, so it's not good to
8573 * keep the tree lock
8575 if (path->locks[level] && level > 0) {
8576 btrfs_tree_unlock_rw(eb, path->locks[level]);
8577 path->locks[level] = 0;
8583 * helper to process tree block pointer.
8585 * when wc->stage == DROP_REFERENCE, this function checks
8586 * reference count of the block pointed to. if the block
8587 * is shared and we need update back refs for the subtree
8588 * rooted at the block, this function changes wc->stage to
8589 * UPDATE_BACKREF. if the block is shared and there is no
8590 * need to update back, this function drops the reference
8593 * NOTE: return value 1 means we should stop walking down.
8595 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8596 struct btrfs_root *root,
8597 struct btrfs_path *path,
8598 struct walk_control *wc, int *lookup_info)
8600 struct btrfs_fs_info *fs_info = root->fs_info;
8605 struct btrfs_key key;
8606 struct btrfs_key first_key;
8607 struct extent_buffer *next;
8608 int level = wc->level;
8611 bool need_account = false;
8613 generation = btrfs_node_ptr_generation(path->nodes[level],
8614 path->slots[level]);
8616 * if the lower level block was created before the snapshot
8617 * was created, we know there is no need to update back refs
8620 if (wc->stage == UPDATE_BACKREF &&
8621 generation <= root->root_key.offset) {
8626 bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8627 btrfs_node_key_to_cpu(path->nodes[level], &first_key,
8628 path->slots[level]);
8629 blocksize = fs_info->nodesize;
8631 next = find_extent_buffer(fs_info, bytenr);
8633 next = btrfs_find_create_tree_block(fs_info, bytenr);
8635 return PTR_ERR(next);
8637 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8641 btrfs_tree_lock(next);
8642 btrfs_set_lock_blocking(next);
8644 ret = btrfs_lookup_extent_info(trans, fs_info, bytenr, level - 1, 1,
8645 &wc->refs[level - 1],
8646 &wc->flags[level - 1]);
8650 if (unlikely(wc->refs[level - 1] == 0)) {
8651 btrfs_err(fs_info, "Missing references.");
8657 if (wc->stage == DROP_REFERENCE) {
8658 if (wc->refs[level - 1] > 1) {
8659 need_account = true;
8661 (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8664 if (!wc->update_ref ||
8665 generation <= root->root_key.offset)
8668 btrfs_node_key_to_cpu(path->nodes[level], &key,
8669 path->slots[level]);
8670 ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8674 wc->stage = UPDATE_BACKREF;
8675 wc->shared_level = level - 1;
8679 (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8683 if (!btrfs_buffer_uptodate(next, generation, 0)) {
8684 btrfs_tree_unlock(next);
8685 free_extent_buffer(next);
8691 if (reada && level == 1)
8692 reada_walk_down(trans, root, wc, path);
8693 next = read_tree_block(fs_info, bytenr, generation, level - 1,
8696 return PTR_ERR(next);
8697 } else if (!extent_buffer_uptodate(next)) {
8698 free_extent_buffer(next);
8701 btrfs_tree_lock(next);
8702 btrfs_set_lock_blocking(next);
8706 ASSERT(level == btrfs_header_level(next));
8707 if (level != btrfs_header_level(next)) {
8708 btrfs_err(root->fs_info, "mismatched level");
8712 path->nodes[level] = next;
8713 path->slots[level] = 0;
8714 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8720 wc->refs[level - 1] = 0;
8721 wc->flags[level - 1] = 0;
8722 if (wc->stage == DROP_REFERENCE) {
8723 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8724 parent = path->nodes[level]->start;
8726 ASSERT(root->root_key.objectid ==
8727 btrfs_header_owner(path->nodes[level]));
8728 if (root->root_key.objectid !=
8729 btrfs_header_owner(path->nodes[level])) {
8730 btrfs_err(root->fs_info,
8731 "mismatched block owner");
8739 ret = btrfs_qgroup_trace_subtree(trans, next,
8740 generation, level - 1);
8742 btrfs_err_rl(fs_info,
8743 "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8747 ret = btrfs_free_extent(trans, root, bytenr, blocksize,
8748 parent, root->root_key.objectid,
8758 btrfs_tree_unlock(next);
8759 free_extent_buffer(next);
8765 * helper to process tree block while walking up the tree.
8767 * when wc->stage == DROP_REFERENCE, this function drops
8768 * reference count on the block.
8770 * when wc->stage == UPDATE_BACKREF, this function changes
8771 * wc->stage back to DROP_REFERENCE if we changed wc->stage
8772 * to UPDATE_BACKREF previously while processing the block.
8774 * NOTE: return value 1 means we should stop walking up.
8776 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8777 struct btrfs_root *root,
8778 struct btrfs_path *path,
8779 struct walk_control *wc)
8781 struct btrfs_fs_info *fs_info = root->fs_info;
8783 int level = wc->level;
8784 struct extent_buffer *eb = path->nodes[level];
8787 if (wc->stage == UPDATE_BACKREF) {
8788 BUG_ON(wc->shared_level < level);
8789 if (level < wc->shared_level)
8792 ret = find_next_key(path, level + 1, &wc->update_progress);
8796 wc->stage = DROP_REFERENCE;
8797 wc->shared_level = -1;
8798 path->slots[level] = 0;
8801 * check reference count again if the block isn't locked.
8802 * we should start walking down the tree again if reference
8805 if (!path->locks[level]) {
8807 btrfs_tree_lock(eb);
8808 btrfs_set_lock_blocking(eb);
8809 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8811 ret = btrfs_lookup_extent_info(trans, fs_info,
8812 eb->start, level, 1,
8816 btrfs_tree_unlock_rw(eb, path->locks[level]);
8817 path->locks[level] = 0;
8820 BUG_ON(wc->refs[level] == 0);
8821 if (wc->refs[level] == 1) {
8822 btrfs_tree_unlock_rw(eb, path->locks[level]);
8823 path->locks[level] = 0;
8829 /* wc->stage == DROP_REFERENCE */
8830 BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8832 if (wc->refs[level] == 1) {
8834 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8835 ret = btrfs_dec_ref(trans, root, eb, 1);
8837 ret = btrfs_dec_ref(trans, root, eb, 0);
8838 BUG_ON(ret); /* -ENOMEM */
8839 ret = btrfs_qgroup_trace_leaf_items(trans, eb);
8841 btrfs_err_rl(fs_info,
8842 "error %d accounting leaf items. Quota is out of sync, rescan required.",
8846 /* make block locked assertion in clean_tree_block happy */
8847 if (!path->locks[level] &&
8848 btrfs_header_generation(eb) == trans->transid) {
8849 btrfs_tree_lock(eb);
8850 btrfs_set_lock_blocking(eb);
8851 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8853 clean_tree_block(fs_info, eb);
8856 if (eb == root->node) {
8857 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8859 else if (root->root_key.objectid != btrfs_header_owner(eb))
8860 goto owner_mismatch;
8862 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8863 parent = path->nodes[level + 1]->start;
8864 else if (root->root_key.objectid !=
8865 btrfs_header_owner(path->nodes[level + 1]))
8866 goto owner_mismatch;
8869 btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8871 wc->refs[level] = 0;
8872 wc->flags[level] = 0;
8876 btrfs_err_rl(fs_info, "unexpected tree owner, have %llu expect %llu",
8877 btrfs_header_owner(eb), root->root_key.objectid);
8881 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8882 struct btrfs_root *root,
8883 struct btrfs_path *path,
8884 struct walk_control *wc)
8886 int level = wc->level;
8887 int lookup_info = 1;
8890 while (level >= 0) {
8891 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8898 if (path->slots[level] >=
8899 btrfs_header_nritems(path->nodes[level]))
8902 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8904 path->slots[level]++;
8913 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8914 struct btrfs_root *root,
8915 struct btrfs_path *path,
8916 struct walk_control *wc, int max_level)
8918 int level = wc->level;
8921 path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8922 while (level < max_level && path->nodes[level]) {
8924 if (path->slots[level] + 1 <
8925 btrfs_header_nritems(path->nodes[level])) {
8926 path->slots[level]++;
8929 ret = walk_up_proc(trans, root, path, wc);
8935 if (path->locks[level]) {
8936 btrfs_tree_unlock_rw(path->nodes[level],
8937 path->locks[level]);
8938 path->locks[level] = 0;
8940 free_extent_buffer(path->nodes[level]);
8941 path->nodes[level] = NULL;
8949 * drop a subvolume tree.
8951 * this function traverses the tree freeing any blocks that only
8952 * referenced by the tree.
8954 * when a shared tree block is found. this function decreases its
8955 * reference count by one. if update_ref is true, this function
8956 * also make sure backrefs for the shared block and all lower level
8957 * blocks are properly updated.
8959 * If called with for_reloc == 0, may exit early with -EAGAIN
8961 int btrfs_drop_snapshot(struct btrfs_root *root,
8962 struct btrfs_block_rsv *block_rsv, int update_ref,
8965 struct btrfs_fs_info *fs_info = root->fs_info;
8966 struct btrfs_path *path;
8967 struct btrfs_trans_handle *trans;
8968 struct btrfs_root *tree_root = fs_info->tree_root;
8969 struct btrfs_root_item *root_item = &root->root_item;
8970 struct walk_control *wc;
8971 struct btrfs_key key;
8975 bool root_dropped = false;
8977 btrfs_debug(fs_info, "Drop subvolume %llu", root->root_key.objectid);
8979 path = btrfs_alloc_path();
8985 wc = kzalloc(sizeof(*wc), GFP_NOFS);
8987 btrfs_free_path(path);
8992 trans = btrfs_start_transaction(tree_root, 0);
8993 if (IS_ERR(trans)) {
8994 err = PTR_ERR(trans);
8999 trans->block_rsv = block_rsv;
9001 if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9002 level = btrfs_header_level(root->node);
9003 path->nodes[level] = btrfs_lock_root_node(root);
9004 btrfs_set_lock_blocking(path->nodes[level]);
9005 path->slots[level] = 0;
9006 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9007 memset(&wc->update_progress, 0,
9008 sizeof(wc->update_progress));
9010 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9011 memcpy(&wc->update_progress, &key,
9012 sizeof(wc->update_progress));
9014 level = root_item->drop_level;
9016 path->lowest_level = level;
9017 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9018 path->lowest_level = 0;
9026 * unlock our path, this is safe because only this
9027 * function is allowed to delete this snapshot
9029 btrfs_unlock_up_safe(path, 0);
9031 level = btrfs_header_level(root->node);
9033 btrfs_tree_lock(path->nodes[level]);
9034 btrfs_set_lock_blocking(path->nodes[level]);
9035 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9037 ret = btrfs_lookup_extent_info(trans, fs_info,
9038 path->nodes[level]->start,
9039 level, 1, &wc->refs[level],
9045 BUG_ON(wc->refs[level] == 0);
9047 if (level == root_item->drop_level)
9050 btrfs_tree_unlock(path->nodes[level]);
9051 path->locks[level] = 0;
9052 WARN_ON(wc->refs[level] != 1);
9058 wc->shared_level = -1;
9059 wc->stage = DROP_REFERENCE;
9060 wc->update_ref = update_ref;
9062 wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9066 ret = walk_down_tree(trans, root, path, wc);
9072 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9079 BUG_ON(wc->stage != DROP_REFERENCE);
9083 if (wc->stage == DROP_REFERENCE) {
9085 btrfs_node_key(path->nodes[level],
9086 &root_item->drop_progress,
9087 path->slots[level]);
9088 root_item->drop_level = level;
9091 BUG_ON(wc->level == 0);
9092 if (btrfs_should_end_transaction(trans) ||
9093 (!for_reloc && btrfs_need_cleaner_sleep(fs_info))) {
9094 ret = btrfs_update_root(trans, tree_root,
9098 btrfs_abort_transaction(trans, ret);
9103 btrfs_end_transaction_throttle(trans);
9104 if (!for_reloc && btrfs_need_cleaner_sleep(fs_info)) {
9105 btrfs_debug(fs_info,
9106 "drop snapshot early exit");
9111 trans = btrfs_start_transaction(tree_root, 0);
9112 if (IS_ERR(trans)) {
9113 err = PTR_ERR(trans);
9117 trans->block_rsv = block_rsv;
9120 btrfs_release_path(path);
9124 ret = btrfs_del_root(trans, &root->root_key);
9126 btrfs_abort_transaction(trans, ret);
9131 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9132 ret = btrfs_find_root(tree_root, &root->root_key, path,
9135 btrfs_abort_transaction(trans, ret);
9138 } else if (ret > 0) {
9139 /* if we fail to delete the orphan item this time
9140 * around, it'll get picked up the next time.
9142 * The most common failure here is just -ENOENT.
9144 btrfs_del_orphan_item(trans, tree_root,
9145 root->root_key.objectid);
9149 if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9150 btrfs_add_dropped_root(trans, root);
9152 free_extent_buffer(root->node);
9153 free_extent_buffer(root->commit_root);
9154 btrfs_put_fs_root(root);
9156 root_dropped = true;
9158 btrfs_end_transaction_throttle(trans);
9161 btrfs_free_path(path);
9164 * So if we need to stop dropping the snapshot for whatever reason we
9165 * need to make sure to add it back to the dead root list so that we
9166 * keep trying to do the work later. This also cleans up roots if we
9167 * don't have it in the radix (like when we recover after a power fail
9168 * or unmount) so we don't leak memory.
9170 if (!for_reloc && !root_dropped)
9171 btrfs_add_dead_root(root);
9172 if (err && err != -EAGAIN)
9173 btrfs_handle_fs_error(fs_info, err, NULL);
9178 * drop subtree rooted at tree block 'node'.
9180 * NOTE: this function will unlock and release tree block 'node'
9181 * only used by relocation code
9183 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9184 struct btrfs_root *root,
9185 struct extent_buffer *node,
9186 struct extent_buffer *parent)
9188 struct btrfs_fs_info *fs_info = root->fs_info;
9189 struct btrfs_path *path;
9190 struct walk_control *wc;
9196 BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9198 path = btrfs_alloc_path();
9202 wc = kzalloc(sizeof(*wc), GFP_NOFS);
9204 btrfs_free_path(path);
9208 btrfs_assert_tree_locked(parent);
9209 parent_level = btrfs_header_level(parent);
9210 extent_buffer_get(parent);
9211 path->nodes[parent_level] = parent;
9212 path->slots[parent_level] = btrfs_header_nritems(parent);
9214 btrfs_assert_tree_locked(node);
9215 level = btrfs_header_level(node);
9216 path->nodes[level] = node;
9217 path->slots[level] = 0;
9218 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9220 wc->refs[parent_level] = 1;
9221 wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9223 wc->shared_level = -1;
9224 wc->stage = DROP_REFERENCE;
9227 wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(fs_info);
9230 wret = walk_down_tree(trans, root, path, wc);
9236 wret = walk_up_tree(trans, root, path, wc, parent_level);
9244 btrfs_free_path(path);
9248 static u64 update_block_group_flags(struct btrfs_fs_info *fs_info, u64 flags)
9254 * if restripe for this chunk_type is on pick target profile and
9255 * return, otherwise do the usual balance
9257 stripped = get_restripe_target(fs_info, flags);
9259 return extended_to_chunk(stripped);
9261 num_devices = fs_info->fs_devices->rw_devices;
9263 stripped = BTRFS_BLOCK_GROUP_RAID0 |
9264 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9265 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9267 if (num_devices == 1) {
9268 stripped |= BTRFS_BLOCK_GROUP_DUP;
9269 stripped = flags & ~stripped;
9271 /* turn raid0 into single device chunks */
9272 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9275 /* turn mirroring into duplication */
9276 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9277 BTRFS_BLOCK_GROUP_RAID10))
9278 return stripped | BTRFS_BLOCK_GROUP_DUP;
9280 /* they already had raid on here, just return */
9281 if (flags & stripped)
9284 stripped |= BTRFS_BLOCK_GROUP_DUP;
9285 stripped = flags & ~stripped;
9287 /* switch duplicated blocks with raid1 */
9288 if (flags & BTRFS_BLOCK_GROUP_DUP)
9289 return stripped | BTRFS_BLOCK_GROUP_RAID1;
9291 /* this is drive concat, leave it alone */
9297 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9299 struct btrfs_space_info *sinfo = cache->space_info;
9301 u64 min_allocable_bytes;
9305 * We need some metadata space and system metadata space for
9306 * allocating chunks in some corner cases until we force to set
9307 * it to be readonly.
9310 (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9312 min_allocable_bytes = SZ_1M;
9314 min_allocable_bytes = 0;
9316 spin_lock(&sinfo->lock);
9317 spin_lock(&cache->lock);
9325 num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9326 cache->bytes_super - btrfs_block_group_used(&cache->item);
9328 if (btrfs_space_info_used(sinfo, true) + num_bytes +
9329 min_allocable_bytes <= sinfo->total_bytes) {
9330 sinfo->bytes_readonly += num_bytes;
9332 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9336 spin_unlock(&cache->lock);
9337 spin_unlock(&sinfo->lock);
9341 int btrfs_inc_block_group_ro(struct btrfs_block_group_cache *cache)
9344 struct btrfs_fs_info *fs_info = cache->fs_info;
9345 struct btrfs_trans_handle *trans;
9350 trans = btrfs_join_transaction(fs_info->extent_root);
9352 return PTR_ERR(trans);
9355 * we're not allowed to set block groups readonly after the dirty
9356 * block groups cache has started writing. If it already started,
9357 * back off and let this transaction commit
9359 mutex_lock(&fs_info->ro_block_group_mutex);
9360 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9361 u64 transid = trans->transid;
9363 mutex_unlock(&fs_info->ro_block_group_mutex);
9364 btrfs_end_transaction(trans);
9366 ret = btrfs_wait_for_commit(fs_info, transid);
9373 * if we are changing raid levels, try to allocate a corresponding
9374 * block group with the new raid level.
9376 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9377 if (alloc_flags != cache->flags) {
9378 ret = do_chunk_alloc(trans, alloc_flags,
9381 * ENOSPC is allowed here, we may have enough space
9382 * already allocated at the new raid level to
9391 ret = inc_block_group_ro(cache, 0);
9394 alloc_flags = get_alloc_profile(fs_info, cache->space_info->flags);
9395 ret = do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9398 ret = inc_block_group_ro(cache, 0);
9400 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9401 alloc_flags = update_block_group_flags(fs_info, cache->flags);
9402 mutex_lock(&fs_info->chunk_mutex);
9403 check_system_chunk(trans, alloc_flags);
9404 mutex_unlock(&fs_info->chunk_mutex);
9406 mutex_unlock(&fs_info->ro_block_group_mutex);
9408 btrfs_end_transaction(trans);
9412 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
9414 u64 alloc_flags = get_alloc_profile(trans->fs_info, type);
9416 return do_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
9420 * helper to account the unused space of all the readonly block group in the
9421 * space_info. takes mirrors into account.
9423 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9425 struct btrfs_block_group_cache *block_group;
9429 /* It's df, we don't care if it's racy */
9430 if (list_empty(&sinfo->ro_bgs))
9433 spin_lock(&sinfo->lock);
9434 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9435 spin_lock(&block_group->lock);
9437 if (!block_group->ro) {
9438 spin_unlock(&block_group->lock);
9442 factor = btrfs_bg_type_to_factor(block_group->flags);
9443 free_bytes += (block_group->key.offset -
9444 btrfs_block_group_used(&block_group->item)) *
9447 spin_unlock(&block_group->lock);
9449 spin_unlock(&sinfo->lock);
9454 void btrfs_dec_block_group_ro(struct btrfs_block_group_cache *cache)
9456 struct btrfs_space_info *sinfo = cache->space_info;
9461 spin_lock(&sinfo->lock);
9462 spin_lock(&cache->lock);
9464 num_bytes = cache->key.offset - cache->reserved -
9465 cache->pinned - cache->bytes_super -
9466 btrfs_block_group_used(&cache->item);
9467 sinfo->bytes_readonly -= num_bytes;
9468 list_del_init(&cache->ro_list);
9470 spin_unlock(&cache->lock);
9471 spin_unlock(&sinfo->lock);
9475 * checks to see if its even possible to relocate this block group.
9477 * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9478 * ok to go ahead and try.
9480 int btrfs_can_relocate(struct btrfs_fs_info *fs_info, u64 bytenr)
9482 struct btrfs_root *root = fs_info->extent_root;
9483 struct btrfs_block_group_cache *block_group;
9484 struct btrfs_space_info *space_info;
9485 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
9486 struct btrfs_device *device;
9487 struct btrfs_trans_handle *trans;
9497 debug = btrfs_test_opt(fs_info, ENOSPC_DEBUG);
9499 block_group = btrfs_lookup_block_group(fs_info, bytenr);
9501 /* odd, couldn't find the block group, leave it alone */
9505 "can't find block group for bytenr %llu",
9510 min_free = btrfs_block_group_used(&block_group->item);
9512 /* no bytes used, we're good */
9516 space_info = block_group->space_info;
9517 spin_lock(&space_info->lock);
9519 full = space_info->full;
9522 * if this is the last block group we have in this space, we can't
9523 * relocate it unless we're able to allocate a new chunk below.
9525 * Otherwise, we need to make sure we have room in the space to handle
9526 * all of the extents from this block group. If we can, we're good
9528 if ((space_info->total_bytes != block_group->key.offset) &&
9529 (btrfs_space_info_used(space_info, false) + min_free <
9530 space_info->total_bytes)) {
9531 spin_unlock(&space_info->lock);
9534 spin_unlock(&space_info->lock);
9537 * ok we don't have enough space, but maybe we have free space on our
9538 * devices to allocate new chunks for relocation, so loop through our
9539 * alloc devices and guess if we have enough space. if this block
9540 * group is going to be restriped, run checks against the target
9541 * profile instead of the current one.
9553 target = get_restripe_target(fs_info, block_group->flags);
9555 index = btrfs_bg_flags_to_raid_index(extended_to_chunk(target));
9558 * this is just a balance, so if we were marked as full
9559 * we know there is no space for a new chunk
9564 "no space to alloc new chunk for block group %llu",
9565 block_group->key.objectid);
9569 index = btrfs_bg_flags_to_raid_index(block_group->flags);
9572 if (index == BTRFS_RAID_RAID10) {
9576 } else if (index == BTRFS_RAID_RAID1) {
9578 } else if (index == BTRFS_RAID_DUP) {
9581 } else if (index == BTRFS_RAID_RAID0) {
9582 dev_min = fs_devices->rw_devices;
9583 min_free = div64_u64(min_free, dev_min);
9586 /* We need to do this so that we can look at pending chunks */
9587 trans = btrfs_join_transaction(root);
9588 if (IS_ERR(trans)) {
9589 ret = PTR_ERR(trans);
9593 mutex_lock(&fs_info->chunk_mutex);
9594 list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9598 * check to make sure we can actually find a chunk with enough
9599 * space to fit our block group in.
9601 if (device->total_bytes > device->bytes_used + min_free &&
9602 !test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state)) {
9603 ret = find_free_dev_extent(trans, device, min_free,
9608 if (dev_nr >= dev_min)
9614 if (debug && ret == -1)
9616 "no space to allocate a new chunk for block group %llu",
9617 block_group->key.objectid);
9618 mutex_unlock(&fs_info->chunk_mutex);
9619 btrfs_end_transaction(trans);
9621 btrfs_put_block_group(block_group);
9625 static int find_first_block_group(struct btrfs_fs_info *fs_info,
9626 struct btrfs_path *path,
9627 struct btrfs_key *key)
9629 struct btrfs_root *root = fs_info->extent_root;
9631 struct btrfs_key found_key;
9632 struct extent_buffer *leaf;
9633 struct btrfs_block_group_item bg;
9637 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9642 slot = path->slots[0];
9643 leaf = path->nodes[0];
9644 if (slot >= btrfs_header_nritems(leaf)) {
9645 ret = btrfs_next_leaf(root, path);
9652 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9654 if (found_key.objectid >= key->objectid &&
9655 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9656 struct extent_map_tree *em_tree;
9657 struct extent_map *em;
9659 em_tree = &root->fs_info->mapping_tree.map_tree;
9660 read_lock(&em_tree->lock);
9661 em = lookup_extent_mapping(em_tree, found_key.objectid,
9663 read_unlock(&em_tree->lock);
9666 "logical %llu len %llu found bg but no related chunk",
9667 found_key.objectid, found_key.offset);
9669 } else if (em->start != found_key.objectid ||
9670 em->len != found_key.offset) {
9672 "block group %llu len %llu mismatch with chunk %llu len %llu",
9673 found_key.objectid, found_key.offset,
9674 em->start, em->len);
9677 read_extent_buffer(leaf, &bg,
9678 btrfs_item_ptr_offset(leaf, slot),
9680 flags = btrfs_block_group_flags(&bg) &
9681 BTRFS_BLOCK_GROUP_TYPE_MASK;
9683 if (flags != (em->map_lookup->type &
9684 BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9686 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
9688 found_key.offset, flags,
9689 (BTRFS_BLOCK_GROUP_TYPE_MASK &
9690 em->map_lookup->type));
9696 free_extent_map(em);
9705 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9707 struct btrfs_block_group_cache *block_group;
9711 struct inode *inode;
9713 block_group = btrfs_lookup_first_block_group(info, last);
9714 while (block_group) {
9715 wait_block_group_cache_done(block_group);
9716 spin_lock(&block_group->lock);
9717 if (block_group->iref)
9719 spin_unlock(&block_group->lock);
9720 block_group = next_block_group(info, block_group);
9729 inode = block_group->inode;
9730 block_group->iref = 0;
9731 block_group->inode = NULL;
9732 spin_unlock(&block_group->lock);
9733 ASSERT(block_group->io_ctl.inode == NULL);
9735 last = block_group->key.objectid + block_group->key.offset;
9736 btrfs_put_block_group(block_group);
9741 * Must be called only after stopping all workers, since we could have block
9742 * group caching kthreads running, and therefore they could race with us if we
9743 * freed the block groups before stopping them.
9745 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9747 struct btrfs_block_group_cache *block_group;
9748 struct btrfs_space_info *space_info;
9749 struct btrfs_caching_control *caching_ctl;
9752 down_write(&info->commit_root_sem);
9753 while (!list_empty(&info->caching_block_groups)) {
9754 caching_ctl = list_entry(info->caching_block_groups.next,
9755 struct btrfs_caching_control, list);
9756 list_del(&caching_ctl->list);
9757 put_caching_control(caching_ctl);
9759 up_write(&info->commit_root_sem);
9761 spin_lock(&info->unused_bgs_lock);
9762 while (!list_empty(&info->unused_bgs)) {
9763 block_group = list_first_entry(&info->unused_bgs,
9764 struct btrfs_block_group_cache,
9766 list_del_init(&block_group->bg_list);
9767 btrfs_put_block_group(block_group);
9769 spin_unlock(&info->unused_bgs_lock);
9771 spin_lock(&info->block_group_cache_lock);
9772 while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9773 block_group = rb_entry(n, struct btrfs_block_group_cache,
9775 rb_erase(&block_group->cache_node,
9776 &info->block_group_cache_tree);
9777 RB_CLEAR_NODE(&block_group->cache_node);
9778 spin_unlock(&info->block_group_cache_lock);
9780 down_write(&block_group->space_info->groups_sem);
9781 list_del(&block_group->list);
9782 up_write(&block_group->space_info->groups_sem);
9785 * We haven't cached this block group, which means we could
9786 * possibly have excluded extents on this block group.
9788 if (block_group->cached == BTRFS_CACHE_NO ||
9789 block_group->cached == BTRFS_CACHE_ERROR)
9790 free_excluded_extents(block_group);
9792 btrfs_remove_free_space_cache(block_group);
9793 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
9794 ASSERT(list_empty(&block_group->dirty_list));
9795 ASSERT(list_empty(&block_group->io_list));
9796 ASSERT(list_empty(&block_group->bg_list));
9797 ASSERT(atomic_read(&block_group->count) == 1);
9798 btrfs_put_block_group(block_group);
9800 spin_lock(&info->block_group_cache_lock);
9802 spin_unlock(&info->block_group_cache_lock);
9804 /* now that all the block groups are freed, go through and
9805 * free all the space_info structs. This is only called during
9806 * the final stages of unmount, and so we know nobody is
9807 * using them. We call synchronize_rcu() once before we start,
9808 * just to be on the safe side.
9812 release_global_block_rsv(info);
9814 while (!list_empty(&info->space_info)) {
9817 space_info = list_entry(info->space_info.next,
9818 struct btrfs_space_info,
9822 * Do not hide this behind enospc_debug, this is actually
9823 * important and indicates a real bug if this happens.
9825 if (WARN_ON(space_info->bytes_pinned > 0 ||
9826 space_info->bytes_reserved > 0 ||
9827 space_info->bytes_may_use > 0))
9828 dump_space_info(info, space_info, 0, 0);
9829 list_del(&space_info->list);
9830 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9831 struct kobject *kobj;
9832 kobj = space_info->block_group_kobjs[i];
9833 space_info->block_group_kobjs[i] = NULL;
9839 kobject_del(&space_info->kobj);
9840 kobject_put(&space_info->kobj);
9845 /* link_block_group will queue up kobjects to add when we're reclaim-safe */
9846 void btrfs_add_raid_kobjects(struct btrfs_fs_info *fs_info)
9848 struct btrfs_space_info *space_info;
9849 struct raid_kobject *rkobj;
9854 spin_lock(&fs_info->pending_raid_kobjs_lock);
9855 list_splice_init(&fs_info->pending_raid_kobjs, &list);
9856 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9858 list_for_each_entry(rkobj, &list, list) {
9859 space_info = __find_space_info(fs_info, rkobj->flags);
9860 index = btrfs_bg_flags_to_raid_index(rkobj->flags);
9862 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9863 "%s", get_raid_name(index));
9865 kobject_put(&rkobj->kobj);
9871 "failed to add kobject for block cache, ignoring");
9874 static void link_block_group(struct btrfs_block_group_cache *cache)
9876 struct btrfs_space_info *space_info = cache->space_info;
9877 struct btrfs_fs_info *fs_info = cache->fs_info;
9878 int index = btrfs_bg_flags_to_raid_index(cache->flags);
9881 down_write(&space_info->groups_sem);
9882 if (list_empty(&space_info->block_groups[index]))
9884 list_add_tail(&cache->list, &space_info->block_groups[index]);
9885 up_write(&space_info->groups_sem);
9888 struct raid_kobject *rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9890 btrfs_warn(cache->fs_info,
9891 "couldn't alloc memory for raid level kobject");
9894 rkobj->flags = cache->flags;
9895 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9897 spin_lock(&fs_info->pending_raid_kobjs_lock);
9898 list_add_tail(&rkobj->list, &fs_info->pending_raid_kobjs);
9899 spin_unlock(&fs_info->pending_raid_kobjs_lock);
9900 space_info->block_group_kobjs[index] = &rkobj->kobj;
9904 static struct btrfs_block_group_cache *
9905 btrfs_create_block_group_cache(struct btrfs_fs_info *fs_info,
9906 u64 start, u64 size)
9908 struct btrfs_block_group_cache *cache;
9910 cache = kzalloc(sizeof(*cache), GFP_NOFS);
9914 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9916 if (!cache->free_space_ctl) {
9921 cache->key.objectid = start;
9922 cache->key.offset = size;
9923 cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9925 cache->fs_info = fs_info;
9926 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
9927 set_free_space_tree_thresholds(cache);
9929 atomic_set(&cache->count, 1);
9930 spin_lock_init(&cache->lock);
9931 init_rwsem(&cache->data_rwsem);
9932 INIT_LIST_HEAD(&cache->list);
9933 INIT_LIST_HEAD(&cache->cluster_list);
9934 INIT_LIST_HEAD(&cache->bg_list);
9935 INIT_LIST_HEAD(&cache->ro_list);
9936 INIT_LIST_HEAD(&cache->dirty_list);
9937 INIT_LIST_HEAD(&cache->io_list);
9938 btrfs_init_free_space_ctl(cache);
9939 atomic_set(&cache->trimming, 0);
9940 mutex_init(&cache->free_space_lock);
9941 btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
9948 * Iterate all chunks and verify that each of them has the corresponding block
9951 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
9953 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
9954 struct extent_map *em;
9955 struct btrfs_block_group_cache *bg;
9960 read_lock(&map_tree->map_tree.lock);
9962 * lookup_extent_mapping will return the first extent map
9963 * intersecting the range, so setting @len to 1 is enough to
9964 * get the first chunk.
9966 em = lookup_extent_mapping(&map_tree->map_tree, start, 1);
9967 read_unlock(&map_tree->map_tree.lock);
9971 bg = btrfs_lookup_block_group(fs_info, em->start);
9974 "chunk start=%llu len=%llu doesn't have corresponding block group",
9975 em->start, em->len);
9977 free_extent_map(em);
9980 if (bg->key.objectid != em->start ||
9981 bg->key.offset != em->len ||
9982 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
9983 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
9985 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
9987 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
9988 bg->key.objectid, bg->key.offset,
9989 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
9991 free_extent_map(em);
9992 btrfs_put_block_group(bg);
9995 start = em->start + em->len;
9996 free_extent_map(em);
9997 btrfs_put_block_group(bg);
10002 int btrfs_read_block_groups(struct btrfs_fs_info *info)
10004 struct btrfs_path *path;
10006 struct btrfs_block_group_cache *cache;
10007 struct btrfs_space_info *space_info;
10008 struct btrfs_key key;
10009 struct btrfs_key found_key;
10010 struct extent_buffer *leaf;
10011 int need_clear = 0;
10016 feature = btrfs_super_incompat_flags(info->super_copy);
10017 mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10021 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10022 path = btrfs_alloc_path();
10025 path->reada = READA_FORWARD;
10027 cache_gen = btrfs_super_cache_generation(info->super_copy);
10028 if (btrfs_test_opt(info, SPACE_CACHE) &&
10029 btrfs_super_generation(info->super_copy) != cache_gen)
10031 if (btrfs_test_opt(info, CLEAR_CACHE))
10035 ret = find_first_block_group(info, path, &key);
10041 leaf = path->nodes[0];
10042 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10044 cache = btrfs_create_block_group_cache(info, found_key.objectid,
10053 * When we mount with old space cache, we need to
10054 * set BTRFS_DC_CLEAR and set dirty flag.
10056 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10057 * truncate the old free space cache inode and
10059 * b) Setting 'dirty flag' makes sure that we flush
10060 * the new space cache info onto disk.
10062 if (btrfs_test_opt(info, SPACE_CACHE))
10063 cache->disk_cache_state = BTRFS_DC_CLEAR;
10066 read_extent_buffer(leaf, &cache->item,
10067 btrfs_item_ptr_offset(leaf, path->slots[0]),
10068 sizeof(cache->item));
10069 cache->flags = btrfs_block_group_flags(&cache->item);
10071 ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10072 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10074 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10075 cache->key.objectid);
10080 key.objectid = found_key.objectid + found_key.offset;
10081 btrfs_release_path(path);
10084 * We need to exclude the super stripes now so that the space
10085 * info has super bytes accounted for, otherwise we'll think
10086 * we have more space than we actually do.
10088 ret = exclude_super_stripes(cache);
10091 * We may have excluded something, so call this just in
10094 free_excluded_extents(cache);
10095 btrfs_put_block_group(cache);
10100 * check for two cases, either we are full, and therefore
10101 * don't need to bother with the caching work since we won't
10102 * find any space, or we are empty, and we can just add all
10103 * the space in and be done with it. This saves us _alot_ of
10104 * time, particularly in the full case.
10106 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10107 cache->last_byte_to_unpin = (u64)-1;
10108 cache->cached = BTRFS_CACHE_FINISHED;
10109 free_excluded_extents(cache);
10110 } else if (btrfs_block_group_used(&cache->item) == 0) {
10111 cache->last_byte_to_unpin = (u64)-1;
10112 cache->cached = BTRFS_CACHE_FINISHED;
10113 add_new_free_space(cache, found_key.objectid,
10114 found_key.objectid +
10116 free_excluded_extents(cache);
10119 ret = btrfs_add_block_group_cache(info, cache);
10121 btrfs_remove_free_space_cache(cache);
10122 btrfs_put_block_group(cache);
10126 trace_btrfs_add_block_group(info, cache, 0);
10127 update_space_info(info, cache->flags, found_key.offset,
10128 btrfs_block_group_used(&cache->item),
10129 cache->bytes_super, &space_info);
10131 cache->space_info = space_info;
10133 link_block_group(cache);
10135 set_avail_alloc_bits(info, cache->flags);
10136 if (btrfs_chunk_readonly(info, cache->key.objectid)) {
10137 inc_block_group_ro(cache, 1);
10138 } else if (btrfs_block_group_used(&cache->item) == 0) {
10139 ASSERT(list_empty(&cache->bg_list));
10140 btrfs_mark_bg_unused(cache);
10144 list_for_each_entry_rcu(space_info, &info->space_info, list) {
10145 if (!(get_alloc_profile(info, space_info->flags) &
10146 (BTRFS_BLOCK_GROUP_RAID10 |
10147 BTRFS_BLOCK_GROUP_RAID1 |
10148 BTRFS_BLOCK_GROUP_RAID5 |
10149 BTRFS_BLOCK_GROUP_RAID6 |
10150 BTRFS_BLOCK_GROUP_DUP)))
10153 * avoid allocating from un-mirrored block group if there are
10154 * mirrored block groups.
10156 list_for_each_entry(cache,
10157 &space_info->block_groups[BTRFS_RAID_RAID0],
10159 inc_block_group_ro(cache, 1);
10160 list_for_each_entry(cache,
10161 &space_info->block_groups[BTRFS_RAID_SINGLE],
10163 inc_block_group_ro(cache, 1);
10166 btrfs_add_raid_kobjects(info);
10167 init_global_block_rsv(info);
10168 ret = check_chunk_block_group_mappings(info);
10170 btrfs_free_path(path);
10174 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
10176 struct btrfs_fs_info *fs_info = trans->fs_info;
10177 struct btrfs_block_group_cache *block_group, *tmp;
10178 struct btrfs_root *extent_root = fs_info->extent_root;
10179 struct btrfs_block_group_item item;
10180 struct btrfs_key key;
10182 bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10184 trans->can_flush_pending_bgs = false;
10185 list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10189 spin_lock(&block_group->lock);
10190 memcpy(&item, &block_group->item, sizeof(item));
10191 memcpy(&key, &block_group->key, sizeof(key));
10192 spin_unlock(&block_group->lock);
10194 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10197 btrfs_abort_transaction(trans, ret);
10198 ret = btrfs_finish_chunk_alloc(trans, key.objectid, key.offset);
10200 btrfs_abort_transaction(trans, ret);
10201 add_block_group_free_space(trans, block_group);
10202 /* already aborted the transaction if it failed. */
10204 list_del_init(&block_group->bg_list);
10206 trans->can_flush_pending_bgs = can_flush_pending_bgs;
10209 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
10210 u64 type, u64 chunk_offset, u64 size)
10212 struct btrfs_fs_info *fs_info = trans->fs_info;
10213 struct btrfs_block_group_cache *cache;
10216 btrfs_set_log_full_commit(fs_info, trans);
10218 cache = btrfs_create_block_group_cache(fs_info, chunk_offset, size);
10222 btrfs_set_block_group_used(&cache->item, bytes_used);
10223 btrfs_set_block_group_chunk_objectid(&cache->item,
10224 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
10225 btrfs_set_block_group_flags(&cache->item, type);
10227 cache->flags = type;
10228 cache->last_byte_to_unpin = (u64)-1;
10229 cache->cached = BTRFS_CACHE_FINISHED;
10230 cache->needs_free_space = 1;
10231 ret = exclude_super_stripes(cache);
10234 * We may have excluded something, so call this just in
10237 free_excluded_extents(cache);
10238 btrfs_put_block_group(cache);
10242 add_new_free_space(cache, chunk_offset, chunk_offset + size);
10244 free_excluded_extents(cache);
10246 #ifdef CONFIG_BTRFS_DEBUG
10247 if (btrfs_should_fragment_free_space(cache)) {
10248 u64 new_bytes_used = size - bytes_used;
10250 bytes_used += new_bytes_used >> 1;
10251 fragment_free_space(cache);
10255 * Ensure the corresponding space_info object is created and
10256 * assigned to our block group. We want our bg to be added to the rbtree
10257 * with its ->space_info set.
10259 cache->space_info = __find_space_info(fs_info, cache->flags);
10260 ASSERT(cache->space_info);
10262 ret = btrfs_add_block_group_cache(fs_info, cache);
10264 btrfs_remove_free_space_cache(cache);
10265 btrfs_put_block_group(cache);
10270 * Now that our block group has its ->space_info set and is inserted in
10271 * the rbtree, update the space info's counters.
10273 trace_btrfs_add_block_group(fs_info, cache, 1);
10274 update_space_info(fs_info, cache->flags, size, bytes_used,
10275 cache->bytes_super, &cache->space_info);
10276 update_global_block_rsv(fs_info);
10278 link_block_group(cache);
10280 list_add_tail(&cache->bg_list, &trans->new_bgs);
10282 set_avail_alloc_bits(fs_info, type);
10286 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10288 u64 extra_flags = chunk_to_extended(flags) &
10289 BTRFS_EXTENDED_PROFILE_MASK;
10291 write_seqlock(&fs_info->profiles_lock);
10292 if (flags & BTRFS_BLOCK_GROUP_DATA)
10293 fs_info->avail_data_alloc_bits &= ~extra_flags;
10294 if (flags & BTRFS_BLOCK_GROUP_METADATA)
10295 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10296 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10297 fs_info->avail_system_alloc_bits &= ~extra_flags;
10298 write_sequnlock(&fs_info->profiles_lock);
10301 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10302 u64 group_start, struct extent_map *em)
10304 struct btrfs_fs_info *fs_info = trans->fs_info;
10305 struct btrfs_root *root = fs_info->extent_root;
10306 struct btrfs_path *path;
10307 struct btrfs_block_group_cache *block_group;
10308 struct btrfs_free_cluster *cluster;
10309 struct btrfs_root *tree_root = fs_info->tree_root;
10310 struct btrfs_key key;
10311 struct inode *inode;
10312 struct kobject *kobj = NULL;
10316 struct btrfs_caching_control *caching_ctl = NULL;
10319 block_group = btrfs_lookup_block_group(fs_info, group_start);
10320 BUG_ON(!block_group);
10321 BUG_ON(!block_group->ro);
10323 trace_btrfs_remove_block_group(block_group);
10325 * Free the reserved super bytes from this block group before
10328 free_excluded_extents(block_group);
10329 btrfs_free_ref_tree_range(fs_info, block_group->key.objectid,
10330 block_group->key.offset);
10332 memcpy(&key, &block_group->key, sizeof(key));
10333 index = btrfs_bg_flags_to_raid_index(block_group->flags);
10334 factor = btrfs_bg_type_to_factor(block_group->flags);
10336 /* make sure this block group isn't part of an allocation cluster */
10337 cluster = &fs_info->data_alloc_cluster;
10338 spin_lock(&cluster->refill_lock);
10339 btrfs_return_cluster_to_free_space(block_group, cluster);
10340 spin_unlock(&cluster->refill_lock);
10343 * make sure this block group isn't part of a metadata
10344 * allocation cluster
10346 cluster = &fs_info->meta_alloc_cluster;
10347 spin_lock(&cluster->refill_lock);
10348 btrfs_return_cluster_to_free_space(block_group, cluster);
10349 spin_unlock(&cluster->refill_lock);
10351 path = btrfs_alloc_path();
10358 * get the inode first so any iput calls done for the io_list
10359 * aren't the final iput (no unlinks allowed now)
10361 inode = lookup_free_space_inode(fs_info, block_group, path);
10363 mutex_lock(&trans->transaction->cache_write_mutex);
10365 * make sure our free spache cache IO is done before remove the
10368 spin_lock(&trans->transaction->dirty_bgs_lock);
10369 if (!list_empty(&block_group->io_list)) {
10370 list_del_init(&block_group->io_list);
10372 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10374 spin_unlock(&trans->transaction->dirty_bgs_lock);
10375 btrfs_wait_cache_io(trans, block_group, path);
10376 btrfs_put_block_group(block_group);
10377 spin_lock(&trans->transaction->dirty_bgs_lock);
10380 if (!list_empty(&block_group->dirty_list)) {
10381 list_del_init(&block_group->dirty_list);
10382 btrfs_put_block_group(block_group);
10384 spin_unlock(&trans->transaction->dirty_bgs_lock);
10385 mutex_unlock(&trans->transaction->cache_write_mutex);
10387 if (!IS_ERR(inode)) {
10388 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10390 btrfs_add_delayed_iput(inode);
10393 clear_nlink(inode);
10394 /* One for the block groups ref */
10395 spin_lock(&block_group->lock);
10396 if (block_group->iref) {
10397 block_group->iref = 0;
10398 block_group->inode = NULL;
10399 spin_unlock(&block_group->lock);
10402 spin_unlock(&block_group->lock);
10404 /* One for our lookup ref */
10405 btrfs_add_delayed_iput(inode);
10408 key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10409 key.offset = block_group->key.objectid;
10412 ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10416 btrfs_release_path(path);
10418 ret = btrfs_del_item(trans, tree_root, path);
10421 btrfs_release_path(path);
10424 spin_lock(&fs_info->block_group_cache_lock);
10425 rb_erase(&block_group->cache_node,
10426 &fs_info->block_group_cache_tree);
10427 RB_CLEAR_NODE(&block_group->cache_node);
10429 if (fs_info->first_logical_byte == block_group->key.objectid)
10430 fs_info->first_logical_byte = (u64)-1;
10431 spin_unlock(&fs_info->block_group_cache_lock);
10433 down_write(&block_group->space_info->groups_sem);
10435 * we must use list_del_init so people can check to see if they
10436 * are still on the list after taking the semaphore
10438 list_del_init(&block_group->list);
10439 if (list_empty(&block_group->space_info->block_groups[index])) {
10440 kobj = block_group->space_info->block_group_kobjs[index];
10441 block_group->space_info->block_group_kobjs[index] = NULL;
10442 clear_avail_alloc_bits(fs_info, block_group->flags);
10444 up_write(&block_group->space_info->groups_sem);
10450 if (block_group->has_caching_ctl)
10451 caching_ctl = get_caching_control(block_group);
10452 if (block_group->cached == BTRFS_CACHE_STARTED)
10453 wait_block_group_cache_done(block_group);
10454 if (block_group->has_caching_ctl) {
10455 down_write(&fs_info->commit_root_sem);
10456 if (!caching_ctl) {
10457 struct btrfs_caching_control *ctl;
10459 list_for_each_entry(ctl,
10460 &fs_info->caching_block_groups, list)
10461 if (ctl->block_group == block_group) {
10463 refcount_inc(&caching_ctl->count);
10468 list_del_init(&caching_ctl->list);
10469 up_write(&fs_info->commit_root_sem);
10471 /* Once for the caching bgs list and once for us. */
10472 put_caching_control(caching_ctl);
10473 put_caching_control(caching_ctl);
10477 spin_lock(&trans->transaction->dirty_bgs_lock);
10478 if (!list_empty(&block_group->dirty_list)) {
10481 if (!list_empty(&block_group->io_list)) {
10484 spin_unlock(&trans->transaction->dirty_bgs_lock);
10485 btrfs_remove_free_space_cache(block_group);
10487 spin_lock(&block_group->space_info->lock);
10488 list_del_init(&block_group->ro_list);
10490 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
10491 WARN_ON(block_group->space_info->total_bytes
10492 < block_group->key.offset);
10493 WARN_ON(block_group->space_info->bytes_readonly
10494 < block_group->key.offset);
10495 WARN_ON(block_group->space_info->disk_total
10496 < block_group->key.offset * factor);
10498 block_group->space_info->total_bytes -= block_group->key.offset;
10499 block_group->space_info->bytes_readonly -= block_group->key.offset;
10500 block_group->space_info->disk_total -= block_group->key.offset * factor;
10502 spin_unlock(&block_group->space_info->lock);
10504 memcpy(&key, &block_group->key, sizeof(key));
10506 mutex_lock(&fs_info->chunk_mutex);
10507 if (!list_empty(&em->list)) {
10508 /* We're in the transaction->pending_chunks list. */
10509 free_extent_map(em);
10511 spin_lock(&block_group->lock);
10512 block_group->removed = 1;
10514 * At this point trimming can't start on this block group, because we
10515 * removed the block group from the tree fs_info->block_group_cache_tree
10516 * so no one can't find it anymore and even if someone already got this
10517 * block group before we removed it from the rbtree, they have already
10518 * incremented block_group->trimming - if they didn't, they won't find
10519 * any free space entries because we already removed them all when we
10520 * called btrfs_remove_free_space_cache().
10522 * And we must not remove the extent map from the fs_info->mapping_tree
10523 * to prevent the same logical address range and physical device space
10524 * ranges from being reused for a new block group. This is because our
10525 * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10526 * completely transactionless, so while it is trimming a range the
10527 * currently running transaction might finish and a new one start,
10528 * allowing for new block groups to be created that can reuse the same
10529 * physical device locations unless we take this special care.
10531 * There may also be an implicit trim operation if the file system
10532 * is mounted with -odiscard. The same protections must remain
10533 * in place until the extents have been discarded completely when
10534 * the transaction commit has completed.
10536 remove_em = (atomic_read(&block_group->trimming) == 0);
10538 * Make sure a trimmer task always sees the em in the pinned_chunks list
10539 * if it sees block_group->removed == 1 (needs to lock block_group->lock
10540 * before checking block_group->removed).
10544 * Our em might be in trans->transaction->pending_chunks which
10545 * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10546 * and so is the fs_info->pinned_chunks list.
10548 * So at this point we must be holding the chunk_mutex to avoid
10549 * any races with chunk allocation (more specifically at
10550 * volumes.c:contains_pending_extent()), to ensure it always
10551 * sees the em, either in the pending_chunks list or in the
10552 * pinned_chunks list.
10554 list_move_tail(&em->list, &fs_info->pinned_chunks);
10556 spin_unlock(&block_group->lock);
10559 struct extent_map_tree *em_tree;
10561 em_tree = &fs_info->mapping_tree.map_tree;
10562 write_lock(&em_tree->lock);
10564 * The em might be in the pending_chunks list, so make sure the
10565 * chunk mutex is locked, since remove_extent_mapping() will
10566 * delete us from that list.
10568 remove_extent_mapping(em_tree, em);
10569 write_unlock(&em_tree->lock);
10570 /* once for the tree */
10571 free_extent_map(em);
10574 mutex_unlock(&fs_info->chunk_mutex);
10576 ret = remove_block_group_free_space(trans, block_group);
10580 btrfs_put_block_group(block_group);
10581 btrfs_put_block_group(block_group);
10583 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10589 ret = btrfs_del_item(trans, root, path);
10591 btrfs_free_path(path);
10595 struct btrfs_trans_handle *
10596 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10597 const u64 chunk_offset)
10599 struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10600 struct extent_map *em;
10601 struct map_lookup *map;
10602 unsigned int num_items;
10604 read_lock(&em_tree->lock);
10605 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10606 read_unlock(&em_tree->lock);
10607 ASSERT(em && em->start == chunk_offset);
10610 * We need to reserve 3 + N units from the metadata space info in order
10611 * to remove a block group (done at btrfs_remove_chunk() and at
10612 * btrfs_remove_block_group()), which are used for:
10614 * 1 unit for adding the free space inode's orphan (located in the tree
10616 * 1 unit for deleting the block group item (located in the extent
10618 * 1 unit for deleting the free space item (located in tree of tree
10620 * N units for deleting N device extent items corresponding to each
10621 * stripe (located in the device tree).
10623 * In order to remove a block group we also need to reserve units in the
10624 * system space info in order to update the chunk tree (update one or
10625 * more device items and remove one chunk item), but this is done at
10626 * btrfs_remove_chunk() through a call to check_system_chunk().
10628 map = em->map_lookup;
10629 num_items = 3 + map->num_stripes;
10630 free_extent_map(em);
10632 return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10637 * Process the unused_bgs list and remove any that don't have any allocated
10638 * space inside of them.
10640 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10642 struct btrfs_block_group_cache *block_group;
10643 struct btrfs_space_info *space_info;
10644 struct btrfs_trans_handle *trans;
10647 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10650 spin_lock(&fs_info->unused_bgs_lock);
10651 while (!list_empty(&fs_info->unused_bgs)) {
10655 block_group = list_first_entry(&fs_info->unused_bgs,
10656 struct btrfs_block_group_cache,
10658 list_del_init(&block_group->bg_list);
10660 space_info = block_group->space_info;
10662 if (ret || btrfs_mixed_space_info(space_info)) {
10663 btrfs_put_block_group(block_group);
10666 spin_unlock(&fs_info->unused_bgs_lock);
10668 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10670 /* Don't want to race with allocators so take the groups_sem */
10671 down_write(&space_info->groups_sem);
10672 spin_lock(&block_group->lock);
10673 if (block_group->reserved || block_group->pinned ||
10674 btrfs_block_group_used(&block_group->item) ||
10676 list_is_singular(&block_group->list)) {
10678 * We want to bail if we made new allocations or have
10679 * outstanding allocations in this block group. We do
10680 * the ro check in case balance is currently acting on
10681 * this block group.
10683 trace_btrfs_skip_unused_block_group(block_group);
10684 spin_unlock(&block_group->lock);
10685 up_write(&space_info->groups_sem);
10688 spin_unlock(&block_group->lock);
10690 /* We don't want to force the issue, only flip if it's ok. */
10691 ret = inc_block_group_ro(block_group, 0);
10692 up_write(&space_info->groups_sem);
10699 * Want to do this before we do anything else so we can recover
10700 * properly if we fail to join the transaction.
10702 trans = btrfs_start_trans_remove_block_group(fs_info,
10703 block_group->key.objectid);
10704 if (IS_ERR(trans)) {
10705 btrfs_dec_block_group_ro(block_group);
10706 ret = PTR_ERR(trans);
10711 * We could have pending pinned extents for this block group,
10712 * just delete them, we don't care about them anymore.
10714 start = block_group->key.objectid;
10715 end = start + block_group->key.offset - 1;
10717 * Hold the unused_bg_unpin_mutex lock to avoid racing with
10718 * btrfs_finish_extent_commit(). If we are at transaction N,
10719 * another task might be running finish_extent_commit() for the
10720 * previous transaction N - 1, and have seen a range belonging
10721 * to the block group in freed_extents[] before we were able to
10722 * clear the whole block group range from freed_extents[]. This
10723 * means that task can lookup for the block group after we
10724 * unpinned it from freed_extents[] and removed it, leading to
10725 * a BUG_ON() at btrfs_unpin_extent_range().
10727 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10728 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10731 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10732 btrfs_dec_block_group_ro(block_group);
10735 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10738 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10739 btrfs_dec_block_group_ro(block_group);
10742 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10744 /* Reset pinned so btrfs_put_block_group doesn't complain */
10745 spin_lock(&space_info->lock);
10746 spin_lock(&block_group->lock);
10748 space_info->bytes_pinned -= block_group->pinned;
10749 space_info->bytes_readonly += block_group->pinned;
10750 percpu_counter_add_batch(&space_info->total_bytes_pinned,
10751 -block_group->pinned,
10752 BTRFS_TOTAL_BYTES_PINNED_BATCH);
10753 block_group->pinned = 0;
10755 spin_unlock(&block_group->lock);
10756 spin_unlock(&space_info->lock);
10758 /* DISCARD can flip during remount */
10759 trimming = btrfs_test_opt(fs_info, DISCARD);
10761 /* Implicit trim during transaction commit. */
10763 btrfs_get_block_group_trimming(block_group);
10766 * Btrfs_remove_chunk will abort the transaction if things go
10769 ret = btrfs_remove_chunk(trans, block_group->key.objectid);
10773 btrfs_put_block_group_trimming(block_group);
10778 * If we're not mounted with -odiscard, we can just forget
10779 * about this block group. Otherwise we'll need to wait
10780 * until transaction commit to do the actual discard.
10783 spin_lock(&fs_info->unused_bgs_lock);
10785 * A concurrent scrub might have added us to the list
10786 * fs_info->unused_bgs, so use a list_move operation
10787 * to add the block group to the deleted_bgs list.
10789 list_move(&block_group->bg_list,
10790 &trans->transaction->deleted_bgs);
10791 spin_unlock(&fs_info->unused_bgs_lock);
10792 btrfs_get_block_group(block_group);
10795 btrfs_end_transaction(trans);
10797 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10798 btrfs_put_block_group(block_group);
10799 spin_lock(&fs_info->unused_bgs_lock);
10801 spin_unlock(&fs_info->unused_bgs_lock);
10804 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10806 struct btrfs_super_block *disk_super;
10812 disk_super = fs_info->super_copy;
10813 if (!btrfs_super_root(disk_super))
10816 features = btrfs_super_incompat_flags(disk_super);
10817 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10820 flags = BTRFS_BLOCK_GROUP_SYSTEM;
10821 ret = create_space_info(fs_info, flags);
10826 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10827 ret = create_space_info(fs_info, flags);
10829 flags = BTRFS_BLOCK_GROUP_METADATA;
10830 ret = create_space_info(fs_info, flags);
10834 flags = BTRFS_BLOCK_GROUP_DATA;
10835 ret = create_space_info(fs_info, flags);
10841 int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
10842 u64 start, u64 end)
10844 return unpin_extent_range(fs_info, start, end, false);
10848 * It used to be that old block groups would be left around forever.
10849 * Iterating over them would be enough to trim unused space. Since we
10850 * now automatically remove them, we also need to iterate over unallocated
10853 * We don't want a transaction for this since the discard may take a
10854 * substantial amount of time. We don't require that a transaction be
10855 * running, but we do need to take a running transaction into account
10856 * to ensure that we're not discarding chunks that were released or
10857 * allocated in the current transaction.
10859 * Holding the chunks lock will prevent other threads from allocating
10860 * or releasing chunks, but it won't prevent a running transaction
10861 * from committing and releasing the memory that the pending chunks
10862 * list head uses. For that, we need to take a reference to the
10863 * transaction and hold the commit root sem. We only need to hold
10864 * it while performing the free space search since we have already
10865 * held back allocations.
10867 static int btrfs_trim_free_extents(struct btrfs_device *device,
10868 u64 minlen, u64 *trimmed)
10870 u64 start = 0, len = 0;
10875 /* Discard not supported = nothing to do. */
10876 if (!blk_queue_discard(bdev_get_queue(device->bdev)))
10879 /* Not writeable = nothing to do. */
10880 if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
10883 /* No free space = nothing to do. */
10884 if (device->total_bytes <= device->bytes_used)
10890 struct btrfs_fs_info *fs_info = device->fs_info;
10891 struct btrfs_transaction *trans;
10894 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10898 ret = down_read_killable(&fs_info->commit_root_sem);
10900 mutex_unlock(&fs_info->chunk_mutex);
10904 spin_lock(&fs_info->trans_lock);
10905 trans = fs_info->running_transaction;
10907 refcount_inc(&trans->use_count);
10908 spin_unlock(&fs_info->trans_lock);
10911 up_read(&fs_info->commit_root_sem);
10913 ret = find_free_dev_extent_start(trans, device, minlen, start,
10916 up_read(&fs_info->commit_root_sem);
10917 btrfs_put_transaction(trans);
10921 mutex_unlock(&fs_info->chunk_mutex);
10922 if (ret == -ENOSPC)
10927 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10928 mutex_unlock(&fs_info->chunk_mutex);
10936 if (fatal_signal_pending(current)) {
10937 ret = -ERESTARTSYS;
10948 * Trim the whole filesystem by:
10949 * 1) trimming the free space in each block group
10950 * 2) trimming the unallocated space on each device
10952 * This will also continue trimming even if a block group or device encounters
10953 * an error. The return value will be the last error, or 0 if nothing bad
10956 int btrfs_trim_fs(struct btrfs_fs_info *fs_info, struct fstrim_range *range)
10958 struct btrfs_block_group_cache *cache = NULL;
10959 struct btrfs_device *device;
10960 struct list_head *devices;
10966 u64 dev_failed = 0;
10971 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10972 for (; cache; cache = next_block_group(fs_info, cache)) {
10973 if (cache->key.objectid >= (range->start + range->len)) {
10974 btrfs_put_block_group(cache);
10978 start = max(range->start, cache->key.objectid);
10979 end = min(range->start + range->len,
10980 cache->key.objectid + cache->key.offset);
10982 if (end - start >= range->minlen) {
10983 if (!block_group_cache_done(cache)) {
10984 ret = cache_block_group(cache, 0);
10990 ret = wait_block_group_cache_done(cache);
10997 ret = btrfs_trim_block_group(cache,
11003 trimmed += group_trimmed;
11013 btrfs_warn(fs_info,
11014 "failed to trim %llu block group(s), last error %d",
11015 bg_failed, bg_ret);
11016 mutex_lock(&fs_info->fs_devices->device_list_mutex);
11017 devices = &fs_info->fs_devices->devices;
11018 list_for_each_entry(device, devices, dev_list) {
11019 ret = btrfs_trim_free_extents(device, range->minlen,
11027 trimmed += group_trimmed;
11029 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
11032 btrfs_warn(fs_info,
11033 "failed to trim %llu device(s), last error %d",
11034 dev_failed, dev_ret);
11035 range->len = trimmed;
11042 * btrfs_{start,end}_write_no_snapshotting() are similar to
11043 * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11044 * data into the page cache through nocow before the subvolume is snapshoted,
11045 * but flush the data into disk after the snapshot creation, or to prevent
11046 * operations while snapshotting is ongoing and that cause the snapshot to be
11047 * inconsistent (writes followed by expanding truncates for example).
11049 void btrfs_end_write_no_snapshotting(struct btrfs_root *root)
11051 percpu_counter_dec(&root->subv_writers->counter);
11052 cond_wake_up(&root->subv_writers->wait);
11055 int btrfs_start_write_no_snapshotting(struct btrfs_root *root)
11057 if (atomic_read(&root->will_be_snapshotted))
11060 percpu_counter_inc(&root->subv_writers->counter);
11062 * Make sure counter is updated before we check for snapshot creation.
11065 if (atomic_read(&root->will_be_snapshotted)) {
11066 btrfs_end_write_no_snapshotting(root);
11072 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11077 ret = btrfs_start_write_no_snapshotting(root);
11080 wait_var_event(&root->will_be_snapshotted,
11081 !atomic_read(&root->will_be_snapshotted));
11085 void btrfs_mark_bg_unused(struct btrfs_block_group_cache *bg)
11087 struct btrfs_fs_info *fs_info = bg->fs_info;
11089 spin_lock(&fs_info->unused_bgs_lock);
11090 if (list_empty(&bg->bg_list)) {
11091 btrfs_get_block_group(bg);
11092 trace_btrfs_add_unused_block_group(bg);
11093 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
11095 spin_unlock(&fs_info->unused_bgs_lock);