xref: /openbmc/linux/fs/btrfs/backref.c (revision 98cc42227a1b9b0fb6e89729d08f87f9356bb846)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2011 STRATO.  All rights reserved.
4  */
5 
6 #include <linux/mm.h>
7 #include <linux/rbtree.h>
8 #include <trace/events/btrfs.h>
9 #include "ctree.h"
10 #include "disk-io.h"
11 #include "backref.h"
12 #include "ulist.h"
13 #include "transaction.h"
14 #include "delayed-ref.h"
15 #include "locking.h"
16 #include "misc.h"
17 #include "tree-mod-log.h"
18 
19 /* Just an arbitrary number so we can be sure this happened */
20 #define BACKREF_FOUND_SHARED 6
21 
22 struct extent_inode_elem {
23 	u64 inum;
24 	u64 offset;
25 	struct extent_inode_elem *next;
26 };
27 
28 static int check_extent_in_eb(const struct btrfs_key *key,
29 			      const struct extent_buffer *eb,
30 			      const struct btrfs_file_extent_item *fi,
31 			      u64 extent_item_pos,
32 			      struct extent_inode_elem **eie,
33 			      bool ignore_offset)
34 {
35 	u64 offset = 0;
36 	struct extent_inode_elem *e;
37 
38 	if (!ignore_offset &&
39 	    !btrfs_file_extent_compression(eb, fi) &&
40 	    !btrfs_file_extent_encryption(eb, fi) &&
41 	    !btrfs_file_extent_other_encoding(eb, fi)) {
42 		u64 data_offset;
43 		u64 data_len;
44 
45 		data_offset = btrfs_file_extent_offset(eb, fi);
46 		data_len = btrfs_file_extent_num_bytes(eb, fi);
47 
48 		if (extent_item_pos < data_offset ||
49 		    extent_item_pos >= data_offset + data_len)
50 			return 1;
51 		offset = extent_item_pos - data_offset;
52 	}
53 
54 	e = kmalloc(sizeof(*e), GFP_NOFS);
55 	if (!e)
56 		return -ENOMEM;
57 
58 	e->next = *eie;
59 	e->inum = key->objectid;
60 	e->offset = key->offset + offset;
61 	*eie = e;
62 
63 	return 0;
64 }
65 
66 static void free_inode_elem_list(struct extent_inode_elem *eie)
67 {
68 	struct extent_inode_elem *eie_next;
69 
70 	for (; eie; eie = eie_next) {
71 		eie_next = eie->next;
72 		kfree(eie);
73 	}
74 }
75 
76 static int find_extent_in_eb(const struct extent_buffer *eb,
77 			     u64 wanted_disk_byte, u64 extent_item_pos,
78 			     struct extent_inode_elem **eie,
79 			     bool ignore_offset)
80 {
81 	u64 disk_byte;
82 	struct btrfs_key key;
83 	struct btrfs_file_extent_item *fi;
84 	int slot;
85 	int nritems;
86 	int extent_type;
87 	int ret;
88 
89 	/*
90 	 * from the shared data ref, we only have the leaf but we need
91 	 * the key. thus, we must look into all items and see that we
92 	 * find one (some) with a reference to our extent item.
93 	 */
94 	nritems = btrfs_header_nritems(eb);
95 	for (slot = 0; slot < nritems; ++slot) {
96 		btrfs_item_key_to_cpu(eb, &key, slot);
97 		if (key.type != BTRFS_EXTENT_DATA_KEY)
98 			continue;
99 		fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
100 		extent_type = btrfs_file_extent_type(eb, fi);
101 		if (extent_type == BTRFS_FILE_EXTENT_INLINE)
102 			continue;
103 		/* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
104 		disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
105 		if (disk_byte != wanted_disk_byte)
106 			continue;
107 
108 		ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
109 		if (ret < 0)
110 			return ret;
111 	}
112 
113 	return 0;
114 }
115 
116 struct preftree {
117 	struct rb_root_cached root;
118 	unsigned int count;
119 };
120 
121 #define PREFTREE_INIT	{ .root = RB_ROOT_CACHED, .count = 0 }
122 
123 struct preftrees {
124 	struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
125 	struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
126 	struct preftree indirect_missing_keys;
127 };
128 
129 /*
130  * Checks for a shared extent during backref search.
131  *
132  * The share_count tracks prelim_refs (direct and indirect) having a
133  * ref->count >0:
134  *  - incremented when a ref->count transitions to >0
135  *  - decremented when a ref->count transitions to <1
136  */
137 struct share_check {
138 	u64 root_objectid;
139 	u64 inum;
140 	int share_count;
141 };
142 
143 static inline int extent_is_shared(struct share_check *sc)
144 {
145 	return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
146 }
147 
148 static struct kmem_cache *btrfs_prelim_ref_cache;
149 
150 int __init btrfs_prelim_ref_init(void)
151 {
152 	btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
153 					sizeof(struct prelim_ref),
154 					0,
155 					SLAB_MEM_SPREAD,
156 					NULL);
157 	if (!btrfs_prelim_ref_cache)
158 		return -ENOMEM;
159 	return 0;
160 }
161 
162 void __cold btrfs_prelim_ref_exit(void)
163 {
164 	kmem_cache_destroy(btrfs_prelim_ref_cache);
165 }
166 
167 static void free_pref(struct prelim_ref *ref)
168 {
169 	kmem_cache_free(btrfs_prelim_ref_cache, ref);
170 }
171 
172 /*
173  * Return 0 when both refs are for the same block (and can be merged).
174  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
175  * indicates a 'higher' block.
176  */
177 static int prelim_ref_compare(struct prelim_ref *ref1,
178 			      struct prelim_ref *ref2)
179 {
180 	if (ref1->level < ref2->level)
181 		return -1;
182 	if (ref1->level > ref2->level)
183 		return 1;
184 	if (ref1->root_id < ref2->root_id)
185 		return -1;
186 	if (ref1->root_id > ref2->root_id)
187 		return 1;
188 	if (ref1->key_for_search.type < ref2->key_for_search.type)
189 		return -1;
190 	if (ref1->key_for_search.type > ref2->key_for_search.type)
191 		return 1;
192 	if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
193 		return -1;
194 	if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
195 		return 1;
196 	if (ref1->key_for_search.offset < ref2->key_for_search.offset)
197 		return -1;
198 	if (ref1->key_for_search.offset > ref2->key_for_search.offset)
199 		return 1;
200 	if (ref1->parent < ref2->parent)
201 		return -1;
202 	if (ref1->parent > ref2->parent)
203 		return 1;
204 
205 	return 0;
206 }
207 
208 static void update_share_count(struct share_check *sc, int oldcount,
209 			       int newcount)
210 {
211 	if ((!sc) || (oldcount == 0 && newcount < 1))
212 		return;
213 
214 	if (oldcount > 0 && newcount < 1)
215 		sc->share_count--;
216 	else if (oldcount < 1 && newcount > 0)
217 		sc->share_count++;
218 }
219 
220 /*
221  * Add @newref to the @root rbtree, merging identical refs.
222  *
223  * Callers should assume that newref has been freed after calling.
224  */
225 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
226 			      struct preftree *preftree,
227 			      struct prelim_ref *newref,
228 			      struct share_check *sc)
229 {
230 	struct rb_root_cached *root;
231 	struct rb_node **p;
232 	struct rb_node *parent = NULL;
233 	struct prelim_ref *ref;
234 	int result;
235 	bool leftmost = true;
236 
237 	root = &preftree->root;
238 	p = &root->rb_root.rb_node;
239 
240 	while (*p) {
241 		parent = *p;
242 		ref = rb_entry(parent, struct prelim_ref, rbnode);
243 		result = prelim_ref_compare(ref, newref);
244 		if (result < 0) {
245 			p = &(*p)->rb_left;
246 		} else if (result > 0) {
247 			p = &(*p)->rb_right;
248 			leftmost = false;
249 		} else {
250 			/* Identical refs, merge them and free @newref */
251 			struct extent_inode_elem *eie = ref->inode_list;
252 
253 			while (eie && eie->next)
254 				eie = eie->next;
255 
256 			if (!eie)
257 				ref->inode_list = newref->inode_list;
258 			else
259 				eie->next = newref->inode_list;
260 			trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
261 						     preftree->count);
262 			/*
263 			 * A delayed ref can have newref->count < 0.
264 			 * The ref->count is updated to follow any
265 			 * BTRFS_[ADD|DROP]_DELAYED_REF actions.
266 			 */
267 			update_share_count(sc, ref->count,
268 					   ref->count + newref->count);
269 			ref->count += newref->count;
270 			free_pref(newref);
271 			return;
272 		}
273 	}
274 
275 	update_share_count(sc, 0, newref->count);
276 	preftree->count++;
277 	trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
278 	rb_link_node(&newref->rbnode, parent, p);
279 	rb_insert_color_cached(&newref->rbnode, root, leftmost);
280 }
281 
282 /*
283  * Release the entire tree.  We don't care about internal consistency so
284  * just free everything and then reset the tree root.
285  */
286 static void prelim_release(struct preftree *preftree)
287 {
288 	struct prelim_ref *ref, *next_ref;
289 
290 	rbtree_postorder_for_each_entry_safe(ref, next_ref,
291 					     &preftree->root.rb_root, rbnode)
292 		free_pref(ref);
293 
294 	preftree->root = RB_ROOT_CACHED;
295 	preftree->count = 0;
296 }
297 
298 /*
299  * the rules for all callers of this function are:
300  * - obtaining the parent is the goal
301  * - if you add a key, you must know that it is a correct key
302  * - if you cannot add the parent or a correct key, then we will look into the
303  *   block later to set a correct key
304  *
305  * delayed refs
306  * ============
307  *        backref type | shared | indirect | shared | indirect
308  * information         |   tree |     tree |   data |     data
309  * --------------------+--------+----------+--------+----------
310  *      parent logical |    y   |     -    |    -   |     -
311  *      key to resolve |    -   |     y    |    y   |     y
312  *  tree block logical |    -   |     -    |    -   |     -
313  *  root for resolving |    y   |     y    |    y   |     y
314  *
315  * - column 1:       we've the parent -> done
316  * - column 2, 3, 4: we use the key to find the parent
317  *
318  * on disk refs (inline or keyed)
319  * ==============================
320  *        backref type | shared | indirect | shared | indirect
321  * information         |   tree |     tree |   data |     data
322  * --------------------+--------+----------+--------+----------
323  *      parent logical |    y   |     -    |    y   |     -
324  *      key to resolve |    -   |     -    |    -   |     y
325  *  tree block logical |    y   |     y    |    y   |     y
326  *  root for resolving |    -   |     y    |    y   |     y
327  *
328  * - column 1, 3: we've the parent -> done
329  * - column 2:    we take the first key from the block to find the parent
330  *                (see add_missing_keys)
331  * - column 4:    we use the key to find the parent
332  *
333  * additional information that's available but not required to find the parent
334  * block might help in merging entries to gain some speed.
335  */
336 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
337 			  struct preftree *preftree, u64 root_id,
338 			  const struct btrfs_key *key, int level, u64 parent,
339 			  u64 wanted_disk_byte, int count,
340 			  struct share_check *sc, gfp_t gfp_mask)
341 {
342 	struct prelim_ref *ref;
343 
344 	if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
345 		return 0;
346 
347 	ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
348 	if (!ref)
349 		return -ENOMEM;
350 
351 	ref->root_id = root_id;
352 	if (key)
353 		ref->key_for_search = *key;
354 	else
355 		memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
356 
357 	ref->inode_list = NULL;
358 	ref->level = level;
359 	ref->count = count;
360 	ref->parent = parent;
361 	ref->wanted_disk_byte = wanted_disk_byte;
362 	prelim_ref_insert(fs_info, preftree, ref, sc);
363 	return extent_is_shared(sc);
364 }
365 
366 /* direct refs use root == 0, key == NULL */
367 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
368 			  struct preftrees *preftrees, int level, u64 parent,
369 			  u64 wanted_disk_byte, int count,
370 			  struct share_check *sc, gfp_t gfp_mask)
371 {
372 	return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
373 			      parent, wanted_disk_byte, count, sc, gfp_mask);
374 }
375 
376 /* indirect refs use parent == 0 */
377 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
378 			    struct preftrees *preftrees, u64 root_id,
379 			    const struct btrfs_key *key, int level,
380 			    u64 wanted_disk_byte, int count,
381 			    struct share_check *sc, gfp_t gfp_mask)
382 {
383 	struct preftree *tree = &preftrees->indirect;
384 
385 	if (!key)
386 		tree = &preftrees->indirect_missing_keys;
387 	return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
388 			      wanted_disk_byte, count, sc, gfp_mask);
389 }
390 
391 static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
392 {
393 	struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
394 	struct rb_node *parent = NULL;
395 	struct prelim_ref *ref = NULL;
396 	struct prelim_ref target = {};
397 	int result;
398 
399 	target.parent = bytenr;
400 
401 	while (*p) {
402 		parent = *p;
403 		ref = rb_entry(parent, struct prelim_ref, rbnode);
404 		result = prelim_ref_compare(ref, &target);
405 
406 		if (result < 0)
407 			p = &(*p)->rb_left;
408 		else if (result > 0)
409 			p = &(*p)->rb_right;
410 		else
411 			return 1;
412 	}
413 	return 0;
414 }
415 
416 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
417 			   struct ulist *parents,
418 			   struct preftrees *preftrees, struct prelim_ref *ref,
419 			   int level, u64 time_seq, const u64 *extent_item_pos,
420 			   bool ignore_offset)
421 {
422 	int ret = 0;
423 	int slot;
424 	struct extent_buffer *eb;
425 	struct btrfs_key key;
426 	struct btrfs_key *key_for_search = &ref->key_for_search;
427 	struct btrfs_file_extent_item *fi;
428 	struct extent_inode_elem *eie = NULL, *old = NULL;
429 	u64 disk_byte;
430 	u64 wanted_disk_byte = ref->wanted_disk_byte;
431 	u64 count = 0;
432 	u64 data_offset;
433 
434 	if (level != 0) {
435 		eb = path->nodes[level];
436 		ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
437 		if (ret < 0)
438 			return ret;
439 		return 0;
440 	}
441 
442 	/*
443 	 * 1. We normally enter this function with the path already pointing to
444 	 *    the first item to check. But sometimes, we may enter it with
445 	 *    slot == nritems.
446 	 * 2. We are searching for normal backref but bytenr of this leaf
447 	 *    matches shared data backref
448 	 * 3. The leaf owner is not equal to the root we are searching
449 	 *
450 	 * For these cases, go to the next leaf before we continue.
451 	 */
452 	eb = path->nodes[0];
453 	if (path->slots[0] >= btrfs_header_nritems(eb) ||
454 	    is_shared_data_backref(preftrees, eb->start) ||
455 	    ref->root_id != btrfs_header_owner(eb)) {
456 		if (time_seq == BTRFS_SEQ_LAST)
457 			ret = btrfs_next_leaf(root, path);
458 		else
459 			ret = btrfs_next_old_leaf(root, path, time_seq);
460 	}
461 
462 	while (!ret && count < ref->count) {
463 		eb = path->nodes[0];
464 		slot = path->slots[0];
465 
466 		btrfs_item_key_to_cpu(eb, &key, slot);
467 
468 		if (key.objectid != key_for_search->objectid ||
469 		    key.type != BTRFS_EXTENT_DATA_KEY)
470 			break;
471 
472 		/*
473 		 * We are searching for normal backref but bytenr of this leaf
474 		 * matches shared data backref, OR
475 		 * the leaf owner is not equal to the root we are searching for
476 		 */
477 		if (slot == 0 &&
478 		    (is_shared_data_backref(preftrees, eb->start) ||
479 		     ref->root_id != btrfs_header_owner(eb))) {
480 			if (time_seq == BTRFS_SEQ_LAST)
481 				ret = btrfs_next_leaf(root, path);
482 			else
483 				ret = btrfs_next_old_leaf(root, path, time_seq);
484 			continue;
485 		}
486 		fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
487 		disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
488 		data_offset = btrfs_file_extent_offset(eb, fi);
489 
490 		if (disk_byte == wanted_disk_byte) {
491 			eie = NULL;
492 			old = NULL;
493 			if (ref->key_for_search.offset == key.offset - data_offset)
494 				count++;
495 			else
496 				goto next;
497 			if (extent_item_pos) {
498 				ret = check_extent_in_eb(&key, eb, fi,
499 						*extent_item_pos,
500 						&eie, ignore_offset);
501 				if (ret < 0)
502 					break;
503 			}
504 			if (ret > 0)
505 				goto next;
506 			ret = ulist_add_merge_ptr(parents, eb->start,
507 						  eie, (void **)&old, GFP_NOFS);
508 			if (ret < 0)
509 				break;
510 			if (!ret && extent_item_pos) {
511 				while (old->next)
512 					old = old->next;
513 				old->next = eie;
514 			}
515 			eie = NULL;
516 		}
517 next:
518 		if (time_seq == BTRFS_SEQ_LAST)
519 			ret = btrfs_next_item(root, path);
520 		else
521 			ret = btrfs_next_old_item(root, path, time_seq);
522 	}
523 
524 	if (ret > 0)
525 		ret = 0;
526 	else if (ret < 0)
527 		free_inode_elem_list(eie);
528 	return ret;
529 }
530 
531 /*
532  * resolve an indirect backref in the form (root_id, key, level)
533  * to a logical address
534  */
535 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
536 				struct btrfs_path *path, u64 time_seq,
537 				struct preftrees *preftrees,
538 				struct prelim_ref *ref, struct ulist *parents,
539 				const u64 *extent_item_pos, bool ignore_offset)
540 {
541 	struct btrfs_root *root;
542 	struct extent_buffer *eb;
543 	int ret = 0;
544 	int root_level;
545 	int level = ref->level;
546 	struct btrfs_key search_key = ref->key_for_search;
547 
548 	/*
549 	 * If we're search_commit_root we could possibly be holding locks on
550 	 * other tree nodes.  This happens when qgroups does backref walks when
551 	 * adding new delayed refs.  To deal with this we need to look in cache
552 	 * for the root, and if we don't find it then we need to search the
553 	 * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
554 	 * here.
555 	 */
556 	if (path->search_commit_root)
557 		root = btrfs_get_fs_root_commit_root(fs_info, path, ref->root_id);
558 	else
559 		root = btrfs_get_fs_root(fs_info, ref->root_id, false);
560 	if (IS_ERR(root)) {
561 		ret = PTR_ERR(root);
562 		goto out_free;
563 	}
564 
565 	if (!path->search_commit_root &&
566 	    test_bit(BTRFS_ROOT_DELETING, &root->state)) {
567 		ret = -ENOENT;
568 		goto out;
569 	}
570 
571 	if (btrfs_is_testing(fs_info)) {
572 		ret = -ENOENT;
573 		goto out;
574 	}
575 
576 	if (path->search_commit_root)
577 		root_level = btrfs_header_level(root->commit_root);
578 	else if (time_seq == BTRFS_SEQ_LAST)
579 		root_level = btrfs_header_level(root->node);
580 	else
581 		root_level = btrfs_old_root_level(root, time_seq);
582 
583 	if (root_level + 1 == level)
584 		goto out;
585 
586 	/*
587 	 * We can often find data backrefs with an offset that is too large
588 	 * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
589 	 * subtracting a file's offset with the data offset of its
590 	 * corresponding extent data item. This can happen for example in the
591 	 * clone ioctl.
592 	 *
593 	 * So if we detect such case we set the search key's offset to zero to
594 	 * make sure we will find the matching file extent item at
595 	 * add_all_parents(), otherwise we will miss it because the offset
596 	 * taken form the backref is much larger then the offset of the file
597 	 * extent item. This can make us scan a very large number of file
598 	 * extent items, but at least it will not make us miss any.
599 	 *
600 	 * This is an ugly workaround for a behaviour that should have never
601 	 * existed, but it does and a fix for the clone ioctl would touch a lot
602 	 * of places, cause backwards incompatibility and would not fix the
603 	 * problem for extents cloned with older kernels.
604 	 */
605 	if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
606 	    search_key.offset >= LLONG_MAX)
607 		search_key.offset = 0;
608 	path->lowest_level = level;
609 	if (time_seq == BTRFS_SEQ_LAST)
610 		ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
611 	else
612 		ret = btrfs_search_old_slot(root, &search_key, path, time_seq);
613 
614 	btrfs_debug(fs_info,
615 		"search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
616 		 ref->root_id, level, ref->count, ret,
617 		 ref->key_for_search.objectid, ref->key_for_search.type,
618 		 ref->key_for_search.offset);
619 	if (ret < 0)
620 		goto out;
621 
622 	eb = path->nodes[level];
623 	while (!eb) {
624 		if (WARN_ON(!level)) {
625 			ret = 1;
626 			goto out;
627 		}
628 		level--;
629 		eb = path->nodes[level];
630 	}
631 
632 	ret = add_all_parents(root, path, parents, preftrees, ref, level,
633 			      time_seq, extent_item_pos, ignore_offset);
634 out:
635 	btrfs_put_root(root);
636 out_free:
637 	path->lowest_level = 0;
638 	btrfs_release_path(path);
639 	return ret;
640 }
641 
642 static struct extent_inode_elem *
643 unode_aux_to_inode_list(struct ulist_node *node)
644 {
645 	if (!node)
646 		return NULL;
647 	return (struct extent_inode_elem *)(uintptr_t)node->aux;
648 }
649 
650 /*
651  * We maintain three separate rbtrees: one for direct refs, one for
652  * indirect refs which have a key, and one for indirect refs which do not
653  * have a key. Each tree does merge on insertion.
654  *
655  * Once all of the references are located, we iterate over the tree of
656  * indirect refs with missing keys. An appropriate key is located and
657  * the ref is moved onto the tree for indirect refs. After all missing
658  * keys are thus located, we iterate over the indirect ref tree, resolve
659  * each reference, and then insert the resolved reference onto the
660  * direct tree (merging there too).
661  *
662  * New backrefs (i.e., for parent nodes) are added to the appropriate
663  * rbtree as they are encountered. The new backrefs are subsequently
664  * resolved as above.
665  */
666 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
667 				 struct btrfs_path *path, u64 time_seq,
668 				 struct preftrees *preftrees,
669 				 const u64 *extent_item_pos,
670 				 struct share_check *sc, bool ignore_offset)
671 {
672 	int err;
673 	int ret = 0;
674 	struct ulist *parents;
675 	struct ulist_node *node;
676 	struct ulist_iterator uiter;
677 	struct rb_node *rnode;
678 
679 	parents = ulist_alloc(GFP_NOFS);
680 	if (!parents)
681 		return -ENOMEM;
682 
683 	/*
684 	 * We could trade memory usage for performance here by iterating
685 	 * the tree, allocating new refs for each insertion, and then
686 	 * freeing the entire indirect tree when we're done.  In some test
687 	 * cases, the tree can grow quite large (~200k objects).
688 	 */
689 	while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
690 		struct prelim_ref *ref;
691 
692 		ref = rb_entry(rnode, struct prelim_ref, rbnode);
693 		if (WARN(ref->parent,
694 			 "BUG: direct ref found in indirect tree")) {
695 			ret = -EINVAL;
696 			goto out;
697 		}
698 
699 		rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
700 		preftrees->indirect.count--;
701 
702 		if (ref->count == 0) {
703 			free_pref(ref);
704 			continue;
705 		}
706 
707 		if (sc && sc->root_objectid &&
708 		    ref->root_id != sc->root_objectid) {
709 			free_pref(ref);
710 			ret = BACKREF_FOUND_SHARED;
711 			goto out;
712 		}
713 		err = resolve_indirect_ref(fs_info, path, time_seq, preftrees,
714 					   ref, parents, extent_item_pos,
715 					   ignore_offset);
716 		/*
717 		 * we can only tolerate ENOENT,otherwise,we should catch error
718 		 * and return directly.
719 		 */
720 		if (err == -ENOENT) {
721 			prelim_ref_insert(fs_info, &preftrees->direct, ref,
722 					  NULL);
723 			continue;
724 		} else if (err) {
725 			free_pref(ref);
726 			ret = err;
727 			goto out;
728 		}
729 
730 		/* we put the first parent into the ref at hand */
731 		ULIST_ITER_INIT(&uiter);
732 		node = ulist_next(parents, &uiter);
733 		ref->parent = node ? node->val : 0;
734 		ref->inode_list = unode_aux_to_inode_list(node);
735 
736 		/* Add a prelim_ref(s) for any other parent(s). */
737 		while ((node = ulist_next(parents, &uiter))) {
738 			struct prelim_ref *new_ref;
739 
740 			new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
741 						   GFP_NOFS);
742 			if (!new_ref) {
743 				free_pref(ref);
744 				ret = -ENOMEM;
745 				goto out;
746 			}
747 			memcpy(new_ref, ref, sizeof(*ref));
748 			new_ref->parent = node->val;
749 			new_ref->inode_list = unode_aux_to_inode_list(node);
750 			prelim_ref_insert(fs_info, &preftrees->direct,
751 					  new_ref, NULL);
752 		}
753 
754 		/*
755 		 * Now it's a direct ref, put it in the direct tree. We must
756 		 * do this last because the ref could be merged/freed here.
757 		 */
758 		prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
759 
760 		ulist_reinit(parents);
761 		cond_resched();
762 	}
763 out:
764 	ulist_free(parents);
765 	return ret;
766 }
767 
768 /*
769  * read tree blocks and add keys where required.
770  */
771 static int add_missing_keys(struct btrfs_fs_info *fs_info,
772 			    struct preftrees *preftrees, bool lock)
773 {
774 	struct prelim_ref *ref;
775 	struct extent_buffer *eb;
776 	struct preftree *tree = &preftrees->indirect_missing_keys;
777 	struct rb_node *node;
778 
779 	while ((node = rb_first_cached(&tree->root))) {
780 		ref = rb_entry(node, struct prelim_ref, rbnode);
781 		rb_erase_cached(node, &tree->root);
782 
783 		BUG_ON(ref->parent);	/* should not be a direct ref */
784 		BUG_ON(ref->key_for_search.type);
785 		BUG_ON(!ref->wanted_disk_byte);
786 
787 		eb = read_tree_block(fs_info, ref->wanted_disk_byte,
788 				     ref->root_id, 0, ref->level - 1, NULL);
789 		if (IS_ERR(eb)) {
790 			free_pref(ref);
791 			return PTR_ERR(eb);
792 		} else if (!extent_buffer_uptodate(eb)) {
793 			free_pref(ref);
794 			free_extent_buffer(eb);
795 			return -EIO;
796 		}
797 		if (lock)
798 			btrfs_tree_read_lock(eb);
799 		if (btrfs_header_level(eb) == 0)
800 			btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
801 		else
802 			btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
803 		if (lock)
804 			btrfs_tree_read_unlock(eb);
805 		free_extent_buffer(eb);
806 		prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
807 		cond_resched();
808 	}
809 	return 0;
810 }
811 
812 /*
813  * add all currently queued delayed refs from this head whose seq nr is
814  * smaller or equal that seq to the list
815  */
816 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
817 			    struct btrfs_delayed_ref_head *head, u64 seq,
818 			    struct preftrees *preftrees, struct share_check *sc)
819 {
820 	struct btrfs_delayed_ref_node *node;
821 	struct btrfs_delayed_extent_op *extent_op = head->extent_op;
822 	struct btrfs_key key;
823 	struct btrfs_key tmp_op_key;
824 	struct rb_node *n;
825 	int count;
826 	int ret = 0;
827 
828 	if (extent_op && extent_op->update_key)
829 		btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
830 
831 	spin_lock(&head->lock);
832 	for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
833 		node = rb_entry(n, struct btrfs_delayed_ref_node,
834 				ref_node);
835 		if (node->seq > seq)
836 			continue;
837 
838 		switch (node->action) {
839 		case BTRFS_ADD_DELAYED_EXTENT:
840 		case BTRFS_UPDATE_DELAYED_HEAD:
841 			WARN_ON(1);
842 			continue;
843 		case BTRFS_ADD_DELAYED_REF:
844 			count = node->ref_mod;
845 			break;
846 		case BTRFS_DROP_DELAYED_REF:
847 			count = node->ref_mod * -1;
848 			break;
849 		default:
850 			BUG();
851 		}
852 		switch (node->type) {
853 		case BTRFS_TREE_BLOCK_REF_KEY: {
854 			/* NORMAL INDIRECT METADATA backref */
855 			struct btrfs_delayed_tree_ref *ref;
856 
857 			ref = btrfs_delayed_node_to_tree_ref(node);
858 			ret = add_indirect_ref(fs_info, preftrees, ref->root,
859 					       &tmp_op_key, ref->level + 1,
860 					       node->bytenr, count, sc,
861 					       GFP_ATOMIC);
862 			break;
863 		}
864 		case BTRFS_SHARED_BLOCK_REF_KEY: {
865 			/* SHARED DIRECT METADATA backref */
866 			struct btrfs_delayed_tree_ref *ref;
867 
868 			ref = btrfs_delayed_node_to_tree_ref(node);
869 
870 			ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
871 					     ref->parent, node->bytenr, count,
872 					     sc, GFP_ATOMIC);
873 			break;
874 		}
875 		case BTRFS_EXTENT_DATA_REF_KEY: {
876 			/* NORMAL INDIRECT DATA backref */
877 			struct btrfs_delayed_data_ref *ref;
878 			ref = btrfs_delayed_node_to_data_ref(node);
879 
880 			key.objectid = ref->objectid;
881 			key.type = BTRFS_EXTENT_DATA_KEY;
882 			key.offset = ref->offset;
883 
884 			/*
885 			 * Found a inum that doesn't match our known inum, we
886 			 * know it's shared.
887 			 */
888 			if (sc && sc->inum && ref->objectid != sc->inum) {
889 				ret = BACKREF_FOUND_SHARED;
890 				goto out;
891 			}
892 
893 			ret = add_indirect_ref(fs_info, preftrees, ref->root,
894 					       &key, 0, node->bytenr, count, sc,
895 					       GFP_ATOMIC);
896 			break;
897 		}
898 		case BTRFS_SHARED_DATA_REF_KEY: {
899 			/* SHARED DIRECT FULL backref */
900 			struct btrfs_delayed_data_ref *ref;
901 
902 			ref = btrfs_delayed_node_to_data_ref(node);
903 
904 			ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
905 					     node->bytenr, count, sc,
906 					     GFP_ATOMIC);
907 			break;
908 		}
909 		default:
910 			WARN_ON(1);
911 		}
912 		/*
913 		 * We must ignore BACKREF_FOUND_SHARED until all delayed
914 		 * refs have been checked.
915 		 */
916 		if (ret && (ret != BACKREF_FOUND_SHARED))
917 			break;
918 	}
919 	if (!ret)
920 		ret = extent_is_shared(sc);
921 out:
922 	spin_unlock(&head->lock);
923 	return ret;
924 }
925 
926 /*
927  * add all inline backrefs for bytenr to the list
928  *
929  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
930  */
931 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
932 			   struct btrfs_path *path, u64 bytenr,
933 			   int *info_level, struct preftrees *preftrees,
934 			   struct share_check *sc)
935 {
936 	int ret = 0;
937 	int slot;
938 	struct extent_buffer *leaf;
939 	struct btrfs_key key;
940 	struct btrfs_key found_key;
941 	unsigned long ptr;
942 	unsigned long end;
943 	struct btrfs_extent_item *ei;
944 	u64 flags;
945 	u64 item_size;
946 
947 	/*
948 	 * enumerate all inline refs
949 	 */
950 	leaf = path->nodes[0];
951 	slot = path->slots[0];
952 
953 	item_size = btrfs_item_size(leaf, slot);
954 	BUG_ON(item_size < sizeof(*ei));
955 
956 	ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
957 	flags = btrfs_extent_flags(leaf, ei);
958 	btrfs_item_key_to_cpu(leaf, &found_key, slot);
959 
960 	ptr = (unsigned long)(ei + 1);
961 	end = (unsigned long)ei + item_size;
962 
963 	if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
964 	    flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
965 		struct btrfs_tree_block_info *info;
966 
967 		info = (struct btrfs_tree_block_info *)ptr;
968 		*info_level = btrfs_tree_block_level(leaf, info);
969 		ptr += sizeof(struct btrfs_tree_block_info);
970 		BUG_ON(ptr > end);
971 	} else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
972 		*info_level = found_key.offset;
973 	} else {
974 		BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
975 	}
976 
977 	while (ptr < end) {
978 		struct btrfs_extent_inline_ref *iref;
979 		u64 offset;
980 		int type;
981 
982 		iref = (struct btrfs_extent_inline_ref *)ptr;
983 		type = btrfs_get_extent_inline_ref_type(leaf, iref,
984 							BTRFS_REF_TYPE_ANY);
985 		if (type == BTRFS_REF_TYPE_INVALID)
986 			return -EUCLEAN;
987 
988 		offset = btrfs_extent_inline_ref_offset(leaf, iref);
989 
990 		switch (type) {
991 		case BTRFS_SHARED_BLOCK_REF_KEY:
992 			ret = add_direct_ref(fs_info, preftrees,
993 					     *info_level + 1, offset,
994 					     bytenr, 1, NULL, GFP_NOFS);
995 			break;
996 		case BTRFS_SHARED_DATA_REF_KEY: {
997 			struct btrfs_shared_data_ref *sdref;
998 			int count;
999 
1000 			sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1001 			count = btrfs_shared_data_ref_count(leaf, sdref);
1002 
1003 			ret = add_direct_ref(fs_info, preftrees, 0, offset,
1004 					     bytenr, count, sc, GFP_NOFS);
1005 			break;
1006 		}
1007 		case BTRFS_TREE_BLOCK_REF_KEY:
1008 			ret = add_indirect_ref(fs_info, preftrees, offset,
1009 					       NULL, *info_level + 1,
1010 					       bytenr, 1, NULL, GFP_NOFS);
1011 			break;
1012 		case BTRFS_EXTENT_DATA_REF_KEY: {
1013 			struct btrfs_extent_data_ref *dref;
1014 			int count;
1015 			u64 root;
1016 
1017 			dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1018 			count = btrfs_extent_data_ref_count(leaf, dref);
1019 			key.objectid = btrfs_extent_data_ref_objectid(leaf,
1020 								      dref);
1021 			key.type = BTRFS_EXTENT_DATA_KEY;
1022 			key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1023 
1024 			if (sc && sc->inum && key.objectid != sc->inum) {
1025 				ret = BACKREF_FOUND_SHARED;
1026 				break;
1027 			}
1028 
1029 			root = btrfs_extent_data_ref_root(leaf, dref);
1030 
1031 			ret = add_indirect_ref(fs_info, preftrees, root,
1032 					       &key, 0, bytenr, count,
1033 					       sc, GFP_NOFS);
1034 			break;
1035 		}
1036 		default:
1037 			WARN_ON(1);
1038 		}
1039 		if (ret)
1040 			return ret;
1041 		ptr += btrfs_extent_inline_ref_size(type);
1042 	}
1043 
1044 	return 0;
1045 }
1046 
1047 /*
1048  * add all non-inline backrefs for bytenr to the list
1049  *
1050  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1051  */
1052 static int add_keyed_refs(struct btrfs_root *extent_root,
1053 			  struct btrfs_path *path, u64 bytenr,
1054 			  int info_level, struct preftrees *preftrees,
1055 			  struct share_check *sc)
1056 {
1057 	struct btrfs_fs_info *fs_info = extent_root->fs_info;
1058 	int ret;
1059 	int slot;
1060 	struct extent_buffer *leaf;
1061 	struct btrfs_key key;
1062 
1063 	while (1) {
1064 		ret = btrfs_next_item(extent_root, path);
1065 		if (ret < 0)
1066 			break;
1067 		if (ret) {
1068 			ret = 0;
1069 			break;
1070 		}
1071 
1072 		slot = path->slots[0];
1073 		leaf = path->nodes[0];
1074 		btrfs_item_key_to_cpu(leaf, &key, slot);
1075 
1076 		if (key.objectid != bytenr)
1077 			break;
1078 		if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1079 			continue;
1080 		if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1081 			break;
1082 
1083 		switch (key.type) {
1084 		case BTRFS_SHARED_BLOCK_REF_KEY:
1085 			/* SHARED DIRECT METADATA backref */
1086 			ret = add_direct_ref(fs_info, preftrees,
1087 					     info_level + 1, key.offset,
1088 					     bytenr, 1, NULL, GFP_NOFS);
1089 			break;
1090 		case BTRFS_SHARED_DATA_REF_KEY: {
1091 			/* SHARED DIRECT FULL backref */
1092 			struct btrfs_shared_data_ref *sdref;
1093 			int count;
1094 
1095 			sdref = btrfs_item_ptr(leaf, slot,
1096 					      struct btrfs_shared_data_ref);
1097 			count = btrfs_shared_data_ref_count(leaf, sdref);
1098 			ret = add_direct_ref(fs_info, preftrees, 0,
1099 					     key.offset, bytenr, count,
1100 					     sc, GFP_NOFS);
1101 			break;
1102 		}
1103 		case BTRFS_TREE_BLOCK_REF_KEY:
1104 			/* NORMAL INDIRECT METADATA backref */
1105 			ret = add_indirect_ref(fs_info, preftrees, key.offset,
1106 					       NULL, info_level + 1, bytenr,
1107 					       1, NULL, GFP_NOFS);
1108 			break;
1109 		case BTRFS_EXTENT_DATA_REF_KEY: {
1110 			/* NORMAL INDIRECT DATA backref */
1111 			struct btrfs_extent_data_ref *dref;
1112 			int count;
1113 			u64 root;
1114 
1115 			dref = btrfs_item_ptr(leaf, slot,
1116 					      struct btrfs_extent_data_ref);
1117 			count = btrfs_extent_data_ref_count(leaf, dref);
1118 			key.objectid = btrfs_extent_data_ref_objectid(leaf,
1119 								      dref);
1120 			key.type = BTRFS_EXTENT_DATA_KEY;
1121 			key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1122 
1123 			if (sc && sc->inum && key.objectid != sc->inum) {
1124 				ret = BACKREF_FOUND_SHARED;
1125 				break;
1126 			}
1127 
1128 			root = btrfs_extent_data_ref_root(leaf, dref);
1129 			ret = add_indirect_ref(fs_info, preftrees, root,
1130 					       &key, 0, bytenr, count,
1131 					       sc, GFP_NOFS);
1132 			break;
1133 		}
1134 		default:
1135 			WARN_ON(1);
1136 		}
1137 		if (ret)
1138 			return ret;
1139 
1140 	}
1141 
1142 	return ret;
1143 }
1144 
1145 /*
1146  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1147  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1148  * indirect refs to their parent bytenr.
1149  * When roots are found, they're added to the roots list
1150  *
1151  * If time_seq is set to BTRFS_SEQ_LAST, it will not search delayed_refs, and
1152  * behave much like trans == NULL case, the difference only lies in it will not
1153  * commit root.
1154  * The special case is for qgroup to search roots in commit_transaction().
1155  *
1156  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1157  * shared extent is detected.
1158  *
1159  * Otherwise this returns 0 for success and <0 for an error.
1160  *
1161  * If ignore_offset is set to false, only extent refs whose offsets match
1162  * extent_item_pos are returned.  If true, every extent ref is returned
1163  * and extent_item_pos is ignored.
1164  *
1165  * FIXME some caching might speed things up
1166  */
1167 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1168 			     struct btrfs_fs_info *fs_info, u64 bytenr,
1169 			     u64 time_seq, struct ulist *refs,
1170 			     struct ulist *roots, const u64 *extent_item_pos,
1171 			     struct share_check *sc, bool ignore_offset)
1172 {
1173 	struct btrfs_root *root = fs_info->extent_root;
1174 	struct btrfs_key key;
1175 	struct btrfs_path *path;
1176 	struct btrfs_delayed_ref_root *delayed_refs = NULL;
1177 	struct btrfs_delayed_ref_head *head;
1178 	int info_level = 0;
1179 	int ret;
1180 	struct prelim_ref *ref;
1181 	struct rb_node *node;
1182 	struct extent_inode_elem *eie = NULL;
1183 	struct preftrees preftrees = {
1184 		.direct = PREFTREE_INIT,
1185 		.indirect = PREFTREE_INIT,
1186 		.indirect_missing_keys = PREFTREE_INIT
1187 	};
1188 
1189 	key.objectid = bytenr;
1190 	key.offset = (u64)-1;
1191 	if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1192 		key.type = BTRFS_METADATA_ITEM_KEY;
1193 	else
1194 		key.type = BTRFS_EXTENT_ITEM_KEY;
1195 
1196 	path = btrfs_alloc_path();
1197 	if (!path)
1198 		return -ENOMEM;
1199 	if (!trans) {
1200 		path->search_commit_root = 1;
1201 		path->skip_locking = 1;
1202 	}
1203 
1204 	if (time_seq == BTRFS_SEQ_LAST)
1205 		path->skip_locking = 1;
1206 
1207 	/*
1208 	 * grab both a lock on the path and a lock on the delayed ref head.
1209 	 * We need both to get a consistent picture of how the refs look
1210 	 * at a specified point in time
1211 	 */
1212 again:
1213 	head = NULL;
1214 
1215 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1216 	if (ret < 0)
1217 		goto out;
1218 	BUG_ON(ret == 0);
1219 
1220 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1221 	if (trans && likely(trans->type != __TRANS_DUMMY) &&
1222 	    time_seq != BTRFS_SEQ_LAST) {
1223 #else
1224 	if (trans && time_seq != BTRFS_SEQ_LAST) {
1225 #endif
1226 		/*
1227 		 * look if there are updates for this ref queued and lock the
1228 		 * head
1229 		 */
1230 		delayed_refs = &trans->transaction->delayed_refs;
1231 		spin_lock(&delayed_refs->lock);
1232 		head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1233 		if (head) {
1234 			if (!mutex_trylock(&head->mutex)) {
1235 				refcount_inc(&head->refs);
1236 				spin_unlock(&delayed_refs->lock);
1237 
1238 				btrfs_release_path(path);
1239 
1240 				/*
1241 				 * Mutex was contended, block until it's
1242 				 * released and try again
1243 				 */
1244 				mutex_lock(&head->mutex);
1245 				mutex_unlock(&head->mutex);
1246 				btrfs_put_delayed_ref_head(head);
1247 				goto again;
1248 			}
1249 			spin_unlock(&delayed_refs->lock);
1250 			ret = add_delayed_refs(fs_info, head, time_seq,
1251 					       &preftrees, sc);
1252 			mutex_unlock(&head->mutex);
1253 			if (ret)
1254 				goto out;
1255 		} else {
1256 			spin_unlock(&delayed_refs->lock);
1257 		}
1258 	}
1259 
1260 	if (path->slots[0]) {
1261 		struct extent_buffer *leaf;
1262 		int slot;
1263 
1264 		path->slots[0]--;
1265 		leaf = path->nodes[0];
1266 		slot = path->slots[0];
1267 		btrfs_item_key_to_cpu(leaf, &key, slot);
1268 		if (key.objectid == bytenr &&
1269 		    (key.type == BTRFS_EXTENT_ITEM_KEY ||
1270 		     key.type == BTRFS_METADATA_ITEM_KEY)) {
1271 			ret = add_inline_refs(fs_info, path, bytenr,
1272 					      &info_level, &preftrees, sc);
1273 			if (ret)
1274 				goto out;
1275 			ret = add_keyed_refs(root, path, bytenr, info_level,
1276 					     &preftrees, sc);
1277 			if (ret)
1278 				goto out;
1279 		}
1280 	}
1281 
1282 	btrfs_release_path(path);
1283 
1284 	ret = add_missing_keys(fs_info, &preftrees, path->skip_locking == 0);
1285 	if (ret)
1286 		goto out;
1287 
1288 	WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1289 
1290 	ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1291 				    extent_item_pos, sc, ignore_offset);
1292 	if (ret)
1293 		goto out;
1294 
1295 	WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1296 
1297 	/*
1298 	 * This walks the tree of merged and resolved refs. Tree blocks are
1299 	 * read in as needed. Unique entries are added to the ulist, and
1300 	 * the list of found roots is updated.
1301 	 *
1302 	 * We release the entire tree in one go before returning.
1303 	 */
1304 	node = rb_first_cached(&preftrees.direct.root);
1305 	while (node) {
1306 		ref = rb_entry(node, struct prelim_ref, rbnode);
1307 		node = rb_next(&ref->rbnode);
1308 		/*
1309 		 * ref->count < 0 can happen here if there are delayed
1310 		 * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1311 		 * prelim_ref_insert() relies on this when merging
1312 		 * identical refs to keep the overall count correct.
1313 		 * prelim_ref_insert() will merge only those refs
1314 		 * which compare identically.  Any refs having
1315 		 * e.g. different offsets would not be merged,
1316 		 * and would retain their original ref->count < 0.
1317 		 */
1318 		if (roots && ref->count && ref->root_id && ref->parent == 0) {
1319 			if (sc && sc->root_objectid &&
1320 			    ref->root_id != sc->root_objectid) {
1321 				ret = BACKREF_FOUND_SHARED;
1322 				goto out;
1323 			}
1324 
1325 			/* no parent == root of tree */
1326 			ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1327 			if (ret < 0)
1328 				goto out;
1329 		}
1330 		if (ref->count && ref->parent) {
1331 			if (extent_item_pos && !ref->inode_list &&
1332 			    ref->level == 0) {
1333 				struct extent_buffer *eb;
1334 
1335 				eb = read_tree_block(fs_info, ref->parent, 0,
1336 						     0, ref->level, NULL);
1337 				if (IS_ERR(eb)) {
1338 					ret = PTR_ERR(eb);
1339 					goto out;
1340 				} else if (!extent_buffer_uptodate(eb)) {
1341 					free_extent_buffer(eb);
1342 					ret = -EIO;
1343 					goto out;
1344 				}
1345 
1346 				if (!path->skip_locking)
1347 					btrfs_tree_read_lock(eb);
1348 				ret = find_extent_in_eb(eb, bytenr,
1349 							*extent_item_pos, &eie, ignore_offset);
1350 				if (!path->skip_locking)
1351 					btrfs_tree_read_unlock(eb);
1352 				free_extent_buffer(eb);
1353 				if (ret < 0)
1354 					goto out;
1355 				ref->inode_list = eie;
1356 			}
1357 			ret = ulist_add_merge_ptr(refs, ref->parent,
1358 						  ref->inode_list,
1359 						  (void **)&eie, GFP_NOFS);
1360 			if (ret < 0)
1361 				goto out;
1362 			if (!ret && extent_item_pos) {
1363 				/*
1364 				 * we've recorded that parent, so we must extend
1365 				 * its inode list here
1366 				 */
1367 				BUG_ON(!eie);
1368 				while (eie->next)
1369 					eie = eie->next;
1370 				eie->next = ref->inode_list;
1371 			}
1372 			eie = NULL;
1373 		}
1374 		cond_resched();
1375 	}
1376 
1377 out:
1378 	btrfs_free_path(path);
1379 
1380 	prelim_release(&preftrees.direct);
1381 	prelim_release(&preftrees.indirect);
1382 	prelim_release(&preftrees.indirect_missing_keys);
1383 
1384 	if (ret < 0)
1385 		free_inode_elem_list(eie);
1386 	return ret;
1387 }
1388 
1389 static void free_leaf_list(struct ulist *blocks)
1390 {
1391 	struct ulist_node *node = NULL;
1392 	struct extent_inode_elem *eie;
1393 	struct ulist_iterator uiter;
1394 
1395 	ULIST_ITER_INIT(&uiter);
1396 	while ((node = ulist_next(blocks, &uiter))) {
1397 		if (!node->aux)
1398 			continue;
1399 		eie = unode_aux_to_inode_list(node);
1400 		free_inode_elem_list(eie);
1401 		node->aux = 0;
1402 	}
1403 
1404 	ulist_free(blocks);
1405 }
1406 
1407 /*
1408  * Finds all leafs with a reference to the specified combination of bytenr and
1409  * offset. key_list_head will point to a list of corresponding keys (caller must
1410  * free each list element). The leafs will be stored in the leafs ulist, which
1411  * must be freed with ulist_free.
1412  *
1413  * returns 0 on success, <0 on error
1414  */
1415 int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1416 			 struct btrfs_fs_info *fs_info, u64 bytenr,
1417 			 u64 time_seq, struct ulist **leafs,
1418 			 const u64 *extent_item_pos, bool ignore_offset)
1419 {
1420 	int ret;
1421 
1422 	*leafs = ulist_alloc(GFP_NOFS);
1423 	if (!*leafs)
1424 		return -ENOMEM;
1425 
1426 	ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1427 				*leafs, NULL, extent_item_pos, NULL, ignore_offset);
1428 	if (ret < 0 && ret != -ENOENT) {
1429 		free_leaf_list(*leafs);
1430 		return ret;
1431 	}
1432 
1433 	return 0;
1434 }
1435 
1436 /*
1437  * walk all backrefs for a given extent to find all roots that reference this
1438  * extent. Walking a backref means finding all extents that reference this
1439  * extent and in turn walk the backrefs of those, too. Naturally this is a
1440  * recursive process, but here it is implemented in an iterative fashion: We
1441  * find all referencing extents for the extent in question and put them on a
1442  * list. In turn, we find all referencing extents for those, further appending
1443  * to the list. The way we iterate the list allows adding more elements after
1444  * the current while iterating. The process stops when we reach the end of the
1445  * list. Found roots are added to the roots list.
1446  *
1447  * returns 0 on success, < 0 on error.
1448  */
1449 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1450 				     struct btrfs_fs_info *fs_info, u64 bytenr,
1451 				     u64 time_seq, struct ulist **roots,
1452 				     bool ignore_offset)
1453 {
1454 	struct ulist *tmp;
1455 	struct ulist_node *node = NULL;
1456 	struct ulist_iterator uiter;
1457 	int ret;
1458 
1459 	tmp = ulist_alloc(GFP_NOFS);
1460 	if (!tmp)
1461 		return -ENOMEM;
1462 	*roots = ulist_alloc(GFP_NOFS);
1463 	if (!*roots) {
1464 		ulist_free(tmp);
1465 		return -ENOMEM;
1466 	}
1467 
1468 	ULIST_ITER_INIT(&uiter);
1469 	while (1) {
1470 		ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1471 					tmp, *roots, NULL, NULL, ignore_offset);
1472 		if (ret < 0 && ret != -ENOENT) {
1473 			ulist_free(tmp);
1474 			ulist_free(*roots);
1475 			*roots = NULL;
1476 			return ret;
1477 		}
1478 		node = ulist_next(tmp, &uiter);
1479 		if (!node)
1480 			break;
1481 		bytenr = node->val;
1482 		cond_resched();
1483 	}
1484 
1485 	ulist_free(tmp);
1486 	return 0;
1487 }
1488 
1489 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1490 			 struct btrfs_fs_info *fs_info, u64 bytenr,
1491 			 u64 time_seq, struct ulist **roots,
1492 			 bool skip_commit_root_sem)
1493 {
1494 	int ret;
1495 
1496 	if (!trans && !skip_commit_root_sem)
1497 		down_read(&fs_info->commit_root_sem);
1498 	ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1499 					time_seq, roots, false);
1500 	if (!trans && !skip_commit_root_sem)
1501 		up_read(&fs_info->commit_root_sem);
1502 	return ret;
1503 }
1504 
1505 /**
1506  * Check if an extent is shared or not
1507  *
1508  * @root:   root inode belongs to
1509  * @inum:   inode number of the inode whose extent we are checking
1510  * @bytenr: logical bytenr of the extent we are checking
1511  * @roots:  list of roots this extent is shared among
1512  * @tmp:    temporary list used for iteration
1513  *
1514  * btrfs_check_shared uses the backref walking code but will short
1515  * circuit as soon as it finds a root or inode that doesn't match the
1516  * one passed in. This provides a significant performance benefit for
1517  * callers (such as fiemap) which want to know whether the extent is
1518  * shared but do not need a ref count.
1519  *
1520  * This attempts to attach to the running transaction in order to account for
1521  * delayed refs, but continues on even when no running transaction exists.
1522  *
1523  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1524  */
1525 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr,
1526 		struct ulist *roots, struct ulist *tmp)
1527 {
1528 	struct btrfs_fs_info *fs_info = root->fs_info;
1529 	struct btrfs_trans_handle *trans;
1530 	struct ulist_iterator uiter;
1531 	struct ulist_node *node;
1532 	struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1533 	int ret = 0;
1534 	struct share_check shared = {
1535 		.root_objectid = root->root_key.objectid,
1536 		.inum = inum,
1537 		.share_count = 0,
1538 	};
1539 
1540 	ulist_init(roots);
1541 	ulist_init(tmp);
1542 
1543 	trans = btrfs_join_transaction_nostart(root);
1544 	if (IS_ERR(trans)) {
1545 		if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1546 			ret = PTR_ERR(trans);
1547 			goto out;
1548 		}
1549 		trans = NULL;
1550 		down_read(&fs_info->commit_root_sem);
1551 	} else {
1552 		btrfs_get_tree_mod_seq(fs_info, &elem);
1553 	}
1554 
1555 	ULIST_ITER_INIT(&uiter);
1556 	while (1) {
1557 		ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1558 					roots, NULL, &shared, false);
1559 		if (ret == BACKREF_FOUND_SHARED) {
1560 			/* this is the only condition under which we return 1 */
1561 			ret = 1;
1562 			break;
1563 		}
1564 		if (ret < 0 && ret != -ENOENT)
1565 			break;
1566 		ret = 0;
1567 		node = ulist_next(tmp, &uiter);
1568 		if (!node)
1569 			break;
1570 		bytenr = node->val;
1571 		shared.share_count = 0;
1572 		cond_resched();
1573 	}
1574 
1575 	if (trans) {
1576 		btrfs_put_tree_mod_seq(fs_info, &elem);
1577 		btrfs_end_transaction(trans);
1578 	} else {
1579 		up_read(&fs_info->commit_root_sem);
1580 	}
1581 out:
1582 	ulist_release(roots);
1583 	ulist_release(tmp);
1584 	return ret;
1585 }
1586 
1587 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1588 			  u64 start_off, struct btrfs_path *path,
1589 			  struct btrfs_inode_extref **ret_extref,
1590 			  u64 *found_off)
1591 {
1592 	int ret, slot;
1593 	struct btrfs_key key;
1594 	struct btrfs_key found_key;
1595 	struct btrfs_inode_extref *extref;
1596 	const struct extent_buffer *leaf;
1597 	unsigned long ptr;
1598 
1599 	key.objectid = inode_objectid;
1600 	key.type = BTRFS_INODE_EXTREF_KEY;
1601 	key.offset = start_off;
1602 
1603 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1604 	if (ret < 0)
1605 		return ret;
1606 
1607 	while (1) {
1608 		leaf = path->nodes[0];
1609 		slot = path->slots[0];
1610 		if (slot >= btrfs_header_nritems(leaf)) {
1611 			/*
1612 			 * If the item at offset is not found,
1613 			 * btrfs_search_slot will point us to the slot
1614 			 * where it should be inserted. In our case
1615 			 * that will be the slot directly before the
1616 			 * next INODE_REF_KEY_V2 item. In the case
1617 			 * that we're pointing to the last slot in a
1618 			 * leaf, we must move one leaf over.
1619 			 */
1620 			ret = btrfs_next_leaf(root, path);
1621 			if (ret) {
1622 				if (ret >= 1)
1623 					ret = -ENOENT;
1624 				break;
1625 			}
1626 			continue;
1627 		}
1628 
1629 		btrfs_item_key_to_cpu(leaf, &found_key, slot);
1630 
1631 		/*
1632 		 * Check that we're still looking at an extended ref key for
1633 		 * this particular objectid. If we have different
1634 		 * objectid or type then there are no more to be found
1635 		 * in the tree and we can exit.
1636 		 */
1637 		ret = -ENOENT;
1638 		if (found_key.objectid != inode_objectid)
1639 			break;
1640 		if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1641 			break;
1642 
1643 		ret = 0;
1644 		ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1645 		extref = (struct btrfs_inode_extref *)ptr;
1646 		*ret_extref = extref;
1647 		if (found_off)
1648 			*found_off = found_key.offset;
1649 		break;
1650 	}
1651 
1652 	return ret;
1653 }
1654 
1655 /*
1656  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1657  * Elements of the path are separated by '/' and the path is guaranteed to be
1658  * 0-terminated. the path is only given within the current file system.
1659  * Therefore, it never starts with a '/'. the caller is responsible to provide
1660  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1661  * the start point of the resulting string is returned. this pointer is within
1662  * dest, normally.
1663  * in case the path buffer would overflow, the pointer is decremented further
1664  * as if output was written to the buffer, though no more output is actually
1665  * generated. that way, the caller can determine how much space would be
1666  * required for the path to fit into the buffer. in that case, the returned
1667  * value will be smaller than dest. callers must check this!
1668  */
1669 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1670 			u32 name_len, unsigned long name_off,
1671 			struct extent_buffer *eb_in, u64 parent,
1672 			char *dest, u32 size)
1673 {
1674 	int slot;
1675 	u64 next_inum;
1676 	int ret;
1677 	s64 bytes_left = ((s64)size) - 1;
1678 	struct extent_buffer *eb = eb_in;
1679 	struct btrfs_key found_key;
1680 	struct btrfs_inode_ref *iref;
1681 
1682 	if (bytes_left >= 0)
1683 		dest[bytes_left] = '\0';
1684 
1685 	while (1) {
1686 		bytes_left -= name_len;
1687 		if (bytes_left >= 0)
1688 			read_extent_buffer(eb, dest + bytes_left,
1689 					   name_off, name_len);
1690 		if (eb != eb_in) {
1691 			if (!path->skip_locking)
1692 				btrfs_tree_read_unlock(eb);
1693 			free_extent_buffer(eb);
1694 		}
1695 		ret = btrfs_find_item(fs_root, path, parent, 0,
1696 				BTRFS_INODE_REF_KEY, &found_key);
1697 		if (ret > 0)
1698 			ret = -ENOENT;
1699 		if (ret)
1700 			break;
1701 
1702 		next_inum = found_key.offset;
1703 
1704 		/* regular exit ahead */
1705 		if (parent == next_inum)
1706 			break;
1707 
1708 		slot = path->slots[0];
1709 		eb = path->nodes[0];
1710 		/* make sure we can use eb after releasing the path */
1711 		if (eb != eb_in) {
1712 			path->nodes[0] = NULL;
1713 			path->locks[0] = 0;
1714 		}
1715 		btrfs_release_path(path);
1716 		iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1717 
1718 		name_len = btrfs_inode_ref_name_len(eb, iref);
1719 		name_off = (unsigned long)(iref + 1);
1720 
1721 		parent = next_inum;
1722 		--bytes_left;
1723 		if (bytes_left >= 0)
1724 			dest[bytes_left] = '/';
1725 	}
1726 
1727 	btrfs_release_path(path);
1728 
1729 	if (ret)
1730 		return ERR_PTR(ret);
1731 
1732 	return dest + bytes_left;
1733 }
1734 
1735 /*
1736  * this makes the path point to (logical EXTENT_ITEM *)
1737  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1738  * tree blocks and <0 on error.
1739  */
1740 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1741 			struct btrfs_path *path, struct btrfs_key *found_key,
1742 			u64 *flags_ret)
1743 {
1744 	int ret;
1745 	u64 flags;
1746 	u64 size = 0;
1747 	u32 item_size;
1748 	const struct extent_buffer *eb;
1749 	struct btrfs_extent_item *ei;
1750 	struct btrfs_key key;
1751 
1752 	if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1753 		key.type = BTRFS_METADATA_ITEM_KEY;
1754 	else
1755 		key.type = BTRFS_EXTENT_ITEM_KEY;
1756 	key.objectid = logical;
1757 	key.offset = (u64)-1;
1758 
1759 	ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1760 	if (ret < 0)
1761 		return ret;
1762 
1763 	ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1764 	if (ret) {
1765 		if (ret > 0)
1766 			ret = -ENOENT;
1767 		return ret;
1768 	}
1769 	btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1770 	if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1771 		size = fs_info->nodesize;
1772 	else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1773 		size = found_key->offset;
1774 
1775 	if (found_key->objectid > logical ||
1776 	    found_key->objectid + size <= logical) {
1777 		btrfs_debug(fs_info,
1778 			"logical %llu is not within any extent", logical);
1779 		return -ENOENT;
1780 	}
1781 
1782 	eb = path->nodes[0];
1783 	item_size = btrfs_item_size(eb, path->slots[0]);
1784 	BUG_ON(item_size < sizeof(*ei));
1785 
1786 	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1787 	flags = btrfs_extent_flags(eb, ei);
1788 
1789 	btrfs_debug(fs_info,
1790 		"logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1791 		 logical, logical - found_key->objectid, found_key->objectid,
1792 		 found_key->offset, flags, item_size);
1793 
1794 	WARN_ON(!flags_ret);
1795 	if (flags_ret) {
1796 		if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1797 			*flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1798 		else if (flags & BTRFS_EXTENT_FLAG_DATA)
1799 			*flags_ret = BTRFS_EXTENT_FLAG_DATA;
1800 		else
1801 			BUG();
1802 		return 0;
1803 	}
1804 
1805 	return -EIO;
1806 }
1807 
1808 /*
1809  * helper function to iterate extent inline refs. ptr must point to a 0 value
1810  * for the first call and may be modified. it is used to track state.
1811  * if more refs exist, 0 is returned and the next call to
1812  * get_extent_inline_ref must pass the modified ptr parameter to get the
1813  * next ref. after the last ref was processed, 1 is returned.
1814  * returns <0 on error
1815  */
1816 static int get_extent_inline_ref(unsigned long *ptr,
1817 				 const struct extent_buffer *eb,
1818 				 const struct btrfs_key *key,
1819 				 const struct btrfs_extent_item *ei,
1820 				 u32 item_size,
1821 				 struct btrfs_extent_inline_ref **out_eiref,
1822 				 int *out_type)
1823 {
1824 	unsigned long end;
1825 	u64 flags;
1826 	struct btrfs_tree_block_info *info;
1827 
1828 	if (!*ptr) {
1829 		/* first call */
1830 		flags = btrfs_extent_flags(eb, ei);
1831 		if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1832 			if (key->type == BTRFS_METADATA_ITEM_KEY) {
1833 				/* a skinny metadata extent */
1834 				*out_eiref =
1835 				     (struct btrfs_extent_inline_ref *)(ei + 1);
1836 			} else {
1837 				WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1838 				info = (struct btrfs_tree_block_info *)(ei + 1);
1839 				*out_eiref =
1840 				   (struct btrfs_extent_inline_ref *)(info + 1);
1841 			}
1842 		} else {
1843 			*out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1844 		}
1845 		*ptr = (unsigned long)*out_eiref;
1846 		if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1847 			return -ENOENT;
1848 	}
1849 
1850 	end = (unsigned long)ei + item_size;
1851 	*out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1852 	*out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1853 						     BTRFS_REF_TYPE_ANY);
1854 	if (*out_type == BTRFS_REF_TYPE_INVALID)
1855 		return -EUCLEAN;
1856 
1857 	*ptr += btrfs_extent_inline_ref_size(*out_type);
1858 	WARN_ON(*ptr > end);
1859 	if (*ptr == end)
1860 		return 1; /* last */
1861 
1862 	return 0;
1863 }
1864 
1865 /*
1866  * reads the tree block backref for an extent. tree level and root are returned
1867  * through out_level and out_root. ptr must point to a 0 value for the first
1868  * call and may be modified (see get_extent_inline_ref comment).
1869  * returns 0 if data was provided, 1 if there was no more data to provide or
1870  * <0 on error.
1871  */
1872 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1873 			    struct btrfs_key *key, struct btrfs_extent_item *ei,
1874 			    u32 item_size, u64 *out_root, u8 *out_level)
1875 {
1876 	int ret;
1877 	int type;
1878 	struct btrfs_extent_inline_ref *eiref;
1879 
1880 	if (*ptr == (unsigned long)-1)
1881 		return 1;
1882 
1883 	while (1) {
1884 		ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1885 					      &eiref, &type);
1886 		if (ret < 0)
1887 			return ret;
1888 
1889 		if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1890 		    type == BTRFS_SHARED_BLOCK_REF_KEY)
1891 			break;
1892 
1893 		if (ret == 1)
1894 			return 1;
1895 	}
1896 
1897 	/* we can treat both ref types equally here */
1898 	*out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1899 
1900 	if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1901 		struct btrfs_tree_block_info *info;
1902 
1903 		info = (struct btrfs_tree_block_info *)(ei + 1);
1904 		*out_level = btrfs_tree_block_level(eb, info);
1905 	} else {
1906 		ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1907 		*out_level = (u8)key->offset;
1908 	}
1909 
1910 	if (ret == 1)
1911 		*ptr = (unsigned long)-1;
1912 
1913 	return 0;
1914 }
1915 
1916 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1917 			     struct extent_inode_elem *inode_list,
1918 			     u64 root, u64 extent_item_objectid,
1919 			     iterate_extent_inodes_t *iterate, void *ctx)
1920 {
1921 	struct extent_inode_elem *eie;
1922 	int ret = 0;
1923 
1924 	for (eie = inode_list; eie; eie = eie->next) {
1925 		btrfs_debug(fs_info,
1926 			    "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1927 			    extent_item_objectid, eie->inum,
1928 			    eie->offset, root);
1929 		ret = iterate(eie->inum, eie->offset, root, ctx);
1930 		if (ret) {
1931 			btrfs_debug(fs_info,
1932 				    "stopping iteration for %llu due to ret=%d",
1933 				    extent_item_objectid, ret);
1934 			break;
1935 		}
1936 	}
1937 
1938 	return ret;
1939 }
1940 
1941 /*
1942  * calls iterate() for every inode that references the extent identified by
1943  * the given parameters.
1944  * when the iterator function returns a non-zero value, iteration stops.
1945  */
1946 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1947 				u64 extent_item_objectid, u64 extent_item_pos,
1948 				int search_commit_root,
1949 				iterate_extent_inodes_t *iterate, void *ctx,
1950 				bool ignore_offset)
1951 {
1952 	int ret;
1953 	struct btrfs_trans_handle *trans = NULL;
1954 	struct ulist *refs = NULL;
1955 	struct ulist *roots = NULL;
1956 	struct ulist_node *ref_node = NULL;
1957 	struct ulist_node *root_node = NULL;
1958 	struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
1959 	struct ulist_iterator ref_uiter;
1960 	struct ulist_iterator root_uiter;
1961 
1962 	btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1963 			extent_item_objectid);
1964 
1965 	if (!search_commit_root) {
1966 		trans = btrfs_attach_transaction(fs_info->extent_root);
1967 		if (IS_ERR(trans)) {
1968 			if (PTR_ERR(trans) != -ENOENT &&
1969 			    PTR_ERR(trans) != -EROFS)
1970 				return PTR_ERR(trans);
1971 			trans = NULL;
1972 		}
1973 	}
1974 
1975 	if (trans)
1976 		btrfs_get_tree_mod_seq(fs_info, &seq_elem);
1977 	else
1978 		down_read(&fs_info->commit_root_sem);
1979 
1980 	ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1981 				   seq_elem.seq, &refs,
1982 				   &extent_item_pos, ignore_offset);
1983 	if (ret)
1984 		goto out;
1985 
1986 	ULIST_ITER_INIT(&ref_uiter);
1987 	while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1988 		ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1989 						seq_elem.seq, &roots,
1990 						ignore_offset);
1991 		if (ret)
1992 			break;
1993 		ULIST_ITER_INIT(&root_uiter);
1994 		while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1995 			btrfs_debug(fs_info,
1996 				    "root %llu references leaf %llu, data list %#llx",
1997 				    root_node->val, ref_node->val,
1998 				    ref_node->aux);
1999 			ret = iterate_leaf_refs(fs_info,
2000 						(struct extent_inode_elem *)
2001 						(uintptr_t)ref_node->aux,
2002 						root_node->val,
2003 						extent_item_objectid,
2004 						iterate, ctx);
2005 		}
2006 		ulist_free(roots);
2007 	}
2008 
2009 	free_leaf_list(refs);
2010 out:
2011 	if (trans) {
2012 		btrfs_put_tree_mod_seq(fs_info, &seq_elem);
2013 		btrfs_end_transaction(trans);
2014 	} else {
2015 		up_read(&fs_info->commit_root_sem);
2016 	}
2017 
2018 	return ret;
2019 }
2020 
2021 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2022 				struct btrfs_path *path,
2023 				iterate_extent_inodes_t *iterate, void *ctx,
2024 				bool ignore_offset)
2025 {
2026 	int ret;
2027 	u64 extent_item_pos;
2028 	u64 flags = 0;
2029 	struct btrfs_key found_key;
2030 	int search_commit_root = path->search_commit_root;
2031 
2032 	ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2033 	btrfs_release_path(path);
2034 	if (ret < 0)
2035 		return ret;
2036 	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2037 		return -EINVAL;
2038 
2039 	extent_item_pos = logical - found_key.objectid;
2040 	ret = iterate_extent_inodes(fs_info, found_key.objectid,
2041 					extent_item_pos, search_commit_root,
2042 					iterate, ctx, ignore_offset);
2043 
2044 	return ret;
2045 }
2046 
2047 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2048 			      struct extent_buffer *eb, void *ctx);
2049 
2050 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2051 			      struct btrfs_path *path,
2052 			      iterate_irefs_t *iterate, void *ctx)
2053 {
2054 	int ret = 0;
2055 	int slot;
2056 	u32 cur;
2057 	u32 len;
2058 	u32 name_len;
2059 	u64 parent = 0;
2060 	int found = 0;
2061 	struct extent_buffer *eb;
2062 	struct btrfs_inode_ref *iref;
2063 	struct btrfs_key found_key;
2064 
2065 	while (!ret) {
2066 		ret = btrfs_find_item(fs_root, path, inum,
2067 				parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2068 				&found_key);
2069 
2070 		if (ret < 0)
2071 			break;
2072 		if (ret) {
2073 			ret = found ? 0 : -ENOENT;
2074 			break;
2075 		}
2076 		++found;
2077 
2078 		parent = found_key.offset;
2079 		slot = path->slots[0];
2080 		eb = btrfs_clone_extent_buffer(path->nodes[0]);
2081 		if (!eb) {
2082 			ret = -ENOMEM;
2083 			break;
2084 		}
2085 		btrfs_release_path(path);
2086 
2087 		iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2088 
2089 		for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2090 			name_len = btrfs_inode_ref_name_len(eb, iref);
2091 			/* path must be released before calling iterate()! */
2092 			btrfs_debug(fs_root->fs_info,
2093 				"following ref at offset %u for inode %llu in tree %llu",
2094 				cur, found_key.objectid,
2095 				fs_root->root_key.objectid);
2096 			ret = iterate(parent, name_len,
2097 				      (unsigned long)(iref + 1), eb, ctx);
2098 			if (ret)
2099 				break;
2100 			len = sizeof(*iref) + name_len;
2101 			iref = (struct btrfs_inode_ref *)((char *)iref + len);
2102 		}
2103 		free_extent_buffer(eb);
2104 	}
2105 
2106 	btrfs_release_path(path);
2107 
2108 	return ret;
2109 }
2110 
2111 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2112 				 struct btrfs_path *path,
2113 				 iterate_irefs_t *iterate, void *ctx)
2114 {
2115 	int ret;
2116 	int slot;
2117 	u64 offset = 0;
2118 	u64 parent;
2119 	int found = 0;
2120 	struct extent_buffer *eb;
2121 	struct btrfs_inode_extref *extref;
2122 	u32 item_size;
2123 	u32 cur_offset;
2124 	unsigned long ptr;
2125 
2126 	while (1) {
2127 		ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2128 					    &offset);
2129 		if (ret < 0)
2130 			break;
2131 		if (ret) {
2132 			ret = found ? 0 : -ENOENT;
2133 			break;
2134 		}
2135 		++found;
2136 
2137 		slot = path->slots[0];
2138 		eb = btrfs_clone_extent_buffer(path->nodes[0]);
2139 		if (!eb) {
2140 			ret = -ENOMEM;
2141 			break;
2142 		}
2143 		btrfs_release_path(path);
2144 
2145 		item_size = btrfs_item_size(eb, slot);
2146 		ptr = btrfs_item_ptr_offset(eb, slot);
2147 		cur_offset = 0;
2148 
2149 		while (cur_offset < item_size) {
2150 			u32 name_len;
2151 
2152 			extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2153 			parent = btrfs_inode_extref_parent(eb, extref);
2154 			name_len = btrfs_inode_extref_name_len(eb, extref);
2155 			ret = iterate(parent, name_len,
2156 				      (unsigned long)&extref->name, eb, ctx);
2157 			if (ret)
2158 				break;
2159 
2160 			cur_offset += btrfs_inode_extref_name_len(eb, extref);
2161 			cur_offset += sizeof(*extref);
2162 		}
2163 		free_extent_buffer(eb);
2164 
2165 		offset++;
2166 	}
2167 
2168 	btrfs_release_path(path);
2169 
2170 	return ret;
2171 }
2172 
2173 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2174 			 struct btrfs_path *path, iterate_irefs_t *iterate,
2175 			 void *ctx)
2176 {
2177 	int ret;
2178 	int found_refs = 0;
2179 
2180 	ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2181 	if (!ret)
2182 		++found_refs;
2183 	else if (ret != -ENOENT)
2184 		return ret;
2185 
2186 	ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2187 	if (ret == -ENOENT && found_refs)
2188 		return 0;
2189 
2190 	return ret;
2191 }
2192 
2193 /*
2194  * returns 0 if the path could be dumped (probably truncated)
2195  * returns <0 in case of an error
2196  */
2197 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2198 			 struct extent_buffer *eb, void *ctx)
2199 {
2200 	struct inode_fs_paths *ipath = ctx;
2201 	char *fspath;
2202 	char *fspath_min;
2203 	int i = ipath->fspath->elem_cnt;
2204 	const int s_ptr = sizeof(char *);
2205 	u32 bytes_left;
2206 
2207 	bytes_left = ipath->fspath->bytes_left > s_ptr ?
2208 					ipath->fspath->bytes_left - s_ptr : 0;
2209 
2210 	fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2211 	fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2212 				   name_off, eb, inum, fspath_min, bytes_left);
2213 	if (IS_ERR(fspath))
2214 		return PTR_ERR(fspath);
2215 
2216 	if (fspath > fspath_min) {
2217 		ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2218 		++ipath->fspath->elem_cnt;
2219 		ipath->fspath->bytes_left = fspath - fspath_min;
2220 	} else {
2221 		++ipath->fspath->elem_missed;
2222 		ipath->fspath->bytes_missing += fspath_min - fspath;
2223 		ipath->fspath->bytes_left = 0;
2224 	}
2225 
2226 	return 0;
2227 }
2228 
2229 /*
2230  * this dumps all file system paths to the inode into the ipath struct, provided
2231  * is has been created large enough. each path is zero-terminated and accessed
2232  * from ipath->fspath->val[i].
2233  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2234  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2235  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2236  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2237  * have been needed to return all paths.
2238  */
2239 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2240 {
2241 	return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2242 			     inode_to_path, ipath);
2243 }
2244 
2245 struct btrfs_data_container *init_data_container(u32 total_bytes)
2246 {
2247 	struct btrfs_data_container *data;
2248 	size_t alloc_bytes;
2249 
2250 	alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2251 	data = kvmalloc(alloc_bytes, GFP_KERNEL);
2252 	if (!data)
2253 		return ERR_PTR(-ENOMEM);
2254 
2255 	if (total_bytes >= sizeof(*data)) {
2256 		data->bytes_left = total_bytes - sizeof(*data);
2257 		data->bytes_missing = 0;
2258 	} else {
2259 		data->bytes_missing = sizeof(*data) - total_bytes;
2260 		data->bytes_left = 0;
2261 	}
2262 
2263 	data->elem_cnt = 0;
2264 	data->elem_missed = 0;
2265 
2266 	return data;
2267 }
2268 
2269 /*
2270  * allocates space to return multiple file system paths for an inode.
2271  * total_bytes to allocate are passed, note that space usable for actual path
2272  * information will be total_bytes - sizeof(struct inode_fs_paths).
2273  * the returned pointer must be freed with free_ipath() in the end.
2274  */
2275 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2276 					struct btrfs_path *path)
2277 {
2278 	struct inode_fs_paths *ifp;
2279 	struct btrfs_data_container *fspath;
2280 
2281 	fspath = init_data_container(total_bytes);
2282 	if (IS_ERR(fspath))
2283 		return ERR_CAST(fspath);
2284 
2285 	ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2286 	if (!ifp) {
2287 		kvfree(fspath);
2288 		return ERR_PTR(-ENOMEM);
2289 	}
2290 
2291 	ifp->btrfs_path = path;
2292 	ifp->fspath = fspath;
2293 	ifp->fs_root = fs_root;
2294 
2295 	return ifp;
2296 }
2297 
2298 void free_ipath(struct inode_fs_paths *ipath)
2299 {
2300 	if (!ipath)
2301 		return;
2302 	kvfree(ipath->fspath);
2303 	kfree(ipath);
2304 }
2305 
2306 struct btrfs_backref_iter *btrfs_backref_iter_alloc(
2307 		struct btrfs_fs_info *fs_info, gfp_t gfp_flag)
2308 {
2309 	struct btrfs_backref_iter *ret;
2310 
2311 	ret = kzalloc(sizeof(*ret), gfp_flag);
2312 	if (!ret)
2313 		return NULL;
2314 
2315 	ret->path = btrfs_alloc_path();
2316 	if (!ret->path) {
2317 		kfree(ret);
2318 		return NULL;
2319 	}
2320 
2321 	/* Current backref iterator only supports iteration in commit root */
2322 	ret->path->search_commit_root = 1;
2323 	ret->path->skip_locking = 1;
2324 	ret->fs_info = fs_info;
2325 
2326 	return ret;
2327 }
2328 
2329 int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2330 {
2331 	struct btrfs_fs_info *fs_info = iter->fs_info;
2332 	struct btrfs_path *path = iter->path;
2333 	struct btrfs_extent_item *ei;
2334 	struct btrfs_key key;
2335 	int ret;
2336 
2337 	key.objectid = bytenr;
2338 	key.type = BTRFS_METADATA_ITEM_KEY;
2339 	key.offset = (u64)-1;
2340 	iter->bytenr = bytenr;
2341 
2342 	ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
2343 	if (ret < 0)
2344 		return ret;
2345 	if (ret == 0) {
2346 		ret = -EUCLEAN;
2347 		goto release;
2348 	}
2349 	if (path->slots[0] == 0) {
2350 		WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2351 		ret = -EUCLEAN;
2352 		goto release;
2353 	}
2354 	path->slots[0]--;
2355 
2356 	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2357 	if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2358 	     key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2359 		ret = -ENOENT;
2360 		goto release;
2361 	}
2362 	memcpy(&iter->cur_key, &key, sizeof(key));
2363 	iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2364 						    path->slots[0]);
2365 	iter->end_ptr = (u32)(iter->item_ptr +
2366 			btrfs_item_size(path->nodes[0], path->slots[0]));
2367 	ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2368 			    struct btrfs_extent_item);
2369 
2370 	/*
2371 	 * Only support iteration on tree backref yet.
2372 	 *
2373 	 * This is an extra precaution for non skinny-metadata, where
2374 	 * EXTENT_ITEM is also used for tree blocks, that we can only use
2375 	 * extent flags to determine if it's a tree block.
2376 	 */
2377 	if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2378 		ret = -ENOTSUPP;
2379 		goto release;
2380 	}
2381 	iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2382 
2383 	/* If there is no inline backref, go search for keyed backref */
2384 	if (iter->cur_ptr >= iter->end_ptr) {
2385 		ret = btrfs_next_item(fs_info->extent_root, path);
2386 
2387 		/* No inline nor keyed ref */
2388 		if (ret > 0) {
2389 			ret = -ENOENT;
2390 			goto release;
2391 		}
2392 		if (ret < 0)
2393 			goto release;
2394 
2395 		btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2396 				path->slots[0]);
2397 		if (iter->cur_key.objectid != bytenr ||
2398 		    (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2399 		     iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2400 			ret = -ENOENT;
2401 			goto release;
2402 		}
2403 		iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2404 							   path->slots[0]);
2405 		iter->item_ptr = iter->cur_ptr;
2406 		iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2407 				      path->nodes[0], path->slots[0]));
2408 	}
2409 
2410 	return 0;
2411 release:
2412 	btrfs_backref_iter_release(iter);
2413 	return ret;
2414 }
2415 
2416 /*
2417  * Go to the next backref item of current bytenr, can be either inlined or
2418  * keyed.
2419  *
2420  * Caller needs to check whether it's inline ref or not by iter->cur_key.
2421  *
2422  * Return 0 if we get next backref without problem.
2423  * Return >0 if there is no extra backref for this bytenr.
2424  * Return <0 if there is something wrong happened.
2425  */
2426 int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2427 {
2428 	struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2429 	struct btrfs_path *path = iter->path;
2430 	struct btrfs_extent_inline_ref *iref;
2431 	int ret;
2432 	u32 size;
2433 
2434 	if (btrfs_backref_iter_is_inline_ref(iter)) {
2435 		/* We're still inside the inline refs */
2436 		ASSERT(iter->cur_ptr < iter->end_ptr);
2437 
2438 		if (btrfs_backref_has_tree_block_info(iter)) {
2439 			/* First tree block info */
2440 			size = sizeof(struct btrfs_tree_block_info);
2441 		} else {
2442 			/* Use inline ref type to determine the size */
2443 			int type;
2444 
2445 			iref = (struct btrfs_extent_inline_ref *)
2446 				((unsigned long)iter->cur_ptr);
2447 			type = btrfs_extent_inline_ref_type(eb, iref);
2448 
2449 			size = btrfs_extent_inline_ref_size(type);
2450 		}
2451 		iter->cur_ptr += size;
2452 		if (iter->cur_ptr < iter->end_ptr)
2453 			return 0;
2454 
2455 		/* All inline items iterated, fall through */
2456 	}
2457 
2458 	/* We're at keyed items, there is no inline item, go to the next one */
2459 	ret = btrfs_next_item(iter->fs_info->extent_root, iter->path);
2460 	if (ret)
2461 		return ret;
2462 
2463 	btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2464 	if (iter->cur_key.objectid != iter->bytenr ||
2465 	    (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2466 	     iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2467 		return 1;
2468 	iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2469 					path->slots[0]);
2470 	iter->cur_ptr = iter->item_ptr;
2471 	iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
2472 						path->slots[0]);
2473 	return 0;
2474 }
2475 
2476 void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2477 			      struct btrfs_backref_cache *cache, int is_reloc)
2478 {
2479 	int i;
2480 
2481 	cache->rb_root = RB_ROOT;
2482 	for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2483 		INIT_LIST_HEAD(&cache->pending[i]);
2484 	INIT_LIST_HEAD(&cache->changed);
2485 	INIT_LIST_HEAD(&cache->detached);
2486 	INIT_LIST_HEAD(&cache->leaves);
2487 	INIT_LIST_HEAD(&cache->pending_edge);
2488 	INIT_LIST_HEAD(&cache->useless_node);
2489 	cache->fs_info = fs_info;
2490 	cache->is_reloc = is_reloc;
2491 }
2492 
2493 struct btrfs_backref_node *btrfs_backref_alloc_node(
2494 		struct btrfs_backref_cache *cache, u64 bytenr, int level)
2495 {
2496 	struct btrfs_backref_node *node;
2497 
2498 	ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2499 	node = kzalloc(sizeof(*node), GFP_NOFS);
2500 	if (!node)
2501 		return node;
2502 
2503 	INIT_LIST_HEAD(&node->list);
2504 	INIT_LIST_HEAD(&node->upper);
2505 	INIT_LIST_HEAD(&node->lower);
2506 	RB_CLEAR_NODE(&node->rb_node);
2507 	cache->nr_nodes++;
2508 	node->level = level;
2509 	node->bytenr = bytenr;
2510 
2511 	return node;
2512 }
2513 
2514 struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2515 		struct btrfs_backref_cache *cache)
2516 {
2517 	struct btrfs_backref_edge *edge;
2518 
2519 	edge = kzalloc(sizeof(*edge), GFP_NOFS);
2520 	if (edge)
2521 		cache->nr_edges++;
2522 	return edge;
2523 }
2524 
2525 /*
2526  * Drop the backref node from cache, also cleaning up all its
2527  * upper edges and any uncached nodes in the path.
2528  *
2529  * This cleanup happens bottom up, thus the node should either
2530  * be the lowest node in the cache or a detached node.
2531  */
2532 void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2533 				struct btrfs_backref_node *node)
2534 {
2535 	struct btrfs_backref_node *upper;
2536 	struct btrfs_backref_edge *edge;
2537 
2538 	if (!node)
2539 		return;
2540 
2541 	BUG_ON(!node->lowest && !node->detached);
2542 	while (!list_empty(&node->upper)) {
2543 		edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2544 				  list[LOWER]);
2545 		upper = edge->node[UPPER];
2546 		list_del(&edge->list[LOWER]);
2547 		list_del(&edge->list[UPPER]);
2548 		btrfs_backref_free_edge(cache, edge);
2549 
2550 		/*
2551 		 * Add the node to leaf node list if no other child block
2552 		 * cached.
2553 		 */
2554 		if (list_empty(&upper->lower)) {
2555 			list_add_tail(&upper->lower, &cache->leaves);
2556 			upper->lowest = 1;
2557 		}
2558 	}
2559 
2560 	btrfs_backref_drop_node(cache, node);
2561 }
2562 
2563 /*
2564  * Release all nodes/edges from current cache
2565  */
2566 void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
2567 {
2568 	struct btrfs_backref_node *node;
2569 	int i;
2570 
2571 	while (!list_empty(&cache->detached)) {
2572 		node = list_entry(cache->detached.next,
2573 				  struct btrfs_backref_node, list);
2574 		btrfs_backref_cleanup_node(cache, node);
2575 	}
2576 
2577 	while (!list_empty(&cache->leaves)) {
2578 		node = list_entry(cache->leaves.next,
2579 				  struct btrfs_backref_node, lower);
2580 		btrfs_backref_cleanup_node(cache, node);
2581 	}
2582 
2583 	cache->last_trans = 0;
2584 
2585 	for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2586 		ASSERT(list_empty(&cache->pending[i]));
2587 	ASSERT(list_empty(&cache->pending_edge));
2588 	ASSERT(list_empty(&cache->useless_node));
2589 	ASSERT(list_empty(&cache->changed));
2590 	ASSERT(list_empty(&cache->detached));
2591 	ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
2592 	ASSERT(!cache->nr_nodes);
2593 	ASSERT(!cache->nr_edges);
2594 }
2595 
2596 /*
2597  * Handle direct tree backref
2598  *
2599  * Direct tree backref means, the backref item shows its parent bytenr
2600  * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
2601  *
2602  * @ref_key:	The converted backref key.
2603  *		For keyed backref, it's the item key.
2604  *		For inlined backref, objectid is the bytenr,
2605  *		type is btrfs_inline_ref_type, offset is
2606  *		btrfs_inline_ref_offset.
2607  */
2608 static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
2609 				      struct btrfs_key *ref_key,
2610 				      struct btrfs_backref_node *cur)
2611 {
2612 	struct btrfs_backref_edge *edge;
2613 	struct btrfs_backref_node *upper;
2614 	struct rb_node *rb_node;
2615 
2616 	ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
2617 
2618 	/* Only reloc root uses backref pointing to itself */
2619 	if (ref_key->objectid == ref_key->offset) {
2620 		struct btrfs_root *root;
2621 
2622 		cur->is_reloc_root = 1;
2623 		/* Only reloc backref cache cares about a specific root */
2624 		if (cache->is_reloc) {
2625 			root = find_reloc_root(cache->fs_info, cur->bytenr);
2626 			if (!root)
2627 				return -ENOENT;
2628 			cur->root = root;
2629 		} else {
2630 			/*
2631 			 * For generic purpose backref cache, reloc root node
2632 			 * is useless.
2633 			 */
2634 			list_add(&cur->list, &cache->useless_node);
2635 		}
2636 		return 0;
2637 	}
2638 
2639 	edge = btrfs_backref_alloc_edge(cache);
2640 	if (!edge)
2641 		return -ENOMEM;
2642 
2643 	rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
2644 	if (!rb_node) {
2645 		/* Parent node not yet cached */
2646 		upper = btrfs_backref_alloc_node(cache, ref_key->offset,
2647 					   cur->level + 1);
2648 		if (!upper) {
2649 			btrfs_backref_free_edge(cache, edge);
2650 			return -ENOMEM;
2651 		}
2652 
2653 		/*
2654 		 *  Backrefs for the upper level block isn't cached, add the
2655 		 *  block to pending list
2656 		 */
2657 		list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2658 	} else {
2659 		/* Parent node already cached */
2660 		upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
2661 		ASSERT(upper->checked);
2662 		INIT_LIST_HEAD(&edge->list[UPPER]);
2663 	}
2664 	btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
2665 	return 0;
2666 }
2667 
2668 /*
2669  * Handle indirect tree backref
2670  *
2671  * Indirect tree backref means, we only know which tree the node belongs to.
2672  * We still need to do a tree search to find out the parents. This is for
2673  * TREE_BLOCK_REF backref (keyed or inlined).
2674  *
2675  * @ref_key:	The same as @ref_key in  handle_direct_tree_backref()
2676  * @tree_key:	The first key of this tree block.
2677  * @path:	A clean (released) path, to avoid allocating path every time
2678  *		the function get called.
2679  */
2680 static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
2681 					struct btrfs_path *path,
2682 					struct btrfs_key *ref_key,
2683 					struct btrfs_key *tree_key,
2684 					struct btrfs_backref_node *cur)
2685 {
2686 	struct btrfs_fs_info *fs_info = cache->fs_info;
2687 	struct btrfs_backref_node *upper;
2688 	struct btrfs_backref_node *lower;
2689 	struct btrfs_backref_edge *edge;
2690 	struct extent_buffer *eb;
2691 	struct btrfs_root *root;
2692 	struct rb_node *rb_node;
2693 	int level;
2694 	bool need_check = true;
2695 	int ret;
2696 
2697 	root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
2698 	if (IS_ERR(root))
2699 		return PTR_ERR(root);
2700 	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2701 		cur->cowonly = 1;
2702 
2703 	if (btrfs_root_level(&root->root_item) == cur->level) {
2704 		/* Tree root */
2705 		ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
2706 		/*
2707 		 * For reloc backref cache, we may ignore reloc root.  But for
2708 		 * general purpose backref cache, we can't rely on
2709 		 * btrfs_should_ignore_reloc_root() as it may conflict with
2710 		 * current running relocation and lead to missing root.
2711 		 *
2712 		 * For general purpose backref cache, reloc root detection is
2713 		 * completely relying on direct backref (key->offset is parent
2714 		 * bytenr), thus only do such check for reloc cache.
2715 		 */
2716 		if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
2717 			btrfs_put_root(root);
2718 			list_add(&cur->list, &cache->useless_node);
2719 		} else {
2720 			cur->root = root;
2721 		}
2722 		return 0;
2723 	}
2724 
2725 	level = cur->level + 1;
2726 
2727 	/* Search the tree to find parent blocks referring to the block */
2728 	path->search_commit_root = 1;
2729 	path->skip_locking = 1;
2730 	path->lowest_level = level;
2731 	ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
2732 	path->lowest_level = 0;
2733 	if (ret < 0) {
2734 		btrfs_put_root(root);
2735 		return ret;
2736 	}
2737 	if (ret > 0 && path->slots[level] > 0)
2738 		path->slots[level]--;
2739 
2740 	eb = path->nodes[level];
2741 	if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
2742 		btrfs_err(fs_info,
2743 "couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
2744 			  cur->bytenr, level - 1, root->root_key.objectid,
2745 			  tree_key->objectid, tree_key->type, tree_key->offset);
2746 		btrfs_put_root(root);
2747 		ret = -ENOENT;
2748 		goto out;
2749 	}
2750 	lower = cur;
2751 
2752 	/* Add all nodes and edges in the path */
2753 	for (; level < BTRFS_MAX_LEVEL; level++) {
2754 		if (!path->nodes[level]) {
2755 			ASSERT(btrfs_root_bytenr(&root->root_item) ==
2756 			       lower->bytenr);
2757 			/* Same as previous should_ignore_reloc_root() call */
2758 			if (btrfs_should_ignore_reloc_root(root) &&
2759 			    cache->is_reloc) {
2760 				btrfs_put_root(root);
2761 				list_add(&lower->list, &cache->useless_node);
2762 			} else {
2763 				lower->root = root;
2764 			}
2765 			break;
2766 		}
2767 
2768 		edge = btrfs_backref_alloc_edge(cache);
2769 		if (!edge) {
2770 			btrfs_put_root(root);
2771 			ret = -ENOMEM;
2772 			goto out;
2773 		}
2774 
2775 		eb = path->nodes[level];
2776 		rb_node = rb_simple_search(&cache->rb_root, eb->start);
2777 		if (!rb_node) {
2778 			upper = btrfs_backref_alloc_node(cache, eb->start,
2779 							 lower->level + 1);
2780 			if (!upper) {
2781 				btrfs_put_root(root);
2782 				btrfs_backref_free_edge(cache, edge);
2783 				ret = -ENOMEM;
2784 				goto out;
2785 			}
2786 			upper->owner = btrfs_header_owner(eb);
2787 			if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2788 				upper->cowonly = 1;
2789 
2790 			/*
2791 			 * If we know the block isn't shared we can avoid
2792 			 * checking its backrefs.
2793 			 */
2794 			if (btrfs_block_can_be_shared(root, eb))
2795 				upper->checked = 0;
2796 			else
2797 				upper->checked = 1;
2798 
2799 			/*
2800 			 * Add the block to pending list if we need to check its
2801 			 * backrefs, we only do this once while walking up a
2802 			 * tree as we will catch anything else later on.
2803 			 */
2804 			if (!upper->checked && need_check) {
2805 				need_check = false;
2806 				list_add_tail(&edge->list[UPPER],
2807 					      &cache->pending_edge);
2808 			} else {
2809 				if (upper->checked)
2810 					need_check = true;
2811 				INIT_LIST_HEAD(&edge->list[UPPER]);
2812 			}
2813 		} else {
2814 			upper = rb_entry(rb_node, struct btrfs_backref_node,
2815 					 rb_node);
2816 			ASSERT(upper->checked);
2817 			INIT_LIST_HEAD(&edge->list[UPPER]);
2818 			if (!upper->owner)
2819 				upper->owner = btrfs_header_owner(eb);
2820 		}
2821 		btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
2822 
2823 		if (rb_node) {
2824 			btrfs_put_root(root);
2825 			break;
2826 		}
2827 		lower = upper;
2828 		upper = NULL;
2829 	}
2830 out:
2831 	btrfs_release_path(path);
2832 	return ret;
2833 }
2834 
2835 /*
2836  * Add backref node @cur into @cache.
2837  *
2838  * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
2839  *	 links aren't yet bi-directional. Needs to finish such links.
2840  *	 Use btrfs_backref_finish_upper_links() to finish such linkage.
2841  *
2842  * @path:	Released path for indirect tree backref lookup
2843  * @iter:	Released backref iter for extent tree search
2844  * @node_key:	The first key of the tree block
2845  */
2846 int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
2847 				struct btrfs_path *path,
2848 				struct btrfs_backref_iter *iter,
2849 				struct btrfs_key *node_key,
2850 				struct btrfs_backref_node *cur)
2851 {
2852 	struct btrfs_fs_info *fs_info = cache->fs_info;
2853 	struct btrfs_backref_edge *edge;
2854 	struct btrfs_backref_node *exist;
2855 	int ret;
2856 
2857 	ret = btrfs_backref_iter_start(iter, cur->bytenr);
2858 	if (ret < 0)
2859 		return ret;
2860 	/*
2861 	 * We skip the first btrfs_tree_block_info, as we don't use the key
2862 	 * stored in it, but fetch it from the tree block
2863 	 */
2864 	if (btrfs_backref_has_tree_block_info(iter)) {
2865 		ret = btrfs_backref_iter_next(iter);
2866 		if (ret < 0)
2867 			goto out;
2868 		/* No extra backref? This means the tree block is corrupted */
2869 		if (ret > 0) {
2870 			ret = -EUCLEAN;
2871 			goto out;
2872 		}
2873 	}
2874 	WARN_ON(cur->checked);
2875 	if (!list_empty(&cur->upper)) {
2876 		/*
2877 		 * The backref was added previously when processing backref of
2878 		 * type BTRFS_TREE_BLOCK_REF_KEY
2879 		 */
2880 		ASSERT(list_is_singular(&cur->upper));
2881 		edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
2882 				  list[LOWER]);
2883 		ASSERT(list_empty(&edge->list[UPPER]));
2884 		exist = edge->node[UPPER];
2885 		/*
2886 		 * Add the upper level block to pending list if we need check
2887 		 * its backrefs
2888 		 */
2889 		if (!exist->checked)
2890 			list_add_tail(&edge->list[UPPER], &cache->pending_edge);
2891 	} else {
2892 		exist = NULL;
2893 	}
2894 
2895 	for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
2896 		struct extent_buffer *eb;
2897 		struct btrfs_key key;
2898 		int type;
2899 
2900 		cond_resched();
2901 		eb = btrfs_backref_get_eb(iter);
2902 
2903 		key.objectid = iter->bytenr;
2904 		if (btrfs_backref_iter_is_inline_ref(iter)) {
2905 			struct btrfs_extent_inline_ref *iref;
2906 
2907 			/* Update key for inline backref */
2908 			iref = (struct btrfs_extent_inline_ref *)
2909 				((unsigned long)iter->cur_ptr);
2910 			type = btrfs_get_extent_inline_ref_type(eb, iref,
2911 							BTRFS_REF_TYPE_BLOCK);
2912 			if (type == BTRFS_REF_TYPE_INVALID) {
2913 				ret = -EUCLEAN;
2914 				goto out;
2915 			}
2916 			key.type = type;
2917 			key.offset = btrfs_extent_inline_ref_offset(eb, iref);
2918 		} else {
2919 			key.type = iter->cur_key.type;
2920 			key.offset = iter->cur_key.offset;
2921 		}
2922 
2923 		/*
2924 		 * Parent node found and matches current inline ref, no need to
2925 		 * rebuild this node for this inline ref
2926 		 */
2927 		if (exist &&
2928 		    ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
2929 		      exist->owner == key.offset) ||
2930 		     (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
2931 		      exist->bytenr == key.offset))) {
2932 			exist = NULL;
2933 			continue;
2934 		}
2935 
2936 		/* SHARED_BLOCK_REF means key.offset is the parent bytenr */
2937 		if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2938 			ret = handle_direct_tree_backref(cache, &key, cur);
2939 			if (ret < 0)
2940 				goto out;
2941 			continue;
2942 		} else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
2943 			ret = -EINVAL;
2944 			btrfs_print_v0_err(fs_info);
2945 			btrfs_handle_fs_error(fs_info, ret, NULL);
2946 			goto out;
2947 		} else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
2948 			continue;
2949 		}
2950 
2951 		/*
2952 		 * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
2953 		 * means the root objectid. We need to search the tree to get
2954 		 * its parent bytenr.
2955 		 */
2956 		ret = handle_indirect_tree_backref(cache, path, &key, node_key,
2957 						   cur);
2958 		if (ret < 0)
2959 			goto out;
2960 	}
2961 	ret = 0;
2962 	cur->checked = 1;
2963 	WARN_ON(exist);
2964 out:
2965 	btrfs_backref_iter_release(iter);
2966 	return ret;
2967 }
2968 
2969 /*
2970  * Finish the upwards linkage created by btrfs_backref_add_tree_node()
2971  */
2972 int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
2973 				     struct btrfs_backref_node *start)
2974 {
2975 	struct list_head *useless_node = &cache->useless_node;
2976 	struct btrfs_backref_edge *edge;
2977 	struct rb_node *rb_node;
2978 	LIST_HEAD(pending_edge);
2979 
2980 	ASSERT(start->checked);
2981 
2982 	/* Insert this node to cache if it's not COW-only */
2983 	if (!start->cowonly) {
2984 		rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
2985 					   &start->rb_node);
2986 		if (rb_node)
2987 			btrfs_backref_panic(cache->fs_info, start->bytenr,
2988 					    -EEXIST);
2989 		list_add_tail(&start->lower, &cache->leaves);
2990 	}
2991 
2992 	/*
2993 	 * Use breadth first search to iterate all related edges.
2994 	 *
2995 	 * The starting points are all the edges of this node
2996 	 */
2997 	list_for_each_entry(edge, &start->upper, list[LOWER])
2998 		list_add_tail(&edge->list[UPPER], &pending_edge);
2999 
3000 	while (!list_empty(&pending_edge)) {
3001 		struct btrfs_backref_node *upper;
3002 		struct btrfs_backref_node *lower;
3003 
3004 		edge = list_first_entry(&pending_edge,
3005 				struct btrfs_backref_edge, list[UPPER]);
3006 		list_del_init(&edge->list[UPPER]);
3007 		upper = edge->node[UPPER];
3008 		lower = edge->node[LOWER];
3009 
3010 		/* Parent is detached, no need to keep any edges */
3011 		if (upper->detached) {
3012 			list_del(&edge->list[LOWER]);
3013 			btrfs_backref_free_edge(cache, edge);
3014 
3015 			/* Lower node is orphan, queue for cleanup */
3016 			if (list_empty(&lower->upper))
3017 				list_add(&lower->list, useless_node);
3018 			continue;
3019 		}
3020 
3021 		/*
3022 		 * All new nodes added in current build_backref_tree() haven't
3023 		 * been linked to the cache rb tree.
3024 		 * So if we have upper->rb_node populated, this means a cache
3025 		 * hit. We only need to link the edge, as @upper and all its
3026 		 * parents have already been linked.
3027 		 */
3028 		if (!RB_EMPTY_NODE(&upper->rb_node)) {
3029 			if (upper->lowest) {
3030 				list_del_init(&upper->lower);
3031 				upper->lowest = 0;
3032 			}
3033 
3034 			list_add_tail(&edge->list[UPPER], &upper->lower);
3035 			continue;
3036 		}
3037 
3038 		/* Sanity check, we shouldn't have any unchecked nodes */
3039 		if (!upper->checked) {
3040 			ASSERT(0);
3041 			return -EUCLEAN;
3042 		}
3043 
3044 		/* Sanity check, COW-only node has non-COW-only parent */
3045 		if (start->cowonly != upper->cowonly) {
3046 			ASSERT(0);
3047 			return -EUCLEAN;
3048 		}
3049 
3050 		/* Only cache non-COW-only (subvolume trees) tree blocks */
3051 		if (!upper->cowonly) {
3052 			rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3053 						   &upper->rb_node);
3054 			if (rb_node) {
3055 				btrfs_backref_panic(cache->fs_info,
3056 						upper->bytenr, -EEXIST);
3057 				return -EUCLEAN;
3058 			}
3059 		}
3060 
3061 		list_add_tail(&edge->list[UPPER], &upper->lower);
3062 
3063 		/*
3064 		 * Also queue all the parent edges of this uncached node
3065 		 * to finish the upper linkage
3066 		 */
3067 		list_for_each_entry(edge, &upper->upper, list[LOWER])
3068 			list_add_tail(&edge->list[UPPER], &pending_edge);
3069 	}
3070 	return 0;
3071 }
3072 
3073 void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3074 				 struct btrfs_backref_node *node)
3075 {
3076 	struct btrfs_backref_node *lower;
3077 	struct btrfs_backref_node *upper;
3078 	struct btrfs_backref_edge *edge;
3079 
3080 	while (!list_empty(&cache->useless_node)) {
3081 		lower = list_first_entry(&cache->useless_node,
3082 				   struct btrfs_backref_node, list);
3083 		list_del_init(&lower->list);
3084 	}
3085 	while (!list_empty(&cache->pending_edge)) {
3086 		edge = list_first_entry(&cache->pending_edge,
3087 				struct btrfs_backref_edge, list[UPPER]);
3088 		list_del(&edge->list[UPPER]);
3089 		list_del(&edge->list[LOWER]);
3090 		lower = edge->node[LOWER];
3091 		upper = edge->node[UPPER];
3092 		btrfs_backref_free_edge(cache, edge);
3093 
3094 		/*
3095 		 * Lower is no longer linked to any upper backref nodes and
3096 		 * isn't in the cache, we can free it ourselves.
3097 		 */
3098 		if (list_empty(&lower->upper) &&
3099 		    RB_EMPTY_NODE(&lower->rb_node))
3100 			list_add(&lower->list, &cache->useless_node);
3101 
3102 		if (!RB_EMPTY_NODE(&upper->rb_node))
3103 			continue;
3104 
3105 		/* Add this guy's upper edges to the list to process */
3106 		list_for_each_entry(edge, &upper->upper, list[LOWER])
3107 			list_add_tail(&edge->list[UPPER],
3108 				      &cache->pending_edge);
3109 		if (list_empty(&upper->upper))
3110 			list_add(&upper->list, &cache->useless_node);
3111 	}
3112 
3113 	while (!list_empty(&cache->useless_node)) {
3114 		lower = list_first_entry(&cache->useless_node,
3115 				   struct btrfs_backref_node, list);
3116 		list_del_init(&lower->list);
3117 		if (lower == node)
3118 			node = NULL;
3119 		btrfs_backref_drop_node(cache, lower);
3120 	}
3121 
3122 	btrfs_backref_cleanup_node(cache, node);
3123 	ASSERT(list_empty(&cache->useless_node) &&
3124 	       list_empty(&cache->pending_edge));
3125 }
3126