1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011 STRATO. All rights reserved.
7 #include <linux/rbtree.h>
8 #include <trace/events/btrfs.h>
13 #include "transaction.h"
14 #include "delayed-ref.h"
17 #include "tree-mod-log.h"
19 #include "accessors.h"
20 #include "extent-tree.h"
21 #include "relocation.h"
22 #include "tree-checker.h"
24 /* Just arbitrary numbers so we can be sure one of these happened. */
25 #define BACKREF_FOUND_SHARED 6
26 #define BACKREF_FOUND_NOT_SHARED 7
28 struct extent_inode_elem {
32 struct extent_inode_elem *next;
35 static int check_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
36 const struct btrfs_key *key,
37 const struct extent_buffer *eb,
38 const struct btrfs_file_extent_item *fi,
39 struct extent_inode_elem **eie)
41 const u64 data_len = btrfs_file_extent_num_bytes(eb, fi);
42 u64 offset = key->offset;
43 struct extent_inode_elem *e;
48 if (!ctx->ignore_extent_item_pos &&
49 !btrfs_file_extent_compression(eb, fi) &&
50 !btrfs_file_extent_encryption(eb, fi) &&
51 !btrfs_file_extent_other_encoding(eb, fi)) {
54 data_offset = btrfs_file_extent_offset(eb, fi);
56 if (ctx->extent_item_pos < data_offset ||
57 ctx->extent_item_pos >= data_offset + data_len)
59 offset += ctx->extent_item_pos - data_offset;
62 if (!ctx->indirect_ref_iterator || !ctx->cache_lookup)
65 cached = ctx->cache_lookup(eb->start, ctx->user_ctx, &root_ids,
70 for (int i = 0; i < root_count; i++) {
73 ret = ctx->indirect_ref_iterator(key->objectid, offset,
74 data_len, root_ids[i],
81 e = kmalloc(sizeof(*e), GFP_NOFS);
86 e->inum = key->objectid;
88 e->num_bytes = data_len;
94 static void free_inode_elem_list(struct extent_inode_elem *eie)
96 struct extent_inode_elem *eie_next;
98 for (; eie; eie = eie_next) {
104 static int find_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
105 const struct extent_buffer *eb,
106 struct extent_inode_elem **eie)
109 struct btrfs_key key;
110 struct btrfs_file_extent_item *fi;
117 * from the shared data ref, we only have the leaf but we need
118 * the key. thus, we must look into all items and see that we
119 * find one (some) with a reference to our extent item.
121 nritems = btrfs_header_nritems(eb);
122 for (slot = 0; slot < nritems; ++slot) {
123 btrfs_item_key_to_cpu(eb, &key, slot);
124 if (key.type != BTRFS_EXTENT_DATA_KEY)
126 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
127 extent_type = btrfs_file_extent_type(eb, fi);
128 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
130 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
131 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
132 if (disk_byte != ctx->bytenr)
135 ret = check_extent_in_eb(ctx, &key, eb, fi, eie);
136 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
144 struct rb_root_cached root;
148 #define PREFTREE_INIT { .root = RB_ROOT_CACHED, .count = 0 }
151 struct preftree direct; /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
152 struct preftree indirect; /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
153 struct preftree indirect_missing_keys;
157 * Checks for a shared extent during backref search.
159 * The share_count tracks prelim_refs (direct and indirect) having a
161 * - incremented when a ref->count transitions to >0
162 * - decremented when a ref->count transitions to <1
165 struct btrfs_backref_share_check_ctx *ctx;
166 struct btrfs_root *root;
171 * Counts number of inodes that refer to an extent (different inodes in
172 * the same root or different roots) that we could find. The sharedness
173 * check typically stops once this counter gets greater than 1, so it
174 * may not reflect the total number of inodes.
178 * The number of times we found our inode refers to the data extent we
179 * are determining the sharedness. In other words, how many file extent
180 * items we could find for our inode that point to our target data
181 * extent. The value we get here after finishing the extent sharedness
182 * check may be smaller than reality, but if it ends up being greater
183 * than 1, then we know for sure the inode has multiple file extent
184 * items that point to our inode, and we can safely assume it's useful
185 * to cache the sharedness check result.
188 bool have_delayed_delete_refs;
191 static inline int extent_is_shared(struct share_check *sc)
193 return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
196 static struct kmem_cache *btrfs_prelim_ref_cache;
198 int __init btrfs_prelim_ref_init(void)
200 btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
201 sizeof(struct prelim_ref),
205 if (!btrfs_prelim_ref_cache)
210 void __cold btrfs_prelim_ref_exit(void)
212 kmem_cache_destroy(btrfs_prelim_ref_cache);
215 static void free_pref(struct prelim_ref *ref)
217 kmem_cache_free(btrfs_prelim_ref_cache, ref);
221 * Return 0 when both refs are for the same block (and can be merged).
222 * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
223 * indicates a 'higher' block.
225 static int prelim_ref_compare(struct prelim_ref *ref1,
226 struct prelim_ref *ref2)
228 if (ref1->level < ref2->level)
230 if (ref1->level > ref2->level)
232 if (ref1->root_id < ref2->root_id)
234 if (ref1->root_id > ref2->root_id)
236 if (ref1->key_for_search.type < ref2->key_for_search.type)
238 if (ref1->key_for_search.type > ref2->key_for_search.type)
240 if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
242 if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
244 if (ref1->key_for_search.offset < ref2->key_for_search.offset)
246 if (ref1->key_for_search.offset > ref2->key_for_search.offset)
248 if (ref1->parent < ref2->parent)
250 if (ref1->parent > ref2->parent)
256 static void update_share_count(struct share_check *sc, int oldcount,
257 int newcount, struct prelim_ref *newref)
259 if ((!sc) || (oldcount == 0 && newcount < 1))
262 if (oldcount > 0 && newcount < 1)
264 else if (oldcount < 1 && newcount > 0)
267 if (newref->root_id == sc->root->root_key.objectid &&
268 newref->wanted_disk_byte == sc->data_bytenr &&
269 newref->key_for_search.objectid == sc->inum)
270 sc->self_ref_count += newref->count;
274 * Add @newref to the @root rbtree, merging identical refs.
276 * Callers should assume that newref has been freed after calling.
278 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
279 struct preftree *preftree,
280 struct prelim_ref *newref,
281 struct share_check *sc)
283 struct rb_root_cached *root;
285 struct rb_node *parent = NULL;
286 struct prelim_ref *ref;
288 bool leftmost = true;
290 root = &preftree->root;
291 p = &root->rb_root.rb_node;
295 ref = rb_entry(parent, struct prelim_ref, rbnode);
296 result = prelim_ref_compare(ref, newref);
299 } else if (result > 0) {
303 /* Identical refs, merge them and free @newref */
304 struct extent_inode_elem *eie = ref->inode_list;
306 while (eie && eie->next)
310 ref->inode_list = newref->inode_list;
312 eie->next = newref->inode_list;
313 trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
316 * A delayed ref can have newref->count < 0.
317 * The ref->count is updated to follow any
318 * BTRFS_[ADD|DROP]_DELAYED_REF actions.
320 update_share_count(sc, ref->count,
321 ref->count + newref->count, newref);
322 ref->count += newref->count;
328 update_share_count(sc, 0, newref->count, newref);
330 trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
331 rb_link_node(&newref->rbnode, parent, p);
332 rb_insert_color_cached(&newref->rbnode, root, leftmost);
336 * Release the entire tree. We don't care about internal consistency so
337 * just free everything and then reset the tree root.
339 static void prelim_release(struct preftree *preftree)
341 struct prelim_ref *ref, *next_ref;
343 rbtree_postorder_for_each_entry_safe(ref, next_ref,
344 &preftree->root.rb_root, rbnode) {
345 free_inode_elem_list(ref->inode_list);
349 preftree->root = RB_ROOT_CACHED;
354 * the rules for all callers of this function are:
355 * - obtaining the parent is the goal
356 * - if you add a key, you must know that it is a correct key
357 * - if you cannot add the parent or a correct key, then we will look into the
358 * block later to set a correct key
362 * backref type | shared | indirect | shared | indirect
363 * information | tree | tree | data | data
364 * --------------------+--------+----------+--------+----------
365 * parent logical | y | - | - | -
366 * key to resolve | - | y | y | y
367 * tree block logical | - | - | - | -
368 * root for resolving | y | y | y | y
370 * - column 1: we've the parent -> done
371 * - column 2, 3, 4: we use the key to find the parent
373 * on disk refs (inline or keyed)
374 * ==============================
375 * backref type | shared | indirect | shared | indirect
376 * information | tree | tree | data | data
377 * --------------------+--------+----------+--------+----------
378 * parent logical | y | - | y | -
379 * key to resolve | - | - | - | y
380 * tree block logical | y | y | y | y
381 * root for resolving | - | y | y | y
383 * - column 1, 3: we've the parent -> done
384 * - column 2: we take the first key from the block to find the parent
385 * (see add_missing_keys)
386 * - column 4: we use the key to find the parent
388 * additional information that's available but not required to find the parent
389 * block might help in merging entries to gain some speed.
391 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
392 struct preftree *preftree, u64 root_id,
393 const struct btrfs_key *key, int level, u64 parent,
394 u64 wanted_disk_byte, int count,
395 struct share_check *sc, gfp_t gfp_mask)
397 struct prelim_ref *ref;
399 if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
402 ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
406 ref->root_id = root_id;
408 ref->key_for_search = *key;
410 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
412 ref->inode_list = NULL;
415 ref->parent = parent;
416 ref->wanted_disk_byte = wanted_disk_byte;
417 prelim_ref_insert(fs_info, preftree, ref, sc);
418 return extent_is_shared(sc);
421 /* direct refs use root == 0, key == NULL */
422 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
423 struct preftrees *preftrees, int level, u64 parent,
424 u64 wanted_disk_byte, int count,
425 struct share_check *sc, gfp_t gfp_mask)
427 return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
428 parent, wanted_disk_byte, count, sc, gfp_mask);
431 /* indirect refs use parent == 0 */
432 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
433 struct preftrees *preftrees, u64 root_id,
434 const struct btrfs_key *key, int level,
435 u64 wanted_disk_byte, int count,
436 struct share_check *sc, gfp_t gfp_mask)
438 struct preftree *tree = &preftrees->indirect;
441 tree = &preftrees->indirect_missing_keys;
442 return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
443 wanted_disk_byte, count, sc, gfp_mask);
446 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
448 struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
449 struct rb_node *parent = NULL;
450 struct prelim_ref *ref = NULL;
451 struct prelim_ref target = {};
454 target.parent = bytenr;
458 ref = rb_entry(parent, struct prelim_ref, rbnode);
459 result = prelim_ref_compare(ref, &target);
471 static int add_all_parents(struct btrfs_backref_walk_ctx *ctx,
472 struct btrfs_root *root, struct btrfs_path *path,
473 struct ulist *parents,
474 struct preftrees *preftrees, struct prelim_ref *ref,
479 struct extent_buffer *eb;
480 struct btrfs_key key;
481 struct btrfs_key *key_for_search = &ref->key_for_search;
482 struct btrfs_file_extent_item *fi;
483 struct extent_inode_elem *eie = NULL, *old = NULL;
485 u64 wanted_disk_byte = ref->wanted_disk_byte;
491 eb = path->nodes[level];
492 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
499 * 1. We normally enter this function with the path already pointing to
500 * the first item to check. But sometimes, we may enter it with
502 * 2. We are searching for normal backref but bytenr of this leaf
503 * matches shared data backref
504 * 3. The leaf owner is not equal to the root we are searching
506 * For these cases, go to the next leaf before we continue.
509 if (path->slots[0] >= btrfs_header_nritems(eb) ||
510 is_shared_data_backref(preftrees, eb->start) ||
511 ref->root_id != btrfs_header_owner(eb)) {
512 if (ctx->time_seq == BTRFS_SEQ_LAST)
513 ret = btrfs_next_leaf(root, path);
515 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
518 while (!ret && count < ref->count) {
520 slot = path->slots[0];
522 btrfs_item_key_to_cpu(eb, &key, slot);
524 if (key.objectid != key_for_search->objectid ||
525 key.type != BTRFS_EXTENT_DATA_KEY)
529 * We are searching for normal backref but bytenr of this leaf
530 * matches shared data backref, OR
531 * the leaf owner is not equal to the root we are searching for
534 (is_shared_data_backref(preftrees, eb->start) ||
535 ref->root_id != btrfs_header_owner(eb))) {
536 if (ctx->time_seq == BTRFS_SEQ_LAST)
537 ret = btrfs_next_leaf(root, path);
539 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
542 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
543 type = btrfs_file_extent_type(eb, fi);
544 if (type == BTRFS_FILE_EXTENT_INLINE)
546 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
547 data_offset = btrfs_file_extent_offset(eb, fi);
549 if (disk_byte == wanted_disk_byte) {
552 if (ref->key_for_search.offset == key.offset - data_offset)
556 if (!ctx->skip_inode_ref_list) {
557 ret = check_extent_in_eb(ctx, &key, eb, fi, &eie);
558 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
564 ret = ulist_add_merge_ptr(parents, eb->start,
565 eie, (void **)&old, GFP_NOFS);
568 if (!ret && !ctx->skip_inode_ref_list) {
576 if (ctx->time_seq == BTRFS_SEQ_LAST)
577 ret = btrfs_next_item(root, path);
579 ret = btrfs_next_old_item(root, path, ctx->time_seq);
582 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
583 free_inode_elem_list(eie);
591 * resolve an indirect backref in the form (root_id, key, level)
592 * to a logical address
594 static int resolve_indirect_ref(struct btrfs_backref_walk_ctx *ctx,
595 struct btrfs_path *path,
596 struct preftrees *preftrees,
597 struct prelim_ref *ref, struct ulist *parents)
599 struct btrfs_root *root;
600 struct extent_buffer *eb;
603 int level = ref->level;
604 struct btrfs_key search_key = ref->key_for_search;
607 * If we're search_commit_root we could possibly be holding locks on
608 * other tree nodes. This happens when qgroups does backref walks when
609 * adding new delayed refs. To deal with this we need to look in cache
610 * for the root, and if we don't find it then we need to search the
611 * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
614 if (path->search_commit_root)
615 root = btrfs_get_fs_root_commit_root(ctx->fs_info, path, ref->root_id);
617 root = btrfs_get_fs_root(ctx->fs_info, ref->root_id, false);
623 if (!path->search_commit_root &&
624 test_bit(BTRFS_ROOT_DELETING, &root->state)) {
629 if (btrfs_is_testing(ctx->fs_info)) {
634 if (path->search_commit_root)
635 root_level = btrfs_header_level(root->commit_root);
636 else if (ctx->time_seq == BTRFS_SEQ_LAST)
637 root_level = btrfs_header_level(root->node);
639 root_level = btrfs_old_root_level(root, ctx->time_seq);
641 if (root_level + 1 == level)
645 * We can often find data backrefs with an offset that is too large
646 * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
647 * subtracting a file's offset with the data offset of its
648 * corresponding extent data item. This can happen for example in the
651 * So if we detect such case we set the search key's offset to zero to
652 * make sure we will find the matching file extent item at
653 * add_all_parents(), otherwise we will miss it because the offset
654 * taken form the backref is much larger then the offset of the file
655 * extent item. This can make us scan a very large number of file
656 * extent items, but at least it will not make us miss any.
658 * This is an ugly workaround for a behaviour that should have never
659 * existed, but it does and a fix for the clone ioctl would touch a lot
660 * of places, cause backwards incompatibility and would not fix the
661 * problem for extents cloned with older kernels.
663 if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
664 search_key.offset >= LLONG_MAX)
665 search_key.offset = 0;
666 path->lowest_level = level;
667 if (ctx->time_seq == BTRFS_SEQ_LAST)
668 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
670 ret = btrfs_search_old_slot(root, &search_key, path, ctx->time_seq);
672 btrfs_debug(ctx->fs_info,
673 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
674 ref->root_id, level, ref->count, ret,
675 ref->key_for_search.objectid, ref->key_for_search.type,
676 ref->key_for_search.offset);
680 eb = path->nodes[level];
682 if (WARN_ON(!level)) {
687 eb = path->nodes[level];
690 ret = add_all_parents(ctx, root, path, parents, preftrees, ref, level);
692 btrfs_put_root(root);
694 path->lowest_level = 0;
695 btrfs_release_path(path);
699 static struct extent_inode_elem *
700 unode_aux_to_inode_list(struct ulist_node *node)
704 return (struct extent_inode_elem *)(uintptr_t)node->aux;
707 static void free_leaf_list(struct ulist *ulist)
709 struct ulist_node *node;
710 struct ulist_iterator uiter;
712 ULIST_ITER_INIT(&uiter);
713 while ((node = ulist_next(ulist, &uiter)))
714 free_inode_elem_list(unode_aux_to_inode_list(node));
720 * We maintain three separate rbtrees: one for direct refs, one for
721 * indirect refs which have a key, and one for indirect refs which do not
722 * have a key. Each tree does merge on insertion.
724 * Once all of the references are located, we iterate over the tree of
725 * indirect refs with missing keys. An appropriate key is located and
726 * the ref is moved onto the tree for indirect refs. After all missing
727 * keys are thus located, we iterate over the indirect ref tree, resolve
728 * each reference, and then insert the resolved reference onto the
729 * direct tree (merging there too).
731 * New backrefs (i.e., for parent nodes) are added to the appropriate
732 * rbtree as they are encountered. The new backrefs are subsequently
735 static int resolve_indirect_refs(struct btrfs_backref_walk_ctx *ctx,
736 struct btrfs_path *path,
737 struct preftrees *preftrees,
738 struct share_check *sc)
742 struct ulist *parents;
743 struct ulist_node *node;
744 struct ulist_iterator uiter;
745 struct rb_node *rnode;
747 parents = ulist_alloc(GFP_NOFS);
752 * We could trade memory usage for performance here by iterating
753 * the tree, allocating new refs for each insertion, and then
754 * freeing the entire indirect tree when we're done. In some test
755 * cases, the tree can grow quite large (~200k objects).
757 while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
758 struct prelim_ref *ref;
760 ref = rb_entry(rnode, struct prelim_ref, rbnode);
761 if (WARN(ref->parent,
762 "BUG: direct ref found in indirect tree")) {
767 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
768 preftrees->indirect.count--;
770 if (ref->count == 0) {
775 if (sc && ref->root_id != sc->root->root_key.objectid) {
777 ret = BACKREF_FOUND_SHARED;
780 err = resolve_indirect_ref(ctx, path, preftrees, ref, parents);
782 * we can only tolerate ENOENT,otherwise,we should catch error
783 * and return directly.
785 if (err == -ENOENT) {
786 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref,
795 /* we put the first parent into the ref at hand */
796 ULIST_ITER_INIT(&uiter);
797 node = ulist_next(parents, &uiter);
798 ref->parent = node ? node->val : 0;
799 ref->inode_list = unode_aux_to_inode_list(node);
801 /* Add a prelim_ref(s) for any other parent(s). */
802 while ((node = ulist_next(parents, &uiter))) {
803 struct prelim_ref *new_ref;
805 new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
812 memcpy(new_ref, ref, sizeof(*ref));
813 new_ref->parent = node->val;
814 new_ref->inode_list = unode_aux_to_inode_list(node);
815 prelim_ref_insert(ctx->fs_info, &preftrees->direct,
820 * Now it's a direct ref, put it in the direct tree. We must
821 * do this last because the ref could be merged/freed here.
823 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref, NULL);
825 ulist_reinit(parents);
830 * We may have inode lists attached to refs in the parents ulist, so we
831 * must free them before freeing the ulist and its refs.
833 free_leaf_list(parents);
838 * read tree blocks and add keys where required.
840 static int add_missing_keys(struct btrfs_fs_info *fs_info,
841 struct preftrees *preftrees, bool lock)
843 struct prelim_ref *ref;
844 struct extent_buffer *eb;
845 struct preftree *tree = &preftrees->indirect_missing_keys;
846 struct rb_node *node;
848 while ((node = rb_first_cached(&tree->root))) {
849 struct btrfs_tree_parent_check check = { 0 };
851 ref = rb_entry(node, struct prelim_ref, rbnode);
852 rb_erase_cached(node, &tree->root);
854 BUG_ON(ref->parent); /* should not be a direct ref */
855 BUG_ON(ref->key_for_search.type);
856 BUG_ON(!ref->wanted_disk_byte);
858 check.level = ref->level - 1;
859 check.owner_root = ref->root_id;
861 eb = read_tree_block(fs_info, ref->wanted_disk_byte, &check);
866 if (!extent_buffer_uptodate(eb)) {
868 free_extent_buffer(eb);
873 btrfs_tree_read_lock(eb);
874 if (btrfs_header_level(eb) == 0)
875 btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
877 btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
879 btrfs_tree_read_unlock(eb);
880 free_extent_buffer(eb);
881 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
888 * add all currently queued delayed refs from this head whose seq nr is
889 * smaller or equal that seq to the list
891 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
892 struct btrfs_delayed_ref_head *head, u64 seq,
893 struct preftrees *preftrees, struct share_check *sc)
895 struct btrfs_delayed_ref_node *node;
896 struct btrfs_key key;
901 spin_lock(&head->lock);
902 for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
903 node = rb_entry(n, struct btrfs_delayed_ref_node,
908 switch (node->action) {
909 case BTRFS_ADD_DELAYED_EXTENT:
910 case BTRFS_UPDATE_DELAYED_HEAD:
913 case BTRFS_ADD_DELAYED_REF:
914 count = node->ref_mod;
916 case BTRFS_DROP_DELAYED_REF:
917 count = node->ref_mod * -1;
922 switch (node->type) {
923 case BTRFS_TREE_BLOCK_REF_KEY: {
924 /* NORMAL INDIRECT METADATA backref */
925 struct btrfs_delayed_tree_ref *ref;
926 struct btrfs_key *key_ptr = NULL;
928 if (head->extent_op && head->extent_op->update_key) {
929 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
933 ref = btrfs_delayed_node_to_tree_ref(node);
934 ret = add_indirect_ref(fs_info, preftrees, ref->root,
935 key_ptr, ref->level + 1,
936 node->bytenr, count, sc,
940 case BTRFS_SHARED_BLOCK_REF_KEY: {
941 /* SHARED DIRECT METADATA backref */
942 struct btrfs_delayed_tree_ref *ref;
944 ref = btrfs_delayed_node_to_tree_ref(node);
946 ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
947 ref->parent, node->bytenr, count,
951 case BTRFS_EXTENT_DATA_REF_KEY: {
952 /* NORMAL INDIRECT DATA backref */
953 struct btrfs_delayed_data_ref *ref;
954 ref = btrfs_delayed_node_to_data_ref(node);
956 key.objectid = ref->objectid;
957 key.type = BTRFS_EXTENT_DATA_KEY;
958 key.offset = ref->offset;
961 * If we have a share check context and a reference for
962 * another inode, we can't exit immediately. This is
963 * because even if this is a BTRFS_ADD_DELAYED_REF
964 * reference we may find next a BTRFS_DROP_DELAYED_REF
965 * which cancels out this ADD reference.
967 * If this is a DROP reference and there was no previous
968 * ADD reference, then we need to signal that when we
969 * process references from the extent tree (through
970 * add_inline_refs() and add_keyed_refs()), we should
971 * not exit early if we find a reference for another
972 * inode, because one of the delayed DROP references
973 * may cancel that reference in the extent tree.
976 sc->have_delayed_delete_refs = true;
978 ret = add_indirect_ref(fs_info, preftrees, ref->root,
979 &key, 0, node->bytenr, count, sc,
983 case BTRFS_SHARED_DATA_REF_KEY: {
984 /* SHARED DIRECT FULL backref */
985 struct btrfs_delayed_data_ref *ref;
987 ref = btrfs_delayed_node_to_data_ref(node);
989 ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
990 node->bytenr, count, sc,
998 * We must ignore BACKREF_FOUND_SHARED until all delayed
999 * refs have been checked.
1001 if (ret && (ret != BACKREF_FOUND_SHARED))
1005 ret = extent_is_shared(sc);
1007 spin_unlock(&head->lock);
1012 * add all inline backrefs for bytenr to the list
1014 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1016 static int add_inline_refs(struct btrfs_backref_walk_ctx *ctx,
1017 struct btrfs_path *path,
1018 int *info_level, struct preftrees *preftrees,
1019 struct share_check *sc)
1023 struct extent_buffer *leaf;
1024 struct btrfs_key key;
1025 struct btrfs_key found_key;
1028 struct btrfs_extent_item *ei;
1033 * enumerate all inline refs
1035 leaf = path->nodes[0];
1036 slot = path->slots[0];
1038 item_size = btrfs_item_size(leaf, slot);
1039 BUG_ON(item_size < sizeof(*ei));
1041 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
1043 if (ctx->check_extent_item) {
1044 ret = ctx->check_extent_item(ctx->bytenr, ei, leaf, ctx->user_ctx);
1049 flags = btrfs_extent_flags(leaf, ei);
1050 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1052 ptr = (unsigned long)(ei + 1);
1053 end = (unsigned long)ei + item_size;
1055 if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
1056 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1057 struct btrfs_tree_block_info *info;
1059 info = (struct btrfs_tree_block_info *)ptr;
1060 *info_level = btrfs_tree_block_level(leaf, info);
1061 ptr += sizeof(struct btrfs_tree_block_info);
1063 } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1064 *info_level = found_key.offset;
1066 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1070 struct btrfs_extent_inline_ref *iref;
1074 iref = (struct btrfs_extent_inline_ref *)ptr;
1075 type = btrfs_get_extent_inline_ref_type(leaf, iref,
1076 BTRFS_REF_TYPE_ANY);
1077 if (type == BTRFS_REF_TYPE_INVALID)
1080 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1083 case BTRFS_SHARED_BLOCK_REF_KEY:
1084 ret = add_direct_ref(ctx->fs_info, preftrees,
1085 *info_level + 1, offset,
1086 ctx->bytenr, 1, NULL, GFP_NOFS);
1088 case BTRFS_SHARED_DATA_REF_KEY: {
1089 struct btrfs_shared_data_ref *sdref;
1092 sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1093 count = btrfs_shared_data_ref_count(leaf, sdref);
1095 ret = add_direct_ref(ctx->fs_info, preftrees, 0, offset,
1096 ctx->bytenr, count, sc, GFP_NOFS);
1099 case BTRFS_TREE_BLOCK_REF_KEY:
1100 ret = add_indirect_ref(ctx->fs_info, preftrees, offset,
1101 NULL, *info_level + 1,
1102 ctx->bytenr, 1, NULL, GFP_NOFS);
1104 case BTRFS_EXTENT_DATA_REF_KEY: {
1105 struct btrfs_extent_data_ref *dref;
1109 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1110 count = btrfs_extent_data_ref_count(leaf, dref);
1111 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1113 key.type = BTRFS_EXTENT_DATA_KEY;
1114 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1116 if (sc && key.objectid != sc->inum &&
1117 !sc->have_delayed_delete_refs) {
1118 ret = BACKREF_FOUND_SHARED;
1122 root = btrfs_extent_data_ref_root(leaf, dref);
1124 if (!ctx->skip_data_ref ||
1125 !ctx->skip_data_ref(root, key.objectid, key.offset,
1127 ret = add_indirect_ref(ctx->fs_info, preftrees,
1128 root, &key, 0, ctx->bytenr,
1129 count, sc, GFP_NOFS);
1137 ptr += btrfs_extent_inline_ref_size(type);
1144 * add all non-inline backrefs for bytenr to the list
1146 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1148 static int add_keyed_refs(struct btrfs_backref_walk_ctx *ctx,
1149 struct btrfs_root *extent_root,
1150 struct btrfs_path *path,
1151 int info_level, struct preftrees *preftrees,
1152 struct share_check *sc)
1154 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1157 struct extent_buffer *leaf;
1158 struct btrfs_key key;
1161 ret = btrfs_next_item(extent_root, path);
1169 slot = path->slots[0];
1170 leaf = path->nodes[0];
1171 btrfs_item_key_to_cpu(leaf, &key, slot);
1173 if (key.objectid != ctx->bytenr)
1175 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1177 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1181 case BTRFS_SHARED_BLOCK_REF_KEY:
1182 /* SHARED DIRECT METADATA backref */
1183 ret = add_direct_ref(fs_info, preftrees,
1184 info_level + 1, key.offset,
1185 ctx->bytenr, 1, NULL, GFP_NOFS);
1187 case BTRFS_SHARED_DATA_REF_KEY: {
1188 /* SHARED DIRECT FULL backref */
1189 struct btrfs_shared_data_ref *sdref;
1192 sdref = btrfs_item_ptr(leaf, slot,
1193 struct btrfs_shared_data_ref);
1194 count = btrfs_shared_data_ref_count(leaf, sdref);
1195 ret = add_direct_ref(fs_info, preftrees, 0,
1196 key.offset, ctx->bytenr, count,
1200 case BTRFS_TREE_BLOCK_REF_KEY:
1201 /* NORMAL INDIRECT METADATA backref */
1202 ret = add_indirect_ref(fs_info, preftrees, key.offset,
1203 NULL, info_level + 1, ctx->bytenr,
1206 case BTRFS_EXTENT_DATA_REF_KEY: {
1207 /* NORMAL INDIRECT DATA backref */
1208 struct btrfs_extent_data_ref *dref;
1212 dref = btrfs_item_ptr(leaf, slot,
1213 struct btrfs_extent_data_ref);
1214 count = btrfs_extent_data_ref_count(leaf, dref);
1215 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1217 key.type = BTRFS_EXTENT_DATA_KEY;
1218 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1220 if (sc && key.objectid != sc->inum &&
1221 !sc->have_delayed_delete_refs) {
1222 ret = BACKREF_FOUND_SHARED;
1226 root = btrfs_extent_data_ref_root(leaf, dref);
1228 if (!ctx->skip_data_ref ||
1229 !ctx->skip_data_ref(root, key.objectid, key.offset,
1231 ret = add_indirect_ref(fs_info, preftrees, root,
1232 &key, 0, ctx->bytenr,
1233 count, sc, GFP_NOFS);
1248 * The caller has joined a transaction or is holding a read lock on the
1249 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1250 * snapshot field changing while updating or checking the cache.
1252 static bool lookup_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1253 struct btrfs_root *root,
1254 u64 bytenr, int level, bool *is_shared)
1256 const struct btrfs_fs_info *fs_info = root->fs_info;
1257 struct btrfs_backref_shared_cache_entry *entry;
1259 if (!current->journal_info)
1260 lockdep_assert_held(&fs_info->commit_root_sem);
1262 if (!ctx->use_path_cache)
1265 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1269 * Level -1 is used for the data extent, which is not reliable to cache
1270 * because its reference count can increase or decrease without us
1271 * realizing. We cache results only for extent buffers that lead from
1272 * the root node down to the leaf with the file extent item.
1276 entry = &ctx->path_cache_entries[level];
1278 /* Unused cache entry or being used for some other extent buffer. */
1279 if (entry->bytenr != bytenr)
1283 * We cached a false result, but the last snapshot generation of the
1284 * root changed, so we now have a snapshot. Don't trust the result.
1286 if (!entry->is_shared &&
1287 entry->gen != btrfs_root_last_snapshot(&root->root_item))
1291 * If we cached a true result and the last generation used for dropping
1292 * a root changed, we can not trust the result, because the dropped root
1293 * could be a snapshot sharing this extent buffer.
1295 if (entry->is_shared &&
1296 entry->gen != btrfs_get_last_root_drop_gen(fs_info))
1299 *is_shared = entry->is_shared;
1301 * If the node at this level is shared, than all nodes below are also
1302 * shared. Currently some of the nodes below may be marked as not shared
1303 * because we have just switched from one leaf to another, and switched
1304 * also other nodes above the leaf and below the current level, so mark
1308 for (int i = 0; i < level; i++) {
1309 ctx->path_cache_entries[i].is_shared = true;
1310 ctx->path_cache_entries[i].gen = entry->gen;
1318 * The caller has joined a transaction or is holding a read lock on the
1319 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1320 * snapshot field changing while updating or checking the cache.
1322 static void store_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1323 struct btrfs_root *root,
1324 u64 bytenr, int level, bool is_shared)
1326 const struct btrfs_fs_info *fs_info = root->fs_info;
1327 struct btrfs_backref_shared_cache_entry *entry;
1330 if (!current->journal_info)
1331 lockdep_assert_held(&fs_info->commit_root_sem);
1333 if (!ctx->use_path_cache)
1336 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1340 * Level -1 is used for the data extent, which is not reliable to cache
1341 * because its reference count can increase or decrease without us
1342 * realizing. We cache results only for extent buffers that lead from
1343 * the root node down to the leaf with the file extent item.
1348 gen = btrfs_get_last_root_drop_gen(fs_info);
1350 gen = btrfs_root_last_snapshot(&root->root_item);
1352 entry = &ctx->path_cache_entries[level];
1353 entry->bytenr = bytenr;
1354 entry->is_shared = is_shared;
1358 * If we found an extent buffer is shared, set the cache result for all
1359 * extent buffers below it to true. As nodes in the path are COWed,
1360 * their sharedness is moved to their children, and if a leaf is COWed,
1361 * then the sharedness of a data extent becomes direct, the refcount of
1362 * data extent is increased in the extent item at the extent tree.
1365 for (int i = 0; i < level; i++) {
1366 entry = &ctx->path_cache_entries[i];
1367 entry->is_shared = is_shared;
1374 * this adds all existing backrefs (inline backrefs, backrefs and delayed
1375 * refs) for the given bytenr to the refs list, merges duplicates and resolves
1376 * indirect refs to their parent bytenr.
1377 * When roots are found, they're added to the roots list
1379 * @ctx: Backref walking context object, must be not NULL.
1380 * @sc: If !NULL, then immediately return BACKREF_FOUND_SHARED when a
1381 * shared extent is detected.
1383 * Otherwise this returns 0 for success and <0 for an error.
1385 * FIXME some caching might speed things up
1387 static int find_parent_nodes(struct btrfs_backref_walk_ctx *ctx,
1388 struct share_check *sc)
1390 struct btrfs_root *root = btrfs_extent_root(ctx->fs_info, ctx->bytenr);
1391 struct btrfs_key key;
1392 struct btrfs_path *path;
1393 struct btrfs_delayed_ref_root *delayed_refs = NULL;
1394 struct btrfs_delayed_ref_head *head;
1397 struct prelim_ref *ref;
1398 struct rb_node *node;
1399 struct extent_inode_elem *eie = NULL;
1400 struct preftrees preftrees = {
1401 .direct = PREFTREE_INIT,
1402 .indirect = PREFTREE_INIT,
1403 .indirect_missing_keys = PREFTREE_INIT
1406 /* Roots ulist is not needed when using a sharedness check context. */
1408 ASSERT(ctx->roots == NULL);
1410 key.objectid = ctx->bytenr;
1411 key.offset = (u64)-1;
1412 if (btrfs_fs_incompat(ctx->fs_info, SKINNY_METADATA))
1413 key.type = BTRFS_METADATA_ITEM_KEY;
1415 key.type = BTRFS_EXTENT_ITEM_KEY;
1417 path = btrfs_alloc_path();
1421 path->search_commit_root = 1;
1422 path->skip_locking = 1;
1425 if (ctx->time_seq == BTRFS_SEQ_LAST)
1426 path->skip_locking = 1;
1431 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1435 /* This shouldn't happen, indicates a bug or fs corruption. */
1441 if (ctx->trans && likely(ctx->trans->type != __TRANS_DUMMY) &&
1442 ctx->time_seq != BTRFS_SEQ_LAST) {
1444 * We have a specific time_seq we care about and trans which
1445 * means we have the path lock, we need to grab the ref head and
1446 * lock it so we have a consistent view of the refs at the given
1449 delayed_refs = &ctx->trans->transaction->delayed_refs;
1450 spin_lock(&delayed_refs->lock);
1451 head = btrfs_find_delayed_ref_head(delayed_refs, ctx->bytenr);
1453 if (!mutex_trylock(&head->mutex)) {
1454 refcount_inc(&head->refs);
1455 spin_unlock(&delayed_refs->lock);
1457 btrfs_release_path(path);
1460 * Mutex was contended, block until it's
1461 * released and try again
1463 mutex_lock(&head->mutex);
1464 mutex_unlock(&head->mutex);
1465 btrfs_put_delayed_ref_head(head);
1468 spin_unlock(&delayed_refs->lock);
1469 ret = add_delayed_refs(ctx->fs_info, head, ctx->time_seq,
1471 mutex_unlock(&head->mutex);
1475 spin_unlock(&delayed_refs->lock);
1479 if (path->slots[0]) {
1480 struct extent_buffer *leaf;
1484 leaf = path->nodes[0];
1485 slot = path->slots[0];
1486 btrfs_item_key_to_cpu(leaf, &key, slot);
1487 if (key.objectid == ctx->bytenr &&
1488 (key.type == BTRFS_EXTENT_ITEM_KEY ||
1489 key.type == BTRFS_METADATA_ITEM_KEY)) {
1490 ret = add_inline_refs(ctx, path, &info_level,
1494 ret = add_keyed_refs(ctx, root, path, info_level,
1502 * If we have a share context and we reached here, it means the extent
1503 * is not directly shared (no multiple reference items for it),
1504 * otherwise we would have exited earlier with a return value of
1505 * BACKREF_FOUND_SHARED after processing delayed references or while
1506 * processing inline or keyed references from the extent tree.
1507 * The extent may however be indirectly shared through shared subtrees
1508 * as a result from creating snapshots, so we determine below what is
1509 * its parent node, in case we are dealing with a metadata extent, or
1510 * what's the leaf (or leaves), from a fs tree, that has a file extent
1511 * item pointing to it in case we are dealing with a data extent.
1513 ASSERT(extent_is_shared(sc) == 0);
1516 * If we are here for a data extent and we have a share_check structure
1517 * it means the data extent is not directly shared (does not have
1518 * multiple reference items), so we have to check if a path in the fs
1519 * tree (going from the root node down to the leaf that has the file
1520 * extent item pointing to the data extent) is shared, that is, if any
1521 * of the extent buffers in the path is referenced by other trees.
1523 if (sc && ctx->bytenr == sc->data_bytenr) {
1525 * If our data extent is from a generation more recent than the
1526 * last generation used to snapshot the root, then we know that
1527 * it can not be shared through subtrees, so we can skip
1528 * resolving indirect references, there's no point in
1529 * determining the extent buffers for the path from the fs tree
1530 * root node down to the leaf that has the file extent item that
1531 * points to the data extent.
1533 if (sc->data_extent_gen >
1534 btrfs_root_last_snapshot(&sc->root->root_item)) {
1535 ret = BACKREF_FOUND_NOT_SHARED;
1540 * If we are only determining if a data extent is shared or not
1541 * and the corresponding file extent item is located in the same
1542 * leaf as the previous file extent item, we can skip resolving
1543 * indirect references for a data extent, since the fs tree path
1544 * is the same (same leaf, so same path). We skip as long as the
1545 * cached result for the leaf is valid and only if there's only
1546 * one file extent item pointing to the data extent, because in
1547 * the case of multiple file extent items, they may be located
1548 * in different leaves and therefore we have multiple paths.
1550 if (sc->ctx->curr_leaf_bytenr == sc->ctx->prev_leaf_bytenr &&
1551 sc->self_ref_count == 1) {
1555 cached = lookup_backref_shared_cache(sc->ctx, sc->root,
1556 sc->ctx->curr_leaf_bytenr,
1560 ret = BACKREF_FOUND_SHARED;
1562 ret = BACKREF_FOUND_NOT_SHARED;
1568 btrfs_release_path(path);
1570 ret = add_missing_keys(ctx->fs_info, &preftrees, path->skip_locking == 0);
1574 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1576 ret = resolve_indirect_refs(ctx, path, &preftrees, sc);
1580 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1583 * This walks the tree of merged and resolved refs. Tree blocks are
1584 * read in as needed. Unique entries are added to the ulist, and
1585 * the list of found roots is updated.
1587 * We release the entire tree in one go before returning.
1589 node = rb_first_cached(&preftrees.direct.root);
1591 ref = rb_entry(node, struct prelim_ref, rbnode);
1592 node = rb_next(&ref->rbnode);
1594 * ref->count < 0 can happen here if there are delayed
1595 * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1596 * prelim_ref_insert() relies on this when merging
1597 * identical refs to keep the overall count correct.
1598 * prelim_ref_insert() will merge only those refs
1599 * which compare identically. Any refs having
1600 * e.g. different offsets would not be merged,
1601 * and would retain their original ref->count < 0.
1603 if (ctx->roots && ref->count && ref->root_id && ref->parent == 0) {
1604 /* no parent == root of tree */
1605 ret = ulist_add(ctx->roots, ref->root_id, 0, GFP_NOFS);
1609 if (ref->count && ref->parent) {
1610 if (!ctx->skip_inode_ref_list && !ref->inode_list &&
1612 struct btrfs_tree_parent_check check = { 0 };
1613 struct extent_buffer *eb;
1615 check.level = ref->level;
1617 eb = read_tree_block(ctx->fs_info, ref->parent,
1623 if (!extent_buffer_uptodate(eb)) {
1624 free_extent_buffer(eb);
1629 if (!path->skip_locking)
1630 btrfs_tree_read_lock(eb);
1631 ret = find_extent_in_eb(ctx, eb, &eie);
1632 if (!path->skip_locking)
1633 btrfs_tree_read_unlock(eb);
1634 free_extent_buffer(eb);
1635 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1638 ref->inode_list = eie;
1640 * We transferred the list ownership to the ref,
1641 * so set to NULL to avoid a double free in case
1642 * an error happens after this.
1646 ret = ulist_add_merge_ptr(ctx->refs, ref->parent,
1648 (void **)&eie, GFP_NOFS);
1651 if (!ret && !ctx->skip_inode_ref_list) {
1653 * We've recorded that parent, so we must extend
1654 * its inode list here.
1656 * However if there was corruption we may not
1657 * have found an eie, return an error in this
1667 eie->next = ref->inode_list;
1671 * We have transferred the inode list ownership from
1672 * this ref to the ref we added to the 'refs' ulist.
1673 * So set this ref's inode list to NULL to avoid
1674 * use-after-free when our caller uses it or double
1675 * frees in case an error happens before we return.
1677 ref->inode_list = NULL;
1683 btrfs_free_path(path);
1685 prelim_release(&preftrees.direct);
1686 prelim_release(&preftrees.indirect);
1687 prelim_release(&preftrees.indirect_missing_keys);
1689 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
1690 free_inode_elem_list(eie);
1695 * Finds all leaves with a reference to the specified combination of
1696 * @ctx->bytenr and @ctx->extent_item_pos. The bytenr of the found leaves are
1697 * added to the ulist at @ctx->refs, and that ulist is allocated by this
1698 * function. The caller should free the ulist with free_leaf_list() if
1699 * @ctx->ignore_extent_item_pos is false, otherwise a fimple ulist_free() is
1702 * Returns 0 on success and < 0 on error. On error @ctx->refs is not allocated.
1704 int btrfs_find_all_leafs(struct btrfs_backref_walk_ctx *ctx)
1708 ASSERT(ctx->refs == NULL);
1710 ctx->refs = ulist_alloc(GFP_NOFS);
1714 ret = find_parent_nodes(ctx, NULL);
1715 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1716 (ret < 0 && ret != -ENOENT)) {
1717 free_leaf_list(ctx->refs);
1726 * Walk all backrefs for a given extent to find all roots that reference this
1727 * extent. Walking a backref means finding all extents that reference this
1728 * extent and in turn walk the backrefs of those, too. Naturally this is a
1729 * recursive process, but here it is implemented in an iterative fashion: We
1730 * find all referencing extents for the extent in question and put them on a
1731 * list. In turn, we find all referencing extents for those, further appending
1732 * to the list. The way we iterate the list allows adding more elements after
1733 * the current while iterating. The process stops when we reach the end of the
1736 * Found roots are added to @ctx->roots, which is allocated by this function if
1737 * it points to NULL, in which case the caller is responsible for freeing it
1738 * after it's not needed anymore.
1739 * This function requires @ctx->refs to be NULL, as it uses it for allocating a
1740 * ulist to do temporary work, and frees it before returning.
1742 * Returns 0 on success, < 0 on error.
1744 static int btrfs_find_all_roots_safe(struct btrfs_backref_walk_ctx *ctx)
1746 const u64 orig_bytenr = ctx->bytenr;
1747 const bool orig_skip_inode_ref_list = ctx->skip_inode_ref_list;
1748 bool roots_ulist_allocated = false;
1749 struct ulist_iterator uiter;
1752 ASSERT(ctx->refs == NULL);
1754 ctx->refs = ulist_alloc(GFP_NOFS);
1759 ctx->roots = ulist_alloc(GFP_NOFS);
1761 ulist_free(ctx->refs);
1765 roots_ulist_allocated = true;
1768 ctx->skip_inode_ref_list = true;
1770 ULIST_ITER_INIT(&uiter);
1772 struct ulist_node *node;
1774 ret = find_parent_nodes(ctx, NULL);
1775 if (ret < 0 && ret != -ENOENT) {
1776 if (roots_ulist_allocated) {
1777 ulist_free(ctx->roots);
1783 node = ulist_next(ctx->refs, &uiter);
1786 ctx->bytenr = node->val;
1790 ulist_free(ctx->refs);
1792 ctx->bytenr = orig_bytenr;
1793 ctx->skip_inode_ref_list = orig_skip_inode_ref_list;
1798 int btrfs_find_all_roots(struct btrfs_backref_walk_ctx *ctx,
1799 bool skip_commit_root_sem)
1803 if (!ctx->trans && !skip_commit_root_sem)
1804 down_read(&ctx->fs_info->commit_root_sem);
1805 ret = btrfs_find_all_roots_safe(ctx);
1806 if (!ctx->trans && !skip_commit_root_sem)
1807 up_read(&ctx->fs_info->commit_root_sem);
1811 struct btrfs_backref_share_check_ctx *btrfs_alloc_backref_share_check_ctx(void)
1813 struct btrfs_backref_share_check_ctx *ctx;
1815 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1819 ulist_init(&ctx->refs);
1824 void btrfs_free_backref_share_ctx(struct btrfs_backref_share_check_ctx *ctx)
1829 ulist_release(&ctx->refs);
1834 * Check if a data extent is shared or not.
1836 * @inode: The inode whose extent we are checking.
1837 * @bytenr: Logical bytenr of the extent we are checking.
1838 * @extent_gen: Generation of the extent (file extent item) or 0 if it is
1840 * @ctx: A backref sharedness check context.
1842 * btrfs_is_data_extent_shared uses the backref walking code but will short
1843 * circuit as soon as it finds a root or inode that doesn't match the
1844 * one passed in. This provides a significant performance benefit for
1845 * callers (such as fiemap) which want to know whether the extent is
1846 * shared but do not need a ref count.
1848 * This attempts to attach to the running transaction in order to account for
1849 * delayed refs, but continues on even when no running transaction exists.
1851 * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1853 int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
1855 struct btrfs_backref_share_check_ctx *ctx)
1857 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
1858 struct btrfs_root *root = inode->root;
1859 struct btrfs_fs_info *fs_info = root->fs_info;
1860 struct btrfs_trans_handle *trans;
1861 struct ulist_iterator uiter;
1862 struct ulist_node *node;
1863 struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1865 struct share_check shared = {
1868 .inum = btrfs_ino(inode),
1869 .data_bytenr = bytenr,
1870 .data_extent_gen = extent_gen,
1872 .self_ref_count = 0,
1873 .have_delayed_delete_refs = false,
1877 bool leaf_is_shared;
1879 for (int i = 0; i < BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE; i++) {
1880 if (ctx->prev_extents_cache[i].bytenr == bytenr)
1881 return ctx->prev_extents_cache[i].is_shared;
1884 ulist_init(&ctx->refs);
1886 trans = btrfs_join_transaction_nostart(root);
1887 if (IS_ERR(trans)) {
1888 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1889 ret = PTR_ERR(trans);
1893 down_read(&fs_info->commit_root_sem);
1895 btrfs_get_tree_mod_seq(fs_info, &elem);
1896 walk_ctx.time_seq = elem.seq;
1899 ctx->use_path_cache = true;
1902 * We may have previously determined that the current leaf is shared.
1903 * If it is, then we have a data extent that is shared due to a shared
1904 * subtree (caused by snapshotting) and we don't need to check for data
1905 * backrefs. If the leaf is not shared, then we must do backref walking
1906 * to determine if the data extent is shared through reflinks.
1908 leaf_cached = lookup_backref_shared_cache(ctx, root,
1909 ctx->curr_leaf_bytenr, 0,
1911 if (leaf_cached && leaf_is_shared) {
1916 walk_ctx.skip_inode_ref_list = true;
1917 walk_ctx.trans = trans;
1918 walk_ctx.fs_info = fs_info;
1919 walk_ctx.refs = &ctx->refs;
1921 /* -1 means we are in the bytenr of the data extent. */
1923 ULIST_ITER_INIT(&uiter);
1925 const unsigned long prev_ref_count = ctx->refs.nnodes;
1927 walk_ctx.bytenr = bytenr;
1928 ret = find_parent_nodes(&walk_ctx, &shared);
1929 if (ret == BACKREF_FOUND_SHARED ||
1930 ret == BACKREF_FOUND_NOT_SHARED) {
1931 /* If shared must return 1, otherwise return 0. */
1932 ret = (ret == BACKREF_FOUND_SHARED) ? 1 : 0;
1934 store_backref_shared_cache(ctx, root, bytenr,
1938 if (ret < 0 && ret != -ENOENT)
1943 * More than one extent buffer (bytenr) may have been added to
1944 * the ctx->refs ulist, in which case we have to check multiple
1945 * tree paths in case the first one is not shared, so we can not
1946 * use the path cache which is made for a single path. Multiple
1947 * extent buffers at the current level happen when:
1949 * 1) level -1, the data extent: If our data extent was not
1950 * directly shared (without multiple reference items), then
1951 * it might have a single reference item with a count > 1 for
1952 * the same offset, which means there are 2 (or more) file
1953 * extent items that point to the data extent - this happens
1954 * when a file extent item needs to be split and then one
1955 * item gets moved to another leaf due to a b+tree leaf split
1956 * when inserting some item. In this case the file extent
1957 * items may be located in different leaves and therefore
1958 * some of the leaves may be referenced through shared
1959 * subtrees while others are not. Since our extent buffer
1960 * cache only works for a single path (by far the most common
1961 * case and simpler to deal with), we can not use it if we
1962 * have multiple leaves (which implies multiple paths).
1964 * 2) level >= 0, a tree node/leaf: We can have a mix of direct
1965 * and indirect references on a b+tree node/leaf, so we have
1966 * to check multiple paths, and the extent buffer (the
1967 * current bytenr) may be shared or not. One example is
1968 * during relocation as we may get a shared tree block ref
1969 * (direct ref) and a non-shared tree block ref (indirect
1970 * ref) for the same node/leaf.
1972 if ((ctx->refs.nnodes - prev_ref_count) > 1)
1973 ctx->use_path_cache = false;
1976 store_backref_shared_cache(ctx, root, bytenr,
1978 node = ulist_next(&ctx->refs, &uiter);
1982 if (ctx->use_path_cache) {
1987 cached = lookup_backref_shared_cache(ctx, root, bytenr,
1990 ret = (is_shared ? 1 : 0);
1994 shared.share_count = 0;
1995 shared.have_delayed_delete_refs = false;
2000 * If the path cache is disabled, then it means at some tree level we
2001 * got multiple parents due to a mix of direct and indirect backrefs or
2002 * multiple leaves with file extent items pointing to the same data
2003 * extent. We have to invalidate the cache and cache only the sharedness
2004 * result for the levels where we got only one node/reference.
2006 if (!ctx->use_path_cache) {
2010 if (ret >= 0 && level >= 0) {
2011 bytenr = ctx->path_cache_entries[level].bytenr;
2012 ctx->use_path_cache = true;
2013 store_backref_shared_cache(ctx, root, bytenr, level, ret);
2017 for ( ; i < BTRFS_MAX_LEVEL; i++)
2018 ctx->path_cache_entries[i].bytenr = 0;
2022 * Cache the sharedness result for the data extent if we know our inode
2023 * has more than 1 file extent item that refers to the data extent.
2025 if (ret >= 0 && shared.self_ref_count > 1) {
2026 int slot = ctx->prev_extents_cache_slot;
2028 ctx->prev_extents_cache[slot].bytenr = shared.data_bytenr;
2029 ctx->prev_extents_cache[slot].is_shared = (ret == 1);
2031 slot = (slot + 1) % BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE;
2032 ctx->prev_extents_cache_slot = slot;
2037 btrfs_put_tree_mod_seq(fs_info, &elem);
2038 btrfs_end_transaction(trans);
2040 up_read(&fs_info->commit_root_sem);
2043 ulist_release(&ctx->refs);
2044 ctx->prev_leaf_bytenr = ctx->curr_leaf_bytenr;
2049 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
2050 u64 start_off, struct btrfs_path *path,
2051 struct btrfs_inode_extref **ret_extref,
2055 struct btrfs_key key;
2056 struct btrfs_key found_key;
2057 struct btrfs_inode_extref *extref;
2058 const struct extent_buffer *leaf;
2061 key.objectid = inode_objectid;
2062 key.type = BTRFS_INODE_EXTREF_KEY;
2063 key.offset = start_off;
2065 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2070 leaf = path->nodes[0];
2071 slot = path->slots[0];
2072 if (slot >= btrfs_header_nritems(leaf)) {
2074 * If the item at offset is not found,
2075 * btrfs_search_slot will point us to the slot
2076 * where it should be inserted. In our case
2077 * that will be the slot directly before the
2078 * next INODE_REF_KEY_V2 item. In the case
2079 * that we're pointing to the last slot in a
2080 * leaf, we must move one leaf over.
2082 ret = btrfs_next_leaf(root, path);
2091 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2094 * Check that we're still looking at an extended ref key for
2095 * this particular objectid. If we have different
2096 * objectid or type then there are no more to be found
2097 * in the tree and we can exit.
2100 if (found_key.objectid != inode_objectid)
2102 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
2106 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2107 extref = (struct btrfs_inode_extref *)ptr;
2108 *ret_extref = extref;
2110 *found_off = found_key.offset;
2118 * this iterates to turn a name (from iref/extref) into a full filesystem path.
2119 * Elements of the path are separated by '/' and the path is guaranteed to be
2120 * 0-terminated. the path is only given within the current file system.
2121 * Therefore, it never starts with a '/'. the caller is responsible to provide
2122 * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
2123 * the start point of the resulting string is returned. this pointer is within
2125 * in case the path buffer would overflow, the pointer is decremented further
2126 * as if output was written to the buffer, though no more output is actually
2127 * generated. that way, the caller can determine how much space would be
2128 * required for the path to fit into the buffer. in that case, the returned
2129 * value will be smaller than dest. callers must check this!
2131 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
2132 u32 name_len, unsigned long name_off,
2133 struct extent_buffer *eb_in, u64 parent,
2134 char *dest, u32 size)
2139 s64 bytes_left = ((s64)size) - 1;
2140 struct extent_buffer *eb = eb_in;
2141 struct btrfs_key found_key;
2142 struct btrfs_inode_ref *iref;
2144 if (bytes_left >= 0)
2145 dest[bytes_left] = '\0';
2148 bytes_left -= name_len;
2149 if (bytes_left >= 0)
2150 read_extent_buffer(eb, dest + bytes_left,
2151 name_off, name_len);
2153 if (!path->skip_locking)
2154 btrfs_tree_read_unlock(eb);
2155 free_extent_buffer(eb);
2157 ret = btrfs_find_item(fs_root, path, parent, 0,
2158 BTRFS_INODE_REF_KEY, &found_key);
2164 next_inum = found_key.offset;
2166 /* regular exit ahead */
2167 if (parent == next_inum)
2170 slot = path->slots[0];
2171 eb = path->nodes[0];
2172 /* make sure we can use eb after releasing the path */
2174 path->nodes[0] = NULL;
2177 btrfs_release_path(path);
2178 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2180 name_len = btrfs_inode_ref_name_len(eb, iref);
2181 name_off = (unsigned long)(iref + 1);
2185 if (bytes_left >= 0)
2186 dest[bytes_left] = '/';
2189 btrfs_release_path(path);
2192 return ERR_PTR(ret);
2194 return dest + bytes_left;
2198 * this makes the path point to (logical EXTENT_ITEM *)
2199 * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
2200 * tree blocks and <0 on error.
2202 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
2203 struct btrfs_path *path, struct btrfs_key *found_key,
2206 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
2211 const struct extent_buffer *eb;
2212 struct btrfs_extent_item *ei;
2213 struct btrfs_key key;
2215 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2216 key.type = BTRFS_METADATA_ITEM_KEY;
2218 key.type = BTRFS_EXTENT_ITEM_KEY;
2219 key.objectid = logical;
2220 key.offset = (u64)-1;
2222 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2226 ret = btrfs_previous_extent_item(extent_root, path, 0);
2232 btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
2233 if (found_key->type == BTRFS_METADATA_ITEM_KEY)
2234 size = fs_info->nodesize;
2235 else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
2236 size = found_key->offset;
2238 if (found_key->objectid > logical ||
2239 found_key->objectid + size <= logical) {
2240 btrfs_debug(fs_info,
2241 "logical %llu is not within any extent", logical);
2245 eb = path->nodes[0];
2246 item_size = btrfs_item_size(eb, path->slots[0]);
2247 BUG_ON(item_size < sizeof(*ei));
2249 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
2250 flags = btrfs_extent_flags(eb, ei);
2252 btrfs_debug(fs_info,
2253 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
2254 logical, logical - found_key->objectid, found_key->objectid,
2255 found_key->offset, flags, item_size);
2257 WARN_ON(!flags_ret);
2259 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2260 *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
2261 else if (flags & BTRFS_EXTENT_FLAG_DATA)
2262 *flags_ret = BTRFS_EXTENT_FLAG_DATA;
2272 * helper function to iterate extent inline refs. ptr must point to a 0 value
2273 * for the first call and may be modified. it is used to track state.
2274 * if more refs exist, 0 is returned and the next call to
2275 * get_extent_inline_ref must pass the modified ptr parameter to get the
2276 * next ref. after the last ref was processed, 1 is returned.
2277 * returns <0 on error
2279 static int get_extent_inline_ref(unsigned long *ptr,
2280 const struct extent_buffer *eb,
2281 const struct btrfs_key *key,
2282 const struct btrfs_extent_item *ei,
2284 struct btrfs_extent_inline_ref **out_eiref,
2289 struct btrfs_tree_block_info *info;
2293 flags = btrfs_extent_flags(eb, ei);
2294 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2295 if (key->type == BTRFS_METADATA_ITEM_KEY) {
2296 /* a skinny metadata extent */
2298 (struct btrfs_extent_inline_ref *)(ei + 1);
2300 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
2301 info = (struct btrfs_tree_block_info *)(ei + 1);
2303 (struct btrfs_extent_inline_ref *)(info + 1);
2306 *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
2308 *ptr = (unsigned long)*out_eiref;
2309 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
2313 end = (unsigned long)ei + item_size;
2314 *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
2315 *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
2316 BTRFS_REF_TYPE_ANY);
2317 if (*out_type == BTRFS_REF_TYPE_INVALID)
2320 *ptr += btrfs_extent_inline_ref_size(*out_type);
2321 WARN_ON(*ptr > end);
2323 return 1; /* last */
2329 * reads the tree block backref for an extent. tree level and root are returned
2330 * through out_level and out_root. ptr must point to a 0 value for the first
2331 * call and may be modified (see get_extent_inline_ref comment).
2332 * returns 0 if data was provided, 1 if there was no more data to provide or
2335 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
2336 struct btrfs_key *key, struct btrfs_extent_item *ei,
2337 u32 item_size, u64 *out_root, u8 *out_level)
2341 struct btrfs_extent_inline_ref *eiref;
2343 if (*ptr == (unsigned long)-1)
2347 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
2352 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
2353 type == BTRFS_SHARED_BLOCK_REF_KEY)
2360 /* we can treat both ref types equally here */
2361 *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
2363 if (key->type == BTRFS_EXTENT_ITEM_KEY) {
2364 struct btrfs_tree_block_info *info;
2366 info = (struct btrfs_tree_block_info *)(ei + 1);
2367 *out_level = btrfs_tree_block_level(eb, info);
2369 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
2370 *out_level = (u8)key->offset;
2374 *ptr = (unsigned long)-1;
2379 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
2380 struct extent_inode_elem *inode_list,
2381 u64 root, u64 extent_item_objectid,
2382 iterate_extent_inodes_t *iterate, void *ctx)
2384 struct extent_inode_elem *eie;
2387 for (eie = inode_list; eie; eie = eie->next) {
2388 btrfs_debug(fs_info,
2389 "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
2390 extent_item_objectid, eie->inum,
2392 ret = iterate(eie->inum, eie->offset, eie->num_bytes, root, ctx);
2394 btrfs_debug(fs_info,
2395 "stopping iteration for %llu due to ret=%d",
2396 extent_item_objectid, ret);
2405 * calls iterate() for every inode that references the extent identified by
2406 * the given parameters.
2407 * when the iterator function returns a non-zero value, iteration stops.
2409 int iterate_extent_inodes(struct btrfs_backref_walk_ctx *ctx,
2410 bool search_commit_root,
2411 iterate_extent_inodes_t *iterate, void *user_ctx)
2415 struct ulist_node *ref_node;
2416 struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
2417 struct ulist_iterator ref_uiter;
2419 btrfs_debug(ctx->fs_info, "resolving all inodes for extent %llu",
2422 ASSERT(ctx->trans == NULL);
2423 ASSERT(ctx->roots == NULL);
2425 if (!search_commit_root) {
2426 struct btrfs_trans_handle *trans;
2428 trans = btrfs_attach_transaction(ctx->fs_info->tree_root);
2429 if (IS_ERR(trans)) {
2430 if (PTR_ERR(trans) != -ENOENT &&
2431 PTR_ERR(trans) != -EROFS)
2432 return PTR_ERR(trans);
2439 btrfs_get_tree_mod_seq(ctx->fs_info, &seq_elem);
2440 ctx->time_seq = seq_elem.seq;
2442 down_read(&ctx->fs_info->commit_root_sem);
2445 ret = btrfs_find_all_leafs(ctx);
2451 ULIST_ITER_INIT(&ref_uiter);
2452 while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2453 const u64 leaf_bytenr = ref_node->val;
2454 struct ulist_node *root_node;
2455 struct ulist_iterator root_uiter;
2456 struct extent_inode_elem *inode_list;
2458 inode_list = (struct extent_inode_elem *)(uintptr_t)ref_node->aux;
2460 if (ctx->cache_lookup) {
2461 const u64 *root_ids;
2465 cached = ctx->cache_lookup(leaf_bytenr, ctx->user_ctx,
2466 &root_ids, &root_count);
2468 for (int i = 0; i < root_count; i++) {
2469 ret = iterate_leaf_refs(ctx->fs_info,
2483 ctx->roots = ulist_alloc(GFP_NOFS);
2490 ctx->bytenr = leaf_bytenr;
2491 ret = btrfs_find_all_roots_safe(ctx);
2495 if (ctx->cache_store)
2496 ctx->cache_store(leaf_bytenr, ctx->roots, ctx->user_ctx);
2498 ULIST_ITER_INIT(&root_uiter);
2499 while (!ret && (root_node = ulist_next(ctx->roots, &root_uiter))) {
2500 btrfs_debug(ctx->fs_info,
2501 "root %llu references leaf %llu, data list %#llx",
2502 root_node->val, ref_node->val,
2504 ret = iterate_leaf_refs(ctx->fs_info, inode_list,
2505 root_node->val, ctx->bytenr,
2508 ulist_reinit(ctx->roots);
2511 free_leaf_list(refs);
2514 btrfs_put_tree_mod_seq(ctx->fs_info, &seq_elem);
2515 btrfs_end_transaction(ctx->trans);
2518 up_read(&ctx->fs_info->commit_root_sem);
2521 ulist_free(ctx->roots);
2524 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP)
2530 static int build_ino_list(u64 inum, u64 offset, u64 num_bytes, u64 root, void *ctx)
2532 struct btrfs_data_container *inodes = ctx;
2533 const size_t c = 3 * sizeof(u64);
2535 if (inodes->bytes_left >= c) {
2536 inodes->bytes_left -= c;
2537 inodes->val[inodes->elem_cnt] = inum;
2538 inodes->val[inodes->elem_cnt + 1] = offset;
2539 inodes->val[inodes->elem_cnt + 2] = root;
2540 inodes->elem_cnt += 3;
2542 inodes->bytes_missing += c - inodes->bytes_left;
2543 inodes->bytes_left = 0;
2544 inodes->elem_missed += 3;
2550 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2551 struct btrfs_path *path,
2552 void *ctx, bool ignore_offset)
2554 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
2557 struct btrfs_key found_key;
2558 int search_commit_root = path->search_commit_root;
2560 ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2561 btrfs_release_path(path);
2564 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2567 walk_ctx.bytenr = found_key.objectid;
2569 walk_ctx.ignore_extent_item_pos = true;
2571 walk_ctx.extent_item_pos = logical - found_key.objectid;
2572 walk_ctx.fs_info = fs_info;
2574 return iterate_extent_inodes(&walk_ctx, search_commit_root,
2575 build_ino_list, ctx);
2578 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2579 struct extent_buffer *eb, struct inode_fs_paths *ipath);
2581 static int iterate_inode_refs(u64 inum, struct inode_fs_paths *ipath)
2590 struct btrfs_root *fs_root = ipath->fs_root;
2591 struct btrfs_path *path = ipath->btrfs_path;
2592 struct extent_buffer *eb;
2593 struct btrfs_inode_ref *iref;
2594 struct btrfs_key found_key;
2597 ret = btrfs_find_item(fs_root, path, inum,
2598 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2604 ret = found ? 0 : -ENOENT;
2609 parent = found_key.offset;
2610 slot = path->slots[0];
2611 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2616 btrfs_release_path(path);
2618 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2620 for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2621 name_len = btrfs_inode_ref_name_len(eb, iref);
2622 /* path must be released before calling iterate()! */
2623 btrfs_debug(fs_root->fs_info,
2624 "following ref at offset %u for inode %llu in tree %llu",
2625 cur, found_key.objectid,
2626 fs_root->root_key.objectid);
2627 ret = inode_to_path(parent, name_len,
2628 (unsigned long)(iref + 1), eb, ipath);
2631 len = sizeof(*iref) + name_len;
2632 iref = (struct btrfs_inode_ref *)((char *)iref + len);
2634 free_extent_buffer(eb);
2637 btrfs_release_path(path);
2642 static int iterate_inode_extrefs(u64 inum, struct inode_fs_paths *ipath)
2649 struct btrfs_root *fs_root = ipath->fs_root;
2650 struct btrfs_path *path = ipath->btrfs_path;
2651 struct extent_buffer *eb;
2652 struct btrfs_inode_extref *extref;
2658 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2663 ret = found ? 0 : -ENOENT;
2668 slot = path->slots[0];
2669 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2674 btrfs_release_path(path);
2676 item_size = btrfs_item_size(eb, slot);
2677 ptr = btrfs_item_ptr_offset(eb, slot);
2680 while (cur_offset < item_size) {
2683 extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2684 parent = btrfs_inode_extref_parent(eb, extref);
2685 name_len = btrfs_inode_extref_name_len(eb, extref);
2686 ret = inode_to_path(parent, name_len,
2687 (unsigned long)&extref->name, eb, ipath);
2691 cur_offset += btrfs_inode_extref_name_len(eb, extref);
2692 cur_offset += sizeof(*extref);
2694 free_extent_buffer(eb);
2699 btrfs_release_path(path);
2705 * returns 0 if the path could be dumped (probably truncated)
2706 * returns <0 in case of an error
2708 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2709 struct extent_buffer *eb, struct inode_fs_paths *ipath)
2713 int i = ipath->fspath->elem_cnt;
2714 const int s_ptr = sizeof(char *);
2717 bytes_left = ipath->fspath->bytes_left > s_ptr ?
2718 ipath->fspath->bytes_left - s_ptr : 0;
2720 fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2721 fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2722 name_off, eb, inum, fspath_min, bytes_left);
2724 return PTR_ERR(fspath);
2726 if (fspath > fspath_min) {
2727 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2728 ++ipath->fspath->elem_cnt;
2729 ipath->fspath->bytes_left = fspath - fspath_min;
2731 ++ipath->fspath->elem_missed;
2732 ipath->fspath->bytes_missing += fspath_min - fspath;
2733 ipath->fspath->bytes_left = 0;
2740 * this dumps all file system paths to the inode into the ipath struct, provided
2741 * is has been created large enough. each path is zero-terminated and accessed
2742 * from ipath->fspath->val[i].
2743 * when it returns, there are ipath->fspath->elem_cnt number of paths available
2744 * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2745 * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2746 * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2747 * have been needed to return all paths.
2749 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2754 ret = iterate_inode_refs(inum, ipath);
2757 else if (ret != -ENOENT)
2760 ret = iterate_inode_extrefs(inum, ipath);
2761 if (ret == -ENOENT && found_refs)
2767 struct btrfs_data_container *init_data_container(u32 total_bytes)
2769 struct btrfs_data_container *data;
2772 alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2773 data = kvmalloc(alloc_bytes, GFP_KERNEL);
2775 return ERR_PTR(-ENOMEM);
2777 if (total_bytes >= sizeof(*data)) {
2778 data->bytes_left = total_bytes - sizeof(*data);
2779 data->bytes_missing = 0;
2781 data->bytes_missing = sizeof(*data) - total_bytes;
2782 data->bytes_left = 0;
2786 data->elem_missed = 0;
2792 * allocates space to return multiple file system paths for an inode.
2793 * total_bytes to allocate are passed, note that space usable for actual path
2794 * information will be total_bytes - sizeof(struct inode_fs_paths).
2795 * the returned pointer must be freed with free_ipath() in the end.
2797 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2798 struct btrfs_path *path)
2800 struct inode_fs_paths *ifp;
2801 struct btrfs_data_container *fspath;
2803 fspath = init_data_container(total_bytes);
2805 return ERR_CAST(fspath);
2807 ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2810 return ERR_PTR(-ENOMEM);
2813 ifp->btrfs_path = path;
2814 ifp->fspath = fspath;
2815 ifp->fs_root = fs_root;
2820 void free_ipath(struct inode_fs_paths *ipath)
2824 kvfree(ipath->fspath);
2828 struct btrfs_backref_iter *btrfs_backref_iter_alloc(struct btrfs_fs_info *fs_info)
2830 struct btrfs_backref_iter *ret;
2832 ret = kzalloc(sizeof(*ret), GFP_NOFS);
2836 ret->path = btrfs_alloc_path();
2842 /* Current backref iterator only supports iteration in commit root */
2843 ret->path->search_commit_root = 1;
2844 ret->path->skip_locking = 1;
2845 ret->fs_info = fs_info;
2850 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2852 struct btrfs_fs_info *fs_info = iter->fs_info;
2853 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2854 struct btrfs_path *path = iter->path;
2855 struct btrfs_extent_item *ei;
2856 struct btrfs_key key;
2859 key.objectid = bytenr;
2860 key.type = BTRFS_METADATA_ITEM_KEY;
2861 key.offset = (u64)-1;
2862 iter->bytenr = bytenr;
2864 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2871 if (path->slots[0] == 0) {
2872 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2878 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2879 if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2880 key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2884 memcpy(&iter->cur_key, &key, sizeof(key));
2885 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2887 iter->end_ptr = (u32)(iter->item_ptr +
2888 btrfs_item_size(path->nodes[0], path->slots[0]));
2889 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2890 struct btrfs_extent_item);
2893 * Only support iteration on tree backref yet.
2895 * This is an extra precaution for non skinny-metadata, where
2896 * EXTENT_ITEM is also used for tree blocks, that we can only use
2897 * extent flags to determine if it's a tree block.
2899 if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2903 iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2905 /* If there is no inline backref, go search for keyed backref */
2906 if (iter->cur_ptr >= iter->end_ptr) {
2907 ret = btrfs_next_item(extent_root, path);
2909 /* No inline nor keyed ref */
2917 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2919 if (iter->cur_key.objectid != bytenr ||
2920 (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2921 iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2925 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2927 iter->item_ptr = iter->cur_ptr;
2928 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2929 path->nodes[0], path->slots[0]));
2934 btrfs_backref_iter_release(iter);
2939 * Go to the next backref item of current bytenr, can be either inlined or
2942 * Caller needs to check whether it's inline ref or not by iter->cur_key.
2944 * Return 0 if we get next backref without problem.
2945 * Return >0 if there is no extra backref for this bytenr.
2946 * Return <0 if there is something wrong happened.
2948 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2950 struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2951 struct btrfs_root *extent_root;
2952 struct btrfs_path *path = iter->path;
2953 struct btrfs_extent_inline_ref *iref;
2957 if (btrfs_backref_iter_is_inline_ref(iter)) {
2958 /* We're still inside the inline refs */
2959 ASSERT(iter->cur_ptr < iter->end_ptr);
2961 if (btrfs_backref_has_tree_block_info(iter)) {
2962 /* First tree block info */
2963 size = sizeof(struct btrfs_tree_block_info);
2965 /* Use inline ref type to determine the size */
2968 iref = (struct btrfs_extent_inline_ref *)
2969 ((unsigned long)iter->cur_ptr);
2970 type = btrfs_extent_inline_ref_type(eb, iref);
2972 size = btrfs_extent_inline_ref_size(type);
2974 iter->cur_ptr += size;
2975 if (iter->cur_ptr < iter->end_ptr)
2978 /* All inline items iterated, fall through */
2981 /* We're at keyed items, there is no inline item, go to the next one */
2982 extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
2983 ret = btrfs_next_item(extent_root, iter->path);
2987 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2988 if (iter->cur_key.objectid != iter->bytenr ||
2989 (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2990 iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2992 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2994 iter->cur_ptr = iter->item_ptr;
2995 iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
3000 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
3001 struct btrfs_backref_cache *cache, int is_reloc)
3005 cache->rb_root = RB_ROOT;
3006 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3007 INIT_LIST_HEAD(&cache->pending[i]);
3008 INIT_LIST_HEAD(&cache->changed);
3009 INIT_LIST_HEAD(&cache->detached);
3010 INIT_LIST_HEAD(&cache->leaves);
3011 INIT_LIST_HEAD(&cache->pending_edge);
3012 INIT_LIST_HEAD(&cache->useless_node);
3013 cache->fs_info = fs_info;
3014 cache->is_reloc = is_reloc;
3017 struct btrfs_backref_node *btrfs_backref_alloc_node(
3018 struct btrfs_backref_cache *cache, u64 bytenr, int level)
3020 struct btrfs_backref_node *node;
3022 ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
3023 node = kzalloc(sizeof(*node), GFP_NOFS);
3027 INIT_LIST_HEAD(&node->list);
3028 INIT_LIST_HEAD(&node->upper);
3029 INIT_LIST_HEAD(&node->lower);
3030 RB_CLEAR_NODE(&node->rb_node);
3032 node->level = level;
3033 node->bytenr = bytenr;
3038 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
3039 struct btrfs_backref_cache *cache)
3041 struct btrfs_backref_edge *edge;
3043 edge = kzalloc(sizeof(*edge), GFP_NOFS);
3050 * Drop the backref node from cache, also cleaning up all its
3051 * upper edges and any uncached nodes in the path.
3053 * This cleanup happens bottom up, thus the node should either
3054 * be the lowest node in the cache or a detached node.
3056 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
3057 struct btrfs_backref_node *node)
3059 struct btrfs_backref_node *upper;
3060 struct btrfs_backref_edge *edge;
3065 BUG_ON(!node->lowest && !node->detached);
3066 while (!list_empty(&node->upper)) {
3067 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
3069 upper = edge->node[UPPER];
3070 list_del(&edge->list[LOWER]);
3071 list_del(&edge->list[UPPER]);
3072 btrfs_backref_free_edge(cache, edge);
3075 * Add the node to leaf node list if no other child block
3078 if (list_empty(&upper->lower)) {
3079 list_add_tail(&upper->lower, &cache->leaves);
3084 btrfs_backref_drop_node(cache, node);
3088 * Release all nodes/edges from current cache
3090 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
3092 struct btrfs_backref_node *node;
3095 while (!list_empty(&cache->detached)) {
3096 node = list_entry(cache->detached.next,
3097 struct btrfs_backref_node, list);
3098 btrfs_backref_cleanup_node(cache, node);
3101 while (!list_empty(&cache->leaves)) {
3102 node = list_entry(cache->leaves.next,
3103 struct btrfs_backref_node, lower);
3104 btrfs_backref_cleanup_node(cache, node);
3107 cache->last_trans = 0;
3109 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3110 ASSERT(list_empty(&cache->pending[i]));
3111 ASSERT(list_empty(&cache->pending_edge));
3112 ASSERT(list_empty(&cache->useless_node));
3113 ASSERT(list_empty(&cache->changed));
3114 ASSERT(list_empty(&cache->detached));
3115 ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
3116 ASSERT(!cache->nr_nodes);
3117 ASSERT(!cache->nr_edges);
3121 * Handle direct tree backref
3123 * Direct tree backref means, the backref item shows its parent bytenr
3124 * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
3126 * @ref_key: The converted backref key.
3127 * For keyed backref, it's the item key.
3128 * For inlined backref, objectid is the bytenr,
3129 * type is btrfs_inline_ref_type, offset is
3130 * btrfs_inline_ref_offset.
3132 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
3133 struct btrfs_key *ref_key,
3134 struct btrfs_backref_node *cur)
3136 struct btrfs_backref_edge *edge;
3137 struct btrfs_backref_node *upper;
3138 struct rb_node *rb_node;
3140 ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
3142 /* Only reloc root uses backref pointing to itself */
3143 if (ref_key->objectid == ref_key->offset) {
3144 struct btrfs_root *root;
3146 cur->is_reloc_root = 1;
3147 /* Only reloc backref cache cares about a specific root */
3148 if (cache->is_reloc) {
3149 root = find_reloc_root(cache->fs_info, cur->bytenr);
3155 * For generic purpose backref cache, reloc root node
3158 list_add(&cur->list, &cache->useless_node);
3163 edge = btrfs_backref_alloc_edge(cache);
3167 rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
3169 /* Parent node not yet cached */
3170 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
3173 btrfs_backref_free_edge(cache, edge);
3178 * Backrefs for the upper level block isn't cached, add the
3179 * block to pending list
3181 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3183 /* Parent node already cached */
3184 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
3185 ASSERT(upper->checked);
3186 INIT_LIST_HEAD(&edge->list[UPPER]);
3188 btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
3193 * Handle indirect tree backref
3195 * Indirect tree backref means, we only know which tree the node belongs to.
3196 * We still need to do a tree search to find out the parents. This is for
3197 * TREE_BLOCK_REF backref (keyed or inlined).
3199 * @trans: Transaction handle.
3200 * @ref_key: The same as @ref_key in handle_direct_tree_backref()
3201 * @tree_key: The first key of this tree block.
3202 * @path: A clean (released) path, to avoid allocating path every time
3203 * the function get called.
3205 static int handle_indirect_tree_backref(struct btrfs_trans_handle *trans,
3206 struct btrfs_backref_cache *cache,
3207 struct btrfs_path *path,
3208 struct btrfs_key *ref_key,
3209 struct btrfs_key *tree_key,
3210 struct btrfs_backref_node *cur)
3212 struct btrfs_fs_info *fs_info = cache->fs_info;
3213 struct btrfs_backref_node *upper;
3214 struct btrfs_backref_node *lower;
3215 struct btrfs_backref_edge *edge;
3216 struct extent_buffer *eb;
3217 struct btrfs_root *root;
3218 struct rb_node *rb_node;
3220 bool need_check = true;
3223 root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
3225 return PTR_ERR(root);
3226 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3229 if (btrfs_root_level(&root->root_item) == cur->level) {
3231 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
3233 * For reloc backref cache, we may ignore reloc root. But for
3234 * general purpose backref cache, we can't rely on
3235 * btrfs_should_ignore_reloc_root() as it may conflict with
3236 * current running relocation and lead to missing root.
3238 * For general purpose backref cache, reloc root detection is
3239 * completely relying on direct backref (key->offset is parent
3240 * bytenr), thus only do such check for reloc cache.
3242 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
3243 btrfs_put_root(root);
3244 list_add(&cur->list, &cache->useless_node);
3251 level = cur->level + 1;
3253 /* Search the tree to find parent blocks referring to the block */
3254 path->search_commit_root = 1;
3255 path->skip_locking = 1;
3256 path->lowest_level = level;
3257 ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
3258 path->lowest_level = 0;
3260 btrfs_put_root(root);
3263 if (ret > 0 && path->slots[level] > 0)
3264 path->slots[level]--;
3266 eb = path->nodes[level];
3267 if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
3269 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
3270 cur->bytenr, level - 1, root->root_key.objectid,
3271 tree_key->objectid, tree_key->type, tree_key->offset);
3272 btrfs_put_root(root);
3278 /* Add all nodes and edges in the path */
3279 for (; level < BTRFS_MAX_LEVEL; level++) {
3280 if (!path->nodes[level]) {
3281 ASSERT(btrfs_root_bytenr(&root->root_item) ==
3283 /* Same as previous should_ignore_reloc_root() call */
3284 if (btrfs_should_ignore_reloc_root(root) &&
3286 btrfs_put_root(root);
3287 list_add(&lower->list, &cache->useless_node);
3294 edge = btrfs_backref_alloc_edge(cache);
3296 btrfs_put_root(root);
3301 eb = path->nodes[level];
3302 rb_node = rb_simple_search(&cache->rb_root, eb->start);
3304 upper = btrfs_backref_alloc_node(cache, eb->start,
3307 btrfs_put_root(root);
3308 btrfs_backref_free_edge(cache, edge);
3312 upper->owner = btrfs_header_owner(eb);
3313 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3317 * If we know the block isn't shared we can avoid
3318 * checking its backrefs.
3320 if (btrfs_block_can_be_shared(trans, root, eb))
3326 * Add the block to pending list if we need to check its
3327 * backrefs, we only do this once while walking up a
3328 * tree as we will catch anything else later on.
3330 if (!upper->checked && need_check) {
3332 list_add_tail(&edge->list[UPPER],
3333 &cache->pending_edge);
3337 INIT_LIST_HEAD(&edge->list[UPPER]);
3340 upper = rb_entry(rb_node, struct btrfs_backref_node,
3342 ASSERT(upper->checked);
3343 INIT_LIST_HEAD(&edge->list[UPPER]);
3345 upper->owner = btrfs_header_owner(eb);
3347 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
3350 btrfs_put_root(root);
3357 btrfs_release_path(path);
3362 * Add backref node @cur into @cache.
3364 * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
3365 * links aren't yet bi-directional. Needs to finish such links.
3366 * Use btrfs_backref_finish_upper_links() to finish such linkage.
3368 * @trans: Transaction handle.
3369 * @path: Released path for indirect tree backref lookup
3370 * @iter: Released backref iter for extent tree search
3371 * @node_key: The first key of the tree block
3373 int btrfs_backref_add_tree_node(struct btrfs_trans_handle *trans,
3374 struct btrfs_backref_cache *cache,
3375 struct btrfs_path *path,
3376 struct btrfs_backref_iter *iter,
3377 struct btrfs_key *node_key,
3378 struct btrfs_backref_node *cur)
3380 struct btrfs_backref_edge *edge;
3381 struct btrfs_backref_node *exist;
3384 ret = btrfs_backref_iter_start(iter, cur->bytenr);
3388 * We skip the first btrfs_tree_block_info, as we don't use the key
3389 * stored in it, but fetch it from the tree block
3391 if (btrfs_backref_has_tree_block_info(iter)) {
3392 ret = btrfs_backref_iter_next(iter);
3395 /* No extra backref? This means the tree block is corrupted */
3401 WARN_ON(cur->checked);
3402 if (!list_empty(&cur->upper)) {
3404 * The backref was added previously when processing backref of
3405 * type BTRFS_TREE_BLOCK_REF_KEY
3407 ASSERT(list_is_singular(&cur->upper));
3408 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
3410 ASSERT(list_empty(&edge->list[UPPER]));
3411 exist = edge->node[UPPER];
3413 * Add the upper level block to pending list if we need check
3416 if (!exist->checked)
3417 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3422 for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
3423 struct extent_buffer *eb;
3424 struct btrfs_key key;
3428 eb = btrfs_backref_get_eb(iter);
3430 key.objectid = iter->bytenr;
3431 if (btrfs_backref_iter_is_inline_ref(iter)) {
3432 struct btrfs_extent_inline_ref *iref;
3434 /* Update key for inline backref */
3435 iref = (struct btrfs_extent_inline_ref *)
3436 ((unsigned long)iter->cur_ptr);
3437 type = btrfs_get_extent_inline_ref_type(eb, iref,
3438 BTRFS_REF_TYPE_BLOCK);
3439 if (type == BTRFS_REF_TYPE_INVALID) {
3444 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3446 key.type = iter->cur_key.type;
3447 key.offset = iter->cur_key.offset;
3451 * Parent node found and matches current inline ref, no need to
3452 * rebuild this node for this inline ref
3455 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
3456 exist->owner == key.offset) ||
3457 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
3458 exist->bytenr == key.offset))) {
3463 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
3464 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3465 ret = handle_direct_tree_backref(cache, &key, cur);
3468 } else if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
3470 * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref
3471 * offset means the root objectid. We need to search
3472 * the tree to get its parent bytenr.
3474 ret = handle_indirect_tree_backref(trans, cache, path,
3475 &key, node_key, cur);
3480 * Unrecognized tree backref items (if it can pass tree-checker)
3488 btrfs_backref_iter_release(iter);
3493 * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3495 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3496 struct btrfs_backref_node *start)
3498 struct list_head *useless_node = &cache->useless_node;
3499 struct btrfs_backref_edge *edge;
3500 struct rb_node *rb_node;
3501 LIST_HEAD(pending_edge);
3503 ASSERT(start->checked);
3505 /* Insert this node to cache if it's not COW-only */
3506 if (!start->cowonly) {
3507 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
3510 btrfs_backref_panic(cache->fs_info, start->bytenr,
3512 list_add_tail(&start->lower, &cache->leaves);
3516 * Use breadth first search to iterate all related edges.
3518 * The starting points are all the edges of this node
3520 list_for_each_entry(edge, &start->upper, list[LOWER])
3521 list_add_tail(&edge->list[UPPER], &pending_edge);
3523 while (!list_empty(&pending_edge)) {
3524 struct btrfs_backref_node *upper;
3525 struct btrfs_backref_node *lower;
3527 edge = list_first_entry(&pending_edge,
3528 struct btrfs_backref_edge, list[UPPER]);
3529 list_del_init(&edge->list[UPPER]);
3530 upper = edge->node[UPPER];
3531 lower = edge->node[LOWER];
3533 /* Parent is detached, no need to keep any edges */
3534 if (upper->detached) {
3535 list_del(&edge->list[LOWER]);
3536 btrfs_backref_free_edge(cache, edge);
3538 /* Lower node is orphan, queue for cleanup */
3539 if (list_empty(&lower->upper))
3540 list_add(&lower->list, useless_node);
3545 * All new nodes added in current build_backref_tree() haven't
3546 * been linked to the cache rb tree.
3547 * So if we have upper->rb_node populated, this means a cache
3548 * hit. We only need to link the edge, as @upper and all its
3549 * parents have already been linked.
3551 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3552 if (upper->lowest) {
3553 list_del_init(&upper->lower);
3557 list_add_tail(&edge->list[UPPER], &upper->lower);
3561 /* Sanity check, we shouldn't have any unchecked nodes */
3562 if (!upper->checked) {
3567 /* Sanity check, COW-only node has non-COW-only parent */
3568 if (start->cowonly != upper->cowonly) {
3573 /* Only cache non-COW-only (subvolume trees) tree blocks */
3574 if (!upper->cowonly) {
3575 rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3578 btrfs_backref_panic(cache->fs_info,
3579 upper->bytenr, -EEXIST);
3584 list_add_tail(&edge->list[UPPER], &upper->lower);
3587 * Also queue all the parent edges of this uncached node
3588 * to finish the upper linkage
3590 list_for_each_entry(edge, &upper->upper, list[LOWER])
3591 list_add_tail(&edge->list[UPPER], &pending_edge);
3596 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3597 struct btrfs_backref_node *node)
3599 struct btrfs_backref_node *lower;
3600 struct btrfs_backref_node *upper;
3601 struct btrfs_backref_edge *edge;
3603 while (!list_empty(&cache->useless_node)) {
3604 lower = list_first_entry(&cache->useless_node,
3605 struct btrfs_backref_node, list);
3606 list_del_init(&lower->list);
3608 while (!list_empty(&cache->pending_edge)) {
3609 edge = list_first_entry(&cache->pending_edge,
3610 struct btrfs_backref_edge, list[UPPER]);
3611 list_del(&edge->list[UPPER]);
3612 list_del(&edge->list[LOWER]);
3613 lower = edge->node[LOWER];
3614 upper = edge->node[UPPER];
3615 btrfs_backref_free_edge(cache, edge);
3618 * Lower is no longer linked to any upper backref nodes and
3619 * isn't in the cache, we can free it ourselves.
3621 if (list_empty(&lower->upper) &&
3622 RB_EMPTY_NODE(&lower->rb_node))
3623 list_add(&lower->list, &cache->useless_node);
3625 if (!RB_EMPTY_NODE(&upper->rb_node))
3628 /* Add this guy's upper edges to the list to process */
3629 list_for_each_entry(edge, &upper->upper, list[LOWER])
3630 list_add_tail(&edge->list[UPPER],
3631 &cache->pending_edge);
3632 if (list_empty(&upper->upper))
3633 list_add(&upper->list, &cache->useless_node);
3636 while (!list_empty(&cache->useless_node)) {
3637 lower = list_first_entry(&cache->useless_node,
3638 struct btrfs_backref_node, list);
3639 list_del_init(&lower->list);
3642 btrfs_backref_drop_node(cache, lower);
3645 btrfs_backref_cleanup_node(cache, node);
3646 ASSERT(list_empty(&cache->useless_node) &&
3647 list_empty(&cache->pending_edge));