Loading...
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#include "fs.h"
19#include "accessors.h"
20#include "extent-tree.h"
21#include "relocation.h"
22#include "tree-checker.h"
23
24/* Just arbitrary numbers so we can be sure one of these happened. */
25#define BACKREF_FOUND_SHARED 6
26#define BACKREF_FOUND_NOT_SHARED 7
27
28struct extent_inode_elem {
29 u64 inum;
30 u64 offset;
31 u64 num_bytes;
32 struct extent_inode_elem *next;
33};
34
35static int check_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
36 const struct btrfs_key *key,
37 const struct extent_buffer *eb,
38 const struct btrfs_file_extent_item *fi,
39 struct extent_inode_elem **eie)
40{
41 const u64 data_len = btrfs_file_extent_num_bytes(eb, fi);
42 u64 offset = key->offset;
43 struct extent_inode_elem *e;
44 const u64 *root_ids;
45 int root_count;
46 bool cached;
47
48 if (!ctx->ignore_extent_item_pos &&
49 !btrfs_file_extent_compression(eb, fi) &&
50 !btrfs_file_extent_encryption(eb, fi) &&
51 !btrfs_file_extent_other_encoding(eb, fi)) {
52 u64 data_offset;
53
54 data_offset = btrfs_file_extent_offset(eb, fi);
55
56 if (ctx->extent_item_pos < data_offset ||
57 ctx->extent_item_pos >= data_offset + data_len)
58 return 1;
59 offset += ctx->extent_item_pos - data_offset;
60 }
61
62 if (!ctx->indirect_ref_iterator || !ctx->cache_lookup)
63 goto add_inode_elem;
64
65 cached = ctx->cache_lookup(eb->start, ctx->user_ctx, &root_ids,
66 &root_count);
67 if (!cached)
68 goto add_inode_elem;
69
70 for (int i = 0; i < root_count; i++) {
71 int ret;
72
73 ret = ctx->indirect_ref_iterator(key->objectid, offset,
74 data_len, root_ids[i],
75 ctx->user_ctx);
76 if (ret)
77 return ret;
78 }
79
80add_inode_elem:
81 e = kmalloc(sizeof(*e), GFP_NOFS);
82 if (!e)
83 return -ENOMEM;
84
85 e->next = *eie;
86 e->inum = key->objectid;
87 e->offset = offset;
88 e->num_bytes = data_len;
89 *eie = e;
90
91 return 0;
92}
93
94static void free_inode_elem_list(struct extent_inode_elem *eie)
95{
96 struct extent_inode_elem *eie_next;
97
98 for (; eie; eie = eie_next) {
99 eie_next = eie->next;
100 kfree(eie);
101 }
102}
103
104static int find_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
105 const struct extent_buffer *eb,
106 struct extent_inode_elem **eie)
107{
108 u64 disk_byte;
109 struct btrfs_key key;
110 struct btrfs_file_extent_item *fi;
111 int slot;
112 int nritems;
113 int extent_type;
114 int ret;
115
116 /*
117 * from the shared data ref, we only have the leaf but we need
118 * the key. thus, we must look into all items and see that we
119 * find one (some) with a reference to our extent item.
120 */
121 nritems = btrfs_header_nritems(eb);
122 for (slot = 0; slot < nritems; ++slot) {
123 btrfs_item_key_to_cpu(eb, &key, slot);
124 if (key.type != BTRFS_EXTENT_DATA_KEY)
125 continue;
126 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
127 extent_type = btrfs_file_extent_type(eb, fi);
128 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
129 continue;
130 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
131 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
132 if (disk_byte != ctx->bytenr)
133 continue;
134
135 ret = check_extent_in_eb(ctx, &key, eb, fi, eie);
136 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
137 return ret;
138 }
139
140 return 0;
141}
142
143struct preftree {
144 struct rb_root_cached root;
145 unsigned int count;
146};
147
148#define PREFTREE_INIT { .root = RB_ROOT_CACHED, .count = 0 }
149
150struct preftrees {
151 struct preftree direct; /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
152 struct preftree indirect; /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
153 struct preftree indirect_missing_keys;
154};
155
156/*
157 * Checks for a shared extent during backref search.
158 *
159 * The share_count tracks prelim_refs (direct and indirect) having a
160 * ref->count >0:
161 * - incremented when a ref->count transitions to >0
162 * - decremented when a ref->count transitions to <1
163 */
164struct share_check {
165 struct btrfs_backref_share_check_ctx *ctx;
166 struct btrfs_root *root;
167 u64 inum;
168 u64 data_bytenr;
169 u64 data_extent_gen;
170 /*
171 * Counts number of inodes that refer to an extent (different inodes in
172 * the same root or different roots) that we could find. The sharedness
173 * check typically stops once this counter gets greater than 1, so it
174 * may not reflect the total number of inodes.
175 */
176 int share_count;
177 /*
178 * The number of times we found our inode refers to the data extent we
179 * are determining the sharedness. In other words, how many file extent
180 * items we could find for our inode that point to our target data
181 * extent. The value we get here after finishing the extent sharedness
182 * check may be smaller than reality, but if it ends up being greater
183 * than 1, then we know for sure the inode has multiple file extent
184 * items that point to our inode, and we can safely assume it's useful
185 * to cache the sharedness check result.
186 */
187 int self_ref_count;
188 bool have_delayed_delete_refs;
189};
190
191static inline int extent_is_shared(struct share_check *sc)
192{
193 return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
194}
195
196static struct kmem_cache *btrfs_prelim_ref_cache;
197
198int __init btrfs_prelim_ref_init(void)
199{
200 btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
201 sizeof(struct prelim_ref),
202 0,
203 SLAB_MEM_SPREAD,
204 NULL);
205 if (!btrfs_prelim_ref_cache)
206 return -ENOMEM;
207 return 0;
208}
209
210void __cold btrfs_prelim_ref_exit(void)
211{
212 kmem_cache_destroy(btrfs_prelim_ref_cache);
213}
214
215static void free_pref(struct prelim_ref *ref)
216{
217 kmem_cache_free(btrfs_prelim_ref_cache, ref);
218}
219
220/*
221 * Return 0 when both refs are for the same block (and can be merged).
222 * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
223 * indicates a 'higher' block.
224 */
225static int prelim_ref_compare(struct prelim_ref *ref1,
226 struct prelim_ref *ref2)
227{
228 if (ref1->level < ref2->level)
229 return -1;
230 if (ref1->level > ref2->level)
231 return 1;
232 if (ref1->root_id < ref2->root_id)
233 return -1;
234 if (ref1->root_id > ref2->root_id)
235 return 1;
236 if (ref1->key_for_search.type < ref2->key_for_search.type)
237 return -1;
238 if (ref1->key_for_search.type > ref2->key_for_search.type)
239 return 1;
240 if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
241 return -1;
242 if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
243 return 1;
244 if (ref1->key_for_search.offset < ref2->key_for_search.offset)
245 return -1;
246 if (ref1->key_for_search.offset > ref2->key_for_search.offset)
247 return 1;
248 if (ref1->parent < ref2->parent)
249 return -1;
250 if (ref1->parent > ref2->parent)
251 return 1;
252
253 return 0;
254}
255
256static void update_share_count(struct share_check *sc, int oldcount,
257 int newcount, struct prelim_ref *newref)
258{
259 if ((!sc) || (oldcount == 0 && newcount < 1))
260 return;
261
262 if (oldcount > 0 && newcount < 1)
263 sc->share_count--;
264 else if (oldcount < 1 && newcount > 0)
265 sc->share_count++;
266
267 if (newref->root_id == sc->root->root_key.objectid &&
268 newref->wanted_disk_byte == sc->data_bytenr &&
269 newref->key_for_search.objectid == sc->inum)
270 sc->self_ref_count += newref->count;
271}
272
273/*
274 * Add @newref to the @root rbtree, merging identical refs.
275 *
276 * Callers should assume that newref has been freed after calling.
277 */
278static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
279 struct preftree *preftree,
280 struct prelim_ref *newref,
281 struct share_check *sc)
282{
283 struct rb_root_cached *root;
284 struct rb_node **p;
285 struct rb_node *parent = NULL;
286 struct prelim_ref *ref;
287 int result;
288 bool leftmost = true;
289
290 root = &preftree->root;
291 p = &root->rb_root.rb_node;
292
293 while (*p) {
294 parent = *p;
295 ref = rb_entry(parent, struct prelim_ref, rbnode);
296 result = prelim_ref_compare(ref, newref);
297 if (result < 0) {
298 p = &(*p)->rb_left;
299 } else if (result > 0) {
300 p = &(*p)->rb_right;
301 leftmost = false;
302 } else {
303 /* Identical refs, merge them and free @newref */
304 struct extent_inode_elem *eie = ref->inode_list;
305
306 while (eie && eie->next)
307 eie = eie->next;
308
309 if (!eie)
310 ref->inode_list = newref->inode_list;
311 else
312 eie->next = newref->inode_list;
313 trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
314 preftree->count);
315 /*
316 * A delayed ref can have newref->count < 0.
317 * The ref->count is updated to follow any
318 * BTRFS_[ADD|DROP]_DELAYED_REF actions.
319 */
320 update_share_count(sc, ref->count,
321 ref->count + newref->count, newref);
322 ref->count += newref->count;
323 free_pref(newref);
324 return;
325 }
326 }
327
328 update_share_count(sc, 0, newref->count, newref);
329 preftree->count++;
330 trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
331 rb_link_node(&newref->rbnode, parent, p);
332 rb_insert_color_cached(&newref->rbnode, root, leftmost);
333}
334
335/*
336 * Release the entire tree. We don't care about internal consistency so
337 * just free everything and then reset the tree root.
338 */
339static void prelim_release(struct preftree *preftree)
340{
341 struct prelim_ref *ref, *next_ref;
342
343 rbtree_postorder_for_each_entry_safe(ref, next_ref,
344 &preftree->root.rb_root, rbnode) {
345 free_inode_elem_list(ref->inode_list);
346 free_pref(ref);
347 }
348
349 preftree->root = RB_ROOT_CACHED;
350 preftree->count = 0;
351}
352
353/*
354 * the rules for all callers of this function are:
355 * - obtaining the parent is the goal
356 * - if you add a key, you must know that it is a correct key
357 * - if you cannot add the parent or a correct key, then we will look into the
358 * block later to set a correct key
359 *
360 * delayed refs
361 * ============
362 * backref type | shared | indirect | shared | indirect
363 * information | tree | tree | data | data
364 * --------------------+--------+----------+--------+----------
365 * parent logical | y | - | - | -
366 * key to resolve | - | y | y | y
367 * tree block logical | - | - | - | -
368 * root for resolving | y | y | y | y
369 *
370 * - column 1: we've the parent -> done
371 * - column 2, 3, 4: we use the key to find the parent
372 *
373 * on disk refs (inline or keyed)
374 * ==============================
375 * backref type | shared | indirect | shared | indirect
376 * information | tree | tree | data | data
377 * --------------------+--------+----------+--------+----------
378 * parent logical | y | - | y | -
379 * key to resolve | - | - | - | y
380 * tree block logical | y | y | y | y
381 * root for resolving | - | y | y | y
382 *
383 * - column 1, 3: we've the parent -> done
384 * - column 2: we take the first key from the block to find the parent
385 * (see add_missing_keys)
386 * - column 4: we use the key to find the parent
387 *
388 * additional information that's available but not required to find the parent
389 * block might help in merging entries to gain some speed.
390 */
391static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
392 struct preftree *preftree, u64 root_id,
393 const struct btrfs_key *key, int level, u64 parent,
394 u64 wanted_disk_byte, int count,
395 struct share_check *sc, gfp_t gfp_mask)
396{
397 struct prelim_ref *ref;
398
399 if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
400 return 0;
401
402 ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
403 if (!ref)
404 return -ENOMEM;
405
406 ref->root_id = root_id;
407 if (key)
408 ref->key_for_search = *key;
409 else
410 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
411
412 ref->inode_list = NULL;
413 ref->level = level;
414 ref->count = count;
415 ref->parent = parent;
416 ref->wanted_disk_byte = wanted_disk_byte;
417 prelim_ref_insert(fs_info, preftree, ref, sc);
418 return extent_is_shared(sc);
419}
420
421/* direct refs use root == 0, key == NULL */
422static int add_direct_ref(const struct btrfs_fs_info *fs_info,
423 struct preftrees *preftrees, int level, u64 parent,
424 u64 wanted_disk_byte, int count,
425 struct share_check *sc, gfp_t gfp_mask)
426{
427 return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
428 parent, wanted_disk_byte, count, sc, gfp_mask);
429}
430
431/* indirect refs use parent == 0 */
432static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
433 struct preftrees *preftrees, u64 root_id,
434 const struct btrfs_key *key, int level,
435 u64 wanted_disk_byte, int count,
436 struct share_check *sc, gfp_t gfp_mask)
437{
438 struct preftree *tree = &preftrees->indirect;
439
440 if (!key)
441 tree = &preftrees->indirect_missing_keys;
442 return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
443 wanted_disk_byte, count, sc, gfp_mask);
444}
445
446static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
447{
448 struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
449 struct rb_node *parent = NULL;
450 struct prelim_ref *ref = NULL;
451 struct prelim_ref target = {};
452 int result;
453
454 target.parent = bytenr;
455
456 while (*p) {
457 parent = *p;
458 ref = rb_entry(parent, struct prelim_ref, rbnode);
459 result = prelim_ref_compare(ref, &target);
460
461 if (result < 0)
462 p = &(*p)->rb_left;
463 else if (result > 0)
464 p = &(*p)->rb_right;
465 else
466 return 1;
467 }
468 return 0;
469}
470
471static int add_all_parents(struct btrfs_backref_walk_ctx *ctx,
472 struct btrfs_root *root, struct btrfs_path *path,
473 struct ulist *parents,
474 struct preftrees *preftrees, struct prelim_ref *ref,
475 int level)
476{
477 int ret = 0;
478 int slot;
479 struct extent_buffer *eb;
480 struct btrfs_key key;
481 struct btrfs_key *key_for_search = &ref->key_for_search;
482 struct btrfs_file_extent_item *fi;
483 struct extent_inode_elem *eie = NULL, *old = NULL;
484 u64 disk_byte;
485 u64 wanted_disk_byte = ref->wanted_disk_byte;
486 u64 count = 0;
487 u64 data_offset;
488 u8 type;
489
490 if (level != 0) {
491 eb = path->nodes[level];
492 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
493 if (ret < 0)
494 return ret;
495 return 0;
496 }
497
498 /*
499 * 1. We normally enter this function with the path already pointing to
500 * the first item to check. But sometimes, we may enter it with
501 * slot == nritems.
502 * 2. We are searching for normal backref but bytenr of this leaf
503 * matches shared data backref
504 * 3. The leaf owner is not equal to the root we are searching
505 *
506 * For these cases, go to the next leaf before we continue.
507 */
508 eb = path->nodes[0];
509 if (path->slots[0] >= btrfs_header_nritems(eb) ||
510 is_shared_data_backref(preftrees, eb->start) ||
511 ref->root_id != btrfs_header_owner(eb)) {
512 if (ctx->time_seq == BTRFS_SEQ_LAST)
513 ret = btrfs_next_leaf(root, path);
514 else
515 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
516 }
517
518 while (!ret && count < ref->count) {
519 eb = path->nodes[0];
520 slot = path->slots[0];
521
522 btrfs_item_key_to_cpu(eb, &key, slot);
523
524 if (key.objectid != key_for_search->objectid ||
525 key.type != BTRFS_EXTENT_DATA_KEY)
526 break;
527
528 /*
529 * We are searching for normal backref but bytenr of this leaf
530 * matches shared data backref, OR
531 * the leaf owner is not equal to the root we are searching for
532 */
533 if (slot == 0 &&
534 (is_shared_data_backref(preftrees, eb->start) ||
535 ref->root_id != btrfs_header_owner(eb))) {
536 if (ctx->time_seq == BTRFS_SEQ_LAST)
537 ret = btrfs_next_leaf(root, path);
538 else
539 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
540 continue;
541 }
542 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
543 type = btrfs_file_extent_type(eb, fi);
544 if (type == BTRFS_FILE_EXTENT_INLINE)
545 goto next;
546 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
547 data_offset = btrfs_file_extent_offset(eb, fi);
548
549 if (disk_byte == wanted_disk_byte) {
550 eie = NULL;
551 old = NULL;
552 if (ref->key_for_search.offset == key.offset - data_offset)
553 count++;
554 else
555 goto next;
556 if (!ctx->skip_inode_ref_list) {
557 ret = check_extent_in_eb(ctx, &key, eb, fi, &eie);
558 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
559 ret < 0)
560 break;
561 }
562 if (ret > 0)
563 goto next;
564 ret = ulist_add_merge_ptr(parents, eb->start,
565 eie, (void **)&old, GFP_NOFS);
566 if (ret < 0)
567 break;
568 if (!ret && !ctx->skip_inode_ref_list) {
569 while (old->next)
570 old = old->next;
571 old->next = eie;
572 }
573 eie = NULL;
574 }
575next:
576 if (ctx->time_seq == BTRFS_SEQ_LAST)
577 ret = btrfs_next_item(root, path);
578 else
579 ret = btrfs_next_old_item(root, path, ctx->time_seq);
580 }
581
582 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
583 free_inode_elem_list(eie);
584 else if (ret > 0)
585 ret = 0;
586
587 return ret;
588}
589
590/*
591 * resolve an indirect backref in the form (root_id, key, level)
592 * to a logical address
593 */
594static int resolve_indirect_ref(struct btrfs_backref_walk_ctx *ctx,
595 struct btrfs_path *path,
596 struct preftrees *preftrees,
597 struct prelim_ref *ref, struct ulist *parents)
598{
599 struct btrfs_root *root;
600 struct extent_buffer *eb;
601 int ret = 0;
602 int root_level;
603 int level = ref->level;
604 struct btrfs_key search_key = ref->key_for_search;
605
606 /*
607 * If we're search_commit_root we could possibly be holding locks on
608 * other tree nodes. This happens when qgroups does backref walks when
609 * adding new delayed refs. To deal with this we need to look in cache
610 * for the root, and if we don't find it then we need to search the
611 * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
612 * here.
613 */
614 if (path->search_commit_root)
615 root = btrfs_get_fs_root_commit_root(ctx->fs_info, path, ref->root_id);
616 else
617 root = btrfs_get_fs_root(ctx->fs_info, ref->root_id, false);
618 if (IS_ERR(root)) {
619 ret = PTR_ERR(root);
620 goto out_free;
621 }
622
623 if (!path->search_commit_root &&
624 test_bit(BTRFS_ROOT_DELETING, &root->state)) {
625 ret = -ENOENT;
626 goto out;
627 }
628
629 if (btrfs_is_testing(ctx->fs_info)) {
630 ret = -ENOENT;
631 goto out;
632 }
633
634 if (path->search_commit_root)
635 root_level = btrfs_header_level(root->commit_root);
636 else if (ctx->time_seq == BTRFS_SEQ_LAST)
637 root_level = btrfs_header_level(root->node);
638 else
639 root_level = btrfs_old_root_level(root, ctx->time_seq);
640
641 if (root_level + 1 == level)
642 goto out;
643
644 /*
645 * We can often find data backrefs with an offset that is too large
646 * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
647 * subtracting a file's offset with the data offset of its
648 * corresponding extent data item. This can happen for example in the
649 * clone ioctl.
650 *
651 * So if we detect such case we set the search key's offset to zero to
652 * make sure we will find the matching file extent item at
653 * add_all_parents(), otherwise we will miss it because the offset
654 * taken form the backref is much larger then the offset of the file
655 * extent item. This can make us scan a very large number of file
656 * extent items, but at least it will not make us miss any.
657 *
658 * This is an ugly workaround for a behaviour that should have never
659 * existed, but it does and a fix for the clone ioctl would touch a lot
660 * of places, cause backwards incompatibility and would not fix the
661 * problem for extents cloned with older kernels.
662 */
663 if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
664 search_key.offset >= LLONG_MAX)
665 search_key.offset = 0;
666 path->lowest_level = level;
667 if (ctx->time_seq == BTRFS_SEQ_LAST)
668 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
669 else
670 ret = btrfs_search_old_slot(root, &search_key, path, ctx->time_seq);
671
672 btrfs_debug(ctx->fs_info,
673 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
674 ref->root_id, level, ref->count, ret,
675 ref->key_for_search.objectid, ref->key_for_search.type,
676 ref->key_for_search.offset);
677 if (ret < 0)
678 goto out;
679
680 eb = path->nodes[level];
681 while (!eb) {
682 if (WARN_ON(!level)) {
683 ret = 1;
684 goto out;
685 }
686 level--;
687 eb = path->nodes[level];
688 }
689
690 ret = add_all_parents(ctx, root, path, parents, preftrees, ref, level);
691out:
692 btrfs_put_root(root);
693out_free:
694 path->lowest_level = 0;
695 btrfs_release_path(path);
696 return ret;
697}
698
699static struct extent_inode_elem *
700unode_aux_to_inode_list(struct ulist_node *node)
701{
702 if (!node)
703 return NULL;
704 return (struct extent_inode_elem *)(uintptr_t)node->aux;
705}
706
707static void free_leaf_list(struct ulist *ulist)
708{
709 struct ulist_node *node;
710 struct ulist_iterator uiter;
711
712 ULIST_ITER_INIT(&uiter);
713 while ((node = ulist_next(ulist, &uiter)))
714 free_inode_elem_list(unode_aux_to_inode_list(node));
715
716 ulist_free(ulist);
717}
718
719/*
720 * We maintain three separate rbtrees: one for direct refs, one for
721 * indirect refs which have a key, and one for indirect refs which do not
722 * have a key. Each tree does merge on insertion.
723 *
724 * Once all of the references are located, we iterate over the tree of
725 * indirect refs with missing keys. An appropriate key is located and
726 * the ref is moved onto the tree for indirect refs. After all missing
727 * keys are thus located, we iterate over the indirect ref tree, resolve
728 * each reference, and then insert the resolved reference onto the
729 * direct tree (merging there too).
730 *
731 * New backrefs (i.e., for parent nodes) are added to the appropriate
732 * rbtree as they are encountered. The new backrefs are subsequently
733 * resolved as above.
734 */
735static int resolve_indirect_refs(struct btrfs_backref_walk_ctx *ctx,
736 struct btrfs_path *path,
737 struct preftrees *preftrees,
738 struct share_check *sc)
739{
740 int err;
741 int ret = 0;
742 struct ulist *parents;
743 struct ulist_node *node;
744 struct ulist_iterator uiter;
745 struct rb_node *rnode;
746
747 parents = ulist_alloc(GFP_NOFS);
748 if (!parents)
749 return -ENOMEM;
750
751 /*
752 * We could trade memory usage for performance here by iterating
753 * the tree, allocating new refs for each insertion, and then
754 * freeing the entire indirect tree when we're done. In some test
755 * cases, the tree can grow quite large (~200k objects).
756 */
757 while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
758 struct prelim_ref *ref;
759
760 ref = rb_entry(rnode, struct prelim_ref, rbnode);
761 if (WARN(ref->parent,
762 "BUG: direct ref found in indirect tree")) {
763 ret = -EINVAL;
764 goto out;
765 }
766
767 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
768 preftrees->indirect.count--;
769
770 if (ref->count == 0) {
771 free_pref(ref);
772 continue;
773 }
774
775 if (sc && ref->root_id != sc->root->root_key.objectid) {
776 free_pref(ref);
777 ret = BACKREF_FOUND_SHARED;
778 goto out;
779 }
780 err = resolve_indirect_ref(ctx, path, preftrees, ref, parents);
781 /*
782 * we can only tolerate ENOENT,otherwise,we should catch error
783 * and return directly.
784 */
785 if (err == -ENOENT) {
786 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref,
787 NULL);
788 continue;
789 } else if (err) {
790 free_pref(ref);
791 ret = err;
792 goto out;
793 }
794
795 /* we put the first parent into the ref at hand */
796 ULIST_ITER_INIT(&uiter);
797 node = ulist_next(parents, &uiter);
798 ref->parent = node ? node->val : 0;
799 ref->inode_list = unode_aux_to_inode_list(node);
800
801 /* Add a prelim_ref(s) for any other parent(s). */
802 while ((node = ulist_next(parents, &uiter))) {
803 struct prelim_ref *new_ref;
804
805 new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
806 GFP_NOFS);
807 if (!new_ref) {
808 free_pref(ref);
809 ret = -ENOMEM;
810 goto out;
811 }
812 memcpy(new_ref, ref, sizeof(*ref));
813 new_ref->parent = node->val;
814 new_ref->inode_list = unode_aux_to_inode_list(node);
815 prelim_ref_insert(ctx->fs_info, &preftrees->direct,
816 new_ref, NULL);
817 }
818
819 /*
820 * Now it's a direct ref, put it in the direct tree. We must
821 * do this last because the ref could be merged/freed here.
822 */
823 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref, NULL);
824
825 ulist_reinit(parents);
826 cond_resched();
827 }
828out:
829 /*
830 * We may have inode lists attached to refs in the parents ulist, so we
831 * must free them before freeing the ulist and its refs.
832 */
833 free_leaf_list(parents);
834 return ret;
835}
836
837/*
838 * read tree blocks and add keys where required.
839 */
840static int add_missing_keys(struct btrfs_fs_info *fs_info,
841 struct preftrees *preftrees, bool lock)
842{
843 struct prelim_ref *ref;
844 struct extent_buffer *eb;
845 struct preftree *tree = &preftrees->indirect_missing_keys;
846 struct rb_node *node;
847
848 while ((node = rb_first_cached(&tree->root))) {
849 struct btrfs_tree_parent_check check = { 0 };
850
851 ref = rb_entry(node, struct prelim_ref, rbnode);
852 rb_erase_cached(node, &tree->root);
853
854 BUG_ON(ref->parent); /* should not be a direct ref */
855 BUG_ON(ref->key_for_search.type);
856 BUG_ON(!ref->wanted_disk_byte);
857
858 check.level = ref->level - 1;
859 check.owner_root = ref->root_id;
860
861 eb = read_tree_block(fs_info, ref->wanted_disk_byte, &check);
862 if (IS_ERR(eb)) {
863 free_pref(ref);
864 return PTR_ERR(eb);
865 }
866 if (!extent_buffer_uptodate(eb)) {
867 free_pref(ref);
868 free_extent_buffer(eb);
869 return -EIO;
870 }
871
872 if (lock)
873 btrfs_tree_read_lock(eb);
874 if (btrfs_header_level(eb) == 0)
875 btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
876 else
877 btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
878 if (lock)
879 btrfs_tree_read_unlock(eb);
880 free_extent_buffer(eb);
881 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
882 cond_resched();
883 }
884 return 0;
885}
886
887/*
888 * add all currently queued delayed refs from this head whose seq nr is
889 * smaller or equal that seq to the list
890 */
891static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
892 struct btrfs_delayed_ref_head *head, u64 seq,
893 struct preftrees *preftrees, struct share_check *sc)
894{
895 struct btrfs_delayed_ref_node *node;
896 struct btrfs_key key;
897 struct rb_node *n;
898 int count;
899 int ret = 0;
900
901 spin_lock(&head->lock);
902 for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
903 node = rb_entry(n, struct btrfs_delayed_ref_node,
904 ref_node);
905 if (node->seq > seq)
906 continue;
907
908 switch (node->action) {
909 case BTRFS_ADD_DELAYED_EXTENT:
910 case BTRFS_UPDATE_DELAYED_HEAD:
911 WARN_ON(1);
912 continue;
913 case BTRFS_ADD_DELAYED_REF:
914 count = node->ref_mod;
915 break;
916 case BTRFS_DROP_DELAYED_REF:
917 count = node->ref_mod * -1;
918 break;
919 default:
920 BUG();
921 }
922 switch (node->type) {
923 case BTRFS_TREE_BLOCK_REF_KEY: {
924 /* NORMAL INDIRECT METADATA backref */
925 struct btrfs_delayed_tree_ref *ref;
926 struct btrfs_key *key_ptr = NULL;
927
928 if (head->extent_op && head->extent_op->update_key) {
929 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
930 key_ptr = &key;
931 }
932
933 ref = btrfs_delayed_node_to_tree_ref(node);
934 ret = add_indirect_ref(fs_info, preftrees, ref->root,
935 key_ptr, ref->level + 1,
936 node->bytenr, count, sc,
937 GFP_ATOMIC);
938 break;
939 }
940 case BTRFS_SHARED_BLOCK_REF_KEY: {
941 /* SHARED DIRECT METADATA backref */
942 struct btrfs_delayed_tree_ref *ref;
943
944 ref = btrfs_delayed_node_to_tree_ref(node);
945
946 ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
947 ref->parent, node->bytenr, count,
948 sc, GFP_ATOMIC);
949 break;
950 }
951 case BTRFS_EXTENT_DATA_REF_KEY: {
952 /* NORMAL INDIRECT DATA backref */
953 struct btrfs_delayed_data_ref *ref;
954 ref = btrfs_delayed_node_to_data_ref(node);
955
956 key.objectid = ref->objectid;
957 key.type = BTRFS_EXTENT_DATA_KEY;
958 key.offset = ref->offset;
959
960 /*
961 * If we have a share check context and a reference for
962 * another inode, we can't exit immediately. This is
963 * because even if this is a BTRFS_ADD_DELAYED_REF
964 * reference we may find next a BTRFS_DROP_DELAYED_REF
965 * which cancels out this ADD reference.
966 *
967 * If this is a DROP reference and there was no previous
968 * ADD reference, then we need to signal that when we
969 * process references from the extent tree (through
970 * add_inline_refs() and add_keyed_refs()), we should
971 * not exit early if we find a reference for another
972 * inode, because one of the delayed DROP references
973 * may cancel that reference in the extent tree.
974 */
975 if (sc && count < 0)
976 sc->have_delayed_delete_refs = true;
977
978 ret = add_indirect_ref(fs_info, preftrees, ref->root,
979 &key, 0, node->bytenr, count, sc,
980 GFP_ATOMIC);
981 break;
982 }
983 case BTRFS_SHARED_DATA_REF_KEY: {
984 /* SHARED DIRECT FULL backref */
985 struct btrfs_delayed_data_ref *ref;
986
987 ref = btrfs_delayed_node_to_data_ref(node);
988
989 ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
990 node->bytenr, count, sc,
991 GFP_ATOMIC);
992 break;
993 }
994 default:
995 WARN_ON(1);
996 }
997 /*
998 * We must ignore BACKREF_FOUND_SHARED until all delayed
999 * refs have been checked.
1000 */
1001 if (ret && (ret != BACKREF_FOUND_SHARED))
1002 break;
1003 }
1004 if (!ret)
1005 ret = extent_is_shared(sc);
1006
1007 spin_unlock(&head->lock);
1008 return ret;
1009}
1010
1011/*
1012 * add all inline backrefs for bytenr to the list
1013 *
1014 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1015 */
1016static int add_inline_refs(struct btrfs_backref_walk_ctx *ctx,
1017 struct btrfs_path *path,
1018 int *info_level, struct preftrees *preftrees,
1019 struct share_check *sc)
1020{
1021 int ret = 0;
1022 int slot;
1023 struct extent_buffer *leaf;
1024 struct btrfs_key key;
1025 struct btrfs_key found_key;
1026 unsigned long ptr;
1027 unsigned long end;
1028 struct btrfs_extent_item *ei;
1029 u64 flags;
1030 u64 item_size;
1031
1032 /*
1033 * enumerate all inline refs
1034 */
1035 leaf = path->nodes[0];
1036 slot = path->slots[0];
1037
1038 item_size = btrfs_item_size(leaf, slot);
1039 BUG_ON(item_size < sizeof(*ei));
1040
1041 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
1042
1043 if (ctx->check_extent_item) {
1044 ret = ctx->check_extent_item(ctx->bytenr, ei, leaf, ctx->user_ctx);
1045 if (ret)
1046 return ret;
1047 }
1048
1049 flags = btrfs_extent_flags(leaf, ei);
1050 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1051
1052 ptr = (unsigned long)(ei + 1);
1053 end = (unsigned long)ei + item_size;
1054
1055 if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
1056 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1057 struct btrfs_tree_block_info *info;
1058
1059 info = (struct btrfs_tree_block_info *)ptr;
1060 *info_level = btrfs_tree_block_level(leaf, info);
1061 ptr += sizeof(struct btrfs_tree_block_info);
1062 BUG_ON(ptr > end);
1063 } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1064 *info_level = found_key.offset;
1065 } else {
1066 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1067 }
1068
1069 while (ptr < end) {
1070 struct btrfs_extent_inline_ref *iref;
1071 u64 offset;
1072 int type;
1073
1074 iref = (struct btrfs_extent_inline_ref *)ptr;
1075 type = btrfs_get_extent_inline_ref_type(leaf, iref,
1076 BTRFS_REF_TYPE_ANY);
1077 if (type == BTRFS_REF_TYPE_INVALID)
1078 return -EUCLEAN;
1079
1080 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1081
1082 switch (type) {
1083 case BTRFS_SHARED_BLOCK_REF_KEY:
1084 ret = add_direct_ref(ctx->fs_info, preftrees,
1085 *info_level + 1, offset,
1086 ctx->bytenr, 1, NULL, GFP_NOFS);
1087 break;
1088 case BTRFS_SHARED_DATA_REF_KEY: {
1089 struct btrfs_shared_data_ref *sdref;
1090 int count;
1091
1092 sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1093 count = btrfs_shared_data_ref_count(leaf, sdref);
1094
1095 ret = add_direct_ref(ctx->fs_info, preftrees, 0, offset,
1096 ctx->bytenr, count, sc, GFP_NOFS);
1097 break;
1098 }
1099 case BTRFS_TREE_BLOCK_REF_KEY:
1100 ret = add_indirect_ref(ctx->fs_info, preftrees, offset,
1101 NULL, *info_level + 1,
1102 ctx->bytenr, 1, NULL, GFP_NOFS);
1103 break;
1104 case BTRFS_EXTENT_DATA_REF_KEY: {
1105 struct btrfs_extent_data_ref *dref;
1106 int count;
1107 u64 root;
1108
1109 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1110 count = btrfs_extent_data_ref_count(leaf, dref);
1111 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1112 dref);
1113 key.type = BTRFS_EXTENT_DATA_KEY;
1114 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1115
1116 if (sc && key.objectid != sc->inum &&
1117 !sc->have_delayed_delete_refs) {
1118 ret = BACKREF_FOUND_SHARED;
1119 break;
1120 }
1121
1122 root = btrfs_extent_data_ref_root(leaf, dref);
1123
1124 if (!ctx->skip_data_ref ||
1125 !ctx->skip_data_ref(root, key.objectid, key.offset,
1126 ctx->user_ctx))
1127 ret = add_indirect_ref(ctx->fs_info, preftrees,
1128 root, &key, 0, ctx->bytenr,
1129 count, sc, GFP_NOFS);
1130 break;
1131 }
1132 case BTRFS_EXTENT_OWNER_REF_KEY:
1133 ASSERT(btrfs_fs_incompat(ctx->fs_info, SIMPLE_QUOTA));
1134 break;
1135 default:
1136 WARN_ON(1);
1137 }
1138 if (ret)
1139 return ret;
1140 ptr += btrfs_extent_inline_ref_size(type);
1141 }
1142
1143 return 0;
1144}
1145
1146/*
1147 * add all non-inline backrefs for bytenr to the list
1148 *
1149 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1150 */
1151static int add_keyed_refs(struct btrfs_backref_walk_ctx *ctx,
1152 struct btrfs_root *extent_root,
1153 struct btrfs_path *path,
1154 int info_level, struct preftrees *preftrees,
1155 struct share_check *sc)
1156{
1157 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1158 int ret;
1159 int slot;
1160 struct extent_buffer *leaf;
1161 struct btrfs_key key;
1162
1163 while (1) {
1164 ret = btrfs_next_item(extent_root, path);
1165 if (ret < 0)
1166 break;
1167 if (ret) {
1168 ret = 0;
1169 break;
1170 }
1171
1172 slot = path->slots[0];
1173 leaf = path->nodes[0];
1174 btrfs_item_key_to_cpu(leaf, &key, slot);
1175
1176 if (key.objectid != ctx->bytenr)
1177 break;
1178 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1179 continue;
1180 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1181 break;
1182
1183 switch (key.type) {
1184 case BTRFS_SHARED_BLOCK_REF_KEY:
1185 /* SHARED DIRECT METADATA backref */
1186 ret = add_direct_ref(fs_info, preftrees,
1187 info_level + 1, key.offset,
1188 ctx->bytenr, 1, NULL, GFP_NOFS);
1189 break;
1190 case BTRFS_SHARED_DATA_REF_KEY: {
1191 /* SHARED DIRECT FULL backref */
1192 struct btrfs_shared_data_ref *sdref;
1193 int count;
1194
1195 sdref = btrfs_item_ptr(leaf, slot,
1196 struct btrfs_shared_data_ref);
1197 count = btrfs_shared_data_ref_count(leaf, sdref);
1198 ret = add_direct_ref(fs_info, preftrees, 0,
1199 key.offset, ctx->bytenr, count,
1200 sc, GFP_NOFS);
1201 break;
1202 }
1203 case BTRFS_TREE_BLOCK_REF_KEY:
1204 /* NORMAL INDIRECT METADATA backref */
1205 ret = add_indirect_ref(fs_info, preftrees, key.offset,
1206 NULL, info_level + 1, ctx->bytenr,
1207 1, NULL, GFP_NOFS);
1208 break;
1209 case BTRFS_EXTENT_DATA_REF_KEY: {
1210 /* NORMAL INDIRECT DATA backref */
1211 struct btrfs_extent_data_ref *dref;
1212 int count;
1213 u64 root;
1214
1215 dref = btrfs_item_ptr(leaf, slot,
1216 struct btrfs_extent_data_ref);
1217 count = btrfs_extent_data_ref_count(leaf, dref);
1218 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1219 dref);
1220 key.type = BTRFS_EXTENT_DATA_KEY;
1221 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1222
1223 if (sc && key.objectid != sc->inum &&
1224 !sc->have_delayed_delete_refs) {
1225 ret = BACKREF_FOUND_SHARED;
1226 break;
1227 }
1228
1229 root = btrfs_extent_data_ref_root(leaf, dref);
1230
1231 if (!ctx->skip_data_ref ||
1232 !ctx->skip_data_ref(root, key.objectid, key.offset,
1233 ctx->user_ctx))
1234 ret = add_indirect_ref(fs_info, preftrees, root,
1235 &key, 0, ctx->bytenr,
1236 count, sc, GFP_NOFS);
1237 break;
1238 }
1239 default:
1240 WARN_ON(1);
1241 }
1242 if (ret)
1243 return ret;
1244
1245 }
1246
1247 return ret;
1248}
1249
1250/*
1251 * The caller has joined a transaction or is holding a read lock on the
1252 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1253 * snapshot field changing while updating or checking the cache.
1254 */
1255static bool lookup_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1256 struct btrfs_root *root,
1257 u64 bytenr, int level, bool *is_shared)
1258{
1259 const struct btrfs_fs_info *fs_info = root->fs_info;
1260 struct btrfs_backref_shared_cache_entry *entry;
1261
1262 if (!current->journal_info)
1263 lockdep_assert_held(&fs_info->commit_root_sem);
1264
1265 if (!ctx->use_path_cache)
1266 return false;
1267
1268 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1269 return false;
1270
1271 /*
1272 * Level -1 is used for the data extent, which is not reliable to cache
1273 * because its reference count can increase or decrease without us
1274 * realizing. We cache results only for extent buffers that lead from
1275 * the root node down to the leaf with the file extent item.
1276 */
1277 ASSERT(level >= 0);
1278
1279 entry = &ctx->path_cache_entries[level];
1280
1281 /* Unused cache entry or being used for some other extent buffer. */
1282 if (entry->bytenr != bytenr)
1283 return false;
1284
1285 /*
1286 * We cached a false result, but the last snapshot generation of the
1287 * root changed, so we now have a snapshot. Don't trust the result.
1288 */
1289 if (!entry->is_shared &&
1290 entry->gen != btrfs_root_last_snapshot(&root->root_item))
1291 return false;
1292
1293 /*
1294 * If we cached a true result and the last generation used for dropping
1295 * a root changed, we can not trust the result, because the dropped root
1296 * could be a snapshot sharing this extent buffer.
1297 */
1298 if (entry->is_shared &&
1299 entry->gen != btrfs_get_last_root_drop_gen(fs_info))
1300 return false;
1301
1302 *is_shared = entry->is_shared;
1303 /*
1304 * If the node at this level is shared, than all nodes below are also
1305 * shared. Currently some of the nodes below may be marked as not shared
1306 * because we have just switched from one leaf to another, and switched
1307 * also other nodes above the leaf and below the current level, so mark
1308 * them as shared.
1309 */
1310 if (*is_shared) {
1311 for (int i = 0; i < level; i++) {
1312 ctx->path_cache_entries[i].is_shared = true;
1313 ctx->path_cache_entries[i].gen = entry->gen;
1314 }
1315 }
1316
1317 return true;
1318}
1319
1320/*
1321 * The caller has joined a transaction or is holding a read lock on the
1322 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1323 * snapshot field changing while updating or checking the cache.
1324 */
1325static void store_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1326 struct btrfs_root *root,
1327 u64 bytenr, int level, bool is_shared)
1328{
1329 const struct btrfs_fs_info *fs_info = root->fs_info;
1330 struct btrfs_backref_shared_cache_entry *entry;
1331 u64 gen;
1332
1333 if (!current->journal_info)
1334 lockdep_assert_held(&fs_info->commit_root_sem);
1335
1336 if (!ctx->use_path_cache)
1337 return;
1338
1339 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1340 return;
1341
1342 /*
1343 * Level -1 is used for the data extent, which is not reliable to cache
1344 * because its reference count can increase or decrease without us
1345 * realizing. We cache results only for extent buffers that lead from
1346 * the root node down to the leaf with the file extent item.
1347 */
1348 ASSERT(level >= 0);
1349
1350 if (is_shared)
1351 gen = btrfs_get_last_root_drop_gen(fs_info);
1352 else
1353 gen = btrfs_root_last_snapshot(&root->root_item);
1354
1355 entry = &ctx->path_cache_entries[level];
1356 entry->bytenr = bytenr;
1357 entry->is_shared = is_shared;
1358 entry->gen = gen;
1359
1360 /*
1361 * If we found an extent buffer is shared, set the cache result for all
1362 * extent buffers below it to true. As nodes in the path are COWed,
1363 * their sharedness is moved to their children, and if a leaf is COWed,
1364 * then the sharedness of a data extent becomes direct, the refcount of
1365 * data extent is increased in the extent item at the extent tree.
1366 */
1367 if (is_shared) {
1368 for (int i = 0; i < level; i++) {
1369 entry = &ctx->path_cache_entries[i];
1370 entry->is_shared = is_shared;
1371 entry->gen = gen;
1372 }
1373 }
1374}
1375
1376/*
1377 * this adds all existing backrefs (inline backrefs, backrefs and delayed
1378 * refs) for the given bytenr to the refs list, merges duplicates and resolves
1379 * indirect refs to their parent bytenr.
1380 * When roots are found, they're added to the roots list
1381 *
1382 * @ctx: Backref walking context object, must be not NULL.
1383 * @sc: If !NULL, then immediately return BACKREF_FOUND_SHARED when a
1384 * shared extent is detected.
1385 *
1386 * Otherwise this returns 0 for success and <0 for an error.
1387 *
1388 * FIXME some caching might speed things up
1389 */
1390static int find_parent_nodes(struct btrfs_backref_walk_ctx *ctx,
1391 struct share_check *sc)
1392{
1393 struct btrfs_root *root = btrfs_extent_root(ctx->fs_info, ctx->bytenr);
1394 struct btrfs_key key;
1395 struct btrfs_path *path;
1396 struct btrfs_delayed_ref_root *delayed_refs = NULL;
1397 struct btrfs_delayed_ref_head *head;
1398 int info_level = 0;
1399 int ret;
1400 struct prelim_ref *ref;
1401 struct rb_node *node;
1402 struct extent_inode_elem *eie = NULL;
1403 struct preftrees preftrees = {
1404 .direct = PREFTREE_INIT,
1405 .indirect = PREFTREE_INIT,
1406 .indirect_missing_keys = PREFTREE_INIT
1407 };
1408
1409 /* Roots ulist is not needed when using a sharedness check context. */
1410 if (sc)
1411 ASSERT(ctx->roots == NULL);
1412
1413 key.objectid = ctx->bytenr;
1414 key.offset = (u64)-1;
1415 if (btrfs_fs_incompat(ctx->fs_info, SKINNY_METADATA))
1416 key.type = BTRFS_METADATA_ITEM_KEY;
1417 else
1418 key.type = BTRFS_EXTENT_ITEM_KEY;
1419
1420 path = btrfs_alloc_path();
1421 if (!path)
1422 return -ENOMEM;
1423 if (!ctx->trans) {
1424 path->search_commit_root = 1;
1425 path->skip_locking = 1;
1426 }
1427
1428 if (ctx->time_seq == BTRFS_SEQ_LAST)
1429 path->skip_locking = 1;
1430
1431again:
1432 head = NULL;
1433
1434 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1435 if (ret < 0)
1436 goto out;
1437 if (ret == 0) {
1438 /* This shouldn't happen, indicates a bug or fs corruption. */
1439 ASSERT(ret != 0);
1440 ret = -EUCLEAN;
1441 goto out;
1442 }
1443
1444 if (ctx->trans && likely(ctx->trans->type != __TRANS_DUMMY) &&
1445 ctx->time_seq != BTRFS_SEQ_LAST) {
1446 /*
1447 * We have a specific time_seq we care about and trans which
1448 * means we have the path lock, we need to grab the ref head and
1449 * lock it so we have a consistent view of the refs at the given
1450 * time.
1451 */
1452 delayed_refs = &ctx->trans->transaction->delayed_refs;
1453 spin_lock(&delayed_refs->lock);
1454 head = btrfs_find_delayed_ref_head(delayed_refs, ctx->bytenr);
1455 if (head) {
1456 if (!mutex_trylock(&head->mutex)) {
1457 refcount_inc(&head->refs);
1458 spin_unlock(&delayed_refs->lock);
1459
1460 btrfs_release_path(path);
1461
1462 /*
1463 * Mutex was contended, block until it's
1464 * released and try again
1465 */
1466 mutex_lock(&head->mutex);
1467 mutex_unlock(&head->mutex);
1468 btrfs_put_delayed_ref_head(head);
1469 goto again;
1470 }
1471 spin_unlock(&delayed_refs->lock);
1472 ret = add_delayed_refs(ctx->fs_info, head, ctx->time_seq,
1473 &preftrees, sc);
1474 mutex_unlock(&head->mutex);
1475 if (ret)
1476 goto out;
1477 } else {
1478 spin_unlock(&delayed_refs->lock);
1479 }
1480 }
1481
1482 if (path->slots[0]) {
1483 struct extent_buffer *leaf;
1484 int slot;
1485
1486 path->slots[0]--;
1487 leaf = path->nodes[0];
1488 slot = path->slots[0];
1489 btrfs_item_key_to_cpu(leaf, &key, slot);
1490 if (key.objectid == ctx->bytenr &&
1491 (key.type == BTRFS_EXTENT_ITEM_KEY ||
1492 key.type == BTRFS_METADATA_ITEM_KEY)) {
1493 ret = add_inline_refs(ctx, path, &info_level,
1494 &preftrees, sc);
1495 if (ret)
1496 goto out;
1497 ret = add_keyed_refs(ctx, root, path, info_level,
1498 &preftrees, sc);
1499 if (ret)
1500 goto out;
1501 }
1502 }
1503
1504 /*
1505 * If we have a share context and we reached here, it means the extent
1506 * is not directly shared (no multiple reference items for it),
1507 * otherwise we would have exited earlier with a return value of
1508 * BACKREF_FOUND_SHARED after processing delayed references or while
1509 * processing inline or keyed references from the extent tree.
1510 * The extent may however be indirectly shared through shared subtrees
1511 * as a result from creating snapshots, so we determine below what is
1512 * its parent node, in case we are dealing with a metadata extent, or
1513 * what's the leaf (or leaves), from a fs tree, that has a file extent
1514 * item pointing to it in case we are dealing with a data extent.
1515 */
1516 ASSERT(extent_is_shared(sc) == 0);
1517
1518 /*
1519 * If we are here for a data extent and we have a share_check structure
1520 * it means the data extent is not directly shared (does not have
1521 * multiple reference items), so we have to check if a path in the fs
1522 * tree (going from the root node down to the leaf that has the file
1523 * extent item pointing to the data extent) is shared, that is, if any
1524 * of the extent buffers in the path is referenced by other trees.
1525 */
1526 if (sc && ctx->bytenr == sc->data_bytenr) {
1527 /*
1528 * If our data extent is from a generation more recent than the
1529 * last generation used to snapshot the root, then we know that
1530 * it can not be shared through subtrees, so we can skip
1531 * resolving indirect references, there's no point in
1532 * determining the extent buffers for the path from the fs tree
1533 * root node down to the leaf that has the file extent item that
1534 * points to the data extent.
1535 */
1536 if (sc->data_extent_gen >
1537 btrfs_root_last_snapshot(&sc->root->root_item)) {
1538 ret = BACKREF_FOUND_NOT_SHARED;
1539 goto out;
1540 }
1541
1542 /*
1543 * If we are only determining if a data extent is shared or not
1544 * and the corresponding file extent item is located in the same
1545 * leaf as the previous file extent item, we can skip resolving
1546 * indirect references for a data extent, since the fs tree path
1547 * is the same (same leaf, so same path). We skip as long as the
1548 * cached result for the leaf is valid and only if there's only
1549 * one file extent item pointing to the data extent, because in
1550 * the case of multiple file extent items, they may be located
1551 * in different leaves and therefore we have multiple paths.
1552 */
1553 if (sc->ctx->curr_leaf_bytenr == sc->ctx->prev_leaf_bytenr &&
1554 sc->self_ref_count == 1) {
1555 bool cached;
1556 bool is_shared;
1557
1558 cached = lookup_backref_shared_cache(sc->ctx, sc->root,
1559 sc->ctx->curr_leaf_bytenr,
1560 0, &is_shared);
1561 if (cached) {
1562 if (is_shared)
1563 ret = BACKREF_FOUND_SHARED;
1564 else
1565 ret = BACKREF_FOUND_NOT_SHARED;
1566 goto out;
1567 }
1568 }
1569 }
1570
1571 btrfs_release_path(path);
1572
1573 ret = add_missing_keys(ctx->fs_info, &preftrees, path->skip_locking == 0);
1574 if (ret)
1575 goto out;
1576
1577 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1578
1579 ret = resolve_indirect_refs(ctx, path, &preftrees, sc);
1580 if (ret)
1581 goto out;
1582
1583 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1584
1585 /*
1586 * This walks the tree of merged and resolved refs. Tree blocks are
1587 * read in as needed. Unique entries are added to the ulist, and
1588 * the list of found roots is updated.
1589 *
1590 * We release the entire tree in one go before returning.
1591 */
1592 node = rb_first_cached(&preftrees.direct.root);
1593 while (node) {
1594 ref = rb_entry(node, struct prelim_ref, rbnode);
1595 node = rb_next(&ref->rbnode);
1596 /*
1597 * ref->count < 0 can happen here if there are delayed
1598 * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1599 * prelim_ref_insert() relies on this when merging
1600 * identical refs to keep the overall count correct.
1601 * prelim_ref_insert() will merge only those refs
1602 * which compare identically. Any refs having
1603 * e.g. different offsets would not be merged,
1604 * and would retain their original ref->count < 0.
1605 */
1606 if (ctx->roots && ref->count && ref->root_id && ref->parent == 0) {
1607 /* no parent == root of tree */
1608 ret = ulist_add(ctx->roots, ref->root_id, 0, GFP_NOFS);
1609 if (ret < 0)
1610 goto out;
1611 }
1612 if (ref->count && ref->parent) {
1613 if (!ctx->skip_inode_ref_list && !ref->inode_list &&
1614 ref->level == 0) {
1615 struct btrfs_tree_parent_check check = { 0 };
1616 struct extent_buffer *eb;
1617
1618 check.level = ref->level;
1619
1620 eb = read_tree_block(ctx->fs_info, ref->parent,
1621 &check);
1622 if (IS_ERR(eb)) {
1623 ret = PTR_ERR(eb);
1624 goto out;
1625 }
1626 if (!extent_buffer_uptodate(eb)) {
1627 free_extent_buffer(eb);
1628 ret = -EIO;
1629 goto out;
1630 }
1631
1632 if (!path->skip_locking)
1633 btrfs_tree_read_lock(eb);
1634 ret = find_extent_in_eb(ctx, eb, &eie);
1635 if (!path->skip_locking)
1636 btrfs_tree_read_unlock(eb);
1637 free_extent_buffer(eb);
1638 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1639 ret < 0)
1640 goto out;
1641 ref->inode_list = eie;
1642 /*
1643 * We transferred the list ownership to the ref,
1644 * so set to NULL to avoid a double free in case
1645 * an error happens after this.
1646 */
1647 eie = NULL;
1648 }
1649 ret = ulist_add_merge_ptr(ctx->refs, ref->parent,
1650 ref->inode_list,
1651 (void **)&eie, GFP_NOFS);
1652 if (ret < 0)
1653 goto out;
1654 if (!ret && !ctx->skip_inode_ref_list) {
1655 /*
1656 * We've recorded that parent, so we must extend
1657 * its inode list here.
1658 *
1659 * However if there was corruption we may not
1660 * have found an eie, return an error in this
1661 * case.
1662 */
1663 ASSERT(eie);
1664 if (!eie) {
1665 ret = -EUCLEAN;
1666 goto out;
1667 }
1668 while (eie->next)
1669 eie = eie->next;
1670 eie->next = ref->inode_list;
1671 }
1672 eie = NULL;
1673 /*
1674 * We have transferred the inode list ownership from
1675 * this ref to the ref we added to the 'refs' ulist.
1676 * So set this ref's inode list to NULL to avoid
1677 * use-after-free when our caller uses it or double
1678 * frees in case an error happens before we return.
1679 */
1680 ref->inode_list = NULL;
1681 }
1682 cond_resched();
1683 }
1684
1685out:
1686 btrfs_free_path(path);
1687
1688 prelim_release(&preftrees.direct);
1689 prelim_release(&preftrees.indirect);
1690 prelim_release(&preftrees.indirect_missing_keys);
1691
1692 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
1693 free_inode_elem_list(eie);
1694 return ret;
1695}
1696
1697/*
1698 * Finds all leaves with a reference to the specified combination of
1699 * @ctx->bytenr and @ctx->extent_item_pos. The bytenr of the found leaves are
1700 * added to the ulist at @ctx->refs, and that ulist is allocated by this
1701 * function. The caller should free the ulist with free_leaf_list() if
1702 * @ctx->ignore_extent_item_pos is false, otherwise a fimple ulist_free() is
1703 * enough.
1704 *
1705 * Returns 0 on success and < 0 on error. On error @ctx->refs is not allocated.
1706 */
1707int btrfs_find_all_leafs(struct btrfs_backref_walk_ctx *ctx)
1708{
1709 int ret;
1710
1711 ASSERT(ctx->refs == NULL);
1712
1713 ctx->refs = ulist_alloc(GFP_NOFS);
1714 if (!ctx->refs)
1715 return -ENOMEM;
1716
1717 ret = find_parent_nodes(ctx, NULL);
1718 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1719 (ret < 0 && ret != -ENOENT)) {
1720 free_leaf_list(ctx->refs);
1721 ctx->refs = NULL;
1722 return ret;
1723 }
1724
1725 return 0;
1726}
1727
1728/*
1729 * Walk all backrefs for a given extent to find all roots that reference this
1730 * extent. Walking a backref means finding all extents that reference this
1731 * extent and in turn walk the backrefs of those, too. Naturally this is a
1732 * recursive process, but here it is implemented in an iterative fashion: We
1733 * find all referencing extents for the extent in question and put them on a
1734 * list. In turn, we find all referencing extents for those, further appending
1735 * to the list. The way we iterate the list allows adding more elements after
1736 * the current while iterating. The process stops when we reach the end of the
1737 * list.
1738 *
1739 * Found roots are added to @ctx->roots, which is allocated by this function if
1740 * it points to NULL, in which case the caller is responsible for freeing it
1741 * after it's not needed anymore.
1742 * This function requires @ctx->refs to be NULL, as it uses it for allocating a
1743 * ulist to do temporary work, and frees it before returning.
1744 *
1745 * Returns 0 on success, < 0 on error.
1746 */
1747static int btrfs_find_all_roots_safe(struct btrfs_backref_walk_ctx *ctx)
1748{
1749 const u64 orig_bytenr = ctx->bytenr;
1750 const bool orig_skip_inode_ref_list = ctx->skip_inode_ref_list;
1751 bool roots_ulist_allocated = false;
1752 struct ulist_iterator uiter;
1753 int ret = 0;
1754
1755 ASSERT(ctx->refs == NULL);
1756
1757 ctx->refs = ulist_alloc(GFP_NOFS);
1758 if (!ctx->refs)
1759 return -ENOMEM;
1760
1761 if (!ctx->roots) {
1762 ctx->roots = ulist_alloc(GFP_NOFS);
1763 if (!ctx->roots) {
1764 ulist_free(ctx->refs);
1765 ctx->refs = NULL;
1766 return -ENOMEM;
1767 }
1768 roots_ulist_allocated = true;
1769 }
1770
1771 ctx->skip_inode_ref_list = true;
1772
1773 ULIST_ITER_INIT(&uiter);
1774 while (1) {
1775 struct ulist_node *node;
1776
1777 ret = find_parent_nodes(ctx, NULL);
1778 if (ret < 0 && ret != -ENOENT) {
1779 if (roots_ulist_allocated) {
1780 ulist_free(ctx->roots);
1781 ctx->roots = NULL;
1782 }
1783 break;
1784 }
1785 ret = 0;
1786 node = ulist_next(ctx->refs, &uiter);
1787 if (!node)
1788 break;
1789 ctx->bytenr = node->val;
1790 cond_resched();
1791 }
1792
1793 ulist_free(ctx->refs);
1794 ctx->refs = NULL;
1795 ctx->bytenr = orig_bytenr;
1796 ctx->skip_inode_ref_list = orig_skip_inode_ref_list;
1797
1798 return ret;
1799}
1800
1801int btrfs_find_all_roots(struct btrfs_backref_walk_ctx *ctx,
1802 bool skip_commit_root_sem)
1803{
1804 int ret;
1805
1806 if (!ctx->trans && !skip_commit_root_sem)
1807 down_read(&ctx->fs_info->commit_root_sem);
1808 ret = btrfs_find_all_roots_safe(ctx);
1809 if (!ctx->trans && !skip_commit_root_sem)
1810 up_read(&ctx->fs_info->commit_root_sem);
1811 return ret;
1812}
1813
1814struct btrfs_backref_share_check_ctx *btrfs_alloc_backref_share_check_ctx(void)
1815{
1816 struct btrfs_backref_share_check_ctx *ctx;
1817
1818 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1819 if (!ctx)
1820 return NULL;
1821
1822 ulist_init(&ctx->refs);
1823
1824 return ctx;
1825}
1826
1827void btrfs_free_backref_share_ctx(struct btrfs_backref_share_check_ctx *ctx)
1828{
1829 if (!ctx)
1830 return;
1831
1832 ulist_release(&ctx->refs);
1833 kfree(ctx);
1834}
1835
1836/*
1837 * Check if a data extent is shared or not.
1838 *
1839 * @inode: The inode whose extent we are checking.
1840 * @bytenr: Logical bytenr of the extent we are checking.
1841 * @extent_gen: Generation of the extent (file extent item) or 0 if it is
1842 * not known.
1843 * @ctx: A backref sharedness check context.
1844 *
1845 * btrfs_is_data_extent_shared uses the backref walking code but will short
1846 * circuit as soon as it finds a root or inode that doesn't match the
1847 * one passed in. This provides a significant performance benefit for
1848 * callers (such as fiemap) which want to know whether the extent is
1849 * shared but do not need a ref count.
1850 *
1851 * This attempts to attach to the running transaction in order to account for
1852 * delayed refs, but continues on even when no running transaction exists.
1853 *
1854 * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1855 */
1856int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
1857 u64 extent_gen,
1858 struct btrfs_backref_share_check_ctx *ctx)
1859{
1860 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
1861 struct btrfs_root *root = inode->root;
1862 struct btrfs_fs_info *fs_info = root->fs_info;
1863 struct btrfs_trans_handle *trans;
1864 struct ulist_iterator uiter;
1865 struct ulist_node *node;
1866 struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1867 int ret = 0;
1868 struct share_check shared = {
1869 .ctx = ctx,
1870 .root = root,
1871 .inum = btrfs_ino(inode),
1872 .data_bytenr = bytenr,
1873 .data_extent_gen = extent_gen,
1874 .share_count = 0,
1875 .self_ref_count = 0,
1876 .have_delayed_delete_refs = false,
1877 };
1878 int level;
1879 bool leaf_cached;
1880 bool leaf_is_shared;
1881
1882 for (int i = 0; i < BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE; i++) {
1883 if (ctx->prev_extents_cache[i].bytenr == bytenr)
1884 return ctx->prev_extents_cache[i].is_shared;
1885 }
1886
1887 ulist_init(&ctx->refs);
1888
1889 trans = btrfs_join_transaction_nostart(root);
1890 if (IS_ERR(trans)) {
1891 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1892 ret = PTR_ERR(trans);
1893 goto out;
1894 }
1895 trans = NULL;
1896 down_read(&fs_info->commit_root_sem);
1897 } else {
1898 btrfs_get_tree_mod_seq(fs_info, &elem);
1899 walk_ctx.time_seq = elem.seq;
1900 }
1901
1902 ctx->use_path_cache = true;
1903
1904 /*
1905 * We may have previously determined that the current leaf is shared.
1906 * If it is, then we have a data extent that is shared due to a shared
1907 * subtree (caused by snapshotting) and we don't need to check for data
1908 * backrefs. If the leaf is not shared, then we must do backref walking
1909 * to determine if the data extent is shared through reflinks.
1910 */
1911 leaf_cached = lookup_backref_shared_cache(ctx, root,
1912 ctx->curr_leaf_bytenr, 0,
1913 &leaf_is_shared);
1914 if (leaf_cached && leaf_is_shared) {
1915 ret = 1;
1916 goto out_trans;
1917 }
1918
1919 walk_ctx.skip_inode_ref_list = true;
1920 walk_ctx.trans = trans;
1921 walk_ctx.fs_info = fs_info;
1922 walk_ctx.refs = &ctx->refs;
1923
1924 /* -1 means we are in the bytenr of the data extent. */
1925 level = -1;
1926 ULIST_ITER_INIT(&uiter);
1927 while (1) {
1928 const unsigned long prev_ref_count = ctx->refs.nnodes;
1929
1930 walk_ctx.bytenr = bytenr;
1931 ret = find_parent_nodes(&walk_ctx, &shared);
1932 if (ret == BACKREF_FOUND_SHARED ||
1933 ret == BACKREF_FOUND_NOT_SHARED) {
1934 /* If shared must return 1, otherwise return 0. */
1935 ret = (ret == BACKREF_FOUND_SHARED) ? 1 : 0;
1936 if (level >= 0)
1937 store_backref_shared_cache(ctx, root, bytenr,
1938 level, ret == 1);
1939 break;
1940 }
1941 if (ret < 0 && ret != -ENOENT)
1942 break;
1943 ret = 0;
1944
1945 /*
1946 * More than one extent buffer (bytenr) may have been added to
1947 * the ctx->refs ulist, in which case we have to check multiple
1948 * tree paths in case the first one is not shared, so we can not
1949 * use the path cache which is made for a single path. Multiple
1950 * extent buffers at the current level happen when:
1951 *
1952 * 1) level -1, the data extent: If our data extent was not
1953 * directly shared (without multiple reference items), then
1954 * it might have a single reference item with a count > 1 for
1955 * the same offset, which means there are 2 (or more) file
1956 * extent items that point to the data extent - this happens
1957 * when a file extent item needs to be split and then one
1958 * item gets moved to another leaf due to a b+tree leaf split
1959 * when inserting some item. In this case the file extent
1960 * items may be located in different leaves and therefore
1961 * some of the leaves may be referenced through shared
1962 * subtrees while others are not. Since our extent buffer
1963 * cache only works for a single path (by far the most common
1964 * case and simpler to deal with), we can not use it if we
1965 * have multiple leaves (which implies multiple paths).
1966 *
1967 * 2) level >= 0, a tree node/leaf: We can have a mix of direct
1968 * and indirect references on a b+tree node/leaf, so we have
1969 * to check multiple paths, and the extent buffer (the
1970 * current bytenr) may be shared or not. One example is
1971 * during relocation as we may get a shared tree block ref
1972 * (direct ref) and a non-shared tree block ref (indirect
1973 * ref) for the same node/leaf.
1974 */
1975 if ((ctx->refs.nnodes - prev_ref_count) > 1)
1976 ctx->use_path_cache = false;
1977
1978 if (level >= 0)
1979 store_backref_shared_cache(ctx, root, bytenr,
1980 level, false);
1981 node = ulist_next(&ctx->refs, &uiter);
1982 if (!node)
1983 break;
1984 bytenr = node->val;
1985 if (ctx->use_path_cache) {
1986 bool is_shared;
1987 bool cached;
1988
1989 level++;
1990 cached = lookup_backref_shared_cache(ctx, root, bytenr,
1991 level, &is_shared);
1992 if (cached) {
1993 ret = (is_shared ? 1 : 0);
1994 break;
1995 }
1996 }
1997 shared.share_count = 0;
1998 shared.have_delayed_delete_refs = false;
1999 cond_resched();
2000 }
2001
2002 /*
2003 * If the path cache is disabled, then it means at some tree level we
2004 * got multiple parents due to a mix of direct and indirect backrefs or
2005 * multiple leaves with file extent items pointing to the same data
2006 * extent. We have to invalidate the cache and cache only the sharedness
2007 * result for the levels where we got only one node/reference.
2008 */
2009 if (!ctx->use_path_cache) {
2010 int i = 0;
2011
2012 level--;
2013 if (ret >= 0 && level >= 0) {
2014 bytenr = ctx->path_cache_entries[level].bytenr;
2015 ctx->use_path_cache = true;
2016 store_backref_shared_cache(ctx, root, bytenr, level, ret);
2017 i = level + 1;
2018 }
2019
2020 for ( ; i < BTRFS_MAX_LEVEL; i++)
2021 ctx->path_cache_entries[i].bytenr = 0;
2022 }
2023
2024 /*
2025 * Cache the sharedness result for the data extent if we know our inode
2026 * has more than 1 file extent item that refers to the data extent.
2027 */
2028 if (ret >= 0 && shared.self_ref_count > 1) {
2029 int slot = ctx->prev_extents_cache_slot;
2030
2031 ctx->prev_extents_cache[slot].bytenr = shared.data_bytenr;
2032 ctx->prev_extents_cache[slot].is_shared = (ret == 1);
2033
2034 slot = (slot + 1) % BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE;
2035 ctx->prev_extents_cache_slot = slot;
2036 }
2037
2038out_trans:
2039 if (trans) {
2040 btrfs_put_tree_mod_seq(fs_info, &elem);
2041 btrfs_end_transaction(trans);
2042 } else {
2043 up_read(&fs_info->commit_root_sem);
2044 }
2045out:
2046 ulist_release(&ctx->refs);
2047 ctx->prev_leaf_bytenr = ctx->curr_leaf_bytenr;
2048
2049 return ret;
2050}
2051
2052int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
2053 u64 start_off, struct btrfs_path *path,
2054 struct btrfs_inode_extref **ret_extref,
2055 u64 *found_off)
2056{
2057 int ret, slot;
2058 struct btrfs_key key;
2059 struct btrfs_key found_key;
2060 struct btrfs_inode_extref *extref;
2061 const struct extent_buffer *leaf;
2062 unsigned long ptr;
2063
2064 key.objectid = inode_objectid;
2065 key.type = BTRFS_INODE_EXTREF_KEY;
2066 key.offset = start_off;
2067
2068 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2069 if (ret < 0)
2070 return ret;
2071
2072 while (1) {
2073 leaf = path->nodes[0];
2074 slot = path->slots[0];
2075 if (slot >= btrfs_header_nritems(leaf)) {
2076 /*
2077 * If the item at offset is not found,
2078 * btrfs_search_slot will point us to the slot
2079 * where it should be inserted. In our case
2080 * that will be the slot directly before the
2081 * next INODE_REF_KEY_V2 item. In the case
2082 * that we're pointing to the last slot in a
2083 * leaf, we must move one leaf over.
2084 */
2085 ret = btrfs_next_leaf(root, path);
2086 if (ret) {
2087 if (ret >= 1)
2088 ret = -ENOENT;
2089 break;
2090 }
2091 continue;
2092 }
2093
2094 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2095
2096 /*
2097 * Check that we're still looking at an extended ref key for
2098 * this particular objectid. If we have different
2099 * objectid or type then there are no more to be found
2100 * in the tree and we can exit.
2101 */
2102 ret = -ENOENT;
2103 if (found_key.objectid != inode_objectid)
2104 break;
2105 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
2106 break;
2107
2108 ret = 0;
2109 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2110 extref = (struct btrfs_inode_extref *)ptr;
2111 *ret_extref = extref;
2112 if (found_off)
2113 *found_off = found_key.offset;
2114 break;
2115 }
2116
2117 return ret;
2118}
2119
2120/*
2121 * this iterates to turn a name (from iref/extref) into a full filesystem path.
2122 * Elements of the path are separated by '/' and the path is guaranteed to be
2123 * 0-terminated. the path is only given within the current file system.
2124 * Therefore, it never starts with a '/'. the caller is responsible to provide
2125 * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
2126 * the start point of the resulting string is returned. this pointer is within
2127 * dest, normally.
2128 * in case the path buffer would overflow, the pointer is decremented further
2129 * as if output was written to the buffer, though no more output is actually
2130 * generated. that way, the caller can determine how much space would be
2131 * required for the path to fit into the buffer. in that case, the returned
2132 * value will be smaller than dest. callers must check this!
2133 */
2134char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
2135 u32 name_len, unsigned long name_off,
2136 struct extent_buffer *eb_in, u64 parent,
2137 char *dest, u32 size)
2138{
2139 int slot;
2140 u64 next_inum;
2141 int ret;
2142 s64 bytes_left = ((s64)size) - 1;
2143 struct extent_buffer *eb = eb_in;
2144 struct btrfs_key found_key;
2145 struct btrfs_inode_ref *iref;
2146
2147 if (bytes_left >= 0)
2148 dest[bytes_left] = '\0';
2149
2150 while (1) {
2151 bytes_left -= name_len;
2152 if (bytes_left >= 0)
2153 read_extent_buffer(eb, dest + bytes_left,
2154 name_off, name_len);
2155 if (eb != eb_in) {
2156 if (!path->skip_locking)
2157 btrfs_tree_read_unlock(eb);
2158 free_extent_buffer(eb);
2159 }
2160 ret = btrfs_find_item(fs_root, path, parent, 0,
2161 BTRFS_INODE_REF_KEY, &found_key);
2162 if (ret > 0)
2163 ret = -ENOENT;
2164 if (ret)
2165 break;
2166
2167 next_inum = found_key.offset;
2168
2169 /* regular exit ahead */
2170 if (parent == next_inum)
2171 break;
2172
2173 slot = path->slots[0];
2174 eb = path->nodes[0];
2175 /* make sure we can use eb after releasing the path */
2176 if (eb != eb_in) {
2177 path->nodes[0] = NULL;
2178 path->locks[0] = 0;
2179 }
2180 btrfs_release_path(path);
2181 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2182
2183 name_len = btrfs_inode_ref_name_len(eb, iref);
2184 name_off = (unsigned long)(iref + 1);
2185
2186 parent = next_inum;
2187 --bytes_left;
2188 if (bytes_left >= 0)
2189 dest[bytes_left] = '/';
2190 }
2191
2192 btrfs_release_path(path);
2193
2194 if (ret)
2195 return ERR_PTR(ret);
2196
2197 return dest + bytes_left;
2198}
2199
2200/*
2201 * this makes the path point to (logical EXTENT_ITEM *)
2202 * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
2203 * tree blocks and <0 on error.
2204 */
2205int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
2206 struct btrfs_path *path, struct btrfs_key *found_key,
2207 u64 *flags_ret)
2208{
2209 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
2210 int ret;
2211 u64 flags;
2212 u64 size = 0;
2213 u32 item_size;
2214 const struct extent_buffer *eb;
2215 struct btrfs_extent_item *ei;
2216 struct btrfs_key key;
2217
2218 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2219 key.type = BTRFS_METADATA_ITEM_KEY;
2220 else
2221 key.type = BTRFS_EXTENT_ITEM_KEY;
2222 key.objectid = logical;
2223 key.offset = (u64)-1;
2224
2225 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2226 if (ret < 0)
2227 return ret;
2228
2229 ret = btrfs_previous_extent_item(extent_root, path, 0);
2230 if (ret) {
2231 if (ret > 0)
2232 ret = -ENOENT;
2233 return ret;
2234 }
2235 btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
2236 if (found_key->type == BTRFS_METADATA_ITEM_KEY)
2237 size = fs_info->nodesize;
2238 else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
2239 size = found_key->offset;
2240
2241 if (found_key->objectid > logical ||
2242 found_key->objectid + size <= logical) {
2243 btrfs_debug(fs_info,
2244 "logical %llu is not within any extent", logical);
2245 return -ENOENT;
2246 }
2247
2248 eb = path->nodes[0];
2249 item_size = btrfs_item_size(eb, path->slots[0]);
2250 BUG_ON(item_size < sizeof(*ei));
2251
2252 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
2253 flags = btrfs_extent_flags(eb, ei);
2254
2255 btrfs_debug(fs_info,
2256 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
2257 logical, logical - found_key->objectid, found_key->objectid,
2258 found_key->offset, flags, item_size);
2259
2260 WARN_ON(!flags_ret);
2261 if (flags_ret) {
2262 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2263 *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
2264 else if (flags & BTRFS_EXTENT_FLAG_DATA)
2265 *flags_ret = BTRFS_EXTENT_FLAG_DATA;
2266 else
2267 BUG();
2268 return 0;
2269 }
2270
2271 return -EIO;
2272}
2273
2274/*
2275 * helper function to iterate extent inline refs. ptr must point to a 0 value
2276 * for the first call and may be modified. it is used to track state.
2277 * if more refs exist, 0 is returned and the next call to
2278 * get_extent_inline_ref must pass the modified ptr parameter to get the
2279 * next ref. after the last ref was processed, 1 is returned.
2280 * returns <0 on error
2281 */
2282static int get_extent_inline_ref(unsigned long *ptr,
2283 const struct extent_buffer *eb,
2284 const struct btrfs_key *key,
2285 const struct btrfs_extent_item *ei,
2286 u32 item_size,
2287 struct btrfs_extent_inline_ref **out_eiref,
2288 int *out_type)
2289{
2290 unsigned long end;
2291 u64 flags;
2292 struct btrfs_tree_block_info *info;
2293
2294 if (!*ptr) {
2295 /* first call */
2296 flags = btrfs_extent_flags(eb, ei);
2297 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2298 if (key->type == BTRFS_METADATA_ITEM_KEY) {
2299 /* a skinny metadata extent */
2300 *out_eiref =
2301 (struct btrfs_extent_inline_ref *)(ei + 1);
2302 } else {
2303 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
2304 info = (struct btrfs_tree_block_info *)(ei + 1);
2305 *out_eiref =
2306 (struct btrfs_extent_inline_ref *)(info + 1);
2307 }
2308 } else {
2309 *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
2310 }
2311 *ptr = (unsigned long)*out_eiref;
2312 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
2313 return -ENOENT;
2314 }
2315
2316 end = (unsigned long)ei + item_size;
2317 *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
2318 *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
2319 BTRFS_REF_TYPE_ANY);
2320 if (*out_type == BTRFS_REF_TYPE_INVALID)
2321 return -EUCLEAN;
2322
2323 *ptr += btrfs_extent_inline_ref_size(*out_type);
2324 WARN_ON(*ptr > end);
2325 if (*ptr == end)
2326 return 1; /* last */
2327
2328 return 0;
2329}
2330
2331/*
2332 * reads the tree block backref for an extent. tree level and root are returned
2333 * through out_level and out_root. ptr must point to a 0 value for the first
2334 * call and may be modified (see get_extent_inline_ref comment).
2335 * returns 0 if data was provided, 1 if there was no more data to provide or
2336 * <0 on error.
2337 */
2338int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
2339 struct btrfs_key *key, struct btrfs_extent_item *ei,
2340 u32 item_size, u64 *out_root, u8 *out_level)
2341{
2342 int ret;
2343 int type;
2344 struct btrfs_extent_inline_ref *eiref;
2345
2346 if (*ptr == (unsigned long)-1)
2347 return 1;
2348
2349 while (1) {
2350 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
2351 &eiref, &type);
2352 if (ret < 0)
2353 return ret;
2354
2355 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
2356 type == BTRFS_SHARED_BLOCK_REF_KEY)
2357 break;
2358
2359 if (ret == 1)
2360 return 1;
2361 }
2362
2363 /* we can treat both ref types equally here */
2364 *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
2365
2366 if (key->type == BTRFS_EXTENT_ITEM_KEY) {
2367 struct btrfs_tree_block_info *info;
2368
2369 info = (struct btrfs_tree_block_info *)(ei + 1);
2370 *out_level = btrfs_tree_block_level(eb, info);
2371 } else {
2372 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
2373 *out_level = (u8)key->offset;
2374 }
2375
2376 if (ret == 1)
2377 *ptr = (unsigned long)-1;
2378
2379 return 0;
2380}
2381
2382static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
2383 struct extent_inode_elem *inode_list,
2384 u64 root, u64 extent_item_objectid,
2385 iterate_extent_inodes_t *iterate, void *ctx)
2386{
2387 struct extent_inode_elem *eie;
2388 int ret = 0;
2389
2390 for (eie = inode_list; eie; eie = eie->next) {
2391 btrfs_debug(fs_info,
2392 "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
2393 extent_item_objectid, eie->inum,
2394 eie->offset, root);
2395 ret = iterate(eie->inum, eie->offset, eie->num_bytes, root, ctx);
2396 if (ret) {
2397 btrfs_debug(fs_info,
2398 "stopping iteration for %llu due to ret=%d",
2399 extent_item_objectid, ret);
2400 break;
2401 }
2402 }
2403
2404 return ret;
2405}
2406
2407/*
2408 * calls iterate() for every inode that references the extent identified by
2409 * the given parameters.
2410 * when the iterator function returns a non-zero value, iteration stops.
2411 */
2412int iterate_extent_inodes(struct btrfs_backref_walk_ctx *ctx,
2413 bool search_commit_root,
2414 iterate_extent_inodes_t *iterate, void *user_ctx)
2415{
2416 int ret;
2417 struct ulist *refs;
2418 struct ulist_node *ref_node;
2419 struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
2420 struct ulist_iterator ref_uiter;
2421
2422 btrfs_debug(ctx->fs_info, "resolving all inodes for extent %llu",
2423 ctx->bytenr);
2424
2425 ASSERT(ctx->trans == NULL);
2426 ASSERT(ctx->roots == NULL);
2427
2428 if (!search_commit_root) {
2429 struct btrfs_trans_handle *trans;
2430
2431 trans = btrfs_attach_transaction(ctx->fs_info->tree_root);
2432 if (IS_ERR(trans)) {
2433 if (PTR_ERR(trans) != -ENOENT &&
2434 PTR_ERR(trans) != -EROFS)
2435 return PTR_ERR(trans);
2436 trans = NULL;
2437 }
2438 ctx->trans = trans;
2439 }
2440
2441 if (ctx->trans) {
2442 btrfs_get_tree_mod_seq(ctx->fs_info, &seq_elem);
2443 ctx->time_seq = seq_elem.seq;
2444 } else {
2445 down_read(&ctx->fs_info->commit_root_sem);
2446 }
2447
2448 ret = btrfs_find_all_leafs(ctx);
2449 if (ret)
2450 goto out;
2451 refs = ctx->refs;
2452 ctx->refs = NULL;
2453
2454 ULIST_ITER_INIT(&ref_uiter);
2455 while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2456 const u64 leaf_bytenr = ref_node->val;
2457 struct ulist_node *root_node;
2458 struct ulist_iterator root_uiter;
2459 struct extent_inode_elem *inode_list;
2460
2461 inode_list = (struct extent_inode_elem *)(uintptr_t)ref_node->aux;
2462
2463 if (ctx->cache_lookup) {
2464 const u64 *root_ids;
2465 int root_count;
2466 bool cached;
2467
2468 cached = ctx->cache_lookup(leaf_bytenr, ctx->user_ctx,
2469 &root_ids, &root_count);
2470 if (cached) {
2471 for (int i = 0; i < root_count; i++) {
2472 ret = iterate_leaf_refs(ctx->fs_info,
2473 inode_list,
2474 root_ids[i],
2475 leaf_bytenr,
2476 iterate,
2477 user_ctx);
2478 if (ret)
2479 break;
2480 }
2481 continue;
2482 }
2483 }
2484
2485 if (!ctx->roots) {
2486 ctx->roots = ulist_alloc(GFP_NOFS);
2487 if (!ctx->roots) {
2488 ret = -ENOMEM;
2489 break;
2490 }
2491 }
2492
2493 ctx->bytenr = leaf_bytenr;
2494 ret = btrfs_find_all_roots_safe(ctx);
2495 if (ret)
2496 break;
2497
2498 if (ctx->cache_store)
2499 ctx->cache_store(leaf_bytenr, ctx->roots, ctx->user_ctx);
2500
2501 ULIST_ITER_INIT(&root_uiter);
2502 while (!ret && (root_node = ulist_next(ctx->roots, &root_uiter))) {
2503 btrfs_debug(ctx->fs_info,
2504 "root %llu references leaf %llu, data list %#llx",
2505 root_node->val, ref_node->val,
2506 ref_node->aux);
2507 ret = iterate_leaf_refs(ctx->fs_info, inode_list,
2508 root_node->val, ctx->bytenr,
2509 iterate, user_ctx);
2510 }
2511 ulist_reinit(ctx->roots);
2512 }
2513
2514 free_leaf_list(refs);
2515out:
2516 if (ctx->trans) {
2517 btrfs_put_tree_mod_seq(ctx->fs_info, &seq_elem);
2518 btrfs_end_transaction(ctx->trans);
2519 ctx->trans = NULL;
2520 } else {
2521 up_read(&ctx->fs_info->commit_root_sem);
2522 }
2523
2524 ulist_free(ctx->roots);
2525 ctx->roots = NULL;
2526
2527 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP)
2528 ret = 0;
2529
2530 return ret;
2531}
2532
2533static int build_ino_list(u64 inum, u64 offset, u64 num_bytes, u64 root, void *ctx)
2534{
2535 struct btrfs_data_container *inodes = ctx;
2536 const size_t c = 3 * sizeof(u64);
2537
2538 if (inodes->bytes_left >= c) {
2539 inodes->bytes_left -= c;
2540 inodes->val[inodes->elem_cnt] = inum;
2541 inodes->val[inodes->elem_cnt + 1] = offset;
2542 inodes->val[inodes->elem_cnt + 2] = root;
2543 inodes->elem_cnt += 3;
2544 } else {
2545 inodes->bytes_missing += c - inodes->bytes_left;
2546 inodes->bytes_left = 0;
2547 inodes->elem_missed += 3;
2548 }
2549
2550 return 0;
2551}
2552
2553int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2554 struct btrfs_path *path,
2555 void *ctx, bool ignore_offset)
2556{
2557 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
2558 int ret;
2559 u64 flags = 0;
2560 struct btrfs_key found_key;
2561 int search_commit_root = path->search_commit_root;
2562
2563 ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2564 btrfs_release_path(path);
2565 if (ret < 0)
2566 return ret;
2567 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2568 return -EINVAL;
2569
2570 walk_ctx.bytenr = found_key.objectid;
2571 if (ignore_offset)
2572 walk_ctx.ignore_extent_item_pos = true;
2573 else
2574 walk_ctx.extent_item_pos = logical - found_key.objectid;
2575 walk_ctx.fs_info = fs_info;
2576
2577 return iterate_extent_inodes(&walk_ctx, search_commit_root,
2578 build_ino_list, ctx);
2579}
2580
2581static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2582 struct extent_buffer *eb, struct inode_fs_paths *ipath);
2583
2584static int iterate_inode_refs(u64 inum, struct inode_fs_paths *ipath)
2585{
2586 int ret = 0;
2587 int slot;
2588 u32 cur;
2589 u32 len;
2590 u32 name_len;
2591 u64 parent = 0;
2592 int found = 0;
2593 struct btrfs_root *fs_root = ipath->fs_root;
2594 struct btrfs_path *path = ipath->btrfs_path;
2595 struct extent_buffer *eb;
2596 struct btrfs_inode_ref *iref;
2597 struct btrfs_key found_key;
2598
2599 while (!ret) {
2600 ret = btrfs_find_item(fs_root, path, inum,
2601 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2602 &found_key);
2603
2604 if (ret < 0)
2605 break;
2606 if (ret) {
2607 ret = found ? 0 : -ENOENT;
2608 break;
2609 }
2610 ++found;
2611
2612 parent = found_key.offset;
2613 slot = path->slots[0];
2614 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2615 if (!eb) {
2616 ret = -ENOMEM;
2617 break;
2618 }
2619 btrfs_release_path(path);
2620
2621 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2622
2623 for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2624 name_len = btrfs_inode_ref_name_len(eb, iref);
2625 /* path must be released before calling iterate()! */
2626 btrfs_debug(fs_root->fs_info,
2627 "following ref at offset %u for inode %llu in tree %llu",
2628 cur, found_key.objectid,
2629 fs_root->root_key.objectid);
2630 ret = inode_to_path(parent, name_len,
2631 (unsigned long)(iref + 1), eb, ipath);
2632 if (ret)
2633 break;
2634 len = sizeof(*iref) + name_len;
2635 iref = (struct btrfs_inode_ref *)((char *)iref + len);
2636 }
2637 free_extent_buffer(eb);
2638 }
2639
2640 btrfs_release_path(path);
2641
2642 return ret;
2643}
2644
2645static int iterate_inode_extrefs(u64 inum, struct inode_fs_paths *ipath)
2646{
2647 int ret;
2648 int slot;
2649 u64 offset = 0;
2650 u64 parent;
2651 int found = 0;
2652 struct btrfs_root *fs_root = ipath->fs_root;
2653 struct btrfs_path *path = ipath->btrfs_path;
2654 struct extent_buffer *eb;
2655 struct btrfs_inode_extref *extref;
2656 u32 item_size;
2657 u32 cur_offset;
2658 unsigned long ptr;
2659
2660 while (1) {
2661 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2662 &offset);
2663 if (ret < 0)
2664 break;
2665 if (ret) {
2666 ret = found ? 0 : -ENOENT;
2667 break;
2668 }
2669 ++found;
2670
2671 slot = path->slots[0];
2672 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2673 if (!eb) {
2674 ret = -ENOMEM;
2675 break;
2676 }
2677 btrfs_release_path(path);
2678
2679 item_size = btrfs_item_size(eb, slot);
2680 ptr = btrfs_item_ptr_offset(eb, slot);
2681 cur_offset = 0;
2682
2683 while (cur_offset < item_size) {
2684 u32 name_len;
2685
2686 extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2687 parent = btrfs_inode_extref_parent(eb, extref);
2688 name_len = btrfs_inode_extref_name_len(eb, extref);
2689 ret = inode_to_path(parent, name_len,
2690 (unsigned long)&extref->name, eb, ipath);
2691 if (ret)
2692 break;
2693
2694 cur_offset += btrfs_inode_extref_name_len(eb, extref);
2695 cur_offset += sizeof(*extref);
2696 }
2697 free_extent_buffer(eb);
2698
2699 offset++;
2700 }
2701
2702 btrfs_release_path(path);
2703
2704 return ret;
2705}
2706
2707/*
2708 * returns 0 if the path could be dumped (probably truncated)
2709 * returns <0 in case of an error
2710 */
2711static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2712 struct extent_buffer *eb, struct inode_fs_paths *ipath)
2713{
2714 char *fspath;
2715 char *fspath_min;
2716 int i = ipath->fspath->elem_cnt;
2717 const int s_ptr = sizeof(char *);
2718 u32 bytes_left;
2719
2720 bytes_left = ipath->fspath->bytes_left > s_ptr ?
2721 ipath->fspath->bytes_left - s_ptr : 0;
2722
2723 fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2724 fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2725 name_off, eb, inum, fspath_min, bytes_left);
2726 if (IS_ERR(fspath))
2727 return PTR_ERR(fspath);
2728
2729 if (fspath > fspath_min) {
2730 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2731 ++ipath->fspath->elem_cnt;
2732 ipath->fspath->bytes_left = fspath - fspath_min;
2733 } else {
2734 ++ipath->fspath->elem_missed;
2735 ipath->fspath->bytes_missing += fspath_min - fspath;
2736 ipath->fspath->bytes_left = 0;
2737 }
2738
2739 return 0;
2740}
2741
2742/*
2743 * this dumps all file system paths to the inode into the ipath struct, provided
2744 * is has been created large enough. each path is zero-terminated and accessed
2745 * from ipath->fspath->val[i].
2746 * when it returns, there are ipath->fspath->elem_cnt number of paths available
2747 * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2748 * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2749 * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2750 * have been needed to return all paths.
2751 */
2752int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2753{
2754 int ret;
2755 int found_refs = 0;
2756
2757 ret = iterate_inode_refs(inum, ipath);
2758 if (!ret)
2759 ++found_refs;
2760 else if (ret != -ENOENT)
2761 return ret;
2762
2763 ret = iterate_inode_extrefs(inum, ipath);
2764 if (ret == -ENOENT && found_refs)
2765 return 0;
2766
2767 return ret;
2768}
2769
2770struct btrfs_data_container *init_data_container(u32 total_bytes)
2771{
2772 struct btrfs_data_container *data;
2773 size_t alloc_bytes;
2774
2775 alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2776 data = kvmalloc(alloc_bytes, GFP_KERNEL);
2777 if (!data)
2778 return ERR_PTR(-ENOMEM);
2779
2780 if (total_bytes >= sizeof(*data)) {
2781 data->bytes_left = total_bytes - sizeof(*data);
2782 data->bytes_missing = 0;
2783 } else {
2784 data->bytes_missing = sizeof(*data) - total_bytes;
2785 data->bytes_left = 0;
2786 }
2787
2788 data->elem_cnt = 0;
2789 data->elem_missed = 0;
2790
2791 return data;
2792}
2793
2794/*
2795 * allocates space to return multiple file system paths for an inode.
2796 * total_bytes to allocate are passed, note that space usable for actual path
2797 * information will be total_bytes - sizeof(struct inode_fs_paths).
2798 * the returned pointer must be freed with free_ipath() in the end.
2799 */
2800struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2801 struct btrfs_path *path)
2802{
2803 struct inode_fs_paths *ifp;
2804 struct btrfs_data_container *fspath;
2805
2806 fspath = init_data_container(total_bytes);
2807 if (IS_ERR(fspath))
2808 return ERR_CAST(fspath);
2809
2810 ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2811 if (!ifp) {
2812 kvfree(fspath);
2813 return ERR_PTR(-ENOMEM);
2814 }
2815
2816 ifp->btrfs_path = path;
2817 ifp->fspath = fspath;
2818 ifp->fs_root = fs_root;
2819
2820 return ifp;
2821}
2822
2823void free_ipath(struct inode_fs_paths *ipath)
2824{
2825 if (!ipath)
2826 return;
2827 kvfree(ipath->fspath);
2828 kfree(ipath);
2829}
2830
2831struct btrfs_backref_iter *btrfs_backref_iter_alloc(struct btrfs_fs_info *fs_info)
2832{
2833 struct btrfs_backref_iter *ret;
2834
2835 ret = kzalloc(sizeof(*ret), GFP_NOFS);
2836 if (!ret)
2837 return NULL;
2838
2839 ret->path = btrfs_alloc_path();
2840 if (!ret->path) {
2841 kfree(ret);
2842 return NULL;
2843 }
2844
2845 /* Current backref iterator only supports iteration in commit root */
2846 ret->path->search_commit_root = 1;
2847 ret->path->skip_locking = 1;
2848 ret->fs_info = fs_info;
2849
2850 return ret;
2851}
2852
2853int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2854{
2855 struct btrfs_fs_info *fs_info = iter->fs_info;
2856 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2857 struct btrfs_path *path = iter->path;
2858 struct btrfs_extent_item *ei;
2859 struct btrfs_key key;
2860 int ret;
2861
2862 key.objectid = bytenr;
2863 key.type = BTRFS_METADATA_ITEM_KEY;
2864 key.offset = (u64)-1;
2865 iter->bytenr = bytenr;
2866
2867 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2868 if (ret < 0)
2869 return ret;
2870 if (ret == 0) {
2871 ret = -EUCLEAN;
2872 goto release;
2873 }
2874 if (path->slots[0] == 0) {
2875 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2876 ret = -EUCLEAN;
2877 goto release;
2878 }
2879 path->slots[0]--;
2880
2881 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2882 if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2883 key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2884 ret = -ENOENT;
2885 goto release;
2886 }
2887 memcpy(&iter->cur_key, &key, sizeof(key));
2888 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2889 path->slots[0]);
2890 iter->end_ptr = (u32)(iter->item_ptr +
2891 btrfs_item_size(path->nodes[0], path->slots[0]));
2892 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2893 struct btrfs_extent_item);
2894
2895 /*
2896 * Only support iteration on tree backref yet.
2897 *
2898 * This is an extra precaution for non skinny-metadata, where
2899 * EXTENT_ITEM is also used for tree blocks, that we can only use
2900 * extent flags to determine if it's a tree block.
2901 */
2902 if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2903 ret = -ENOTSUPP;
2904 goto release;
2905 }
2906 iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2907
2908 /* If there is no inline backref, go search for keyed backref */
2909 if (iter->cur_ptr >= iter->end_ptr) {
2910 ret = btrfs_next_item(extent_root, path);
2911
2912 /* No inline nor keyed ref */
2913 if (ret > 0) {
2914 ret = -ENOENT;
2915 goto release;
2916 }
2917 if (ret < 0)
2918 goto release;
2919
2920 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2921 path->slots[0]);
2922 if (iter->cur_key.objectid != bytenr ||
2923 (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2924 iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2925 ret = -ENOENT;
2926 goto release;
2927 }
2928 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2929 path->slots[0]);
2930 iter->item_ptr = iter->cur_ptr;
2931 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2932 path->nodes[0], path->slots[0]));
2933 }
2934
2935 return 0;
2936release:
2937 btrfs_backref_iter_release(iter);
2938 return ret;
2939}
2940
2941/*
2942 * Go to the next backref item of current bytenr, can be either inlined or
2943 * keyed.
2944 *
2945 * Caller needs to check whether it's inline ref or not by iter->cur_key.
2946 *
2947 * Return 0 if we get next backref without problem.
2948 * Return >0 if there is no extra backref for this bytenr.
2949 * Return <0 if there is something wrong happened.
2950 */
2951int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2952{
2953 struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2954 struct btrfs_root *extent_root;
2955 struct btrfs_path *path = iter->path;
2956 struct btrfs_extent_inline_ref *iref;
2957 int ret;
2958 u32 size;
2959
2960 if (btrfs_backref_iter_is_inline_ref(iter)) {
2961 /* We're still inside the inline refs */
2962 ASSERT(iter->cur_ptr < iter->end_ptr);
2963
2964 if (btrfs_backref_has_tree_block_info(iter)) {
2965 /* First tree block info */
2966 size = sizeof(struct btrfs_tree_block_info);
2967 } else {
2968 /* Use inline ref type to determine the size */
2969 int type;
2970
2971 iref = (struct btrfs_extent_inline_ref *)
2972 ((unsigned long)iter->cur_ptr);
2973 type = btrfs_extent_inline_ref_type(eb, iref);
2974
2975 size = btrfs_extent_inline_ref_size(type);
2976 }
2977 iter->cur_ptr += size;
2978 if (iter->cur_ptr < iter->end_ptr)
2979 return 0;
2980
2981 /* All inline items iterated, fall through */
2982 }
2983
2984 /* We're at keyed items, there is no inline item, go to the next one */
2985 extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
2986 ret = btrfs_next_item(extent_root, iter->path);
2987 if (ret)
2988 return ret;
2989
2990 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2991 if (iter->cur_key.objectid != iter->bytenr ||
2992 (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2993 iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2994 return 1;
2995 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2996 path->slots[0]);
2997 iter->cur_ptr = iter->item_ptr;
2998 iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
2999 path->slots[0]);
3000 return 0;
3001}
3002
3003void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
3004 struct btrfs_backref_cache *cache, bool is_reloc)
3005{
3006 int i;
3007
3008 cache->rb_root = RB_ROOT;
3009 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3010 INIT_LIST_HEAD(&cache->pending[i]);
3011 INIT_LIST_HEAD(&cache->changed);
3012 INIT_LIST_HEAD(&cache->detached);
3013 INIT_LIST_HEAD(&cache->leaves);
3014 INIT_LIST_HEAD(&cache->pending_edge);
3015 INIT_LIST_HEAD(&cache->useless_node);
3016 cache->fs_info = fs_info;
3017 cache->is_reloc = is_reloc;
3018}
3019
3020struct btrfs_backref_node *btrfs_backref_alloc_node(
3021 struct btrfs_backref_cache *cache, u64 bytenr, int level)
3022{
3023 struct btrfs_backref_node *node;
3024
3025 ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
3026 node = kzalloc(sizeof(*node), GFP_NOFS);
3027 if (!node)
3028 return node;
3029
3030 INIT_LIST_HEAD(&node->list);
3031 INIT_LIST_HEAD(&node->upper);
3032 INIT_LIST_HEAD(&node->lower);
3033 RB_CLEAR_NODE(&node->rb_node);
3034 cache->nr_nodes++;
3035 node->level = level;
3036 node->bytenr = bytenr;
3037
3038 return node;
3039}
3040
3041struct btrfs_backref_edge *btrfs_backref_alloc_edge(
3042 struct btrfs_backref_cache *cache)
3043{
3044 struct btrfs_backref_edge *edge;
3045
3046 edge = kzalloc(sizeof(*edge), GFP_NOFS);
3047 if (edge)
3048 cache->nr_edges++;
3049 return edge;
3050}
3051
3052/*
3053 * Drop the backref node from cache, also cleaning up all its
3054 * upper edges and any uncached nodes in the path.
3055 *
3056 * This cleanup happens bottom up, thus the node should either
3057 * be the lowest node in the cache or a detached node.
3058 */
3059void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
3060 struct btrfs_backref_node *node)
3061{
3062 struct btrfs_backref_node *upper;
3063 struct btrfs_backref_edge *edge;
3064
3065 if (!node)
3066 return;
3067
3068 BUG_ON(!node->lowest && !node->detached);
3069 while (!list_empty(&node->upper)) {
3070 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
3071 list[LOWER]);
3072 upper = edge->node[UPPER];
3073 list_del(&edge->list[LOWER]);
3074 list_del(&edge->list[UPPER]);
3075 btrfs_backref_free_edge(cache, edge);
3076
3077 /*
3078 * Add the node to leaf node list if no other child block
3079 * cached.
3080 */
3081 if (list_empty(&upper->lower)) {
3082 list_add_tail(&upper->lower, &cache->leaves);
3083 upper->lowest = 1;
3084 }
3085 }
3086
3087 btrfs_backref_drop_node(cache, node);
3088}
3089
3090/*
3091 * Release all nodes/edges from current cache
3092 */
3093void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
3094{
3095 struct btrfs_backref_node *node;
3096 int i;
3097
3098 while (!list_empty(&cache->detached)) {
3099 node = list_entry(cache->detached.next,
3100 struct btrfs_backref_node, list);
3101 btrfs_backref_cleanup_node(cache, node);
3102 }
3103
3104 while (!list_empty(&cache->leaves)) {
3105 node = list_entry(cache->leaves.next,
3106 struct btrfs_backref_node, lower);
3107 btrfs_backref_cleanup_node(cache, node);
3108 }
3109
3110 cache->last_trans = 0;
3111
3112 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3113 ASSERT(list_empty(&cache->pending[i]));
3114 ASSERT(list_empty(&cache->pending_edge));
3115 ASSERT(list_empty(&cache->useless_node));
3116 ASSERT(list_empty(&cache->changed));
3117 ASSERT(list_empty(&cache->detached));
3118 ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
3119 ASSERT(!cache->nr_nodes);
3120 ASSERT(!cache->nr_edges);
3121}
3122
3123/*
3124 * Handle direct tree backref
3125 *
3126 * Direct tree backref means, the backref item shows its parent bytenr
3127 * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
3128 *
3129 * @ref_key: The converted backref key.
3130 * For keyed backref, it's the item key.
3131 * For inlined backref, objectid is the bytenr,
3132 * type is btrfs_inline_ref_type, offset is
3133 * btrfs_inline_ref_offset.
3134 */
3135static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
3136 struct btrfs_key *ref_key,
3137 struct btrfs_backref_node *cur)
3138{
3139 struct btrfs_backref_edge *edge;
3140 struct btrfs_backref_node *upper;
3141 struct rb_node *rb_node;
3142
3143 ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
3144
3145 /* Only reloc root uses backref pointing to itself */
3146 if (ref_key->objectid == ref_key->offset) {
3147 struct btrfs_root *root;
3148
3149 cur->is_reloc_root = 1;
3150 /* Only reloc backref cache cares about a specific root */
3151 if (cache->is_reloc) {
3152 root = find_reloc_root(cache->fs_info, cur->bytenr);
3153 if (!root)
3154 return -ENOENT;
3155 cur->root = root;
3156 } else {
3157 /*
3158 * For generic purpose backref cache, reloc root node
3159 * is useless.
3160 */
3161 list_add(&cur->list, &cache->useless_node);
3162 }
3163 return 0;
3164 }
3165
3166 edge = btrfs_backref_alloc_edge(cache);
3167 if (!edge)
3168 return -ENOMEM;
3169
3170 rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
3171 if (!rb_node) {
3172 /* Parent node not yet cached */
3173 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
3174 cur->level + 1);
3175 if (!upper) {
3176 btrfs_backref_free_edge(cache, edge);
3177 return -ENOMEM;
3178 }
3179
3180 /*
3181 * Backrefs for the upper level block isn't cached, add the
3182 * block to pending list
3183 */
3184 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3185 } else {
3186 /* Parent node already cached */
3187 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
3188 ASSERT(upper->checked);
3189 INIT_LIST_HEAD(&edge->list[UPPER]);
3190 }
3191 btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
3192 return 0;
3193}
3194
3195/*
3196 * Handle indirect tree backref
3197 *
3198 * Indirect tree backref means, we only know which tree the node belongs to.
3199 * We still need to do a tree search to find out the parents. This is for
3200 * TREE_BLOCK_REF backref (keyed or inlined).
3201 *
3202 * @trans: Transaction handle.
3203 * @ref_key: The same as @ref_key in handle_direct_tree_backref()
3204 * @tree_key: The first key of this tree block.
3205 * @path: A clean (released) path, to avoid allocating path every time
3206 * the function get called.
3207 */
3208static int handle_indirect_tree_backref(struct btrfs_trans_handle *trans,
3209 struct btrfs_backref_cache *cache,
3210 struct btrfs_path *path,
3211 struct btrfs_key *ref_key,
3212 struct btrfs_key *tree_key,
3213 struct btrfs_backref_node *cur)
3214{
3215 struct btrfs_fs_info *fs_info = cache->fs_info;
3216 struct btrfs_backref_node *upper;
3217 struct btrfs_backref_node *lower;
3218 struct btrfs_backref_edge *edge;
3219 struct extent_buffer *eb;
3220 struct btrfs_root *root;
3221 struct rb_node *rb_node;
3222 int level;
3223 bool need_check = true;
3224 int ret;
3225
3226 root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
3227 if (IS_ERR(root))
3228 return PTR_ERR(root);
3229 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3230 cur->cowonly = 1;
3231
3232 if (btrfs_root_level(&root->root_item) == cur->level) {
3233 /* Tree root */
3234 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
3235 /*
3236 * For reloc backref cache, we may ignore reloc root. But for
3237 * general purpose backref cache, we can't rely on
3238 * btrfs_should_ignore_reloc_root() as it may conflict with
3239 * current running relocation and lead to missing root.
3240 *
3241 * For general purpose backref cache, reloc root detection is
3242 * completely relying on direct backref (key->offset is parent
3243 * bytenr), thus only do such check for reloc cache.
3244 */
3245 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
3246 btrfs_put_root(root);
3247 list_add(&cur->list, &cache->useless_node);
3248 } else {
3249 cur->root = root;
3250 }
3251 return 0;
3252 }
3253
3254 level = cur->level + 1;
3255
3256 /* Search the tree to find parent blocks referring to the block */
3257 path->search_commit_root = 1;
3258 path->skip_locking = 1;
3259 path->lowest_level = level;
3260 ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
3261 path->lowest_level = 0;
3262 if (ret < 0) {
3263 btrfs_put_root(root);
3264 return ret;
3265 }
3266 if (ret > 0 && path->slots[level] > 0)
3267 path->slots[level]--;
3268
3269 eb = path->nodes[level];
3270 if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
3271 btrfs_err(fs_info,
3272"couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
3273 cur->bytenr, level - 1, root->root_key.objectid,
3274 tree_key->objectid, tree_key->type, tree_key->offset);
3275 btrfs_put_root(root);
3276 ret = -ENOENT;
3277 goto out;
3278 }
3279 lower = cur;
3280
3281 /* Add all nodes and edges in the path */
3282 for (; level < BTRFS_MAX_LEVEL; level++) {
3283 if (!path->nodes[level]) {
3284 ASSERT(btrfs_root_bytenr(&root->root_item) ==
3285 lower->bytenr);
3286 /* Same as previous should_ignore_reloc_root() call */
3287 if (btrfs_should_ignore_reloc_root(root) &&
3288 cache->is_reloc) {
3289 btrfs_put_root(root);
3290 list_add(&lower->list, &cache->useless_node);
3291 } else {
3292 lower->root = root;
3293 }
3294 break;
3295 }
3296
3297 edge = btrfs_backref_alloc_edge(cache);
3298 if (!edge) {
3299 btrfs_put_root(root);
3300 ret = -ENOMEM;
3301 goto out;
3302 }
3303
3304 eb = path->nodes[level];
3305 rb_node = rb_simple_search(&cache->rb_root, eb->start);
3306 if (!rb_node) {
3307 upper = btrfs_backref_alloc_node(cache, eb->start,
3308 lower->level + 1);
3309 if (!upper) {
3310 btrfs_put_root(root);
3311 btrfs_backref_free_edge(cache, edge);
3312 ret = -ENOMEM;
3313 goto out;
3314 }
3315 upper->owner = btrfs_header_owner(eb);
3316 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3317 upper->cowonly = 1;
3318
3319 /*
3320 * If we know the block isn't shared we can avoid
3321 * checking its backrefs.
3322 */
3323 if (btrfs_block_can_be_shared(trans, root, eb))
3324 upper->checked = 0;
3325 else
3326 upper->checked = 1;
3327
3328 /*
3329 * Add the block to pending list if we need to check its
3330 * backrefs, we only do this once while walking up a
3331 * tree as we will catch anything else later on.
3332 */
3333 if (!upper->checked && need_check) {
3334 need_check = false;
3335 list_add_tail(&edge->list[UPPER],
3336 &cache->pending_edge);
3337 } else {
3338 if (upper->checked)
3339 need_check = true;
3340 INIT_LIST_HEAD(&edge->list[UPPER]);
3341 }
3342 } else {
3343 upper = rb_entry(rb_node, struct btrfs_backref_node,
3344 rb_node);
3345 ASSERT(upper->checked);
3346 INIT_LIST_HEAD(&edge->list[UPPER]);
3347 if (!upper->owner)
3348 upper->owner = btrfs_header_owner(eb);
3349 }
3350 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
3351
3352 if (rb_node) {
3353 btrfs_put_root(root);
3354 break;
3355 }
3356 lower = upper;
3357 upper = NULL;
3358 }
3359out:
3360 btrfs_release_path(path);
3361 return ret;
3362}
3363
3364/*
3365 * Add backref node @cur into @cache.
3366 *
3367 * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
3368 * links aren't yet bi-directional. Needs to finish such links.
3369 * Use btrfs_backref_finish_upper_links() to finish such linkage.
3370 *
3371 * @trans: Transaction handle.
3372 * @path: Released path for indirect tree backref lookup
3373 * @iter: Released backref iter for extent tree search
3374 * @node_key: The first key of the tree block
3375 */
3376int btrfs_backref_add_tree_node(struct btrfs_trans_handle *trans,
3377 struct btrfs_backref_cache *cache,
3378 struct btrfs_path *path,
3379 struct btrfs_backref_iter *iter,
3380 struct btrfs_key *node_key,
3381 struct btrfs_backref_node *cur)
3382{
3383 struct btrfs_backref_edge *edge;
3384 struct btrfs_backref_node *exist;
3385 int ret;
3386
3387 ret = btrfs_backref_iter_start(iter, cur->bytenr);
3388 if (ret < 0)
3389 return ret;
3390 /*
3391 * We skip the first btrfs_tree_block_info, as we don't use the key
3392 * stored in it, but fetch it from the tree block
3393 */
3394 if (btrfs_backref_has_tree_block_info(iter)) {
3395 ret = btrfs_backref_iter_next(iter);
3396 if (ret < 0)
3397 goto out;
3398 /* No extra backref? This means the tree block is corrupted */
3399 if (ret > 0) {
3400 ret = -EUCLEAN;
3401 goto out;
3402 }
3403 }
3404 WARN_ON(cur->checked);
3405 if (!list_empty(&cur->upper)) {
3406 /*
3407 * The backref was added previously when processing backref of
3408 * type BTRFS_TREE_BLOCK_REF_KEY
3409 */
3410 ASSERT(list_is_singular(&cur->upper));
3411 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
3412 list[LOWER]);
3413 ASSERT(list_empty(&edge->list[UPPER]));
3414 exist = edge->node[UPPER];
3415 /*
3416 * Add the upper level block to pending list if we need check
3417 * its backrefs
3418 */
3419 if (!exist->checked)
3420 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3421 } else {
3422 exist = NULL;
3423 }
3424
3425 for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
3426 struct extent_buffer *eb;
3427 struct btrfs_key key;
3428 int type;
3429
3430 cond_resched();
3431 eb = btrfs_backref_get_eb(iter);
3432
3433 key.objectid = iter->bytenr;
3434 if (btrfs_backref_iter_is_inline_ref(iter)) {
3435 struct btrfs_extent_inline_ref *iref;
3436
3437 /* Update key for inline backref */
3438 iref = (struct btrfs_extent_inline_ref *)
3439 ((unsigned long)iter->cur_ptr);
3440 type = btrfs_get_extent_inline_ref_type(eb, iref,
3441 BTRFS_REF_TYPE_BLOCK);
3442 if (type == BTRFS_REF_TYPE_INVALID) {
3443 ret = -EUCLEAN;
3444 goto out;
3445 }
3446 key.type = type;
3447 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3448 } else {
3449 key.type = iter->cur_key.type;
3450 key.offset = iter->cur_key.offset;
3451 }
3452
3453 /*
3454 * Parent node found and matches current inline ref, no need to
3455 * rebuild this node for this inline ref
3456 */
3457 if (exist &&
3458 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
3459 exist->owner == key.offset) ||
3460 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
3461 exist->bytenr == key.offset))) {
3462 exist = NULL;
3463 continue;
3464 }
3465
3466 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
3467 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3468 ret = handle_direct_tree_backref(cache, &key, cur);
3469 if (ret < 0)
3470 goto out;
3471 } else if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
3472 /*
3473 * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref
3474 * offset means the root objectid. We need to search
3475 * the tree to get its parent bytenr.
3476 */
3477 ret = handle_indirect_tree_backref(trans, cache, path,
3478 &key, node_key, cur);
3479 if (ret < 0)
3480 goto out;
3481 }
3482 /*
3483 * Unrecognized tree backref items (if it can pass tree-checker)
3484 * would be ignored.
3485 */
3486 }
3487 ret = 0;
3488 cur->checked = 1;
3489 WARN_ON(exist);
3490out:
3491 btrfs_backref_iter_release(iter);
3492 return ret;
3493}
3494
3495/*
3496 * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3497 */
3498int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3499 struct btrfs_backref_node *start)
3500{
3501 struct list_head *useless_node = &cache->useless_node;
3502 struct btrfs_backref_edge *edge;
3503 struct rb_node *rb_node;
3504 LIST_HEAD(pending_edge);
3505
3506 ASSERT(start->checked);
3507
3508 /* Insert this node to cache if it's not COW-only */
3509 if (!start->cowonly) {
3510 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
3511 &start->rb_node);
3512 if (rb_node)
3513 btrfs_backref_panic(cache->fs_info, start->bytenr,
3514 -EEXIST);
3515 list_add_tail(&start->lower, &cache->leaves);
3516 }
3517
3518 /*
3519 * Use breadth first search to iterate all related edges.
3520 *
3521 * The starting points are all the edges of this node
3522 */
3523 list_for_each_entry(edge, &start->upper, list[LOWER])
3524 list_add_tail(&edge->list[UPPER], &pending_edge);
3525
3526 while (!list_empty(&pending_edge)) {
3527 struct btrfs_backref_node *upper;
3528 struct btrfs_backref_node *lower;
3529
3530 edge = list_first_entry(&pending_edge,
3531 struct btrfs_backref_edge, list[UPPER]);
3532 list_del_init(&edge->list[UPPER]);
3533 upper = edge->node[UPPER];
3534 lower = edge->node[LOWER];
3535
3536 /* Parent is detached, no need to keep any edges */
3537 if (upper->detached) {
3538 list_del(&edge->list[LOWER]);
3539 btrfs_backref_free_edge(cache, edge);
3540
3541 /* Lower node is orphan, queue for cleanup */
3542 if (list_empty(&lower->upper))
3543 list_add(&lower->list, useless_node);
3544 continue;
3545 }
3546
3547 /*
3548 * All new nodes added in current build_backref_tree() haven't
3549 * been linked to the cache rb tree.
3550 * So if we have upper->rb_node populated, this means a cache
3551 * hit. We only need to link the edge, as @upper and all its
3552 * parents have already been linked.
3553 */
3554 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3555 if (upper->lowest) {
3556 list_del_init(&upper->lower);
3557 upper->lowest = 0;
3558 }
3559
3560 list_add_tail(&edge->list[UPPER], &upper->lower);
3561 continue;
3562 }
3563
3564 /* Sanity check, we shouldn't have any unchecked nodes */
3565 if (!upper->checked) {
3566 ASSERT(0);
3567 return -EUCLEAN;
3568 }
3569
3570 /* Sanity check, COW-only node has non-COW-only parent */
3571 if (start->cowonly != upper->cowonly) {
3572 ASSERT(0);
3573 return -EUCLEAN;
3574 }
3575
3576 /* Only cache non-COW-only (subvolume trees) tree blocks */
3577 if (!upper->cowonly) {
3578 rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3579 &upper->rb_node);
3580 if (rb_node) {
3581 btrfs_backref_panic(cache->fs_info,
3582 upper->bytenr, -EEXIST);
3583 return -EUCLEAN;
3584 }
3585 }
3586
3587 list_add_tail(&edge->list[UPPER], &upper->lower);
3588
3589 /*
3590 * Also queue all the parent edges of this uncached node
3591 * to finish the upper linkage
3592 */
3593 list_for_each_entry(edge, &upper->upper, list[LOWER])
3594 list_add_tail(&edge->list[UPPER], &pending_edge);
3595 }
3596 return 0;
3597}
3598
3599void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3600 struct btrfs_backref_node *node)
3601{
3602 struct btrfs_backref_node *lower;
3603 struct btrfs_backref_node *upper;
3604 struct btrfs_backref_edge *edge;
3605
3606 while (!list_empty(&cache->useless_node)) {
3607 lower = list_first_entry(&cache->useless_node,
3608 struct btrfs_backref_node, list);
3609 list_del_init(&lower->list);
3610 }
3611 while (!list_empty(&cache->pending_edge)) {
3612 edge = list_first_entry(&cache->pending_edge,
3613 struct btrfs_backref_edge, list[UPPER]);
3614 list_del(&edge->list[UPPER]);
3615 list_del(&edge->list[LOWER]);
3616 lower = edge->node[LOWER];
3617 upper = edge->node[UPPER];
3618 btrfs_backref_free_edge(cache, edge);
3619
3620 /*
3621 * Lower is no longer linked to any upper backref nodes and
3622 * isn't in the cache, we can free it ourselves.
3623 */
3624 if (list_empty(&lower->upper) &&
3625 RB_EMPTY_NODE(&lower->rb_node))
3626 list_add(&lower->list, &cache->useless_node);
3627
3628 if (!RB_EMPTY_NODE(&upper->rb_node))
3629 continue;
3630
3631 /* Add this guy's upper edges to the list to process */
3632 list_for_each_entry(edge, &upper->upper, list[LOWER])
3633 list_add_tail(&edge->list[UPPER],
3634 &cache->pending_edge);
3635 if (list_empty(&upper->upper))
3636 list_add(&upper->list, &cache->useless_node);
3637 }
3638
3639 while (!list_empty(&cache->useless_node)) {
3640 lower = list_first_entry(&cache->useless_node,
3641 struct btrfs_backref_node, list);
3642 list_del_init(&lower->list);
3643 if (lower == node)
3644 node = NULL;
3645 btrfs_backref_drop_node(cache, lower);
3646 }
3647
3648 btrfs_backref_cleanup_node(cache, node);
3649 ASSERT(list_empty(&cache->useless_node) &&
3650 list_empty(&cache->pending_edge));
3651}
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#include "fs.h"
19#include "accessors.h"
20#include "extent-tree.h"
21#include "relocation.h"
22#include "tree-checker.h"
23
24/* Just arbitrary numbers so we can be sure one of these happened. */
25#define BACKREF_FOUND_SHARED 6
26#define BACKREF_FOUND_NOT_SHARED 7
27
28struct extent_inode_elem {
29 u64 inum;
30 u64 offset;
31 u64 num_bytes;
32 struct extent_inode_elem *next;
33};
34
35static int check_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
36 const struct btrfs_key *key,
37 const struct extent_buffer *eb,
38 const struct btrfs_file_extent_item *fi,
39 struct extent_inode_elem **eie)
40{
41 const u64 data_len = btrfs_file_extent_num_bytes(eb, fi);
42 u64 offset = key->offset;
43 struct extent_inode_elem *e;
44 const u64 *root_ids;
45 int root_count;
46 bool cached;
47
48 if (!ctx->ignore_extent_item_pos &&
49 !btrfs_file_extent_compression(eb, fi) &&
50 !btrfs_file_extent_encryption(eb, fi) &&
51 !btrfs_file_extent_other_encoding(eb, fi)) {
52 u64 data_offset;
53
54 data_offset = btrfs_file_extent_offset(eb, fi);
55
56 if (ctx->extent_item_pos < data_offset ||
57 ctx->extent_item_pos >= data_offset + data_len)
58 return 1;
59 offset += ctx->extent_item_pos - data_offset;
60 }
61
62 if (!ctx->indirect_ref_iterator || !ctx->cache_lookup)
63 goto add_inode_elem;
64
65 cached = ctx->cache_lookup(eb->start, ctx->user_ctx, &root_ids,
66 &root_count);
67 if (!cached)
68 goto add_inode_elem;
69
70 for (int i = 0; i < root_count; i++) {
71 int ret;
72
73 ret = ctx->indirect_ref_iterator(key->objectid, offset,
74 data_len, root_ids[i],
75 ctx->user_ctx);
76 if (ret)
77 return ret;
78 }
79
80add_inode_elem:
81 e = kmalloc(sizeof(*e), GFP_NOFS);
82 if (!e)
83 return -ENOMEM;
84
85 e->next = *eie;
86 e->inum = key->objectid;
87 e->offset = offset;
88 e->num_bytes = data_len;
89 *eie = e;
90
91 return 0;
92}
93
94static void free_inode_elem_list(struct extent_inode_elem *eie)
95{
96 struct extent_inode_elem *eie_next;
97
98 for (; eie; eie = eie_next) {
99 eie_next = eie->next;
100 kfree(eie);
101 }
102}
103
104static int find_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
105 const struct extent_buffer *eb,
106 struct extent_inode_elem **eie)
107{
108 u64 disk_byte;
109 struct btrfs_key key;
110 struct btrfs_file_extent_item *fi;
111 int slot;
112 int nritems;
113 int extent_type;
114 int ret;
115
116 /*
117 * from the shared data ref, we only have the leaf but we need
118 * the key. thus, we must look into all items and see that we
119 * find one (some) with a reference to our extent item.
120 */
121 nritems = btrfs_header_nritems(eb);
122 for (slot = 0; slot < nritems; ++slot) {
123 btrfs_item_key_to_cpu(eb, &key, slot);
124 if (key.type != BTRFS_EXTENT_DATA_KEY)
125 continue;
126 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
127 extent_type = btrfs_file_extent_type(eb, fi);
128 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
129 continue;
130 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
131 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
132 if (disk_byte != ctx->bytenr)
133 continue;
134
135 ret = check_extent_in_eb(ctx, &key, eb, fi, eie);
136 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
137 return ret;
138 }
139
140 return 0;
141}
142
143struct preftree {
144 struct rb_root_cached root;
145 unsigned int count;
146};
147
148#define PREFTREE_INIT { .root = RB_ROOT_CACHED, .count = 0 }
149
150struct preftrees {
151 struct preftree direct; /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
152 struct preftree indirect; /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
153 struct preftree indirect_missing_keys;
154};
155
156/*
157 * Checks for a shared extent during backref search.
158 *
159 * The share_count tracks prelim_refs (direct and indirect) having a
160 * ref->count >0:
161 * - incremented when a ref->count transitions to >0
162 * - decremented when a ref->count transitions to <1
163 */
164struct share_check {
165 struct btrfs_backref_share_check_ctx *ctx;
166 struct btrfs_root *root;
167 u64 inum;
168 u64 data_bytenr;
169 u64 data_extent_gen;
170 /*
171 * Counts number of inodes that refer to an extent (different inodes in
172 * the same root or different roots) that we could find. The sharedness
173 * check typically stops once this counter gets greater than 1, so it
174 * may not reflect the total number of inodes.
175 */
176 int share_count;
177 /*
178 * The number of times we found our inode refers to the data extent we
179 * are determining the sharedness. In other words, how many file extent
180 * items we could find for our inode that point to our target data
181 * extent. The value we get here after finishing the extent sharedness
182 * check may be smaller than reality, but if it ends up being greater
183 * than 1, then we know for sure the inode has multiple file extent
184 * items that point to our inode, and we can safely assume it's useful
185 * to cache the sharedness check result.
186 */
187 int self_ref_count;
188 bool have_delayed_delete_refs;
189};
190
191static inline int extent_is_shared(struct share_check *sc)
192{
193 return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
194}
195
196static struct kmem_cache *btrfs_prelim_ref_cache;
197
198int __init btrfs_prelim_ref_init(void)
199{
200 btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
201 sizeof(struct prelim_ref), 0, 0, NULL);
202 if (!btrfs_prelim_ref_cache)
203 return -ENOMEM;
204 return 0;
205}
206
207void __cold btrfs_prelim_ref_exit(void)
208{
209 kmem_cache_destroy(btrfs_prelim_ref_cache);
210}
211
212static void free_pref(struct prelim_ref *ref)
213{
214 kmem_cache_free(btrfs_prelim_ref_cache, ref);
215}
216
217/*
218 * Return 0 when both refs are for the same block (and can be merged).
219 * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
220 * indicates a 'higher' block.
221 */
222static int prelim_ref_compare(const struct prelim_ref *ref1,
223 const struct prelim_ref *ref2)
224{
225 if (ref1->level < ref2->level)
226 return -1;
227 if (ref1->level > ref2->level)
228 return 1;
229 if (ref1->root_id < ref2->root_id)
230 return -1;
231 if (ref1->root_id > ref2->root_id)
232 return 1;
233 if (ref1->key_for_search.type < ref2->key_for_search.type)
234 return -1;
235 if (ref1->key_for_search.type > ref2->key_for_search.type)
236 return 1;
237 if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
238 return -1;
239 if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
240 return 1;
241 if (ref1->key_for_search.offset < ref2->key_for_search.offset)
242 return -1;
243 if (ref1->key_for_search.offset > ref2->key_for_search.offset)
244 return 1;
245 if (ref1->parent < ref2->parent)
246 return -1;
247 if (ref1->parent > ref2->parent)
248 return 1;
249
250 return 0;
251}
252
253static void update_share_count(struct share_check *sc, int oldcount,
254 int newcount, const struct prelim_ref *newref)
255{
256 if ((!sc) || (oldcount == 0 && newcount < 1))
257 return;
258
259 if (oldcount > 0 && newcount < 1)
260 sc->share_count--;
261 else if (oldcount < 1 && newcount > 0)
262 sc->share_count++;
263
264 if (newref->root_id == btrfs_root_id(sc->root) &&
265 newref->wanted_disk_byte == sc->data_bytenr &&
266 newref->key_for_search.objectid == sc->inum)
267 sc->self_ref_count += newref->count;
268}
269
270/*
271 * Add @newref to the @root rbtree, merging identical refs.
272 *
273 * Callers should assume that newref has been freed after calling.
274 */
275static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
276 struct preftree *preftree,
277 struct prelim_ref *newref,
278 struct share_check *sc)
279{
280 struct rb_root_cached *root;
281 struct rb_node **p;
282 struct rb_node *parent = NULL;
283 struct prelim_ref *ref;
284 int result;
285 bool leftmost = true;
286
287 root = &preftree->root;
288 p = &root->rb_root.rb_node;
289
290 while (*p) {
291 parent = *p;
292 ref = rb_entry(parent, struct prelim_ref, rbnode);
293 result = prelim_ref_compare(ref, newref);
294 if (result < 0) {
295 p = &(*p)->rb_left;
296 } else if (result > 0) {
297 p = &(*p)->rb_right;
298 leftmost = false;
299 } else {
300 /* Identical refs, merge them and free @newref */
301 struct extent_inode_elem *eie = ref->inode_list;
302
303 while (eie && eie->next)
304 eie = eie->next;
305
306 if (!eie)
307 ref->inode_list = newref->inode_list;
308 else
309 eie->next = newref->inode_list;
310 trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
311 preftree->count);
312 /*
313 * A delayed ref can have newref->count < 0.
314 * The ref->count is updated to follow any
315 * BTRFS_[ADD|DROP]_DELAYED_REF actions.
316 */
317 update_share_count(sc, ref->count,
318 ref->count + newref->count, newref);
319 ref->count += newref->count;
320 free_pref(newref);
321 return;
322 }
323 }
324
325 update_share_count(sc, 0, newref->count, newref);
326 preftree->count++;
327 trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
328 rb_link_node(&newref->rbnode, parent, p);
329 rb_insert_color_cached(&newref->rbnode, root, leftmost);
330}
331
332/*
333 * Release the entire tree. We don't care about internal consistency so
334 * just free everything and then reset the tree root.
335 */
336static void prelim_release(struct preftree *preftree)
337{
338 struct prelim_ref *ref, *next_ref;
339
340 rbtree_postorder_for_each_entry_safe(ref, next_ref,
341 &preftree->root.rb_root, rbnode) {
342 free_inode_elem_list(ref->inode_list);
343 free_pref(ref);
344 }
345
346 preftree->root = RB_ROOT_CACHED;
347 preftree->count = 0;
348}
349
350/*
351 * the rules for all callers of this function are:
352 * - obtaining the parent is the goal
353 * - if you add a key, you must know that it is a correct key
354 * - if you cannot add the parent or a correct key, then we will look into the
355 * block later to set a correct key
356 *
357 * delayed refs
358 * ============
359 * backref type | shared | indirect | shared | indirect
360 * information | tree | tree | data | data
361 * --------------------+--------+----------+--------+----------
362 * parent logical | y | - | - | -
363 * key to resolve | - | y | y | y
364 * tree block logical | - | - | - | -
365 * root for resolving | y | y | y | y
366 *
367 * - column 1: we've the parent -> done
368 * - column 2, 3, 4: we use the key to find the parent
369 *
370 * on disk refs (inline or keyed)
371 * ==============================
372 * backref type | shared | indirect | shared | indirect
373 * information | tree | tree | data | data
374 * --------------------+--------+----------+--------+----------
375 * parent logical | y | - | y | -
376 * key to resolve | - | - | - | y
377 * tree block logical | y | y | y | y
378 * root for resolving | - | y | y | y
379 *
380 * - column 1, 3: we've the parent -> done
381 * - column 2: we take the first key from the block to find the parent
382 * (see add_missing_keys)
383 * - column 4: we use the key to find the parent
384 *
385 * additional information that's available but not required to find the parent
386 * block might help in merging entries to gain some speed.
387 */
388static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
389 struct preftree *preftree, u64 root_id,
390 const struct btrfs_key *key, int level, u64 parent,
391 u64 wanted_disk_byte, int count,
392 struct share_check *sc, gfp_t gfp_mask)
393{
394 struct prelim_ref *ref;
395
396 if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
397 return 0;
398
399 ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
400 if (!ref)
401 return -ENOMEM;
402
403 ref->root_id = root_id;
404 if (key)
405 ref->key_for_search = *key;
406 else
407 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
408
409 ref->inode_list = NULL;
410 ref->level = level;
411 ref->count = count;
412 ref->parent = parent;
413 ref->wanted_disk_byte = wanted_disk_byte;
414 prelim_ref_insert(fs_info, preftree, ref, sc);
415 return extent_is_shared(sc);
416}
417
418/* direct refs use root == 0, key == NULL */
419static int add_direct_ref(const struct btrfs_fs_info *fs_info,
420 struct preftrees *preftrees, int level, u64 parent,
421 u64 wanted_disk_byte, int count,
422 struct share_check *sc, gfp_t gfp_mask)
423{
424 return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
425 parent, wanted_disk_byte, count, sc, gfp_mask);
426}
427
428/* indirect refs use parent == 0 */
429static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
430 struct preftrees *preftrees, u64 root_id,
431 const struct btrfs_key *key, int level,
432 u64 wanted_disk_byte, int count,
433 struct share_check *sc, gfp_t gfp_mask)
434{
435 struct preftree *tree = &preftrees->indirect;
436
437 if (!key)
438 tree = &preftrees->indirect_missing_keys;
439 return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
440 wanted_disk_byte, count, sc, gfp_mask);
441}
442
443static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
444{
445 struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
446 struct rb_node *parent = NULL;
447 struct prelim_ref *ref = NULL;
448 struct prelim_ref target = {};
449 int result;
450
451 target.parent = bytenr;
452
453 while (*p) {
454 parent = *p;
455 ref = rb_entry(parent, struct prelim_ref, rbnode);
456 result = prelim_ref_compare(ref, &target);
457
458 if (result < 0)
459 p = &(*p)->rb_left;
460 else if (result > 0)
461 p = &(*p)->rb_right;
462 else
463 return 1;
464 }
465 return 0;
466}
467
468static int add_all_parents(struct btrfs_backref_walk_ctx *ctx,
469 struct btrfs_root *root, struct btrfs_path *path,
470 struct ulist *parents,
471 struct preftrees *preftrees, struct prelim_ref *ref,
472 int level)
473{
474 int ret = 0;
475 int slot;
476 struct extent_buffer *eb;
477 struct btrfs_key key;
478 struct btrfs_key *key_for_search = &ref->key_for_search;
479 struct btrfs_file_extent_item *fi;
480 struct extent_inode_elem *eie = NULL, *old = NULL;
481 u64 disk_byte;
482 u64 wanted_disk_byte = ref->wanted_disk_byte;
483 u64 count = 0;
484 u64 data_offset;
485 u8 type;
486
487 if (level != 0) {
488 eb = path->nodes[level];
489 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
490 if (ret < 0)
491 return ret;
492 return 0;
493 }
494
495 /*
496 * 1. We normally enter this function with the path already pointing to
497 * the first item to check. But sometimes, we may enter it with
498 * slot == nritems.
499 * 2. We are searching for normal backref but bytenr of this leaf
500 * matches shared data backref
501 * 3. The leaf owner is not equal to the root we are searching
502 *
503 * For these cases, go to the next leaf before we continue.
504 */
505 eb = path->nodes[0];
506 if (path->slots[0] >= btrfs_header_nritems(eb) ||
507 is_shared_data_backref(preftrees, eb->start) ||
508 ref->root_id != btrfs_header_owner(eb)) {
509 if (ctx->time_seq == BTRFS_SEQ_LAST)
510 ret = btrfs_next_leaf(root, path);
511 else
512 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
513 }
514
515 while (!ret && count < ref->count) {
516 eb = path->nodes[0];
517 slot = path->slots[0];
518
519 btrfs_item_key_to_cpu(eb, &key, slot);
520
521 if (key.objectid != key_for_search->objectid ||
522 key.type != BTRFS_EXTENT_DATA_KEY)
523 break;
524
525 /*
526 * We are searching for normal backref but bytenr of this leaf
527 * matches shared data backref, OR
528 * the leaf owner is not equal to the root we are searching for
529 */
530 if (slot == 0 &&
531 (is_shared_data_backref(preftrees, eb->start) ||
532 ref->root_id != btrfs_header_owner(eb))) {
533 if (ctx->time_seq == BTRFS_SEQ_LAST)
534 ret = btrfs_next_leaf(root, path);
535 else
536 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
537 continue;
538 }
539 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
540 type = btrfs_file_extent_type(eb, fi);
541 if (type == BTRFS_FILE_EXTENT_INLINE)
542 goto next;
543 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
544 data_offset = btrfs_file_extent_offset(eb, fi);
545
546 if (disk_byte == wanted_disk_byte) {
547 eie = NULL;
548 old = NULL;
549 if (ref->key_for_search.offset == key.offset - data_offset)
550 count++;
551 else
552 goto next;
553 if (!ctx->skip_inode_ref_list) {
554 ret = check_extent_in_eb(ctx, &key, eb, fi, &eie);
555 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
556 ret < 0)
557 break;
558 }
559 if (ret > 0)
560 goto next;
561 ret = ulist_add_merge_ptr(parents, eb->start,
562 eie, (void **)&old, GFP_NOFS);
563 if (ret < 0)
564 break;
565 if (!ret && !ctx->skip_inode_ref_list) {
566 while (old->next)
567 old = old->next;
568 old->next = eie;
569 }
570 eie = NULL;
571 }
572next:
573 if (ctx->time_seq == BTRFS_SEQ_LAST)
574 ret = btrfs_next_item(root, path);
575 else
576 ret = btrfs_next_old_item(root, path, ctx->time_seq);
577 }
578
579 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
580 free_inode_elem_list(eie);
581 else if (ret > 0)
582 ret = 0;
583
584 return ret;
585}
586
587/*
588 * resolve an indirect backref in the form (root_id, key, level)
589 * to a logical address
590 */
591static int resolve_indirect_ref(struct btrfs_backref_walk_ctx *ctx,
592 struct btrfs_path *path,
593 struct preftrees *preftrees,
594 struct prelim_ref *ref, struct ulist *parents)
595{
596 struct btrfs_root *root;
597 struct extent_buffer *eb;
598 int ret = 0;
599 int root_level;
600 int level = ref->level;
601 struct btrfs_key search_key = ref->key_for_search;
602
603 /*
604 * If we're search_commit_root we could possibly be holding locks on
605 * other tree nodes. This happens when qgroups does backref walks when
606 * adding new delayed refs. To deal with this we need to look in cache
607 * for the root, and if we don't find it then we need to search the
608 * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
609 * here.
610 */
611 if (path->search_commit_root)
612 root = btrfs_get_fs_root_commit_root(ctx->fs_info, path, ref->root_id);
613 else
614 root = btrfs_get_fs_root(ctx->fs_info, ref->root_id, false);
615 if (IS_ERR(root)) {
616 ret = PTR_ERR(root);
617 goto out_free;
618 }
619
620 if (!path->search_commit_root &&
621 test_bit(BTRFS_ROOT_DELETING, &root->state)) {
622 ret = -ENOENT;
623 goto out;
624 }
625
626 if (btrfs_is_testing(ctx->fs_info)) {
627 ret = -ENOENT;
628 goto out;
629 }
630
631 if (path->search_commit_root)
632 root_level = btrfs_header_level(root->commit_root);
633 else if (ctx->time_seq == BTRFS_SEQ_LAST)
634 root_level = btrfs_header_level(root->node);
635 else
636 root_level = btrfs_old_root_level(root, ctx->time_seq);
637
638 if (root_level + 1 == level)
639 goto out;
640
641 /*
642 * We can often find data backrefs with an offset that is too large
643 * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
644 * subtracting a file's offset with the data offset of its
645 * corresponding extent data item. This can happen for example in the
646 * clone ioctl.
647 *
648 * So if we detect such case we set the search key's offset to zero to
649 * make sure we will find the matching file extent item at
650 * add_all_parents(), otherwise we will miss it because the offset
651 * taken form the backref is much larger then the offset of the file
652 * extent item. This can make us scan a very large number of file
653 * extent items, but at least it will not make us miss any.
654 *
655 * This is an ugly workaround for a behaviour that should have never
656 * existed, but it does and a fix for the clone ioctl would touch a lot
657 * of places, cause backwards incompatibility and would not fix the
658 * problem for extents cloned with older kernels.
659 */
660 if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
661 search_key.offset >= LLONG_MAX)
662 search_key.offset = 0;
663 path->lowest_level = level;
664 if (ctx->time_seq == BTRFS_SEQ_LAST)
665 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
666 else
667 ret = btrfs_search_old_slot(root, &search_key, path, ctx->time_seq);
668
669 btrfs_debug(ctx->fs_info,
670 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
671 ref->root_id, level, ref->count, ret,
672 ref->key_for_search.objectid, ref->key_for_search.type,
673 ref->key_for_search.offset);
674 if (ret < 0)
675 goto out;
676
677 eb = path->nodes[level];
678 while (!eb) {
679 if (WARN_ON(!level)) {
680 ret = 1;
681 goto out;
682 }
683 level--;
684 eb = path->nodes[level];
685 }
686
687 ret = add_all_parents(ctx, root, path, parents, preftrees, ref, level);
688out:
689 btrfs_put_root(root);
690out_free:
691 path->lowest_level = 0;
692 btrfs_release_path(path);
693 return ret;
694}
695
696static struct extent_inode_elem *
697unode_aux_to_inode_list(struct ulist_node *node)
698{
699 if (!node)
700 return NULL;
701 return (struct extent_inode_elem *)(uintptr_t)node->aux;
702}
703
704static void free_leaf_list(struct ulist *ulist)
705{
706 struct ulist_node *node;
707 struct ulist_iterator uiter;
708
709 ULIST_ITER_INIT(&uiter);
710 while ((node = ulist_next(ulist, &uiter)))
711 free_inode_elem_list(unode_aux_to_inode_list(node));
712
713 ulist_free(ulist);
714}
715
716/*
717 * We maintain three separate rbtrees: one for direct refs, one for
718 * indirect refs which have a key, and one for indirect refs which do not
719 * have a key. Each tree does merge on insertion.
720 *
721 * Once all of the references are located, we iterate over the tree of
722 * indirect refs with missing keys. An appropriate key is located and
723 * the ref is moved onto the tree for indirect refs. After all missing
724 * keys are thus located, we iterate over the indirect ref tree, resolve
725 * each reference, and then insert the resolved reference onto the
726 * direct tree (merging there too).
727 *
728 * New backrefs (i.e., for parent nodes) are added to the appropriate
729 * rbtree as they are encountered. The new backrefs are subsequently
730 * resolved as above.
731 */
732static int resolve_indirect_refs(struct btrfs_backref_walk_ctx *ctx,
733 struct btrfs_path *path,
734 struct preftrees *preftrees,
735 struct share_check *sc)
736{
737 int err;
738 int ret = 0;
739 struct ulist *parents;
740 struct ulist_node *node;
741 struct ulist_iterator uiter;
742 struct rb_node *rnode;
743
744 parents = ulist_alloc(GFP_NOFS);
745 if (!parents)
746 return -ENOMEM;
747
748 /*
749 * We could trade memory usage for performance here by iterating
750 * the tree, allocating new refs for each insertion, and then
751 * freeing the entire indirect tree when we're done. In some test
752 * cases, the tree can grow quite large (~200k objects).
753 */
754 while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
755 struct prelim_ref *ref;
756
757 ref = rb_entry(rnode, struct prelim_ref, rbnode);
758 if (WARN(ref->parent,
759 "BUG: direct ref found in indirect tree")) {
760 ret = -EINVAL;
761 goto out;
762 }
763
764 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
765 preftrees->indirect.count--;
766
767 if (ref->count == 0) {
768 free_pref(ref);
769 continue;
770 }
771
772 if (sc && ref->root_id != btrfs_root_id(sc->root)) {
773 free_pref(ref);
774 ret = BACKREF_FOUND_SHARED;
775 goto out;
776 }
777 err = resolve_indirect_ref(ctx, path, preftrees, ref, parents);
778 /*
779 * we can only tolerate ENOENT,otherwise,we should catch error
780 * and return directly.
781 */
782 if (err == -ENOENT) {
783 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref,
784 NULL);
785 continue;
786 } else if (err) {
787 free_pref(ref);
788 ret = err;
789 goto out;
790 }
791
792 /* we put the first parent into the ref at hand */
793 ULIST_ITER_INIT(&uiter);
794 node = ulist_next(parents, &uiter);
795 ref->parent = node ? node->val : 0;
796 ref->inode_list = unode_aux_to_inode_list(node);
797
798 /* Add a prelim_ref(s) for any other parent(s). */
799 while ((node = ulist_next(parents, &uiter))) {
800 struct prelim_ref *new_ref;
801
802 new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
803 GFP_NOFS);
804 if (!new_ref) {
805 free_pref(ref);
806 ret = -ENOMEM;
807 goto out;
808 }
809 memcpy(new_ref, ref, sizeof(*ref));
810 new_ref->parent = node->val;
811 new_ref->inode_list = unode_aux_to_inode_list(node);
812 prelim_ref_insert(ctx->fs_info, &preftrees->direct,
813 new_ref, NULL);
814 }
815
816 /*
817 * Now it's a direct ref, put it in the direct tree. We must
818 * do this last because the ref could be merged/freed here.
819 */
820 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref, NULL);
821
822 ulist_reinit(parents);
823 cond_resched();
824 }
825out:
826 /*
827 * We may have inode lists attached to refs in the parents ulist, so we
828 * must free them before freeing the ulist and its refs.
829 */
830 free_leaf_list(parents);
831 return ret;
832}
833
834/*
835 * read tree blocks and add keys where required.
836 */
837static int add_missing_keys(struct btrfs_fs_info *fs_info,
838 struct preftrees *preftrees, bool lock)
839{
840 struct prelim_ref *ref;
841 struct extent_buffer *eb;
842 struct preftree *tree = &preftrees->indirect_missing_keys;
843 struct rb_node *node;
844
845 while ((node = rb_first_cached(&tree->root))) {
846 struct btrfs_tree_parent_check check = { 0 };
847
848 ref = rb_entry(node, struct prelim_ref, rbnode);
849 rb_erase_cached(node, &tree->root);
850
851 BUG_ON(ref->parent); /* should not be a direct ref */
852 BUG_ON(ref->key_for_search.type);
853 BUG_ON(!ref->wanted_disk_byte);
854
855 check.level = ref->level - 1;
856 check.owner_root = ref->root_id;
857
858 eb = read_tree_block(fs_info, ref->wanted_disk_byte, &check);
859 if (IS_ERR(eb)) {
860 free_pref(ref);
861 return PTR_ERR(eb);
862 }
863 if (!extent_buffer_uptodate(eb)) {
864 free_pref(ref);
865 free_extent_buffer(eb);
866 return -EIO;
867 }
868
869 if (lock)
870 btrfs_tree_read_lock(eb);
871 if (btrfs_header_level(eb) == 0)
872 btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
873 else
874 btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
875 if (lock)
876 btrfs_tree_read_unlock(eb);
877 free_extent_buffer(eb);
878 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
879 cond_resched();
880 }
881 return 0;
882}
883
884/*
885 * add all currently queued delayed refs from this head whose seq nr is
886 * smaller or equal that seq to the list
887 */
888static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
889 struct btrfs_delayed_ref_head *head, u64 seq,
890 struct preftrees *preftrees, struct share_check *sc)
891{
892 struct btrfs_delayed_ref_node *node;
893 struct btrfs_key key;
894 struct rb_node *n;
895 int count;
896 int ret = 0;
897
898 spin_lock(&head->lock);
899 for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
900 node = rb_entry(n, struct btrfs_delayed_ref_node,
901 ref_node);
902 if (node->seq > seq)
903 continue;
904
905 switch (node->action) {
906 case BTRFS_ADD_DELAYED_EXTENT:
907 case BTRFS_UPDATE_DELAYED_HEAD:
908 WARN_ON(1);
909 continue;
910 case BTRFS_ADD_DELAYED_REF:
911 count = node->ref_mod;
912 break;
913 case BTRFS_DROP_DELAYED_REF:
914 count = node->ref_mod * -1;
915 break;
916 default:
917 BUG();
918 }
919 switch (node->type) {
920 case BTRFS_TREE_BLOCK_REF_KEY: {
921 /* NORMAL INDIRECT METADATA backref */
922 struct btrfs_key *key_ptr = NULL;
923 /* The owner of a tree block ref is the level. */
924 int level = btrfs_delayed_ref_owner(node);
925
926 if (head->extent_op && head->extent_op->update_key) {
927 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
928 key_ptr = &key;
929 }
930
931 ret = add_indirect_ref(fs_info, preftrees, node->ref_root,
932 key_ptr, level + 1, node->bytenr,
933 count, sc, GFP_ATOMIC);
934 break;
935 }
936 case BTRFS_SHARED_BLOCK_REF_KEY: {
937 /*
938 * SHARED DIRECT METADATA backref
939 *
940 * The owner of a tree block ref is the level.
941 */
942 int level = btrfs_delayed_ref_owner(node);
943
944 ret = add_direct_ref(fs_info, preftrees, level + 1,
945 node->parent, node->bytenr, count,
946 sc, GFP_ATOMIC);
947 break;
948 }
949 case BTRFS_EXTENT_DATA_REF_KEY: {
950 /* NORMAL INDIRECT DATA backref */
951 key.objectid = btrfs_delayed_ref_owner(node);
952 key.type = BTRFS_EXTENT_DATA_KEY;
953 key.offset = btrfs_delayed_ref_offset(node);
954
955 /*
956 * If we have a share check context and a reference for
957 * another inode, we can't exit immediately. This is
958 * because even if this is a BTRFS_ADD_DELAYED_REF
959 * reference we may find next a BTRFS_DROP_DELAYED_REF
960 * which cancels out this ADD reference.
961 *
962 * If this is a DROP reference and there was no previous
963 * ADD reference, then we need to signal that when we
964 * process references from the extent tree (through
965 * add_inline_refs() and add_keyed_refs()), we should
966 * not exit early if we find a reference for another
967 * inode, because one of the delayed DROP references
968 * may cancel that reference in the extent tree.
969 */
970 if (sc && count < 0)
971 sc->have_delayed_delete_refs = true;
972
973 ret = add_indirect_ref(fs_info, preftrees, node->ref_root,
974 &key, 0, node->bytenr, count, sc,
975 GFP_ATOMIC);
976 break;
977 }
978 case BTRFS_SHARED_DATA_REF_KEY: {
979 /* SHARED DIRECT FULL backref */
980 ret = add_direct_ref(fs_info, preftrees, 0, node->parent,
981 node->bytenr, count, sc,
982 GFP_ATOMIC);
983 break;
984 }
985 default:
986 WARN_ON(1);
987 }
988 /*
989 * We must ignore BACKREF_FOUND_SHARED until all delayed
990 * refs have been checked.
991 */
992 if (ret && (ret != BACKREF_FOUND_SHARED))
993 break;
994 }
995 if (!ret)
996 ret = extent_is_shared(sc);
997
998 spin_unlock(&head->lock);
999 return ret;
1000}
1001
1002/*
1003 * add all inline backrefs for bytenr to the list
1004 *
1005 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1006 */
1007static int add_inline_refs(struct btrfs_backref_walk_ctx *ctx,
1008 struct btrfs_path *path,
1009 int *info_level, struct preftrees *preftrees,
1010 struct share_check *sc)
1011{
1012 int ret = 0;
1013 int slot;
1014 struct extent_buffer *leaf;
1015 struct btrfs_key key;
1016 struct btrfs_key found_key;
1017 unsigned long ptr;
1018 unsigned long end;
1019 struct btrfs_extent_item *ei;
1020 u64 flags;
1021 u64 item_size;
1022
1023 /*
1024 * enumerate all inline refs
1025 */
1026 leaf = path->nodes[0];
1027 slot = path->slots[0];
1028
1029 item_size = btrfs_item_size(leaf, slot);
1030 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
1031
1032 if (ctx->check_extent_item) {
1033 ret = ctx->check_extent_item(ctx->bytenr, ei, leaf, ctx->user_ctx);
1034 if (ret)
1035 return ret;
1036 }
1037
1038 flags = btrfs_extent_flags(leaf, ei);
1039 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1040
1041 ptr = (unsigned long)(ei + 1);
1042 end = (unsigned long)ei + item_size;
1043
1044 if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
1045 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1046 struct btrfs_tree_block_info *info;
1047
1048 info = (struct btrfs_tree_block_info *)ptr;
1049 *info_level = btrfs_tree_block_level(leaf, info);
1050 ptr += sizeof(struct btrfs_tree_block_info);
1051 BUG_ON(ptr > end);
1052 } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1053 *info_level = found_key.offset;
1054 } else {
1055 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1056 }
1057
1058 while (ptr < end) {
1059 struct btrfs_extent_inline_ref *iref;
1060 u64 offset;
1061 int type;
1062
1063 iref = (struct btrfs_extent_inline_ref *)ptr;
1064 type = btrfs_get_extent_inline_ref_type(leaf, iref,
1065 BTRFS_REF_TYPE_ANY);
1066 if (type == BTRFS_REF_TYPE_INVALID)
1067 return -EUCLEAN;
1068
1069 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1070
1071 switch (type) {
1072 case BTRFS_SHARED_BLOCK_REF_KEY:
1073 ret = add_direct_ref(ctx->fs_info, preftrees,
1074 *info_level + 1, offset,
1075 ctx->bytenr, 1, NULL, GFP_NOFS);
1076 break;
1077 case BTRFS_SHARED_DATA_REF_KEY: {
1078 struct btrfs_shared_data_ref *sdref;
1079 int count;
1080
1081 sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1082 count = btrfs_shared_data_ref_count(leaf, sdref);
1083
1084 ret = add_direct_ref(ctx->fs_info, preftrees, 0, offset,
1085 ctx->bytenr, count, sc, GFP_NOFS);
1086 break;
1087 }
1088 case BTRFS_TREE_BLOCK_REF_KEY:
1089 ret = add_indirect_ref(ctx->fs_info, preftrees, offset,
1090 NULL, *info_level + 1,
1091 ctx->bytenr, 1, NULL, GFP_NOFS);
1092 break;
1093 case BTRFS_EXTENT_DATA_REF_KEY: {
1094 struct btrfs_extent_data_ref *dref;
1095 int count;
1096 u64 root;
1097
1098 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1099 count = btrfs_extent_data_ref_count(leaf, dref);
1100 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1101 dref);
1102 key.type = BTRFS_EXTENT_DATA_KEY;
1103 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1104
1105 if (sc && key.objectid != sc->inum &&
1106 !sc->have_delayed_delete_refs) {
1107 ret = BACKREF_FOUND_SHARED;
1108 break;
1109 }
1110
1111 root = btrfs_extent_data_ref_root(leaf, dref);
1112
1113 if (!ctx->skip_data_ref ||
1114 !ctx->skip_data_ref(root, key.objectid, key.offset,
1115 ctx->user_ctx))
1116 ret = add_indirect_ref(ctx->fs_info, preftrees,
1117 root, &key, 0, ctx->bytenr,
1118 count, sc, GFP_NOFS);
1119 break;
1120 }
1121 case BTRFS_EXTENT_OWNER_REF_KEY:
1122 ASSERT(btrfs_fs_incompat(ctx->fs_info, SIMPLE_QUOTA));
1123 break;
1124 default:
1125 WARN_ON(1);
1126 }
1127 if (ret)
1128 return ret;
1129 ptr += btrfs_extent_inline_ref_size(type);
1130 }
1131
1132 return 0;
1133}
1134
1135/*
1136 * add all non-inline backrefs for bytenr to the list
1137 *
1138 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1139 */
1140static int add_keyed_refs(struct btrfs_backref_walk_ctx *ctx,
1141 struct btrfs_root *extent_root,
1142 struct btrfs_path *path,
1143 int info_level, struct preftrees *preftrees,
1144 struct share_check *sc)
1145{
1146 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1147 int ret;
1148 int slot;
1149 struct extent_buffer *leaf;
1150 struct btrfs_key key;
1151
1152 while (1) {
1153 ret = btrfs_next_item(extent_root, path);
1154 if (ret < 0)
1155 break;
1156 if (ret) {
1157 ret = 0;
1158 break;
1159 }
1160
1161 slot = path->slots[0];
1162 leaf = path->nodes[0];
1163 btrfs_item_key_to_cpu(leaf, &key, slot);
1164
1165 if (key.objectid != ctx->bytenr)
1166 break;
1167 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1168 continue;
1169 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1170 break;
1171
1172 switch (key.type) {
1173 case BTRFS_SHARED_BLOCK_REF_KEY:
1174 /* SHARED DIRECT METADATA backref */
1175 ret = add_direct_ref(fs_info, preftrees,
1176 info_level + 1, key.offset,
1177 ctx->bytenr, 1, NULL, GFP_NOFS);
1178 break;
1179 case BTRFS_SHARED_DATA_REF_KEY: {
1180 /* SHARED DIRECT FULL backref */
1181 struct btrfs_shared_data_ref *sdref;
1182 int count;
1183
1184 sdref = btrfs_item_ptr(leaf, slot,
1185 struct btrfs_shared_data_ref);
1186 count = btrfs_shared_data_ref_count(leaf, sdref);
1187 ret = add_direct_ref(fs_info, preftrees, 0,
1188 key.offset, ctx->bytenr, count,
1189 sc, GFP_NOFS);
1190 break;
1191 }
1192 case BTRFS_TREE_BLOCK_REF_KEY:
1193 /* NORMAL INDIRECT METADATA backref */
1194 ret = add_indirect_ref(fs_info, preftrees, key.offset,
1195 NULL, info_level + 1, ctx->bytenr,
1196 1, NULL, GFP_NOFS);
1197 break;
1198 case BTRFS_EXTENT_DATA_REF_KEY: {
1199 /* NORMAL INDIRECT DATA backref */
1200 struct btrfs_extent_data_ref *dref;
1201 int count;
1202 u64 root;
1203
1204 dref = btrfs_item_ptr(leaf, slot,
1205 struct btrfs_extent_data_ref);
1206 count = btrfs_extent_data_ref_count(leaf, dref);
1207 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1208 dref);
1209 key.type = BTRFS_EXTENT_DATA_KEY;
1210 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1211
1212 if (sc && key.objectid != sc->inum &&
1213 !sc->have_delayed_delete_refs) {
1214 ret = BACKREF_FOUND_SHARED;
1215 break;
1216 }
1217
1218 root = btrfs_extent_data_ref_root(leaf, dref);
1219
1220 if (!ctx->skip_data_ref ||
1221 !ctx->skip_data_ref(root, key.objectid, key.offset,
1222 ctx->user_ctx))
1223 ret = add_indirect_ref(fs_info, preftrees, root,
1224 &key, 0, ctx->bytenr,
1225 count, sc, GFP_NOFS);
1226 break;
1227 }
1228 default:
1229 WARN_ON(1);
1230 }
1231 if (ret)
1232 return ret;
1233
1234 }
1235
1236 return ret;
1237}
1238
1239/*
1240 * The caller has joined a transaction or is holding a read lock on the
1241 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1242 * snapshot field changing while updating or checking the cache.
1243 */
1244static bool lookup_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1245 struct btrfs_root *root,
1246 u64 bytenr, int level, bool *is_shared)
1247{
1248 const struct btrfs_fs_info *fs_info = root->fs_info;
1249 struct btrfs_backref_shared_cache_entry *entry;
1250
1251 if (!current->journal_info)
1252 lockdep_assert_held(&fs_info->commit_root_sem);
1253
1254 if (!ctx->use_path_cache)
1255 return false;
1256
1257 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1258 return false;
1259
1260 /*
1261 * Level -1 is used for the data extent, which is not reliable to cache
1262 * because its reference count can increase or decrease without us
1263 * realizing. We cache results only for extent buffers that lead from
1264 * the root node down to the leaf with the file extent item.
1265 */
1266 ASSERT(level >= 0);
1267
1268 entry = &ctx->path_cache_entries[level];
1269
1270 /* Unused cache entry or being used for some other extent buffer. */
1271 if (entry->bytenr != bytenr)
1272 return false;
1273
1274 /*
1275 * We cached a false result, but the last snapshot generation of the
1276 * root changed, so we now have a snapshot. Don't trust the result.
1277 */
1278 if (!entry->is_shared &&
1279 entry->gen != btrfs_root_last_snapshot(&root->root_item))
1280 return false;
1281
1282 /*
1283 * If we cached a true result and the last generation used for dropping
1284 * a root changed, we can not trust the result, because the dropped root
1285 * could be a snapshot sharing this extent buffer.
1286 */
1287 if (entry->is_shared &&
1288 entry->gen != btrfs_get_last_root_drop_gen(fs_info))
1289 return false;
1290
1291 *is_shared = entry->is_shared;
1292 /*
1293 * If the node at this level is shared, than all nodes below are also
1294 * shared. Currently some of the nodes below may be marked as not shared
1295 * because we have just switched from one leaf to another, and switched
1296 * also other nodes above the leaf and below the current level, so mark
1297 * them as shared.
1298 */
1299 if (*is_shared) {
1300 for (int i = 0; i < level; i++) {
1301 ctx->path_cache_entries[i].is_shared = true;
1302 ctx->path_cache_entries[i].gen = entry->gen;
1303 }
1304 }
1305
1306 return true;
1307}
1308
1309/*
1310 * The caller has joined a transaction or is holding a read lock on the
1311 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1312 * snapshot field changing while updating or checking the cache.
1313 */
1314static void store_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1315 struct btrfs_root *root,
1316 u64 bytenr, int level, bool is_shared)
1317{
1318 const struct btrfs_fs_info *fs_info = root->fs_info;
1319 struct btrfs_backref_shared_cache_entry *entry;
1320 u64 gen;
1321
1322 if (!current->journal_info)
1323 lockdep_assert_held(&fs_info->commit_root_sem);
1324
1325 if (!ctx->use_path_cache)
1326 return;
1327
1328 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1329 return;
1330
1331 /*
1332 * Level -1 is used for the data extent, which is not reliable to cache
1333 * because its reference count can increase or decrease without us
1334 * realizing. We cache results only for extent buffers that lead from
1335 * the root node down to the leaf with the file extent item.
1336 */
1337 ASSERT(level >= 0);
1338
1339 if (is_shared)
1340 gen = btrfs_get_last_root_drop_gen(fs_info);
1341 else
1342 gen = btrfs_root_last_snapshot(&root->root_item);
1343
1344 entry = &ctx->path_cache_entries[level];
1345 entry->bytenr = bytenr;
1346 entry->is_shared = is_shared;
1347 entry->gen = gen;
1348
1349 /*
1350 * If we found an extent buffer is shared, set the cache result for all
1351 * extent buffers below it to true. As nodes in the path are COWed,
1352 * their sharedness is moved to their children, and if a leaf is COWed,
1353 * then the sharedness of a data extent becomes direct, the refcount of
1354 * data extent is increased in the extent item at the extent tree.
1355 */
1356 if (is_shared) {
1357 for (int i = 0; i < level; i++) {
1358 entry = &ctx->path_cache_entries[i];
1359 entry->is_shared = is_shared;
1360 entry->gen = gen;
1361 }
1362 }
1363}
1364
1365/*
1366 * this adds all existing backrefs (inline backrefs, backrefs and delayed
1367 * refs) for the given bytenr to the refs list, merges duplicates and resolves
1368 * indirect refs to their parent bytenr.
1369 * When roots are found, they're added to the roots list
1370 *
1371 * @ctx: Backref walking context object, must be not NULL.
1372 * @sc: If !NULL, then immediately return BACKREF_FOUND_SHARED when a
1373 * shared extent is detected.
1374 *
1375 * Otherwise this returns 0 for success and <0 for an error.
1376 *
1377 * FIXME some caching might speed things up
1378 */
1379static int find_parent_nodes(struct btrfs_backref_walk_ctx *ctx,
1380 struct share_check *sc)
1381{
1382 struct btrfs_root *root = btrfs_extent_root(ctx->fs_info, ctx->bytenr);
1383 struct btrfs_key key;
1384 struct btrfs_path *path;
1385 struct btrfs_delayed_ref_root *delayed_refs = NULL;
1386 struct btrfs_delayed_ref_head *head;
1387 int info_level = 0;
1388 int ret;
1389 struct prelim_ref *ref;
1390 struct rb_node *node;
1391 struct extent_inode_elem *eie = NULL;
1392 struct preftrees preftrees = {
1393 .direct = PREFTREE_INIT,
1394 .indirect = PREFTREE_INIT,
1395 .indirect_missing_keys = PREFTREE_INIT
1396 };
1397
1398 /* Roots ulist is not needed when using a sharedness check context. */
1399 if (sc)
1400 ASSERT(ctx->roots == NULL);
1401
1402 key.objectid = ctx->bytenr;
1403 key.offset = (u64)-1;
1404 if (btrfs_fs_incompat(ctx->fs_info, SKINNY_METADATA))
1405 key.type = BTRFS_METADATA_ITEM_KEY;
1406 else
1407 key.type = BTRFS_EXTENT_ITEM_KEY;
1408
1409 path = btrfs_alloc_path();
1410 if (!path)
1411 return -ENOMEM;
1412 if (!ctx->trans) {
1413 path->search_commit_root = 1;
1414 path->skip_locking = 1;
1415 }
1416
1417 if (ctx->time_seq == BTRFS_SEQ_LAST)
1418 path->skip_locking = 1;
1419
1420again:
1421 head = NULL;
1422
1423 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1424 if (ret < 0)
1425 goto out;
1426 if (ret == 0) {
1427 /*
1428 * Key with offset -1 found, there would have to exist an extent
1429 * item with such offset, but this is out of the valid range.
1430 */
1431 ret = -EUCLEAN;
1432 goto out;
1433 }
1434
1435 if (ctx->trans && likely(ctx->trans->type != __TRANS_DUMMY) &&
1436 ctx->time_seq != BTRFS_SEQ_LAST) {
1437 /*
1438 * We have a specific time_seq we care about and trans which
1439 * means we have the path lock, we need to grab the ref head and
1440 * lock it so we have a consistent view of the refs at the given
1441 * time.
1442 */
1443 delayed_refs = &ctx->trans->transaction->delayed_refs;
1444 spin_lock(&delayed_refs->lock);
1445 head = btrfs_find_delayed_ref_head(ctx->fs_info, delayed_refs,
1446 ctx->bytenr);
1447 if (head) {
1448 if (!mutex_trylock(&head->mutex)) {
1449 refcount_inc(&head->refs);
1450 spin_unlock(&delayed_refs->lock);
1451
1452 btrfs_release_path(path);
1453
1454 /*
1455 * Mutex was contended, block until it's
1456 * released and try again
1457 */
1458 mutex_lock(&head->mutex);
1459 mutex_unlock(&head->mutex);
1460 btrfs_put_delayed_ref_head(head);
1461 goto again;
1462 }
1463 spin_unlock(&delayed_refs->lock);
1464 ret = add_delayed_refs(ctx->fs_info, head, ctx->time_seq,
1465 &preftrees, sc);
1466 mutex_unlock(&head->mutex);
1467 if (ret)
1468 goto out;
1469 } else {
1470 spin_unlock(&delayed_refs->lock);
1471 }
1472 }
1473
1474 if (path->slots[0]) {
1475 struct extent_buffer *leaf;
1476 int slot;
1477
1478 path->slots[0]--;
1479 leaf = path->nodes[0];
1480 slot = path->slots[0];
1481 btrfs_item_key_to_cpu(leaf, &key, slot);
1482 if (key.objectid == ctx->bytenr &&
1483 (key.type == BTRFS_EXTENT_ITEM_KEY ||
1484 key.type == BTRFS_METADATA_ITEM_KEY)) {
1485 ret = add_inline_refs(ctx, path, &info_level,
1486 &preftrees, sc);
1487 if (ret)
1488 goto out;
1489 ret = add_keyed_refs(ctx, root, path, info_level,
1490 &preftrees, sc);
1491 if (ret)
1492 goto out;
1493 }
1494 }
1495
1496 /*
1497 * If we have a share context and we reached here, it means the extent
1498 * is not directly shared (no multiple reference items for it),
1499 * otherwise we would have exited earlier with a return value of
1500 * BACKREF_FOUND_SHARED after processing delayed references or while
1501 * processing inline or keyed references from the extent tree.
1502 * The extent may however be indirectly shared through shared subtrees
1503 * as a result from creating snapshots, so we determine below what is
1504 * its parent node, in case we are dealing with a metadata extent, or
1505 * what's the leaf (or leaves), from a fs tree, that has a file extent
1506 * item pointing to it in case we are dealing with a data extent.
1507 */
1508 ASSERT(extent_is_shared(sc) == 0);
1509
1510 /*
1511 * If we are here for a data extent and we have a share_check structure
1512 * it means the data extent is not directly shared (does not have
1513 * multiple reference items), so we have to check if a path in the fs
1514 * tree (going from the root node down to the leaf that has the file
1515 * extent item pointing to the data extent) is shared, that is, if any
1516 * of the extent buffers in the path is referenced by other trees.
1517 */
1518 if (sc && ctx->bytenr == sc->data_bytenr) {
1519 /*
1520 * If our data extent is from a generation more recent than the
1521 * last generation used to snapshot the root, then we know that
1522 * it can not be shared through subtrees, so we can skip
1523 * resolving indirect references, there's no point in
1524 * determining the extent buffers for the path from the fs tree
1525 * root node down to the leaf that has the file extent item that
1526 * points to the data extent.
1527 */
1528 if (sc->data_extent_gen >
1529 btrfs_root_last_snapshot(&sc->root->root_item)) {
1530 ret = BACKREF_FOUND_NOT_SHARED;
1531 goto out;
1532 }
1533
1534 /*
1535 * If we are only determining if a data extent is shared or not
1536 * and the corresponding file extent item is located in the same
1537 * leaf as the previous file extent item, we can skip resolving
1538 * indirect references for a data extent, since the fs tree path
1539 * is the same (same leaf, so same path). We skip as long as the
1540 * cached result for the leaf is valid and only if there's only
1541 * one file extent item pointing to the data extent, because in
1542 * the case of multiple file extent items, they may be located
1543 * in different leaves and therefore we have multiple paths.
1544 */
1545 if (sc->ctx->curr_leaf_bytenr == sc->ctx->prev_leaf_bytenr &&
1546 sc->self_ref_count == 1) {
1547 bool cached;
1548 bool is_shared;
1549
1550 cached = lookup_backref_shared_cache(sc->ctx, sc->root,
1551 sc->ctx->curr_leaf_bytenr,
1552 0, &is_shared);
1553 if (cached) {
1554 if (is_shared)
1555 ret = BACKREF_FOUND_SHARED;
1556 else
1557 ret = BACKREF_FOUND_NOT_SHARED;
1558 goto out;
1559 }
1560 }
1561 }
1562
1563 btrfs_release_path(path);
1564
1565 ret = add_missing_keys(ctx->fs_info, &preftrees, path->skip_locking == 0);
1566 if (ret)
1567 goto out;
1568
1569 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1570
1571 ret = resolve_indirect_refs(ctx, path, &preftrees, sc);
1572 if (ret)
1573 goto out;
1574
1575 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1576
1577 /*
1578 * This walks the tree of merged and resolved refs. Tree blocks are
1579 * read in as needed. Unique entries are added to the ulist, and
1580 * the list of found roots is updated.
1581 *
1582 * We release the entire tree in one go before returning.
1583 */
1584 node = rb_first_cached(&preftrees.direct.root);
1585 while (node) {
1586 ref = rb_entry(node, struct prelim_ref, rbnode);
1587 node = rb_next(&ref->rbnode);
1588 /*
1589 * ref->count < 0 can happen here if there are delayed
1590 * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1591 * prelim_ref_insert() relies on this when merging
1592 * identical refs to keep the overall count correct.
1593 * prelim_ref_insert() will merge only those refs
1594 * which compare identically. Any refs having
1595 * e.g. different offsets would not be merged,
1596 * and would retain their original ref->count < 0.
1597 */
1598 if (ctx->roots && ref->count && ref->root_id && ref->parent == 0) {
1599 /* no parent == root of tree */
1600 ret = ulist_add(ctx->roots, ref->root_id, 0, GFP_NOFS);
1601 if (ret < 0)
1602 goto out;
1603 }
1604 if (ref->count && ref->parent) {
1605 if (!ctx->skip_inode_ref_list && !ref->inode_list &&
1606 ref->level == 0) {
1607 struct btrfs_tree_parent_check check = { 0 };
1608 struct extent_buffer *eb;
1609
1610 check.level = ref->level;
1611
1612 eb = read_tree_block(ctx->fs_info, ref->parent,
1613 &check);
1614 if (IS_ERR(eb)) {
1615 ret = PTR_ERR(eb);
1616 goto out;
1617 }
1618 if (!extent_buffer_uptodate(eb)) {
1619 free_extent_buffer(eb);
1620 ret = -EIO;
1621 goto out;
1622 }
1623
1624 if (!path->skip_locking)
1625 btrfs_tree_read_lock(eb);
1626 ret = find_extent_in_eb(ctx, eb, &eie);
1627 if (!path->skip_locking)
1628 btrfs_tree_read_unlock(eb);
1629 free_extent_buffer(eb);
1630 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1631 ret < 0)
1632 goto out;
1633 ref->inode_list = eie;
1634 /*
1635 * We transferred the list ownership to the ref,
1636 * so set to NULL to avoid a double free in case
1637 * an error happens after this.
1638 */
1639 eie = NULL;
1640 }
1641 ret = ulist_add_merge_ptr(ctx->refs, ref->parent,
1642 ref->inode_list,
1643 (void **)&eie, GFP_NOFS);
1644 if (ret < 0)
1645 goto out;
1646 if (!ret && !ctx->skip_inode_ref_list) {
1647 /*
1648 * We've recorded that parent, so we must extend
1649 * its inode list here.
1650 *
1651 * However if there was corruption we may not
1652 * have found an eie, return an error in this
1653 * case.
1654 */
1655 ASSERT(eie);
1656 if (!eie) {
1657 ret = -EUCLEAN;
1658 goto out;
1659 }
1660 while (eie->next)
1661 eie = eie->next;
1662 eie->next = ref->inode_list;
1663 }
1664 eie = NULL;
1665 /*
1666 * We have transferred the inode list ownership from
1667 * this ref to the ref we added to the 'refs' ulist.
1668 * So set this ref's inode list to NULL to avoid
1669 * use-after-free when our caller uses it or double
1670 * frees in case an error happens before we return.
1671 */
1672 ref->inode_list = NULL;
1673 }
1674 cond_resched();
1675 }
1676
1677out:
1678 btrfs_free_path(path);
1679
1680 prelim_release(&preftrees.direct);
1681 prelim_release(&preftrees.indirect);
1682 prelim_release(&preftrees.indirect_missing_keys);
1683
1684 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
1685 free_inode_elem_list(eie);
1686 return ret;
1687}
1688
1689/*
1690 * Finds all leaves with a reference to the specified combination of
1691 * @ctx->bytenr and @ctx->extent_item_pos. The bytenr of the found leaves are
1692 * added to the ulist at @ctx->refs, and that ulist is allocated by this
1693 * function. The caller should free the ulist with free_leaf_list() if
1694 * @ctx->ignore_extent_item_pos is false, otherwise a fimple ulist_free() is
1695 * enough.
1696 *
1697 * Returns 0 on success and < 0 on error. On error @ctx->refs is not allocated.
1698 */
1699int btrfs_find_all_leafs(struct btrfs_backref_walk_ctx *ctx)
1700{
1701 int ret;
1702
1703 ASSERT(ctx->refs == NULL);
1704
1705 ctx->refs = ulist_alloc(GFP_NOFS);
1706 if (!ctx->refs)
1707 return -ENOMEM;
1708
1709 ret = find_parent_nodes(ctx, NULL);
1710 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1711 (ret < 0 && ret != -ENOENT)) {
1712 free_leaf_list(ctx->refs);
1713 ctx->refs = NULL;
1714 return ret;
1715 }
1716
1717 return 0;
1718}
1719
1720/*
1721 * Walk all backrefs for a given extent to find all roots that reference this
1722 * extent. Walking a backref means finding all extents that reference this
1723 * extent and in turn walk the backrefs of those, too. Naturally this is a
1724 * recursive process, but here it is implemented in an iterative fashion: We
1725 * find all referencing extents for the extent in question and put them on a
1726 * list. In turn, we find all referencing extents for those, further appending
1727 * to the list. The way we iterate the list allows adding more elements after
1728 * the current while iterating. The process stops when we reach the end of the
1729 * list.
1730 *
1731 * Found roots are added to @ctx->roots, which is allocated by this function if
1732 * it points to NULL, in which case the caller is responsible for freeing it
1733 * after it's not needed anymore.
1734 * This function requires @ctx->refs to be NULL, as it uses it for allocating a
1735 * ulist to do temporary work, and frees it before returning.
1736 *
1737 * Returns 0 on success, < 0 on error.
1738 */
1739static int btrfs_find_all_roots_safe(struct btrfs_backref_walk_ctx *ctx)
1740{
1741 const u64 orig_bytenr = ctx->bytenr;
1742 const bool orig_skip_inode_ref_list = ctx->skip_inode_ref_list;
1743 bool roots_ulist_allocated = false;
1744 struct ulist_iterator uiter;
1745 int ret = 0;
1746
1747 ASSERT(ctx->refs == NULL);
1748
1749 ctx->refs = ulist_alloc(GFP_NOFS);
1750 if (!ctx->refs)
1751 return -ENOMEM;
1752
1753 if (!ctx->roots) {
1754 ctx->roots = ulist_alloc(GFP_NOFS);
1755 if (!ctx->roots) {
1756 ulist_free(ctx->refs);
1757 ctx->refs = NULL;
1758 return -ENOMEM;
1759 }
1760 roots_ulist_allocated = true;
1761 }
1762
1763 ctx->skip_inode_ref_list = true;
1764
1765 ULIST_ITER_INIT(&uiter);
1766 while (1) {
1767 struct ulist_node *node;
1768
1769 ret = find_parent_nodes(ctx, NULL);
1770 if (ret < 0 && ret != -ENOENT) {
1771 if (roots_ulist_allocated) {
1772 ulist_free(ctx->roots);
1773 ctx->roots = NULL;
1774 }
1775 break;
1776 }
1777 ret = 0;
1778 node = ulist_next(ctx->refs, &uiter);
1779 if (!node)
1780 break;
1781 ctx->bytenr = node->val;
1782 cond_resched();
1783 }
1784
1785 ulist_free(ctx->refs);
1786 ctx->refs = NULL;
1787 ctx->bytenr = orig_bytenr;
1788 ctx->skip_inode_ref_list = orig_skip_inode_ref_list;
1789
1790 return ret;
1791}
1792
1793int btrfs_find_all_roots(struct btrfs_backref_walk_ctx *ctx,
1794 bool skip_commit_root_sem)
1795{
1796 int ret;
1797
1798 if (!ctx->trans && !skip_commit_root_sem)
1799 down_read(&ctx->fs_info->commit_root_sem);
1800 ret = btrfs_find_all_roots_safe(ctx);
1801 if (!ctx->trans && !skip_commit_root_sem)
1802 up_read(&ctx->fs_info->commit_root_sem);
1803 return ret;
1804}
1805
1806struct btrfs_backref_share_check_ctx *btrfs_alloc_backref_share_check_ctx(void)
1807{
1808 struct btrfs_backref_share_check_ctx *ctx;
1809
1810 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1811 if (!ctx)
1812 return NULL;
1813
1814 ulist_init(&ctx->refs);
1815
1816 return ctx;
1817}
1818
1819void btrfs_free_backref_share_ctx(struct btrfs_backref_share_check_ctx *ctx)
1820{
1821 if (!ctx)
1822 return;
1823
1824 ulist_release(&ctx->refs);
1825 kfree(ctx);
1826}
1827
1828/*
1829 * Check if a data extent is shared or not.
1830 *
1831 * @inode: The inode whose extent we are checking.
1832 * @bytenr: Logical bytenr of the extent we are checking.
1833 * @extent_gen: Generation of the extent (file extent item) or 0 if it is
1834 * not known.
1835 * @ctx: A backref sharedness check context.
1836 *
1837 * btrfs_is_data_extent_shared uses the backref walking code but will short
1838 * circuit as soon as it finds a root or inode that doesn't match the
1839 * one passed in. This provides a significant performance benefit for
1840 * callers (such as fiemap) which want to know whether the extent is
1841 * shared but do not need a ref count.
1842 *
1843 * This attempts to attach to the running transaction in order to account for
1844 * delayed refs, but continues on even when no running transaction exists.
1845 *
1846 * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1847 */
1848int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
1849 u64 extent_gen,
1850 struct btrfs_backref_share_check_ctx *ctx)
1851{
1852 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
1853 struct btrfs_root *root = inode->root;
1854 struct btrfs_fs_info *fs_info = root->fs_info;
1855 struct btrfs_trans_handle *trans;
1856 struct ulist_iterator uiter;
1857 struct ulist_node *node;
1858 struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1859 int ret = 0;
1860 struct share_check shared = {
1861 .ctx = ctx,
1862 .root = root,
1863 .inum = btrfs_ino(inode),
1864 .data_bytenr = bytenr,
1865 .data_extent_gen = extent_gen,
1866 .share_count = 0,
1867 .self_ref_count = 0,
1868 .have_delayed_delete_refs = false,
1869 };
1870 int level;
1871 bool leaf_cached;
1872 bool leaf_is_shared;
1873
1874 for (int i = 0; i < BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE; i++) {
1875 if (ctx->prev_extents_cache[i].bytenr == bytenr)
1876 return ctx->prev_extents_cache[i].is_shared;
1877 }
1878
1879 ulist_init(&ctx->refs);
1880
1881 trans = btrfs_join_transaction_nostart(root);
1882 if (IS_ERR(trans)) {
1883 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1884 ret = PTR_ERR(trans);
1885 goto out;
1886 }
1887 trans = NULL;
1888 down_read(&fs_info->commit_root_sem);
1889 } else {
1890 btrfs_get_tree_mod_seq(fs_info, &elem);
1891 walk_ctx.time_seq = elem.seq;
1892 }
1893
1894 ctx->use_path_cache = true;
1895
1896 /*
1897 * We may have previously determined that the current leaf is shared.
1898 * If it is, then we have a data extent that is shared due to a shared
1899 * subtree (caused by snapshotting) and we don't need to check for data
1900 * backrefs. If the leaf is not shared, then we must do backref walking
1901 * to determine if the data extent is shared through reflinks.
1902 */
1903 leaf_cached = lookup_backref_shared_cache(ctx, root,
1904 ctx->curr_leaf_bytenr, 0,
1905 &leaf_is_shared);
1906 if (leaf_cached && leaf_is_shared) {
1907 ret = 1;
1908 goto out_trans;
1909 }
1910
1911 walk_ctx.skip_inode_ref_list = true;
1912 walk_ctx.trans = trans;
1913 walk_ctx.fs_info = fs_info;
1914 walk_ctx.refs = &ctx->refs;
1915
1916 /* -1 means we are in the bytenr of the data extent. */
1917 level = -1;
1918 ULIST_ITER_INIT(&uiter);
1919 while (1) {
1920 const unsigned long prev_ref_count = ctx->refs.nnodes;
1921
1922 walk_ctx.bytenr = bytenr;
1923 ret = find_parent_nodes(&walk_ctx, &shared);
1924 if (ret == BACKREF_FOUND_SHARED ||
1925 ret == BACKREF_FOUND_NOT_SHARED) {
1926 /* If shared must return 1, otherwise return 0. */
1927 ret = (ret == BACKREF_FOUND_SHARED) ? 1 : 0;
1928 if (level >= 0)
1929 store_backref_shared_cache(ctx, root, bytenr,
1930 level, ret == 1);
1931 break;
1932 }
1933 if (ret < 0 && ret != -ENOENT)
1934 break;
1935 ret = 0;
1936
1937 /*
1938 * More than one extent buffer (bytenr) may have been added to
1939 * the ctx->refs ulist, in which case we have to check multiple
1940 * tree paths in case the first one is not shared, so we can not
1941 * use the path cache which is made for a single path. Multiple
1942 * extent buffers at the current level happen when:
1943 *
1944 * 1) level -1, the data extent: If our data extent was not
1945 * directly shared (without multiple reference items), then
1946 * it might have a single reference item with a count > 1 for
1947 * the same offset, which means there are 2 (or more) file
1948 * extent items that point to the data extent - this happens
1949 * when a file extent item needs to be split and then one
1950 * item gets moved to another leaf due to a b+tree leaf split
1951 * when inserting some item. In this case the file extent
1952 * items may be located in different leaves and therefore
1953 * some of the leaves may be referenced through shared
1954 * subtrees while others are not. Since our extent buffer
1955 * cache only works for a single path (by far the most common
1956 * case and simpler to deal with), we can not use it if we
1957 * have multiple leaves (which implies multiple paths).
1958 *
1959 * 2) level >= 0, a tree node/leaf: We can have a mix of direct
1960 * and indirect references on a b+tree node/leaf, so we have
1961 * to check multiple paths, and the extent buffer (the
1962 * current bytenr) may be shared or not. One example is
1963 * during relocation as we may get a shared tree block ref
1964 * (direct ref) and a non-shared tree block ref (indirect
1965 * ref) for the same node/leaf.
1966 */
1967 if ((ctx->refs.nnodes - prev_ref_count) > 1)
1968 ctx->use_path_cache = false;
1969
1970 if (level >= 0)
1971 store_backref_shared_cache(ctx, root, bytenr,
1972 level, false);
1973 node = ulist_next(&ctx->refs, &uiter);
1974 if (!node)
1975 break;
1976 bytenr = node->val;
1977 if (ctx->use_path_cache) {
1978 bool is_shared;
1979 bool cached;
1980
1981 level++;
1982 cached = lookup_backref_shared_cache(ctx, root, bytenr,
1983 level, &is_shared);
1984 if (cached) {
1985 ret = (is_shared ? 1 : 0);
1986 break;
1987 }
1988 }
1989 shared.share_count = 0;
1990 shared.have_delayed_delete_refs = false;
1991 cond_resched();
1992 }
1993
1994 /*
1995 * If the path cache is disabled, then it means at some tree level we
1996 * got multiple parents due to a mix of direct and indirect backrefs or
1997 * multiple leaves with file extent items pointing to the same data
1998 * extent. We have to invalidate the cache and cache only the sharedness
1999 * result for the levels where we got only one node/reference.
2000 */
2001 if (!ctx->use_path_cache) {
2002 int i = 0;
2003
2004 level--;
2005 if (ret >= 0 && level >= 0) {
2006 bytenr = ctx->path_cache_entries[level].bytenr;
2007 ctx->use_path_cache = true;
2008 store_backref_shared_cache(ctx, root, bytenr, level, ret);
2009 i = level + 1;
2010 }
2011
2012 for ( ; i < BTRFS_MAX_LEVEL; i++)
2013 ctx->path_cache_entries[i].bytenr = 0;
2014 }
2015
2016 /*
2017 * Cache the sharedness result for the data extent if we know our inode
2018 * has more than 1 file extent item that refers to the data extent.
2019 */
2020 if (ret >= 0 && shared.self_ref_count > 1) {
2021 int slot = ctx->prev_extents_cache_slot;
2022
2023 ctx->prev_extents_cache[slot].bytenr = shared.data_bytenr;
2024 ctx->prev_extents_cache[slot].is_shared = (ret == 1);
2025
2026 slot = (slot + 1) % BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE;
2027 ctx->prev_extents_cache_slot = slot;
2028 }
2029
2030out_trans:
2031 if (trans) {
2032 btrfs_put_tree_mod_seq(fs_info, &elem);
2033 btrfs_end_transaction(trans);
2034 } else {
2035 up_read(&fs_info->commit_root_sem);
2036 }
2037out:
2038 ulist_release(&ctx->refs);
2039 ctx->prev_leaf_bytenr = ctx->curr_leaf_bytenr;
2040
2041 return ret;
2042}
2043
2044int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
2045 u64 start_off, struct btrfs_path *path,
2046 struct btrfs_inode_extref **ret_extref,
2047 u64 *found_off)
2048{
2049 int ret, slot;
2050 struct btrfs_key key;
2051 struct btrfs_key found_key;
2052 struct btrfs_inode_extref *extref;
2053 const struct extent_buffer *leaf;
2054 unsigned long ptr;
2055
2056 key.objectid = inode_objectid;
2057 key.type = BTRFS_INODE_EXTREF_KEY;
2058 key.offset = start_off;
2059
2060 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2061 if (ret < 0)
2062 return ret;
2063
2064 while (1) {
2065 leaf = path->nodes[0];
2066 slot = path->slots[0];
2067 if (slot >= btrfs_header_nritems(leaf)) {
2068 /*
2069 * If the item at offset is not found,
2070 * btrfs_search_slot will point us to the slot
2071 * where it should be inserted. In our case
2072 * that will be the slot directly before the
2073 * next INODE_REF_KEY_V2 item. In the case
2074 * that we're pointing to the last slot in a
2075 * leaf, we must move one leaf over.
2076 */
2077 ret = btrfs_next_leaf(root, path);
2078 if (ret) {
2079 if (ret >= 1)
2080 ret = -ENOENT;
2081 break;
2082 }
2083 continue;
2084 }
2085
2086 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2087
2088 /*
2089 * Check that we're still looking at an extended ref key for
2090 * this particular objectid. If we have different
2091 * objectid or type then there are no more to be found
2092 * in the tree and we can exit.
2093 */
2094 ret = -ENOENT;
2095 if (found_key.objectid != inode_objectid)
2096 break;
2097 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
2098 break;
2099
2100 ret = 0;
2101 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2102 extref = (struct btrfs_inode_extref *)ptr;
2103 *ret_extref = extref;
2104 if (found_off)
2105 *found_off = found_key.offset;
2106 break;
2107 }
2108
2109 return ret;
2110}
2111
2112/*
2113 * this iterates to turn a name (from iref/extref) into a full filesystem path.
2114 * Elements of the path are separated by '/' and the path is guaranteed to be
2115 * 0-terminated. the path is only given within the current file system.
2116 * Therefore, it never starts with a '/'. the caller is responsible to provide
2117 * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
2118 * the start point of the resulting string is returned. this pointer is within
2119 * dest, normally.
2120 * in case the path buffer would overflow, the pointer is decremented further
2121 * as if output was written to the buffer, though no more output is actually
2122 * generated. that way, the caller can determine how much space would be
2123 * required for the path to fit into the buffer. in that case, the returned
2124 * value will be smaller than dest. callers must check this!
2125 */
2126char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
2127 u32 name_len, unsigned long name_off,
2128 struct extent_buffer *eb_in, u64 parent,
2129 char *dest, u32 size)
2130{
2131 int slot;
2132 u64 next_inum;
2133 int ret;
2134 s64 bytes_left = ((s64)size) - 1;
2135 struct extent_buffer *eb = eb_in;
2136 struct btrfs_key found_key;
2137 struct btrfs_inode_ref *iref;
2138
2139 if (bytes_left >= 0)
2140 dest[bytes_left] = '\0';
2141
2142 while (1) {
2143 bytes_left -= name_len;
2144 if (bytes_left >= 0)
2145 read_extent_buffer(eb, dest + bytes_left,
2146 name_off, name_len);
2147 if (eb != eb_in) {
2148 if (!path->skip_locking)
2149 btrfs_tree_read_unlock(eb);
2150 free_extent_buffer(eb);
2151 }
2152 ret = btrfs_find_item(fs_root, path, parent, 0,
2153 BTRFS_INODE_REF_KEY, &found_key);
2154 if (ret > 0)
2155 ret = -ENOENT;
2156 if (ret)
2157 break;
2158
2159 next_inum = found_key.offset;
2160
2161 /* regular exit ahead */
2162 if (parent == next_inum)
2163 break;
2164
2165 slot = path->slots[0];
2166 eb = path->nodes[0];
2167 /* make sure we can use eb after releasing the path */
2168 if (eb != eb_in) {
2169 path->nodes[0] = NULL;
2170 path->locks[0] = 0;
2171 }
2172 btrfs_release_path(path);
2173 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2174
2175 name_len = btrfs_inode_ref_name_len(eb, iref);
2176 name_off = (unsigned long)(iref + 1);
2177
2178 parent = next_inum;
2179 --bytes_left;
2180 if (bytes_left >= 0)
2181 dest[bytes_left] = '/';
2182 }
2183
2184 btrfs_release_path(path);
2185
2186 if (ret)
2187 return ERR_PTR(ret);
2188
2189 return dest + bytes_left;
2190}
2191
2192/*
2193 * this makes the path point to (logical EXTENT_ITEM *)
2194 * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
2195 * tree blocks and <0 on error.
2196 */
2197int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
2198 struct btrfs_path *path, struct btrfs_key *found_key,
2199 u64 *flags_ret)
2200{
2201 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
2202 int ret;
2203 u64 flags;
2204 u64 size = 0;
2205 u32 item_size;
2206 const struct extent_buffer *eb;
2207 struct btrfs_extent_item *ei;
2208 struct btrfs_key key;
2209
2210 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2211 key.type = BTRFS_METADATA_ITEM_KEY;
2212 else
2213 key.type = BTRFS_EXTENT_ITEM_KEY;
2214 key.objectid = logical;
2215 key.offset = (u64)-1;
2216
2217 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2218 if (ret < 0)
2219 return ret;
2220 if (ret == 0) {
2221 /*
2222 * Key with offset -1 found, there would have to exist an extent
2223 * item with such offset, but this is out of the valid range.
2224 */
2225 return -EUCLEAN;
2226 }
2227
2228 ret = btrfs_previous_extent_item(extent_root, path, 0);
2229 if (ret) {
2230 if (ret > 0)
2231 ret = -ENOENT;
2232 return ret;
2233 }
2234 btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
2235 if (found_key->type == BTRFS_METADATA_ITEM_KEY)
2236 size = fs_info->nodesize;
2237 else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
2238 size = found_key->offset;
2239
2240 if (found_key->objectid > logical ||
2241 found_key->objectid + size <= logical) {
2242 btrfs_debug(fs_info,
2243 "logical %llu is not within any extent", logical);
2244 return -ENOENT;
2245 }
2246
2247 eb = path->nodes[0];
2248 item_size = btrfs_item_size(eb, path->slots[0]);
2249
2250 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
2251 flags = btrfs_extent_flags(eb, ei);
2252
2253 btrfs_debug(fs_info,
2254 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
2255 logical, logical - found_key->objectid, found_key->objectid,
2256 found_key->offset, flags, item_size);
2257
2258 WARN_ON(!flags_ret);
2259 if (flags_ret) {
2260 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2261 *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
2262 else if (flags & BTRFS_EXTENT_FLAG_DATA)
2263 *flags_ret = BTRFS_EXTENT_FLAG_DATA;
2264 else
2265 BUG();
2266 return 0;
2267 }
2268
2269 return -EIO;
2270}
2271
2272/*
2273 * helper function to iterate extent inline refs. ptr must point to a 0 value
2274 * for the first call and may be modified. it is used to track state.
2275 * if more refs exist, 0 is returned and the next call to
2276 * get_extent_inline_ref must pass the modified ptr parameter to get the
2277 * next ref. after the last ref was processed, 1 is returned.
2278 * returns <0 on error
2279 */
2280static int get_extent_inline_ref(unsigned long *ptr,
2281 const struct extent_buffer *eb,
2282 const struct btrfs_key *key,
2283 const struct btrfs_extent_item *ei,
2284 u32 item_size,
2285 struct btrfs_extent_inline_ref **out_eiref,
2286 int *out_type)
2287{
2288 unsigned long end;
2289 u64 flags;
2290 struct btrfs_tree_block_info *info;
2291
2292 if (!*ptr) {
2293 /* first call */
2294 flags = btrfs_extent_flags(eb, ei);
2295 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2296 if (key->type == BTRFS_METADATA_ITEM_KEY) {
2297 /* a skinny metadata extent */
2298 *out_eiref =
2299 (struct btrfs_extent_inline_ref *)(ei + 1);
2300 } else {
2301 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
2302 info = (struct btrfs_tree_block_info *)(ei + 1);
2303 *out_eiref =
2304 (struct btrfs_extent_inline_ref *)(info + 1);
2305 }
2306 } else {
2307 *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
2308 }
2309 *ptr = (unsigned long)*out_eiref;
2310 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
2311 return -ENOENT;
2312 }
2313
2314 end = (unsigned long)ei + item_size;
2315 *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
2316 *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
2317 BTRFS_REF_TYPE_ANY);
2318 if (*out_type == BTRFS_REF_TYPE_INVALID)
2319 return -EUCLEAN;
2320
2321 *ptr += btrfs_extent_inline_ref_size(*out_type);
2322 WARN_ON(*ptr > end);
2323 if (*ptr == end)
2324 return 1; /* last */
2325
2326 return 0;
2327}
2328
2329/*
2330 * reads the tree block backref for an extent. tree level and root are returned
2331 * through out_level and out_root. ptr must point to a 0 value for the first
2332 * call and may be modified (see get_extent_inline_ref comment).
2333 * returns 0 if data was provided, 1 if there was no more data to provide or
2334 * <0 on error.
2335 */
2336int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
2337 struct btrfs_key *key, struct btrfs_extent_item *ei,
2338 u32 item_size, u64 *out_root, u8 *out_level)
2339{
2340 int ret;
2341 int type;
2342 struct btrfs_extent_inline_ref *eiref;
2343
2344 if (*ptr == (unsigned long)-1)
2345 return 1;
2346
2347 while (1) {
2348 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
2349 &eiref, &type);
2350 if (ret < 0)
2351 return ret;
2352
2353 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
2354 type == BTRFS_SHARED_BLOCK_REF_KEY)
2355 break;
2356
2357 if (ret == 1)
2358 return 1;
2359 }
2360
2361 /* we can treat both ref types equally here */
2362 *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
2363
2364 if (key->type == BTRFS_EXTENT_ITEM_KEY) {
2365 struct btrfs_tree_block_info *info;
2366
2367 info = (struct btrfs_tree_block_info *)(ei + 1);
2368 *out_level = btrfs_tree_block_level(eb, info);
2369 } else {
2370 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
2371 *out_level = (u8)key->offset;
2372 }
2373
2374 if (ret == 1)
2375 *ptr = (unsigned long)-1;
2376
2377 return 0;
2378}
2379
2380static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
2381 struct extent_inode_elem *inode_list,
2382 u64 root, u64 extent_item_objectid,
2383 iterate_extent_inodes_t *iterate, void *ctx)
2384{
2385 struct extent_inode_elem *eie;
2386 int ret = 0;
2387
2388 for (eie = inode_list; eie; eie = eie->next) {
2389 btrfs_debug(fs_info,
2390 "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
2391 extent_item_objectid, eie->inum,
2392 eie->offset, root);
2393 ret = iterate(eie->inum, eie->offset, eie->num_bytes, root, ctx);
2394 if (ret) {
2395 btrfs_debug(fs_info,
2396 "stopping iteration for %llu due to ret=%d",
2397 extent_item_objectid, ret);
2398 break;
2399 }
2400 }
2401
2402 return ret;
2403}
2404
2405/*
2406 * calls iterate() for every inode that references the extent identified by
2407 * the given parameters.
2408 * when the iterator function returns a non-zero value, iteration stops.
2409 */
2410int iterate_extent_inodes(struct btrfs_backref_walk_ctx *ctx,
2411 bool search_commit_root,
2412 iterate_extent_inodes_t *iterate, void *user_ctx)
2413{
2414 int ret;
2415 struct ulist *refs;
2416 struct ulist_node *ref_node;
2417 struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
2418 struct ulist_iterator ref_uiter;
2419
2420 btrfs_debug(ctx->fs_info, "resolving all inodes for extent %llu",
2421 ctx->bytenr);
2422
2423 ASSERT(ctx->trans == NULL);
2424 ASSERT(ctx->roots == NULL);
2425
2426 if (!search_commit_root) {
2427 struct btrfs_trans_handle *trans;
2428
2429 trans = btrfs_attach_transaction(ctx->fs_info->tree_root);
2430 if (IS_ERR(trans)) {
2431 if (PTR_ERR(trans) != -ENOENT &&
2432 PTR_ERR(trans) != -EROFS)
2433 return PTR_ERR(trans);
2434 trans = NULL;
2435 }
2436 ctx->trans = trans;
2437 }
2438
2439 if (ctx->trans) {
2440 btrfs_get_tree_mod_seq(ctx->fs_info, &seq_elem);
2441 ctx->time_seq = seq_elem.seq;
2442 } else {
2443 down_read(&ctx->fs_info->commit_root_sem);
2444 }
2445
2446 ret = btrfs_find_all_leafs(ctx);
2447 if (ret)
2448 goto out;
2449 refs = ctx->refs;
2450 ctx->refs = NULL;
2451
2452 ULIST_ITER_INIT(&ref_uiter);
2453 while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2454 const u64 leaf_bytenr = ref_node->val;
2455 struct ulist_node *root_node;
2456 struct ulist_iterator root_uiter;
2457 struct extent_inode_elem *inode_list;
2458
2459 inode_list = (struct extent_inode_elem *)(uintptr_t)ref_node->aux;
2460
2461 if (ctx->cache_lookup) {
2462 const u64 *root_ids;
2463 int root_count;
2464 bool cached;
2465
2466 cached = ctx->cache_lookup(leaf_bytenr, ctx->user_ctx,
2467 &root_ids, &root_count);
2468 if (cached) {
2469 for (int i = 0; i < root_count; i++) {
2470 ret = iterate_leaf_refs(ctx->fs_info,
2471 inode_list,
2472 root_ids[i],
2473 leaf_bytenr,
2474 iterate,
2475 user_ctx);
2476 if (ret)
2477 break;
2478 }
2479 continue;
2480 }
2481 }
2482
2483 if (!ctx->roots) {
2484 ctx->roots = ulist_alloc(GFP_NOFS);
2485 if (!ctx->roots) {
2486 ret = -ENOMEM;
2487 break;
2488 }
2489 }
2490
2491 ctx->bytenr = leaf_bytenr;
2492 ret = btrfs_find_all_roots_safe(ctx);
2493 if (ret)
2494 break;
2495
2496 if (ctx->cache_store)
2497 ctx->cache_store(leaf_bytenr, ctx->roots, ctx->user_ctx);
2498
2499 ULIST_ITER_INIT(&root_uiter);
2500 while (!ret && (root_node = ulist_next(ctx->roots, &root_uiter))) {
2501 btrfs_debug(ctx->fs_info,
2502 "root %llu references leaf %llu, data list %#llx",
2503 root_node->val, ref_node->val,
2504 ref_node->aux);
2505 ret = iterate_leaf_refs(ctx->fs_info, inode_list,
2506 root_node->val, ctx->bytenr,
2507 iterate, user_ctx);
2508 }
2509 ulist_reinit(ctx->roots);
2510 }
2511
2512 free_leaf_list(refs);
2513out:
2514 if (ctx->trans) {
2515 btrfs_put_tree_mod_seq(ctx->fs_info, &seq_elem);
2516 btrfs_end_transaction(ctx->trans);
2517 ctx->trans = NULL;
2518 } else {
2519 up_read(&ctx->fs_info->commit_root_sem);
2520 }
2521
2522 ulist_free(ctx->roots);
2523 ctx->roots = NULL;
2524
2525 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP)
2526 ret = 0;
2527
2528 return ret;
2529}
2530
2531static int build_ino_list(u64 inum, u64 offset, u64 num_bytes, u64 root, void *ctx)
2532{
2533 struct btrfs_data_container *inodes = ctx;
2534 const size_t c = 3 * sizeof(u64);
2535
2536 if (inodes->bytes_left >= c) {
2537 inodes->bytes_left -= c;
2538 inodes->val[inodes->elem_cnt] = inum;
2539 inodes->val[inodes->elem_cnt + 1] = offset;
2540 inodes->val[inodes->elem_cnt + 2] = root;
2541 inodes->elem_cnt += 3;
2542 } else {
2543 inodes->bytes_missing += c - inodes->bytes_left;
2544 inodes->bytes_left = 0;
2545 inodes->elem_missed += 3;
2546 }
2547
2548 return 0;
2549}
2550
2551int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2552 struct btrfs_path *path,
2553 void *ctx, bool ignore_offset)
2554{
2555 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
2556 int ret;
2557 u64 flags = 0;
2558 struct btrfs_key found_key;
2559 int search_commit_root = path->search_commit_root;
2560
2561 ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2562 btrfs_release_path(path);
2563 if (ret < 0)
2564 return ret;
2565 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2566 return -EINVAL;
2567
2568 walk_ctx.bytenr = found_key.objectid;
2569 if (ignore_offset)
2570 walk_ctx.ignore_extent_item_pos = true;
2571 else
2572 walk_ctx.extent_item_pos = logical - found_key.objectid;
2573 walk_ctx.fs_info = fs_info;
2574
2575 return iterate_extent_inodes(&walk_ctx, search_commit_root,
2576 build_ino_list, ctx);
2577}
2578
2579static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2580 struct extent_buffer *eb, struct inode_fs_paths *ipath);
2581
2582static int iterate_inode_refs(u64 inum, struct inode_fs_paths *ipath)
2583{
2584 int ret = 0;
2585 int slot;
2586 u32 cur;
2587 u32 len;
2588 u32 name_len;
2589 u64 parent = 0;
2590 int found = 0;
2591 struct btrfs_root *fs_root = ipath->fs_root;
2592 struct btrfs_path *path = ipath->btrfs_path;
2593 struct extent_buffer *eb;
2594 struct btrfs_inode_ref *iref;
2595 struct btrfs_key found_key;
2596
2597 while (!ret) {
2598 ret = btrfs_find_item(fs_root, path, inum,
2599 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2600 &found_key);
2601
2602 if (ret < 0)
2603 break;
2604 if (ret) {
2605 ret = found ? 0 : -ENOENT;
2606 break;
2607 }
2608 ++found;
2609
2610 parent = found_key.offset;
2611 slot = path->slots[0];
2612 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2613 if (!eb) {
2614 ret = -ENOMEM;
2615 break;
2616 }
2617 btrfs_release_path(path);
2618
2619 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2620
2621 for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2622 name_len = btrfs_inode_ref_name_len(eb, iref);
2623 /* path must be released before calling iterate()! */
2624 btrfs_debug(fs_root->fs_info,
2625 "following ref at offset %u for inode %llu in tree %llu",
2626 cur, found_key.objectid,
2627 btrfs_root_id(fs_root));
2628 ret = inode_to_path(parent, name_len,
2629 (unsigned long)(iref + 1), eb, ipath);
2630 if (ret)
2631 break;
2632 len = sizeof(*iref) + name_len;
2633 iref = (struct btrfs_inode_ref *)((char *)iref + len);
2634 }
2635 free_extent_buffer(eb);
2636 }
2637
2638 btrfs_release_path(path);
2639
2640 return ret;
2641}
2642
2643static int iterate_inode_extrefs(u64 inum, struct inode_fs_paths *ipath)
2644{
2645 int ret;
2646 int slot;
2647 u64 offset = 0;
2648 u64 parent;
2649 int found = 0;
2650 struct btrfs_root *fs_root = ipath->fs_root;
2651 struct btrfs_path *path = ipath->btrfs_path;
2652 struct extent_buffer *eb;
2653 struct btrfs_inode_extref *extref;
2654 u32 item_size;
2655 u32 cur_offset;
2656 unsigned long ptr;
2657
2658 while (1) {
2659 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2660 &offset);
2661 if (ret < 0)
2662 break;
2663 if (ret) {
2664 ret = found ? 0 : -ENOENT;
2665 break;
2666 }
2667 ++found;
2668
2669 slot = path->slots[0];
2670 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2671 if (!eb) {
2672 ret = -ENOMEM;
2673 break;
2674 }
2675 btrfs_release_path(path);
2676
2677 item_size = btrfs_item_size(eb, slot);
2678 ptr = btrfs_item_ptr_offset(eb, slot);
2679 cur_offset = 0;
2680
2681 while (cur_offset < item_size) {
2682 u32 name_len;
2683
2684 extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2685 parent = btrfs_inode_extref_parent(eb, extref);
2686 name_len = btrfs_inode_extref_name_len(eb, extref);
2687 ret = inode_to_path(parent, name_len,
2688 (unsigned long)&extref->name, eb, ipath);
2689 if (ret)
2690 break;
2691
2692 cur_offset += btrfs_inode_extref_name_len(eb, extref);
2693 cur_offset += sizeof(*extref);
2694 }
2695 free_extent_buffer(eb);
2696
2697 offset++;
2698 }
2699
2700 btrfs_release_path(path);
2701
2702 return ret;
2703}
2704
2705/*
2706 * returns 0 if the path could be dumped (probably truncated)
2707 * returns <0 in case of an error
2708 */
2709static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2710 struct extent_buffer *eb, struct inode_fs_paths *ipath)
2711{
2712 char *fspath;
2713 char *fspath_min;
2714 int i = ipath->fspath->elem_cnt;
2715 const int s_ptr = sizeof(char *);
2716 u32 bytes_left;
2717
2718 bytes_left = ipath->fspath->bytes_left > s_ptr ?
2719 ipath->fspath->bytes_left - s_ptr : 0;
2720
2721 fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2722 fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2723 name_off, eb, inum, fspath_min, bytes_left);
2724 if (IS_ERR(fspath))
2725 return PTR_ERR(fspath);
2726
2727 if (fspath > fspath_min) {
2728 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2729 ++ipath->fspath->elem_cnt;
2730 ipath->fspath->bytes_left = fspath - fspath_min;
2731 } else {
2732 ++ipath->fspath->elem_missed;
2733 ipath->fspath->bytes_missing += fspath_min - fspath;
2734 ipath->fspath->bytes_left = 0;
2735 }
2736
2737 return 0;
2738}
2739
2740/*
2741 * this dumps all file system paths to the inode into the ipath struct, provided
2742 * is has been created large enough. each path is zero-terminated and accessed
2743 * from ipath->fspath->val[i].
2744 * when it returns, there are ipath->fspath->elem_cnt number of paths available
2745 * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2746 * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2747 * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2748 * have been needed to return all paths.
2749 */
2750int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2751{
2752 int ret;
2753 int found_refs = 0;
2754
2755 ret = iterate_inode_refs(inum, ipath);
2756 if (!ret)
2757 ++found_refs;
2758 else if (ret != -ENOENT)
2759 return ret;
2760
2761 ret = iterate_inode_extrefs(inum, ipath);
2762 if (ret == -ENOENT && found_refs)
2763 return 0;
2764
2765 return ret;
2766}
2767
2768struct btrfs_data_container *init_data_container(u32 total_bytes)
2769{
2770 struct btrfs_data_container *data;
2771 size_t alloc_bytes;
2772
2773 alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2774 data = kvzalloc(alloc_bytes, GFP_KERNEL);
2775 if (!data)
2776 return ERR_PTR(-ENOMEM);
2777
2778 if (total_bytes >= sizeof(*data))
2779 data->bytes_left = total_bytes - sizeof(*data);
2780 else
2781 data->bytes_missing = sizeof(*data) - total_bytes;
2782
2783 return data;
2784}
2785
2786/*
2787 * allocates space to return multiple file system paths for an inode.
2788 * total_bytes to allocate are passed, note that space usable for actual path
2789 * information will be total_bytes - sizeof(struct inode_fs_paths).
2790 * the returned pointer must be freed with free_ipath() in the end.
2791 */
2792struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2793 struct btrfs_path *path)
2794{
2795 struct inode_fs_paths *ifp;
2796 struct btrfs_data_container *fspath;
2797
2798 fspath = init_data_container(total_bytes);
2799 if (IS_ERR(fspath))
2800 return ERR_CAST(fspath);
2801
2802 ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2803 if (!ifp) {
2804 kvfree(fspath);
2805 return ERR_PTR(-ENOMEM);
2806 }
2807
2808 ifp->btrfs_path = path;
2809 ifp->fspath = fspath;
2810 ifp->fs_root = fs_root;
2811
2812 return ifp;
2813}
2814
2815void free_ipath(struct inode_fs_paths *ipath)
2816{
2817 if (!ipath)
2818 return;
2819 kvfree(ipath->fspath);
2820 kfree(ipath);
2821}
2822
2823struct btrfs_backref_iter *btrfs_backref_iter_alloc(struct btrfs_fs_info *fs_info)
2824{
2825 struct btrfs_backref_iter *ret;
2826
2827 ret = kzalloc(sizeof(*ret), GFP_NOFS);
2828 if (!ret)
2829 return NULL;
2830
2831 ret->path = btrfs_alloc_path();
2832 if (!ret->path) {
2833 kfree(ret);
2834 return NULL;
2835 }
2836
2837 /* Current backref iterator only supports iteration in commit root */
2838 ret->path->search_commit_root = 1;
2839 ret->path->skip_locking = 1;
2840 ret->fs_info = fs_info;
2841
2842 return ret;
2843}
2844
2845static void btrfs_backref_iter_release(struct btrfs_backref_iter *iter)
2846{
2847 iter->bytenr = 0;
2848 iter->item_ptr = 0;
2849 iter->cur_ptr = 0;
2850 iter->end_ptr = 0;
2851 btrfs_release_path(iter->path);
2852 memset(&iter->cur_key, 0, sizeof(iter->cur_key));
2853}
2854
2855int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2856{
2857 struct btrfs_fs_info *fs_info = iter->fs_info;
2858 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2859 struct btrfs_path *path = iter->path;
2860 struct btrfs_extent_item *ei;
2861 struct btrfs_key key;
2862 int ret;
2863
2864 key.objectid = bytenr;
2865 key.type = BTRFS_METADATA_ITEM_KEY;
2866 key.offset = (u64)-1;
2867 iter->bytenr = bytenr;
2868
2869 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2870 if (ret < 0)
2871 return ret;
2872 if (ret == 0) {
2873 /*
2874 * Key with offset -1 found, there would have to exist an extent
2875 * item with such offset, but this is out of the valid range.
2876 */
2877 ret = -EUCLEAN;
2878 goto release;
2879 }
2880 if (path->slots[0] == 0) {
2881 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2882 ret = -EUCLEAN;
2883 goto release;
2884 }
2885 path->slots[0]--;
2886
2887 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2888 if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2889 key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2890 ret = -ENOENT;
2891 goto release;
2892 }
2893 memcpy(&iter->cur_key, &key, sizeof(key));
2894 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2895 path->slots[0]);
2896 iter->end_ptr = (u32)(iter->item_ptr +
2897 btrfs_item_size(path->nodes[0], path->slots[0]));
2898 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2899 struct btrfs_extent_item);
2900
2901 /*
2902 * Only support iteration on tree backref yet.
2903 *
2904 * This is an extra precaution for non skinny-metadata, where
2905 * EXTENT_ITEM is also used for tree blocks, that we can only use
2906 * extent flags to determine if it's a tree block.
2907 */
2908 if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2909 ret = -ENOTSUPP;
2910 goto release;
2911 }
2912 iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2913
2914 /* If there is no inline backref, go search for keyed backref */
2915 if (iter->cur_ptr >= iter->end_ptr) {
2916 ret = btrfs_next_item(extent_root, path);
2917
2918 /* No inline nor keyed ref */
2919 if (ret > 0) {
2920 ret = -ENOENT;
2921 goto release;
2922 }
2923 if (ret < 0)
2924 goto release;
2925
2926 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2927 path->slots[0]);
2928 if (iter->cur_key.objectid != bytenr ||
2929 (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2930 iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2931 ret = -ENOENT;
2932 goto release;
2933 }
2934 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2935 path->slots[0]);
2936 iter->item_ptr = iter->cur_ptr;
2937 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2938 path->nodes[0], path->slots[0]));
2939 }
2940
2941 return 0;
2942release:
2943 btrfs_backref_iter_release(iter);
2944 return ret;
2945}
2946
2947static bool btrfs_backref_iter_is_inline_ref(struct btrfs_backref_iter *iter)
2948{
2949 if (iter->cur_key.type == BTRFS_EXTENT_ITEM_KEY ||
2950 iter->cur_key.type == BTRFS_METADATA_ITEM_KEY)
2951 return true;
2952 return false;
2953}
2954
2955/*
2956 * Go to the next backref item of current bytenr, can be either inlined or
2957 * keyed.
2958 *
2959 * Caller needs to check whether it's inline ref or not by iter->cur_key.
2960 *
2961 * Return 0 if we get next backref without problem.
2962 * Return >0 if there is no extra backref for this bytenr.
2963 * Return <0 if there is something wrong happened.
2964 */
2965int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2966{
2967 struct extent_buffer *eb = iter->path->nodes[0];
2968 struct btrfs_root *extent_root;
2969 struct btrfs_path *path = iter->path;
2970 struct btrfs_extent_inline_ref *iref;
2971 int ret;
2972 u32 size;
2973
2974 if (btrfs_backref_iter_is_inline_ref(iter)) {
2975 /* We're still inside the inline refs */
2976 ASSERT(iter->cur_ptr < iter->end_ptr);
2977
2978 if (btrfs_backref_has_tree_block_info(iter)) {
2979 /* First tree block info */
2980 size = sizeof(struct btrfs_tree_block_info);
2981 } else {
2982 /* Use inline ref type to determine the size */
2983 int type;
2984
2985 iref = (struct btrfs_extent_inline_ref *)
2986 ((unsigned long)iter->cur_ptr);
2987 type = btrfs_extent_inline_ref_type(eb, iref);
2988
2989 size = btrfs_extent_inline_ref_size(type);
2990 }
2991 iter->cur_ptr += size;
2992 if (iter->cur_ptr < iter->end_ptr)
2993 return 0;
2994
2995 /* All inline items iterated, fall through */
2996 }
2997
2998 /* We're at keyed items, there is no inline item, go to the next one */
2999 extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
3000 ret = btrfs_next_item(extent_root, iter->path);
3001 if (ret)
3002 return ret;
3003
3004 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
3005 if (iter->cur_key.objectid != iter->bytenr ||
3006 (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
3007 iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
3008 return 1;
3009 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
3010 path->slots[0]);
3011 iter->cur_ptr = iter->item_ptr;
3012 iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
3013 path->slots[0]);
3014 return 0;
3015}
3016
3017void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
3018 struct btrfs_backref_cache *cache, bool is_reloc)
3019{
3020 int i;
3021
3022 cache->rb_root = RB_ROOT;
3023 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3024 INIT_LIST_HEAD(&cache->pending[i]);
3025 INIT_LIST_HEAD(&cache->changed);
3026 INIT_LIST_HEAD(&cache->detached);
3027 INIT_LIST_HEAD(&cache->leaves);
3028 INIT_LIST_HEAD(&cache->pending_edge);
3029 INIT_LIST_HEAD(&cache->useless_node);
3030 cache->fs_info = fs_info;
3031 cache->is_reloc = is_reloc;
3032}
3033
3034struct btrfs_backref_node *btrfs_backref_alloc_node(
3035 struct btrfs_backref_cache *cache, u64 bytenr, int level)
3036{
3037 struct btrfs_backref_node *node;
3038
3039 ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
3040 node = kzalloc(sizeof(*node), GFP_NOFS);
3041 if (!node)
3042 return node;
3043
3044 INIT_LIST_HEAD(&node->list);
3045 INIT_LIST_HEAD(&node->upper);
3046 INIT_LIST_HEAD(&node->lower);
3047 RB_CLEAR_NODE(&node->rb_node);
3048 cache->nr_nodes++;
3049 node->level = level;
3050 node->bytenr = bytenr;
3051
3052 return node;
3053}
3054
3055void btrfs_backref_free_node(struct btrfs_backref_cache *cache,
3056 struct btrfs_backref_node *node)
3057{
3058 if (node) {
3059 ASSERT(list_empty(&node->list));
3060 ASSERT(list_empty(&node->lower));
3061 ASSERT(node->eb == NULL);
3062 cache->nr_nodes--;
3063 btrfs_put_root(node->root);
3064 kfree(node);
3065 }
3066}
3067
3068struct btrfs_backref_edge *btrfs_backref_alloc_edge(
3069 struct btrfs_backref_cache *cache)
3070{
3071 struct btrfs_backref_edge *edge;
3072
3073 edge = kzalloc(sizeof(*edge), GFP_NOFS);
3074 if (edge)
3075 cache->nr_edges++;
3076 return edge;
3077}
3078
3079void btrfs_backref_free_edge(struct btrfs_backref_cache *cache,
3080 struct btrfs_backref_edge *edge)
3081{
3082 if (edge) {
3083 cache->nr_edges--;
3084 kfree(edge);
3085 }
3086}
3087
3088void btrfs_backref_unlock_node_buffer(struct btrfs_backref_node *node)
3089{
3090 if (node->locked) {
3091 btrfs_tree_unlock(node->eb);
3092 node->locked = 0;
3093 }
3094}
3095
3096void btrfs_backref_drop_node_buffer(struct btrfs_backref_node *node)
3097{
3098 if (node->eb) {
3099 btrfs_backref_unlock_node_buffer(node);
3100 free_extent_buffer(node->eb);
3101 node->eb = NULL;
3102 }
3103}
3104
3105/*
3106 * Drop the backref node from cache without cleaning up its children
3107 * edges.
3108 *
3109 * This can only be called on node without parent edges.
3110 * The children edges are still kept as is.
3111 */
3112void btrfs_backref_drop_node(struct btrfs_backref_cache *tree,
3113 struct btrfs_backref_node *node)
3114{
3115 ASSERT(list_empty(&node->upper));
3116
3117 btrfs_backref_drop_node_buffer(node);
3118 list_del_init(&node->list);
3119 list_del_init(&node->lower);
3120 if (!RB_EMPTY_NODE(&node->rb_node))
3121 rb_erase(&node->rb_node, &tree->rb_root);
3122 btrfs_backref_free_node(tree, node);
3123}
3124
3125/*
3126 * Drop the backref node from cache, also cleaning up all its
3127 * upper edges and any uncached nodes in the path.
3128 *
3129 * This cleanup happens bottom up, thus the node should either
3130 * be the lowest node in the cache or a detached node.
3131 */
3132void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
3133 struct btrfs_backref_node *node)
3134{
3135 struct btrfs_backref_node *upper;
3136 struct btrfs_backref_edge *edge;
3137
3138 if (!node)
3139 return;
3140
3141 BUG_ON(!node->lowest && !node->detached);
3142 while (!list_empty(&node->upper)) {
3143 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
3144 list[LOWER]);
3145 upper = edge->node[UPPER];
3146 list_del(&edge->list[LOWER]);
3147 list_del(&edge->list[UPPER]);
3148 btrfs_backref_free_edge(cache, edge);
3149
3150 /*
3151 * Add the node to leaf node list if no other child block
3152 * cached.
3153 */
3154 if (list_empty(&upper->lower)) {
3155 list_add_tail(&upper->lower, &cache->leaves);
3156 upper->lowest = 1;
3157 }
3158 }
3159
3160 btrfs_backref_drop_node(cache, node);
3161}
3162
3163/*
3164 * Release all nodes/edges from current cache
3165 */
3166void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
3167{
3168 struct btrfs_backref_node *node;
3169 int i;
3170
3171 while (!list_empty(&cache->detached)) {
3172 node = list_entry(cache->detached.next,
3173 struct btrfs_backref_node, list);
3174 btrfs_backref_cleanup_node(cache, node);
3175 }
3176
3177 while (!list_empty(&cache->leaves)) {
3178 node = list_entry(cache->leaves.next,
3179 struct btrfs_backref_node, lower);
3180 btrfs_backref_cleanup_node(cache, node);
3181 }
3182
3183 for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
3184 while (!list_empty(&cache->pending[i])) {
3185 node = list_first_entry(&cache->pending[i],
3186 struct btrfs_backref_node,
3187 list);
3188 btrfs_backref_cleanup_node(cache, node);
3189 }
3190 }
3191 ASSERT(list_empty(&cache->pending_edge));
3192 ASSERT(list_empty(&cache->useless_node));
3193 ASSERT(list_empty(&cache->changed));
3194 ASSERT(list_empty(&cache->detached));
3195 ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
3196 ASSERT(!cache->nr_nodes);
3197 ASSERT(!cache->nr_edges);
3198}
3199
3200void btrfs_backref_link_edge(struct btrfs_backref_edge *edge,
3201 struct btrfs_backref_node *lower,
3202 struct btrfs_backref_node *upper,
3203 int link_which)
3204{
3205 ASSERT(upper && lower && upper->level == lower->level + 1);
3206 edge->node[LOWER] = lower;
3207 edge->node[UPPER] = upper;
3208 if (link_which & LINK_LOWER)
3209 list_add_tail(&edge->list[LOWER], &lower->upper);
3210 if (link_which & LINK_UPPER)
3211 list_add_tail(&edge->list[UPPER], &upper->lower);
3212}
3213/*
3214 * Handle direct tree backref
3215 *
3216 * Direct tree backref means, the backref item shows its parent bytenr
3217 * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
3218 *
3219 * @ref_key: The converted backref key.
3220 * For keyed backref, it's the item key.
3221 * For inlined backref, objectid is the bytenr,
3222 * type is btrfs_inline_ref_type, offset is
3223 * btrfs_inline_ref_offset.
3224 */
3225static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
3226 struct btrfs_key *ref_key,
3227 struct btrfs_backref_node *cur)
3228{
3229 struct btrfs_backref_edge *edge;
3230 struct btrfs_backref_node *upper;
3231 struct rb_node *rb_node;
3232
3233 ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
3234
3235 /* Only reloc root uses backref pointing to itself */
3236 if (ref_key->objectid == ref_key->offset) {
3237 struct btrfs_root *root;
3238
3239 cur->is_reloc_root = 1;
3240 /* Only reloc backref cache cares about a specific root */
3241 if (cache->is_reloc) {
3242 root = find_reloc_root(cache->fs_info, cur->bytenr);
3243 if (!root)
3244 return -ENOENT;
3245 cur->root = root;
3246 } else {
3247 /*
3248 * For generic purpose backref cache, reloc root node
3249 * is useless.
3250 */
3251 list_add(&cur->list, &cache->useless_node);
3252 }
3253 return 0;
3254 }
3255
3256 edge = btrfs_backref_alloc_edge(cache);
3257 if (!edge)
3258 return -ENOMEM;
3259
3260 rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
3261 if (!rb_node) {
3262 /* Parent node not yet cached */
3263 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
3264 cur->level + 1);
3265 if (!upper) {
3266 btrfs_backref_free_edge(cache, edge);
3267 return -ENOMEM;
3268 }
3269
3270 /*
3271 * Backrefs for the upper level block isn't cached, add the
3272 * block to pending list
3273 */
3274 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3275 } else {
3276 /* Parent node already cached */
3277 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
3278 ASSERT(upper->checked);
3279 INIT_LIST_HEAD(&edge->list[UPPER]);
3280 }
3281 btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
3282 return 0;
3283}
3284
3285/*
3286 * Handle indirect tree backref
3287 *
3288 * Indirect tree backref means, we only know which tree the node belongs to.
3289 * We still need to do a tree search to find out the parents. This is for
3290 * TREE_BLOCK_REF backref (keyed or inlined).
3291 *
3292 * @trans: Transaction handle.
3293 * @ref_key: The same as @ref_key in handle_direct_tree_backref()
3294 * @tree_key: The first key of this tree block.
3295 * @path: A clean (released) path, to avoid allocating path every time
3296 * the function get called.
3297 */
3298static int handle_indirect_tree_backref(struct btrfs_trans_handle *trans,
3299 struct btrfs_backref_cache *cache,
3300 struct btrfs_path *path,
3301 struct btrfs_key *ref_key,
3302 struct btrfs_key *tree_key,
3303 struct btrfs_backref_node *cur)
3304{
3305 struct btrfs_fs_info *fs_info = cache->fs_info;
3306 struct btrfs_backref_node *upper;
3307 struct btrfs_backref_node *lower;
3308 struct btrfs_backref_edge *edge;
3309 struct extent_buffer *eb;
3310 struct btrfs_root *root;
3311 struct rb_node *rb_node;
3312 int level;
3313 bool need_check = true;
3314 int ret;
3315
3316 root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
3317 if (IS_ERR(root))
3318 return PTR_ERR(root);
3319 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3320 cur->cowonly = 1;
3321
3322 if (btrfs_root_level(&root->root_item) == cur->level) {
3323 /* Tree root */
3324 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
3325 /*
3326 * For reloc backref cache, we may ignore reloc root. But for
3327 * general purpose backref cache, we can't rely on
3328 * btrfs_should_ignore_reloc_root() as it may conflict with
3329 * current running relocation and lead to missing root.
3330 *
3331 * For general purpose backref cache, reloc root detection is
3332 * completely relying on direct backref (key->offset is parent
3333 * bytenr), thus only do such check for reloc cache.
3334 */
3335 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
3336 btrfs_put_root(root);
3337 list_add(&cur->list, &cache->useless_node);
3338 } else {
3339 cur->root = root;
3340 }
3341 return 0;
3342 }
3343
3344 level = cur->level + 1;
3345
3346 /* Search the tree to find parent blocks referring to the block */
3347 path->search_commit_root = 1;
3348 path->skip_locking = 1;
3349 path->lowest_level = level;
3350 ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
3351 path->lowest_level = 0;
3352 if (ret < 0) {
3353 btrfs_put_root(root);
3354 return ret;
3355 }
3356 if (ret > 0 && path->slots[level] > 0)
3357 path->slots[level]--;
3358
3359 eb = path->nodes[level];
3360 if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
3361 btrfs_err(fs_info,
3362"couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
3363 cur->bytenr, level - 1, btrfs_root_id(root),
3364 tree_key->objectid, tree_key->type, tree_key->offset);
3365 btrfs_put_root(root);
3366 ret = -ENOENT;
3367 goto out;
3368 }
3369 lower = cur;
3370
3371 /* Add all nodes and edges in the path */
3372 for (; level < BTRFS_MAX_LEVEL; level++) {
3373 if (!path->nodes[level]) {
3374 ASSERT(btrfs_root_bytenr(&root->root_item) ==
3375 lower->bytenr);
3376 /* Same as previous should_ignore_reloc_root() call */
3377 if (btrfs_should_ignore_reloc_root(root) &&
3378 cache->is_reloc) {
3379 btrfs_put_root(root);
3380 list_add(&lower->list, &cache->useless_node);
3381 } else {
3382 lower->root = root;
3383 }
3384 break;
3385 }
3386
3387 edge = btrfs_backref_alloc_edge(cache);
3388 if (!edge) {
3389 btrfs_put_root(root);
3390 ret = -ENOMEM;
3391 goto out;
3392 }
3393
3394 eb = path->nodes[level];
3395 rb_node = rb_simple_search(&cache->rb_root, eb->start);
3396 if (!rb_node) {
3397 upper = btrfs_backref_alloc_node(cache, eb->start,
3398 lower->level + 1);
3399 if (!upper) {
3400 btrfs_put_root(root);
3401 btrfs_backref_free_edge(cache, edge);
3402 ret = -ENOMEM;
3403 goto out;
3404 }
3405 upper->owner = btrfs_header_owner(eb);
3406 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3407 upper->cowonly = 1;
3408
3409 /*
3410 * If we know the block isn't shared we can avoid
3411 * checking its backrefs.
3412 */
3413 if (btrfs_block_can_be_shared(trans, root, eb))
3414 upper->checked = 0;
3415 else
3416 upper->checked = 1;
3417
3418 /*
3419 * Add the block to pending list if we need to check its
3420 * backrefs, we only do this once while walking up a
3421 * tree as we will catch anything else later on.
3422 */
3423 if (!upper->checked && need_check) {
3424 need_check = false;
3425 list_add_tail(&edge->list[UPPER],
3426 &cache->pending_edge);
3427 } else {
3428 if (upper->checked)
3429 need_check = true;
3430 INIT_LIST_HEAD(&edge->list[UPPER]);
3431 }
3432 } else {
3433 upper = rb_entry(rb_node, struct btrfs_backref_node,
3434 rb_node);
3435 ASSERT(upper->checked);
3436 INIT_LIST_HEAD(&edge->list[UPPER]);
3437 if (!upper->owner)
3438 upper->owner = btrfs_header_owner(eb);
3439 }
3440 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
3441
3442 if (rb_node) {
3443 btrfs_put_root(root);
3444 break;
3445 }
3446 lower = upper;
3447 upper = NULL;
3448 }
3449out:
3450 btrfs_release_path(path);
3451 return ret;
3452}
3453
3454/*
3455 * Add backref node @cur into @cache.
3456 *
3457 * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
3458 * links aren't yet bi-directional. Needs to finish such links.
3459 * Use btrfs_backref_finish_upper_links() to finish such linkage.
3460 *
3461 * @trans: Transaction handle.
3462 * @path: Released path for indirect tree backref lookup
3463 * @iter: Released backref iter for extent tree search
3464 * @node_key: The first key of the tree block
3465 */
3466int btrfs_backref_add_tree_node(struct btrfs_trans_handle *trans,
3467 struct btrfs_backref_cache *cache,
3468 struct btrfs_path *path,
3469 struct btrfs_backref_iter *iter,
3470 struct btrfs_key *node_key,
3471 struct btrfs_backref_node *cur)
3472{
3473 struct btrfs_backref_edge *edge;
3474 struct btrfs_backref_node *exist;
3475 int ret;
3476
3477 ret = btrfs_backref_iter_start(iter, cur->bytenr);
3478 if (ret < 0)
3479 return ret;
3480 /*
3481 * We skip the first btrfs_tree_block_info, as we don't use the key
3482 * stored in it, but fetch it from the tree block
3483 */
3484 if (btrfs_backref_has_tree_block_info(iter)) {
3485 ret = btrfs_backref_iter_next(iter);
3486 if (ret < 0)
3487 goto out;
3488 /* No extra backref? This means the tree block is corrupted */
3489 if (ret > 0) {
3490 ret = -EUCLEAN;
3491 goto out;
3492 }
3493 }
3494 WARN_ON(cur->checked);
3495 if (!list_empty(&cur->upper)) {
3496 /*
3497 * The backref was added previously when processing backref of
3498 * type BTRFS_TREE_BLOCK_REF_KEY
3499 */
3500 ASSERT(list_is_singular(&cur->upper));
3501 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
3502 list[LOWER]);
3503 ASSERT(list_empty(&edge->list[UPPER]));
3504 exist = edge->node[UPPER];
3505 /*
3506 * Add the upper level block to pending list if we need check
3507 * its backrefs
3508 */
3509 if (!exist->checked)
3510 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3511 } else {
3512 exist = NULL;
3513 }
3514
3515 for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
3516 struct extent_buffer *eb;
3517 struct btrfs_key key;
3518 int type;
3519
3520 cond_resched();
3521 eb = iter->path->nodes[0];
3522
3523 key.objectid = iter->bytenr;
3524 if (btrfs_backref_iter_is_inline_ref(iter)) {
3525 struct btrfs_extent_inline_ref *iref;
3526
3527 /* Update key for inline backref */
3528 iref = (struct btrfs_extent_inline_ref *)
3529 ((unsigned long)iter->cur_ptr);
3530 type = btrfs_get_extent_inline_ref_type(eb, iref,
3531 BTRFS_REF_TYPE_BLOCK);
3532 if (type == BTRFS_REF_TYPE_INVALID) {
3533 ret = -EUCLEAN;
3534 goto out;
3535 }
3536 key.type = type;
3537 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3538 } else {
3539 key.type = iter->cur_key.type;
3540 key.offset = iter->cur_key.offset;
3541 }
3542
3543 /*
3544 * Parent node found and matches current inline ref, no need to
3545 * rebuild this node for this inline ref
3546 */
3547 if (exist &&
3548 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
3549 exist->owner == key.offset) ||
3550 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
3551 exist->bytenr == key.offset))) {
3552 exist = NULL;
3553 continue;
3554 }
3555
3556 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
3557 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3558 ret = handle_direct_tree_backref(cache, &key, cur);
3559 if (ret < 0)
3560 goto out;
3561 } else if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
3562 /*
3563 * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref
3564 * offset means the root objectid. We need to search
3565 * the tree to get its parent bytenr.
3566 */
3567 ret = handle_indirect_tree_backref(trans, cache, path,
3568 &key, node_key, cur);
3569 if (ret < 0)
3570 goto out;
3571 }
3572 /*
3573 * Unrecognized tree backref items (if it can pass tree-checker)
3574 * would be ignored.
3575 */
3576 }
3577 ret = 0;
3578 cur->checked = 1;
3579 WARN_ON(exist);
3580out:
3581 btrfs_backref_iter_release(iter);
3582 return ret;
3583}
3584
3585/*
3586 * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3587 */
3588int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3589 struct btrfs_backref_node *start)
3590{
3591 struct list_head *useless_node = &cache->useless_node;
3592 struct btrfs_backref_edge *edge;
3593 struct rb_node *rb_node;
3594 LIST_HEAD(pending_edge);
3595
3596 ASSERT(start->checked);
3597
3598 /* Insert this node to cache if it's not COW-only */
3599 if (!start->cowonly) {
3600 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
3601 &start->rb_node);
3602 if (rb_node)
3603 btrfs_backref_panic(cache->fs_info, start->bytenr,
3604 -EEXIST);
3605 list_add_tail(&start->lower, &cache->leaves);
3606 }
3607
3608 /*
3609 * Use breadth first search to iterate all related edges.
3610 *
3611 * The starting points are all the edges of this node
3612 */
3613 list_for_each_entry(edge, &start->upper, list[LOWER])
3614 list_add_tail(&edge->list[UPPER], &pending_edge);
3615
3616 while (!list_empty(&pending_edge)) {
3617 struct btrfs_backref_node *upper;
3618 struct btrfs_backref_node *lower;
3619
3620 edge = list_first_entry(&pending_edge,
3621 struct btrfs_backref_edge, list[UPPER]);
3622 list_del_init(&edge->list[UPPER]);
3623 upper = edge->node[UPPER];
3624 lower = edge->node[LOWER];
3625
3626 /* Parent is detached, no need to keep any edges */
3627 if (upper->detached) {
3628 list_del(&edge->list[LOWER]);
3629 btrfs_backref_free_edge(cache, edge);
3630
3631 /* Lower node is orphan, queue for cleanup */
3632 if (list_empty(&lower->upper))
3633 list_add(&lower->list, useless_node);
3634 continue;
3635 }
3636
3637 /*
3638 * All new nodes added in current build_backref_tree() haven't
3639 * been linked to the cache rb tree.
3640 * So if we have upper->rb_node populated, this means a cache
3641 * hit. We only need to link the edge, as @upper and all its
3642 * parents have already been linked.
3643 */
3644 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3645 if (upper->lowest) {
3646 list_del_init(&upper->lower);
3647 upper->lowest = 0;
3648 }
3649
3650 list_add_tail(&edge->list[UPPER], &upper->lower);
3651 continue;
3652 }
3653
3654 /* Sanity check, we shouldn't have any unchecked nodes */
3655 if (!upper->checked) {
3656 ASSERT(0);
3657 return -EUCLEAN;
3658 }
3659
3660 /* Sanity check, COW-only node has non-COW-only parent */
3661 if (start->cowonly != upper->cowonly) {
3662 ASSERT(0);
3663 return -EUCLEAN;
3664 }
3665
3666 /* Only cache non-COW-only (subvolume trees) tree blocks */
3667 if (!upper->cowonly) {
3668 rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3669 &upper->rb_node);
3670 if (rb_node) {
3671 btrfs_backref_panic(cache->fs_info,
3672 upper->bytenr, -EEXIST);
3673 return -EUCLEAN;
3674 }
3675 }
3676
3677 list_add_tail(&edge->list[UPPER], &upper->lower);
3678
3679 /*
3680 * Also queue all the parent edges of this uncached node
3681 * to finish the upper linkage
3682 */
3683 list_for_each_entry(edge, &upper->upper, list[LOWER])
3684 list_add_tail(&edge->list[UPPER], &pending_edge);
3685 }
3686 return 0;
3687}
3688
3689void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3690 struct btrfs_backref_node *node)
3691{
3692 struct btrfs_backref_node *lower;
3693 struct btrfs_backref_node *upper;
3694 struct btrfs_backref_edge *edge;
3695
3696 while (!list_empty(&cache->useless_node)) {
3697 lower = list_first_entry(&cache->useless_node,
3698 struct btrfs_backref_node, list);
3699 list_del_init(&lower->list);
3700 }
3701 while (!list_empty(&cache->pending_edge)) {
3702 edge = list_first_entry(&cache->pending_edge,
3703 struct btrfs_backref_edge, list[UPPER]);
3704 list_del(&edge->list[UPPER]);
3705 list_del(&edge->list[LOWER]);
3706 lower = edge->node[LOWER];
3707 upper = edge->node[UPPER];
3708 btrfs_backref_free_edge(cache, edge);
3709
3710 /*
3711 * Lower is no longer linked to any upper backref nodes and
3712 * isn't in the cache, we can free it ourselves.
3713 */
3714 if (list_empty(&lower->upper) &&
3715 RB_EMPTY_NODE(&lower->rb_node))
3716 list_add(&lower->list, &cache->useless_node);
3717
3718 if (!RB_EMPTY_NODE(&upper->rb_node))
3719 continue;
3720
3721 /* Add this guy's upper edges to the list to process */
3722 list_for_each_entry(edge, &upper->upper, list[LOWER])
3723 list_add_tail(&edge->list[UPPER],
3724 &cache->pending_edge);
3725 if (list_empty(&upper->upper))
3726 list_add(&upper->list, &cache->useless_node);
3727 }
3728
3729 while (!list_empty(&cache->useless_node)) {
3730 lower = list_first_entry(&cache->useless_node,
3731 struct btrfs_backref_node, list);
3732 list_del_init(&lower->list);
3733 if (lower == node)
3734 node = NULL;
3735 btrfs_backref_drop_node(cache, lower);
3736 }
3737
3738 btrfs_backref_cleanup_node(cache, node);
3739 ASSERT(list_empty(&cache->useless_node) &&
3740 list_empty(&cache->pending_edge));
3741}