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 (!btrfs_file_extent_compression(eb, fi) &&
49 !btrfs_file_extent_encryption(eb, fi) &&
50 !btrfs_file_extent_other_encoding(eb, fi)) {
51 u64 data_offset;
52
53 data_offset = btrfs_file_extent_offset(eb, fi);
54
55 if (ctx->extent_item_pos < data_offset ||
56 ctx->extent_item_pos >= data_offset + data_len)
57 return 1;
58 offset += ctx->extent_item_pos - data_offset;
59 }
60
61 if (!ctx->indirect_ref_iterator || !ctx->cache_lookup)
62 goto add_inode_elem;
63
64 cached = ctx->cache_lookup(eb->start, ctx->user_ctx, &root_ids,
65 &root_count);
66 if (!cached)
67 goto add_inode_elem;
68
69 for (int i = 0; i < root_count; i++) {
70 int ret;
71
72 ret = ctx->indirect_ref_iterator(key->objectid, offset,
73 data_len, root_ids[i],
74 ctx->user_ctx);
75 if (ret)
76 return ret;
77 }
78
79add_inode_elem:
80 e = kmalloc(sizeof(*e), GFP_NOFS);
81 if (!e)
82 return -ENOMEM;
83
84 e->next = *eie;
85 e->inum = key->objectid;
86 e->offset = offset;
87 e->num_bytes = data_len;
88 *eie = e;
89
90 return 0;
91}
92
93static void free_inode_elem_list(struct extent_inode_elem *eie)
94{
95 struct extent_inode_elem *eie_next;
96
97 for (; eie; eie = eie_next) {
98 eie_next = eie->next;
99 kfree(eie);
100 }
101}
102
103static int find_extent_in_eb(struct btrfs_backref_walk_ctx *ctx,
104 const struct extent_buffer *eb,
105 struct extent_inode_elem **eie)
106{
107 u64 disk_byte;
108 struct btrfs_key key;
109 struct btrfs_file_extent_item *fi;
110 int slot;
111 int nritems;
112 int extent_type;
113 int ret;
114
115 /*
116 * from the shared data ref, we only have the leaf but we need
117 * the key. thus, we must look into all items and see that we
118 * find one (some) with a reference to our extent item.
119 */
120 nritems = btrfs_header_nritems(eb);
121 for (slot = 0; slot < nritems; ++slot) {
122 btrfs_item_key_to_cpu(eb, &key, slot);
123 if (key.type != BTRFS_EXTENT_DATA_KEY)
124 continue;
125 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
126 extent_type = btrfs_file_extent_type(eb, fi);
127 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
128 continue;
129 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
130 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
131 if (disk_byte != ctx->bytenr)
132 continue;
133
134 ret = check_extent_in_eb(ctx, &key, eb, fi, eie);
135 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
136 return ret;
137 }
138
139 return 0;
140}
141
142struct preftree {
143 struct rb_root_cached root;
144 unsigned int count;
145};
146
147#define PREFTREE_INIT { .root = RB_ROOT_CACHED, .count = 0 }
148
149struct preftrees {
150 struct preftree direct; /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
151 struct preftree indirect; /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
152 struct preftree indirect_missing_keys;
153};
154
155/*
156 * Checks for a shared extent during backref search.
157 *
158 * The share_count tracks prelim_refs (direct and indirect) having a
159 * ref->count >0:
160 * - incremented when a ref->count transitions to >0
161 * - decremented when a ref->count transitions to <1
162 */
163struct share_check {
164 struct btrfs_backref_share_check_ctx *ctx;
165 struct btrfs_root *root;
166 u64 inum;
167 u64 data_bytenr;
168 u64 data_extent_gen;
169 /*
170 * Counts number of inodes that refer to an extent (different inodes in
171 * the same root or different roots) that we could find. The sharedness
172 * check typically stops once this counter gets greater than 1, so it
173 * may not reflect the total number of inodes.
174 */
175 int share_count;
176 /*
177 * The number of times we found our inode refers to the data extent we
178 * are determining the sharedness. In other words, how many file extent
179 * items we could find for our inode that point to our target data
180 * extent. The value we get here after finishing the extent sharedness
181 * check may be smaller than reality, but if it ends up being greater
182 * than 1, then we know for sure the inode has multiple file extent
183 * items that point to our inode, and we can safely assume it's useful
184 * to cache the sharedness check result.
185 */
186 int self_ref_count;
187 bool have_delayed_delete_refs;
188};
189
190static inline int extent_is_shared(struct share_check *sc)
191{
192 return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
193}
194
195static struct kmem_cache *btrfs_prelim_ref_cache;
196
197int __init btrfs_prelim_ref_init(void)
198{
199 btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
200 sizeof(struct prelim_ref),
201 0,
202 SLAB_MEM_SPREAD,
203 NULL);
204 if (!btrfs_prelim_ref_cache)
205 return -ENOMEM;
206 return 0;
207}
208
209void __cold btrfs_prelim_ref_exit(void)
210{
211 kmem_cache_destroy(btrfs_prelim_ref_cache);
212}
213
214static void free_pref(struct prelim_ref *ref)
215{
216 kmem_cache_free(btrfs_prelim_ref_cache, ref);
217}
218
219/*
220 * Return 0 when both refs are for the same block (and can be merged).
221 * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
222 * indicates a 'higher' block.
223 */
224static int prelim_ref_compare(struct prelim_ref *ref1,
225 struct prelim_ref *ref2)
226{
227 if (ref1->level < ref2->level)
228 return -1;
229 if (ref1->level > ref2->level)
230 return 1;
231 if (ref1->root_id < ref2->root_id)
232 return -1;
233 if (ref1->root_id > ref2->root_id)
234 return 1;
235 if (ref1->key_for_search.type < ref2->key_for_search.type)
236 return -1;
237 if (ref1->key_for_search.type > ref2->key_for_search.type)
238 return 1;
239 if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
240 return -1;
241 if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
242 return 1;
243 if (ref1->key_for_search.offset < ref2->key_for_search.offset)
244 return -1;
245 if (ref1->key_for_search.offset > ref2->key_for_search.offset)
246 return 1;
247 if (ref1->parent < ref2->parent)
248 return -1;
249 if (ref1->parent > ref2->parent)
250 return 1;
251
252 return 0;
253}
254
255static void update_share_count(struct share_check *sc, int oldcount,
256 int newcount, struct prelim_ref *newref)
257{
258 if ((!sc) || (oldcount == 0 && newcount < 1))
259 return;
260
261 if (oldcount > 0 && newcount < 1)
262 sc->share_count--;
263 else if (oldcount < 1 && newcount > 0)
264 sc->share_count++;
265
266 if (newref->root_id == sc->root->root_key.objectid &&
267 newref->wanted_disk_byte == sc->data_bytenr &&
268 newref->key_for_search.objectid == sc->inum)
269 sc->self_ref_count += newref->count;
270}
271
272/*
273 * Add @newref to the @root rbtree, merging identical refs.
274 *
275 * Callers should assume that newref has been freed after calling.
276 */
277static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
278 struct preftree *preftree,
279 struct prelim_ref *newref,
280 struct share_check *sc)
281{
282 struct rb_root_cached *root;
283 struct rb_node **p;
284 struct rb_node *parent = NULL;
285 struct prelim_ref *ref;
286 int result;
287 bool leftmost = true;
288
289 root = &preftree->root;
290 p = &root->rb_root.rb_node;
291
292 while (*p) {
293 parent = *p;
294 ref = rb_entry(parent, struct prelim_ref, rbnode);
295 result = prelim_ref_compare(ref, newref);
296 if (result < 0) {
297 p = &(*p)->rb_left;
298 } else if (result > 0) {
299 p = &(*p)->rb_right;
300 leftmost = false;
301 } else {
302 /* Identical refs, merge them and free @newref */
303 struct extent_inode_elem *eie = ref->inode_list;
304
305 while (eie && eie->next)
306 eie = eie->next;
307
308 if (!eie)
309 ref->inode_list = newref->inode_list;
310 else
311 eie->next = newref->inode_list;
312 trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
313 preftree->count);
314 /*
315 * A delayed ref can have newref->count < 0.
316 * The ref->count is updated to follow any
317 * BTRFS_[ADD|DROP]_DELAYED_REF actions.
318 */
319 update_share_count(sc, ref->count,
320 ref->count + newref->count, newref);
321 ref->count += newref->count;
322 free_pref(newref);
323 return;
324 }
325 }
326
327 update_share_count(sc, 0, newref->count, newref);
328 preftree->count++;
329 trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
330 rb_link_node(&newref->rbnode, parent, p);
331 rb_insert_color_cached(&newref->rbnode, root, leftmost);
332}
333
334/*
335 * Release the entire tree. We don't care about internal consistency so
336 * just free everything and then reset the tree root.
337 */
338static void prelim_release(struct preftree *preftree)
339{
340 struct prelim_ref *ref, *next_ref;
341
342 rbtree_postorder_for_each_entry_safe(ref, next_ref,
343 &preftree->root.rb_root, rbnode) {
344 free_inode_elem_list(ref->inode_list);
345 free_pref(ref);
346 }
347
348 preftree->root = RB_ROOT_CACHED;
349 preftree->count = 0;
350}
351
352/*
353 * the rules for all callers of this function are:
354 * - obtaining the parent is the goal
355 * - if you add a key, you must know that it is a correct key
356 * - if you cannot add the parent or a correct key, then we will look into the
357 * block later to set a correct key
358 *
359 * delayed refs
360 * ============
361 * backref type | shared | indirect | shared | indirect
362 * information | tree | tree | data | data
363 * --------------------+--------+----------+--------+----------
364 * parent logical | y | - | - | -
365 * key to resolve | - | y | y | y
366 * tree block logical | - | - | - | -
367 * root for resolving | y | y | y | y
368 *
369 * - column 1: we've the parent -> done
370 * - column 2, 3, 4: we use the key to find the parent
371 *
372 * on disk refs (inline or keyed)
373 * ==============================
374 * backref type | shared | indirect | shared | indirect
375 * information | tree | tree | data | data
376 * --------------------+--------+----------+--------+----------
377 * parent logical | y | - | y | -
378 * key to resolve | - | - | - | y
379 * tree block logical | y | y | y | y
380 * root for resolving | - | y | y | y
381 *
382 * - column 1, 3: we've the parent -> done
383 * - column 2: we take the first key from the block to find the parent
384 * (see add_missing_keys)
385 * - column 4: we use the key to find the parent
386 *
387 * additional information that's available but not required to find the parent
388 * block might help in merging entries to gain some speed.
389 */
390static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
391 struct preftree *preftree, u64 root_id,
392 const struct btrfs_key *key, int level, u64 parent,
393 u64 wanted_disk_byte, int count,
394 struct share_check *sc, gfp_t gfp_mask)
395{
396 struct prelim_ref *ref;
397
398 if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
399 return 0;
400
401 ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
402 if (!ref)
403 return -ENOMEM;
404
405 ref->root_id = root_id;
406 if (key)
407 ref->key_for_search = *key;
408 else
409 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
410
411 ref->inode_list = NULL;
412 ref->level = level;
413 ref->count = count;
414 ref->parent = parent;
415 ref->wanted_disk_byte = wanted_disk_byte;
416 prelim_ref_insert(fs_info, preftree, ref, sc);
417 return extent_is_shared(sc);
418}
419
420/* direct refs use root == 0, key == NULL */
421static int add_direct_ref(const struct btrfs_fs_info *fs_info,
422 struct preftrees *preftrees, int level, u64 parent,
423 u64 wanted_disk_byte, int count,
424 struct share_check *sc, gfp_t gfp_mask)
425{
426 return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
427 parent, wanted_disk_byte, count, sc, gfp_mask);
428}
429
430/* indirect refs use parent == 0 */
431static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
432 struct preftrees *preftrees, u64 root_id,
433 const struct btrfs_key *key, int level,
434 u64 wanted_disk_byte, int count,
435 struct share_check *sc, gfp_t gfp_mask)
436{
437 struct preftree *tree = &preftrees->indirect;
438
439 if (!key)
440 tree = &preftrees->indirect_missing_keys;
441 return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
442 wanted_disk_byte, count, sc, gfp_mask);
443}
444
445static int is_shared_data_backref(struct preftrees *preftrees, u64 bytenr)
446{
447 struct rb_node **p = &preftrees->direct.root.rb_root.rb_node;
448 struct rb_node *parent = NULL;
449 struct prelim_ref *ref = NULL;
450 struct prelim_ref target = {};
451 int result;
452
453 target.parent = bytenr;
454
455 while (*p) {
456 parent = *p;
457 ref = rb_entry(parent, struct prelim_ref, rbnode);
458 result = prelim_ref_compare(ref, &target);
459
460 if (result < 0)
461 p = &(*p)->rb_left;
462 else if (result > 0)
463 p = &(*p)->rb_right;
464 else
465 return 1;
466 }
467 return 0;
468}
469
470static int add_all_parents(struct btrfs_backref_walk_ctx *ctx,
471 struct btrfs_root *root, struct btrfs_path *path,
472 struct ulist *parents,
473 struct preftrees *preftrees, struct prelim_ref *ref,
474 int level)
475{
476 int ret = 0;
477 int slot;
478 struct extent_buffer *eb;
479 struct btrfs_key key;
480 struct btrfs_key *key_for_search = &ref->key_for_search;
481 struct btrfs_file_extent_item *fi;
482 struct extent_inode_elem *eie = NULL, *old = NULL;
483 u64 disk_byte;
484 u64 wanted_disk_byte = ref->wanted_disk_byte;
485 u64 count = 0;
486 u64 data_offset;
487 u8 type;
488
489 if (level != 0) {
490 eb = path->nodes[level];
491 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
492 if (ret < 0)
493 return ret;
494 return 0;
495 }
496
497 /*
498 * 1. We normally enter this function with the path already pointing to
499 * the first item to check. But sometimes, we may enter it with
500 * slot == nritems.
501 * 2. We are searching for normal backref but bytenr of this leaf
502 * matches shared data backref
503 * 3. The leaf owner is not equal to the root we are searching
504 *
505 * For these cases, go to the next leaf before we continue.
506 */
507 eb = path->nodes[0];
508 if (path->slots[0] >= btrfs_header_nritems(eb) ||
509 is_shared_data_backref(preftrees, eb->start) ||
510 ref->root_id != btrfs_header_owner(eb)) {
511 if (ctx->time_seq == BTRFS_SEQ_LAST)
512 ret = btrfs_next_leaf(root, path);
513 else
514 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
515 }
516
517 while (!ret && count < ref->count) {
518 eb = path->nodes[0];
519 slot = path->slots[0];
520
521 btrfs_item_key_to_cpu(eb, &key, slot);
522
523 if (key.objectid != key_for_search->objectid ||
524 key.type != BTRFS_EXTENT_DATA_KEY)
525 break;
526
527 /*
528 * We are searching for normal backref but bytenr of this leaf
529 * matches shared data backref, OR
530 * the leaf owner is not equal to the root we are searching for
531 */
532 if (slot == 0 &&
533 (is_shared_data_backref(preftrees, eb->start) ||
534 ref->root_id != btrfs_header_owner(eb))) {
535 if (ctx->time_seq == BTRFS_SEQ_LAST)
536 ret = btrfs_next_leaf(root, path);
537 else
538 ret = btrfs_next_old_leaf(root, path, ctx->time_seq);
539 continue;
540 }
541 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
542 type = btrfs_file_extent_type(eb, fi);
543 if (type == BTRFS_FILE_EXTENT_INLINE)
544 goto next;
545 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
546 data_offset = btrfs_file_extent_offset(eb, fi);
547
548 if (disk_byte == wanted_disk_byte) {
549 eie = NULL;
550 old = NULL;
551 if (ref->key_for_search.offset == key.offset - data_offset)
552 count++;
553 else
554 goto next;
555 if (!ctx->ignore_extent_item_pos) {
556 ret = check_extent_in_eb(ctx, &key, eb, fi, &eie);
557 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
558 ret < 0)
559 break;
560 }
561 if (ret > 0)
562 goto next;
563 ret = ulist_add_merge_ptr(parents, eb->start,
564 eie, (void **)&old, GFP_NOFS);
565 if (ret < 0)
566 break;
567 if (!ret && !ctx->ignore_extent_item_pos) {
568 while (old->next)
569 old = old->next;
570 old->next = eie;
571 }
572 eie = NULL;
573 }
574next:
575 if (ctx->time_seq == BTRFS_SEQ_LAST)
576 ret = btrfs_next_item(root, path);
577 else
578 ret = btrfs_next_old_item(root, path, ctx->time_seq);
579 }
580
581 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
582 free_inode_elem_list(eie);
583 else if (ret > 0)
584 ret = 0;
585
586 return ret;
587}
588
589/*
590 * resolve an indirect backref in the form (root_id, key, level)
591 * to a logical address
592 */
593static int resolve_indirect_ref(struct btrfs_backref_walk_ctx *ctx,
594 struct btrfs_path *path,
595 struct preftrees *preftrees,
596 struct prelim_ref *ref, struct ulist *parents)
597{
598 struct btrfs_root *root;
599 struct extent_buffer *eb;
600 int ret = 0;
601 int root_level;
602 int level = ref->level;
603 struct btrfs_key search_key = ref->key_for_search;
604
605 /*
606 * If we're search_commit_root we could possibly be holding locks on
607 * other tree nodes. This happens when qgroups does backref walks when
608 * adding new delayed refs. To deal with this we need to look in cache
609 * for the root, and if we don't find it then we need to search the
610 * tree_root's commit root, thus the btrfs_get_fs_root_commit_root usage
611 * here.
612 */
613 if (path->search_commit_root)
614 root = btrfs_get_fs_root_commit_root(ctx->fs_info, path, ref->root_id);
615 else
616 root = btrfs_get_fs_root(ctx->fs_info, ref->root_id, false);
617 if (IS_ERR(root)) {
618 ret = PTR_ERR(root);
619 goto out_free;
620 }
621
622 if (!path->search_commit_root &&
623 test_bit(BTRFS_ROOT_DELETING, &root->state)) {
624 ret = -ENOENT;
625 goto out;
626 }
627
628 if (btrfs_is_testing(ctx->fs_info)) {
629 ret = -ENOENT;
630 goto out;
631 }
632
633 if (path->search_commit_root)
634 root_level = btrfs_header_level(root->commit_root);
635 else if (ctx->time_seq == BTRFS_SEQ_LAST)
636 root_level = btrfs_header_level(root->node);
637 else
638 root_level = btrfs_old_root_level(root, ctx->time_seq);
639
640 if (root_level + 1 == level)
641 goto out;
642
643 /*
644 * We can often find data backrefs with an offset that is too large
645 * (>= LLONG_MAX, maximum allowed file offset) due to underflows when
646 * subtracting a file's offset with the data offset of its
647 * corresponding extent data item. This can happen for example in the
648 * clone ioctl.
649 *
650 * So if we detect such case we set the search key's offset to zero to
651 * make sure we will find the matching file extent item at
652 * add_all_parents(), otherwise we will miss it because the offset
653 * taken form the backref is much larger then the offset of the file
654 * extent item. This can make us scan a very large number of file
655 * extent items, but at least it will not make us miss any.
656 *
657 * This is an ugly workaround for a behaviour that should have never
658 * existed, but it does and a fix for the clone ioctl would touch a lot
659 * of places, cause backwards incompatibility and would not fix the
660 * problem for extents cloned with older kernels.
661 */
662 if (search_key.type == BTRFS_EXTENT_DATA_KEY &&
663 search_key.offset >= LLONG_MAX)
664 search_key.offset = 0;
665 path->lowest_level = level;
666 if (ctx->time_seq == BTRFS_SEQ_LAST)
667 ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
668 else
669 ret = btrfs_search_old_slot(root, &search_key, path, ctx->time_seq);
670
671 btrfs_debug(ctx->fs_info,
672 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
673 ref->root_id, level, ref->count, ret,
674 ref->key_for_search.objectid, ref->key_for_search.type,
675 ref->key_for_search.offset);
676 if (ret < 0)
677 goto out;
678
679 eb = path->nodes[level];
680 while (!eb) {
681 if (WARN_ON(!level)) {
682 ret = 1;
683 goto out;
684 }
685 level--;
686 eb = path->nodes[level];
687 }
688
689 ret = add_all_parents(ctx, root, path, parents, preftrees, ref, level);
690out:
691 btrfs_put_root(root);
692out_free:
693 path->lowest_level = 0;
694 btrfs_release_path(path);
695 return ret;
696}
697
698static struct extent_inode_elem *
699unode_aux_to_inode_list(struct ulist_node *node)
700{
701 if (!node)
702 return NULL;
703 return (struct extent_inode_elem *)(uintptr_t)node->aux;
704}
705
706static void free_leaf_list(struct ulist *ulist)
707{
708 struct ulist_node *node;
709 struct ulist_iterator uiter;
710
711 ULIST_ITER_INIT(&uiter);
712 while ((node = ulist_next(ulist, &uiter)))
713 free_inode_elem_list(unode_aux_to_inode_list(node));
714
715 ulist_free(ulist);
716}
717
718/*
719 * We maintain three separate rbtrees: one for direct refs, one for
720 * indirect refs which have a key, and one for indirect refs which do not
721 * have a key. Each tree does merge on insertion.
722 *
723 * Once all of the references are located, we iterate over the tree of
724 * indirect refs with missing keys. An appropriate key is located and
725 * the ref is moved onto the tree for indirect refs. After all missing
726 * keys are thus located, we iterate over the indirect ref tree, resolve
727 * each reference, and then insert the resolved reference onto the
728 * direct tree (merging there too).
729 *
730 * New backrefs (i.e., for parent nodes) are added to the appropriate
731 * rbtree as they are encountered. The new backrefs are subsequently
732 * resolved as above.
733 */
734static int resolve_indirect_refs(struct btrfs_backref_walk_ctx *ctx,
735 struct btrfs_path *path,
736 struct preftrees *preftrees,
737 struct share_check *sc)
738{
739 int err;
740 int ret = 0;
741 struct ulist *parents;
742 struct ulist_node *node;
743 struct ulist_iterator uiter;
744 struct rb_node *rnode;
745
746 parents = ulist_alloc(GFP_NOFS);
747 if (!parents)
748 return -ENOMEM;
749
750 /*
751 * We could trade memory usage for performance here by iterating
752 * the tree, allocating new refs for each insertion, and then
753 * freeing the entire indirect tree when we're done. In some test
754 * cases, the tree can grow quite large (~200k objects).
755 */
756 while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
757 struct prelim_ref *ref;
758
759 ref = rb_entry(rnode, struct prelim_ref, rbnode);
760 if (WARN(ref->parent,
761 "BUG: direct ref found in indirect tree")) {
762 ret = -EINVAL;
763 goto out;
764 }
765
766 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
767 preftrees->indirect.count--;
768
769 if (ref->count == 0) {
770 free_pref(ref);
771 continue;
772 }
773
774 if (sc && ref->root_id != sc->root->root_key.objectid) {
775 free_pref(ref);
776 ret = BACKREF_FOUND_SHARED;
777 goto out;
778 }
779 err = resolve_indirect_ref(ctx, path, preftrees, ref, parents);
780 /*
781 * we can only tolerate ENOENT,otherwise,we should catch error
782 * and return directly.
783 */
784 if (err == -ENOENT) {
785 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref,
786 NULL);
787 continue;
788 } else if (err) {
789 free_pref(ref);
790 ret = err;
791 goto out;
792 }
793
794 /* we put the first parent into the ref at hand */
795 ULIST_ITER_INIT(&uiter);
796 node = ulist_next(parents, &uiter);
797 ref->parent = node ? node->val : 0;
798 ref->inode_list = unode_aux_to_inode_list(node);
799
800 /* Add a prelim_ref(s) for any other parent(s). */
801 while ((node = ulist_next(parents, &uiter))) {
802 struct prelim_ref *new_ref;
803
804 new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
805 GFP_NOFS);
806 if (!new_ref) {
807 free_pref(ref);
808 ret = -ENOMEM;
809 goto out;
810 }
811 memcpy(new_ref, ref, sizeof(*ref));
812 new_ref->parent = node->val;
813 new_ref->inode_list = unode_aux_to_inode_list(node);
814 prelim_ref_insert(ctx->fs_info, &preftrees->direct,
815 new_ref, NULL);
816 }
817
818 /*
819 * Now it's a direct ref, put it in the direct tree. We must
820 * do this last because the ref could be merged/freed here.
821 */
822 prelim_ref_insert(ctx->fs_info, &preftrees->direct, ref, NULL);
823
824 ulist_reinit(parents);
825 cond_resched();
826 }
827out:
828 /*
829 * We may have inode lists attached to refs in the parents ulist, so we
830 * must free them before freeing the ulist and its refs.
831 */
832 free_leaf_list(parents);
833 return ret;
834}
835
836/*
837 * read tree blocks and add keys where required.
838 */
839static int add_missing_keys(struct btrfs_fs_info *fs_info,
840 struct preftrees *preftrees, bool lock)
841{
842 struct prelim_ref *ref;
843 struct extent_buffer *eb;
844 struct preftree *tree = &preftrees->indirect_missing_keys;
845 struct rb_node *node;
846
847 while ((node = rb_first_cached(&tree->root))) {
848 struct btrfs_tree_parent_check check = { 0 };
849
850 ref = rb_entry(node, struct prelim_ref, rbnode);
851 rb_erase_cached(node, &tree->root);
852
853 BUG_ON(ref->parent); /* should not be a direct ref */
854 BUG_ON(ref->key_for_search.type);
855 BUG_ON(!ref->wanted_disk_byte);
856
857 check.level = ref->level - 1;
858 check.owner_root = ref->root_id;
859
860 eb = read_tree_block(fs_info, ref->wanted_disk_byte, &check);
861 if (IS_ERR(eb)) {
862 free_pref(ref);
863 return PTR_ERR(eb);
864 }
865 if (!extent_buffer_uptodate(eb)) {
866 free_pref(ref);
867 free_extent_buffer(eb);
868 return -EIO;
869 }
870
871 if (lock)
872 btrfs_tree_read_lock(eb);
873 if (btrfs_header_level(eb) == 0)
874 btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
875 else
876 btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
877 if (lock)
878 btrfs_tree_read_unlock(eb);
879 free_extent_buffer(eb);
880 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
881 cond_resched();
882 }
883 return 0;
884}
885
886/*
887 * add all currently queued delayed refs from this head whose seq nr is
888 * smaller or equal that seq to the list
889 */
890static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
891 struct btrfs_delayed_ref_head *head, u64 seq,
892 struct preftrees *preftrees, struct share_check *sc)
893{
894 struct btrfs_delayed_ref_node *node;
895 struct btrfs_key key;
896 struct rb_node *n;
897 int count;
898 int ret = 0;
899
900 spin_lock(&head->lock);
901 for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
902 node = rb_entry(n, struct btrfs_delayed_ref_node,
903 ref_node);
904 if (node->seq > seq)
905 continue;
906
907 switch (node->action) {
908 case BTRFS_ADD_DELAYED_EXTENT:
909 case BTRFS_UPDATE_DELAYED_HEAD:
910 WARN_ON(1);
911 continue;
912 case BTRFS_ADD_DELAYED_REF:
913 count = node->ref_mod;
914 break;
915 case BTRFS_DROP_DELAYED_REF:
916 count = node->ref_mod * -1;
917 break;
918 default:
919 BUG();
920 }
921 switch (node->type) {
922 case BTRFS_TREE_BLOCK_REF_KEY: {
923 /* NORMAL INDIRECT METADATA backref */
924 struct btrfs_delayed_tree_ref *ref;
925 struct btrfs_key *key_ptr = NULL;
926
927 if (head->extent_op && head->extent_op->update_key) {
928 btrfs_disk_key_to_cpu(&key, &head->extent_op->key);
929 key_ptr = &key;
930 }
931
932 ref = btrfs_delayed_node_to_tree_ref(node);
933 ret = add_indirect_ref(fs_info, preftrees, ref->root,
934 key_ptr, ref->level + 1,
935 node->bytenr, count, sc,
936 GFP_ATOMIC);
937 break;
938 }
939 case BTRFS_SHARED_BLOCK_REF_KEY: {
940 /* SHARED DIRECT METADATA backref */
941 struct btrfs_delayed_tree_ref *ref;
942
943 ref = btrfs_delayed_node_to_tree_ref(node);
944
945 ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
946 ref->parent, node->bytenr, count,
947 sc, GFP_ATOMIC);
948 break;
949 }
950 case BTRFS_EXTENT_DATA_REF_KEY: {
951 /* NORMAL INDIRECT DATA backref */
952 struct btrfs_delayed_data_ref *ref;
953 ref = btrfs_delayed_node_to_data_ref(node);
954
955 key.objectid = ref->objectid;
956 key.type = BTRFS_EXTENT_DATA_KEY;
957 key.offset = ref->offset;
958
959 /*
960 * If we have a share check context and a reference for
961 * another inode, we can't exit immediately. This is
962 * because even if this is a BTRFS_ADD_DELAYED_REF
963 * reference we may find next a BTRFS_DROP_DELAYED_REF
964 * which cancels out this ADD reference.
965 *
966 * If this is a DROP reference and there was no previous
967 * ADD reference, then we need to signal that when we
968 * process references from the extent tree (through
969 * add_inline_refs() and add_keyed_refs()), we should
970 * not exit early if we find a reference for another
971 * inode, because one of the delayed DROP references
972 * may cancel that reference in the extent tree.
973 */
974 if (sc && count < 0)
975 sc->have_delayed_delete_refs = true;
976
977 ret = add_indirect_ref(fs_info, preftrees, ref->root,
978 &key, 0, node->bytenr, count, sc,
979 GFP_ATOMIC);
980 break;
981 }
982 case BTRFS_SHARED_DATA_REF_KEY: {
983 /* SHARED DIRECT FULL backref */
984 struct btrfs_delayed_data_ref *ref;
985
986 ref = btrfs_delayed_node_to_data_ref(node);
987
988 ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
989 node->bytenr, count, sc,
990 GFP_ATOMIC);
991 break;
992 }
993 default:
994 WARN_ON(1);
995 }
996 /*
997 * We must ignore BACKREF_FOUND_SHARED until all delayed
998 * refs have been checked.
999 */
1000 if (ret && (ret != BACKREF_FOUND_SHARED))
1001 break;
1002 }
1003 if (!ret)
1004 ret = extent_is_shared(sc);
1005
1006 spin_unlock(&head->lock);
1007 return ret;
1008}
1009
1010/*
1011 * add all inline backrefs for bytenr to the list
1012 *
1013 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1014 */
1015static int add_inline_refs(struct btrfs_backref_walk_ctx *ctx,
1016 struct btrfs_path *path,
1017 int *info_level, struct preftrees *preftrees,
1018 struct share_check *sc)
1019{
1020 int ret = 0;
1021 int slot;
1022 struct extent_buffer *leaf;
1023 struct btrfs_key key;
1024 struct btrfs_key found_key;
1025 unsigned long ptr;
1026 unsigned long end;
1027 struct btrfs_extent_item *ei;
1028 u64 flags;
1029 u64 item_size;
1030
1031 /*
1032 * enumerate all inline refs
1033 */
1034 leaf = path->nodes[0];
1035 slot = path->slots[0];
1036
1037 item_size = btrfs_item_size(leaf, slot);
1038 BUG_ON(item_size < sizeof(*ei));
1039
1040 ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
1041
1042 if (ctx->check_extent_item) {
1043 ret = ctx->check_extent_item(ctx->bytenr, ei, leaf, ctx->user_ctx);
1044 if (ret)
1045 return ret;
1046 }
1047
1048 flags = btrfs_extent_flags(leaf, ei);
1049 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1050
1051 ptr = (unsigned long)(ei + 1);
1052 end = (unsigned long)ei + item_size;
1053
1054 if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
1055 flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1056 struct btrfs_tree_block_info *info;
1057
1058 info = (struct btrfs_tree_block_info *)ptr;
1059 *info_level = btrfs_tree_block_level(leaf, info);
1060 ptr += sizeof(struct btrfs_tree_block_info);
1061 BUG_ON(ptr > end);
1062 } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1063 *info_level = found_key.offset;
1064 } else {
1065 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1066 }
1067
1068 while (ptr < end) {
1069 struct btrfs_extent_inline_ref *iref;
1070 u64 offset;
1071 int type;
1072
1073 iref = (struct btrfs_extent_inline_ref *)ptr;
1074 type = btrfs_get_extent_inline_ref_type(leaf, iref,
1075 BTRFS_REF_TYPE_ANY);
1076 if (type == BTRFS_REF_TYPE_INVALID)
1077 return -EUCLEAN;
1078
1079 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1080
1081 switch (type) {
1082 case BTRFS_SHARED_BLOCK_REF_KEY:
1083 ret = add_direct_ref(ctx->fs_info, preftrees,
1084 *info_level + 1, offset,
1085 ctx->bytenr, 1, NULL, GFP_NOFS);
1086 break;
1087 case BTRFS_SHARED_DATA_REF_KEY: {
1088 struct btrfs_shared_data_ref *sdref;
1089 int count;
1090
1091 sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1092 count = btrfs_shared_data_ref_count(leaf, sdref);
1093
1094 ret = add_direct_ref(ctx->fs_info, preftrees, 0, offset,
1095 ctx->bytenr, count, sc, GFP_NOFS);
1096 break;
1097 }
1098 case BTRFS_TREE_BLOCK_REF_KEY:
1099 ret = add_indirect_ref(ctx->fs_info, preftrees, offset,
1100 NULL, *info_level + 1,
1101 ctx->bytenr, 1, NULL, GFP_NOFS);
1102 break;
1103 case BTRFS_EXTENT_DATA_REF_KEY: {
1104 struct btrfs_extent_data_ref *dref;
1105 int count;
1106 u64 root;
1107
1108 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1109 count = btrfs_extent_data_ref_count(leaf, dref);
1110 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1111 dref);
1112 key.type = BTRFS_EXTENT_DATA_KEY;
1113 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1114
1115 if (sc && key.objectid != sc->inum &&
1116 !sc->have_delayed_delete_refs) {
1117 ret = BACKREF_FOUND_SHARED;
1118 break;
1119 }
1120
1121 root = btrfs_extent_data_ref_root(leaf, dref);
1122
1123 if (!ctx->skip_data_ref ||
1124 !ctx->skip_data_ref(root, key.objectid, key.offset,
1125 ctx->user_ctx))
1126 ret = add_indirect_ref(ctx->fs_info, preftrees,
1127 root, &key, 0, ctx->bytenr,
1128 count, sc, GFP_NOFS);
1129 break;
1130 }
1131 default:
1132 WARN_ON(1);
1133 }
1134 if (ret)
1135 return ret;
1136 ptr += btrfs_extent_inline_ref_size(type);
1137 }
1138
1139 return 0;
1140}
1141
1142/*
1143 * add all non-inline backrefs for bytenr to the list
1144 *
1145 * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
1146 */
1147static int add_keyed_refs(struct btrfs_backref_walk_ctx *ctx,
1148 struct btrfs_root *extent_root,
1149 struct btrfs_path *path,
1150 int info_level, struct preftrees *preftrees,
1151 struct share_check *sc)
1152{
1153 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1154 int ret;
1155 int slot;
1156 struct extent_buffer *leaf;
1157 struct btrfs_key key;
1158
1159 while (1) {
1160 ret = btrfs_next_item(extent_root, path);
1161 if (ret < 0)
1162 break;
1163 if (ret) {
1164 ret = 0;
1165 break;
1166 }
1167
1168 slot = path->slots[0];
1169 leaf = path->nodes[0];
1170 btrfs_item_key_to_cpu(leaf, &key, slot);
1171
1172 if (key.objectid != ctx->bytenr)
1173 break;
1174 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1175 continue;
1176 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1177 break;
1178
1179 switch (key.type) {
1180 case BTRFS_SHARED_BLOCK_REF_KEY:
1181 /* SHARED DIRECT METADATA backref */
1182 ret = add_direct_ref(fs_info, preftrees,
1183 info_level + 1, key.offset,
1184 ctx->bytenr, 1, NULL, GFP_NOFS);
1185 break;
1186 case BTRFS_SHARED_DATA_REF_KEY: {
1187 /* SHARED DIRECT FULL backref */
1188 struct btrfs_shared_data_ref *sdref;
1189 int count;
1190
1191 sdref = btrfs_item_ptr(leaf, slot,
1192 struct btrfs_shared_data_ref);
1193 count = btrfs_shared_data_ref_count(leaf, sdref);
1194 ret = add_direct_ref(fs_info, preftrees, 0,
1195 key.offset, ctx->bytenr, count,
1196 sc, GFP_NOFS);
1197 break;
1198 }
1199 case BTRFS_TREE_BLOCK_REF_KEY:
1200 /* NORMAL INDIRECT METADATA backref */
1201 ret = add_indirect_ref(fs_info, preftrees, key.offset,
1202 NULL, info_level + 1, ctx->bytenr,
1203 1, NULL, GFP_NOFS);
1204 break;
1205 case BTRFS_EXTENT_DATA_REF_KEY: {
1206 /* NORMAL INDIRECT DATA backref */
1207 struct btrfs_extent_data_ref *dref;
1208 int count;
1209 u64 root;
1210
1211 dref = btrfs_item_ptr(leaf, slot,
1212 struct btrfs_extent_data_ref);
1213 count = btrfs_extent_data_ref_count(leaf, dref);
1214 key.objectid = btrfs_extent_data_ref_objectid(leaf,
1215 dref);
1216 key.type = BTRFS_EXTENT_DATA_KEY;
1217 key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1218
1219 if (sc && key.objectid != sc->inum &&
1220 !sc->have_delayed_delete_refs) {
1221 ret = BACKREF_FOUND_SHARED;
1222 break;
1223 }
1224
1225 root = btrfs_extent_data_ref_root(leaf, dref);
1226
1227 if (!ctx->skip_data_ref ||
1228 !ctx->skip_data_ref(root, key.objectid, key.offset,
1229 ctx->user_ctx))
1230 ret = add_indirect_ref(fs_info, preftrees, root,
1231 &key, 0, ctx->bytenr,
1232 count, sc, GFP_NOFS);
1233 break;
1234 }
1235 default:
1236 WARN_ON(1);
1237 }
1238 if (ret)
1239 return ret;
1240
1241 }
1242
1243 return ret;
1244}
1245
1246/*
1247 * The caller has joined a transaction or is holding a read lock on the
1248 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1249 * snapshot field changing while updating or checking the cache.
1250 */
1251static bool lookup_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1252 struct btrfs_root *root,
1253 u64 bytenr, int level, bool *is_shared)
1254{
1255 struct btrfs_backref_shared_cache_entry *entry;
1256
1257 if (!ctx->use_path_cache)
1258 return false;
1259
1260 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1261 return false;
1262
1263 /*
1264 * Level -1 is used for the data extent, which is not reliable to cache
1265 * because its reference count can increase or decrease without us
1266 * realizing. We cache results only for extent buffers that lead from
1267 * the root node down to the leaf with the file extent item.
1268 */
1269 ASSERT(level >= 0);
1270
1271 entry = &ctx->path_cache_entries[level];
1272
1273 /* Unused cache entry or being used for some other extent buffer. */
1274 if (entry->bytenr != bytenr)
1275 return false;
1276
1277 /*
1278 * We cached a false result, but the last snapshot generation of the
1279 * root changed, so we now have a snapshot. Don't trust the result.
1280 */
1281 if (!entry->is_shared &&
1282 entry->gen != btrfs_root_last_snapshot(&root->root_item))
1283 return false;
1284
1285 /*
1286 * If we cached a true result and the last generation used for dropping
1287 * a root changed, we can not trust the result, because the dropped root
1288 * could be a snapshot sharing this extent buffer.
1289 */
1290 if (entry->is_shared &&
1291 entry->gen != btrfs_get_last_root_drop_gen(root->fs_info))
1292 return false;
1293
1294 *is_shared = entry->is_shared;
1295 /*
1296 * If the node at this level is shared, than all nodes below are also
1297 * shared. Currently some of the nodes below may be marked as not shared
1298 * because we have just switched from one leaf to another, and switched
1299 * also other nodes above the leaf and below the current level, so mark
1300 * them as shared.
1301 */
1302 if (*is_shared) {
1303 for (int i = 0; i < level; i++) {
1304 ctx->path_cache_entries[i].is_shared = true;
1305 ctx->path_cache_entries[i].gen = entry->gen;
1306 }
1307 }
1308
1309 return true;
1310}
1311
1312/*
1313 * The caller has joined a transaction or is holding a read lock on the
1314 * fs_info->commit_root_sem semaphore, so no need to worry about the root's last
1315 * snapshot field changing while updating or checking the cache.
1316 */
1317static void store_backref_shared_cache(struct btrfs_backref_share_check_ctx *ctx,
1318 struct btrfs_root *root,
1319 u64 bytenr, int level, bool is_shared)
1320{
1321 struct btrfs_backref_shared_cache_entry *entry;
1322 u64 gen;
1323
1324 if (!ctx->use_path_cache)
1325 return;
1326
1327 if (WARN_ON_ONCE(level >= BTRFS_MAX_LEVEL))
1328 return;
1329
1330 /*
1331 * Level -1 is used for the data extent, which is not reliable to cache
1332 * because its reference count can increase or decrease without us
1333 * realizing. We cache results only for extent buffers that lead from
1334 * the root node down to the leaf with the file extent item.
1335 */
1336 ASSERT(level >= 0);
1337
1338 if (is_shared)
1339 gen = btrfs_get_last_root_drop_gen(root->fs_info);
1340 else
1341 gen = btrfs_root_last_snapshot(&root->root_item);
1342
1343 entry = &ctx->path_cache_entries[level];
1344 entry->bytenr = bytenr;
1345 entry->is_shared = is_shared;
1346 entry->gen = gen;
1347
1348 /*
1349 * If we found an extent buffer is shared, set the cache result for all
1350 * extent buffers below it to true. As nodes in the path are COWed,
1351 * their sharedness is moved to their children, and if a leaf is COWed,
1352 * then the sharedness of a data extent becomes direct, the refcount of
1353 * data extent is increased in the extent item at the extent tree.
1354 */
1355 if (is_shared) {
1356 for (int i = 0; i < level; i++) {
1357 entry = &ctx->path_cache_entries[i];
1358 entry->is_shared = is_shared;
1359 entry->gen = gen;
1360 }
1361 }
1362}
1363
1364/*
1365 * this adds all existing backrefs (inline backrefs, backrefs and delayed
1366 * refs) for the given bytenr to the refs list, merges duplicates and resolves
1367 * indirect refs to their parent bytenr.
1368 * When roots are found, they're added to the roots list
1369 *
1370 * @ctx: Backref walking context object, must be not NULL.
1371 * @sc: If !NULL, then immediately return BACKREF_FOUND_SHARED when a
1372 * shared extent is detected.
1373 *
1374 * Otherwise this returns 0 for success and <0 for an error.
1375 *
1376 * FIXME some caching might speed things up
1377 */
1378static int find_parent_nodes(struct btrfs_backref_walk_ctx *ctx,
1379 struct share_check *sc)
1380{
1381 struct btrfs_root *root = btrfs_extent_root(ctx->fs_info, ctx->bytenr);
1382 struct btrfs_key key;
1383 struct btrfs_path *path;
1384 struct btrfs_delayed_ref_root *delayed_refs = NULL;
1385 struct btrfs_delayed_ref_head *head;
1386 int info_level = 0;
1387 int ret;
1388 struct prelim_ref *ref;
1389 struct rb_node *node;
1390 struct extent_inode_elem *eie = NULL;
1391 struct preftrees preftrees = {
1392 .direct = PREFTREE_INIT,
1393 .indirect = PREFTREE_INIT,
1394 .indirect_missing_keys = PREFTREE_INIT
1395 };
1396
1397 /* Roots ulist is not needed when using a sharedness check context. */
1398 if (sc)
1399 ASSERT(ctx->roots == NULL);
1400
1401 key.objectid = ctx->bytenr;
1402 key.offset = (u64)-1;
1403 if (btrfs_fs_incompat(ctx->fs_info, SKINNY_METADATA))
1404 key.type = BTRFS_METADATA_ITEM_KEY;
1405 else
1406 key.type = BTRFS_EXTENT_ITEM_KEY;
1407
1408 path = btrfs_alloc_path();
1409 if (!path)
1410 return -ENOMEM;
1411 if (!ctx->trans) {
1412 path->search_commit_root = 1;
1413 path->skip_locking = 1;
1414 }
1415
1416 if (ctx->time_seq == BTRFS_SEQ_LAST)
1417 path->skip_locking = 1;
1418
1419again:
1420 head = NULL;
1421
1422 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1423 if (ret < 0)
1424 goto out;
1425 if (ret == 0) {
1426 /* This shouldn't happen, indicates a bug or fs corruption. */
1427 ASSERT(ret != 0);
1428 ret = -EUCLEAN;
1429 goto out;
1430 }
1431
1432 if (ctx->trans && likely(ctx->trans->type != __TRANS_DUMMY) &&
1433 ctx->time_seq != BTRFS_SEQ_LAST) {
1434 /*
1435 * We have a specific time_seq we care about and trans which
1436 * means we have the path lock, we need to grab the ref head and
1437 * lock it so we have a consistent view of the refs at the given
1438 * time.
1439 */
1440 delayed_refs = &ctx->trans->transaction->delayed_refs;
1441 spin_lock(&delayed_refs->lock);
1442 head = btrfs_find_delayed_ref_head(delayed_refs, ctx->bytenr);
1443 if (head) {
1444 if (!mutex_trylock(&head->mutex)) {
1445 refcount_inc(&head->refs);
1446 spin_unlock(&delayed_refs->lock);
1447
1448 btrfs_release_path(path);
1449
1450 /*
1451 * Mutex was contended, block until it's
1452 * released and try again
1453 */
1454 mutex_lock(&head->mutex);
1455 mutex_unlock(&head->mutex);
1456 btrfs_put_delayed_ref_head(head);
1457 goto again;
1458 }
1459 spin_unlock(&delayed_refs->lock);
1460 ret = add_delayed_refs(ctx->fs_info, head, ctx->time_seq,
1461 &preftrees, sc);
1462 mutex_unlock(&head->mutex);
1463 if (ret)
1464 goto out;
1465 } else {
1466 spin_unlock(&delayed_refs->lock);
1467 }
1468 }
1469
1470 if (path->slots[0]) {
1471 struct extent_buffer *leaf;
1472 int slot;
1473
1474 path->slots[0]--;
1475 leaf = path->nodes[0];
1476 slot = path->slots[0];
1477 btrfs_item_key_to_cpu(leaf, &key, slot);
1478 if (key.objectid == ctx->bytenr &&
1479 (key.type == BTRFS_EXTENT_ITEM_KEY ||
1480 key.type == BTRFS_METADATA_ITEM_KEY)) {
1481 ret = add_inline_refs(ctx, path, &info_level,
1482 &preftrees, sc);
1483 if (ret)
1484 goto out;
1485 ret = add_keyed_refs(ctx, root, path, info_level,
1486 &preftrees, sc);
1487 if (ret)
1488 goto out;
1489 }
1490 }
1491
1492 /*
1493 * If we have a share context and we reached here, it means the extent
1494 * is not directly shared (no multiple reference items for it),
1495 * otherwise we would have exited earlier with a return value of
1496 * BACKREF_FOUND_SHARED after processing delayed references or while
1497 * processing inline or keyed references from the extent tree.
1498 * The extent may however be indirectly shared through shared subtrees
1499 * as a result from creating snapshots, so we determine below what is
1500 * its parent node, in case we are dealing with a metadata extent, or
1501 * what's the leaf (or leaves), from a fs tree, that has a file extent
1502 * item pointing to it in case we are dealing with a data extent.
1503 */
1504 ASSERT(extent_is_shared(sc) == 0);
1505
1506 /*
1507 * If we are here for a data extent and we have a share_check structure
1508 * it means the data extent is not directly shared (does not have
1509 * multiple reference items), so we have to check if a path in the fs
1510 * tree (going from the root node down to the leaf that has the file
1511 * extent item pointing to the data extent) is shared, that is, if any
1512 * of the extent buffers in the path is referenced by other trees.
1513 */
1514 if (sc && ctx->bytenr == sc->data_bytenr) {
1515 /*
1516 * If our data extent is from a generation more recent than the
1517 * last generation used to snapshot the root, then we know that
1518 * it can not be shared through subtrees, so we can skip
1519 * resolving indirect references, there's no point in
1520 * determining the extent buffers for the path from the fs tree
1521 * root node down to the leaf that has the file extent item that
1522 * points to the data extent.
1523 */
1524 if (sc->data_extent_gen >
1525 btrfs_root_last_snapshot(&sc->root->root_item)) {
1526 ret = BACKREF_FOUND_NOT_SHARED;
1527 goto out;
1528 }
1529
1530 /*
1531 * If we are only determining if a data extent is shared or not
1532 * and the corresponding file extent item is located in the same
1533 * leaf as the previous file extent item, we can skip resolving
1534 * indirect references for a data extent, since the fs tree path
1535 * is the same (same leaf, so same path). We skip as long as the
1536 * cached result for the leaf is valid and only if there's only
1537 * one file extent item pointing to the data extent, because in
1538 * the case of multiple file extent items, they may be located
1539 * in different leaves and therefore we have multiple paths.
1540 */
1541 if (sc->ctx->curr_leaf_bytenr == sc->ctx->prev_leaf_bytenr &&
1542 sc->self_ref_count == 1) {
1543 bool cached;
1544 bool is_shared;
1545
1546 cached = lookup_backref_shared_cache(sc->ctx, sc->root,
1547 sc->ctx->curr_leaf_bytenr,
1548 0, &is_shared);
1549 if (cached) {
1550 if (is_shared)
1551 ret = BACKREF_FOUND_SHARED;
1552 else
1553 ret = BACKREF_FOUND_NOT_SHARED;
1554 goto out;
1555 }
1556 }
1557 }
1558
1559 btrfs_release_path(path);
1560
1561 ret = add_missing_keys(ctx->fs_info, &preftrees, path->skip_locking == 0);
1562 if (ret)
1563 goto out;
1564
1565 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1566
1567 ret = resolve_indirect_refs(ctx, path, &preftrees, sc);
1568 if (ret)
1569 goto out;
1570
1571 WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1572
1573 /*
1574 * This walks the tree of merged and resolved refs. Tree blocks are
1575 * read in as needed. Unique entries are added to the ulist, and
1576 * the list of found roots is updated.
1577 *
1578 * We release the entire tree in one go before returning.
1579 */
1580 node = rb_first_cached(&preftrees.direct.root);
1581 while (node) {
1582 ref = rb_entry(node, struct prelim_ref, rbnode);
1583 node = rb_next(&ref->rbnode);
1584 /*
1585 * ref->count < 0 can happen here if there are delayed
1586 * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1587 * prelim_ref_insert() relies on this when merging
1588 * identical refs to keep the overall count correct.
1589 * prelim_ref_insert() will merge only those refs
1590 * which compare identically. Any refs having
1591 * e.g. different offsets would not be merged,
1592 * and would retain their original ref->count < 0.
1593 */
1594 if (ctx->roots && ref->count && ref->root_id && ref->parent == 0) {
1595 /* no parent == root of tree */
1596 ret = ulist_add(ctx->roots, ref->root_id, 0, GFP_NOFS);
1597 if (ret < 0)
1598 goto out;
1599 }
1600 if (ref->count && ref->parent) {
1601 if (!ctx->ignore_extent_item_pos && !ref->inode_list &&
1602 ref->level == 0) {
1603 struct btrfs_tree_parent_check check = { 0 };
1604 struct extent_buffer *eb;
1605
1606 check.level = ref->level;
1607
1608 eb = read_tree_block(ctx->fs_info, ref->parent,
1609 &check);
1610 if (IS_ERR(eb)) {
1611 ret = PTR_ERR(eb);
1612 goto out;
1613 }
1614 if (!extent_buffer_uptodate(eb)) {
1615 free_extent_buffer(eb);
1616 ret = -EIO;
1617 goto out;
1618 }
1619
1620 if (!path->skip_locking)
1621 btrfs_tree_read_lock(eb);
1622 ret = find_extent_in_eb(ctx, eb, &eie);
1623 if (!path->skip_locking)
1624 btrfs_tree_read_unlock(eb);
1625 free_extent_buffer(eb);
1626 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1627 ret < 0)
1628 goto out;
1629 ref->inode_list = eie;
1630 /*
1631 * We transferred the list ownership to the ref,
1632 * so set to NULL to avoid a double free in case
1633 * an error happens after this.
1634 */
1635 eie = NULL;
1636 }
1637 ret = ulist_add_merge_ptr(ctx->refs, ref->parent,
1638 ref->inode_list,
1639 (void **)&eie, GFP_NOFS);
1640 if (ret < 0)
1641 goto out;
1642 if (!ret && !ctx->ignore_extent_item_pos) {
1643 /*
1644 * We've recorded that parent, so we must extend
1645 * its inode list here.
1646 *
1647 * However if there was corruption we may not
1648 * have found an eie, return an error in this
1649 * case.
1650 */
1651 ASSERT(eie);
1652 if (!eie) {
1653 ret = -EUCLEAN;
1654 goto out;
1655 }
1656 while (eie->next)
1657 eie = eie->next;
1658 eie->next = ref->inode_list;
1659 }
1660 eie = NULL;
1661 /*
1662 * We have transferred the inode list ownership from
1663 * this ref to the ref we added to the 'refs' ulist.
1664 * So set this ref's inode list to NULL to avoid
1665 * use-after-free when our caller uses it or double
1666 * frees in case an error happens before we return.
1667 */
1668 ref->inode_list = NULL;
1669 }
1670 cond_resched();
1671 }
1672
1673out:
1674 btrfs_free_path(path);
1675
1676 prelim_release(&preftrees.direct);
1677 prelim_release(&preftrees.indirect);
1678 prelim_release(&preftrees.indirect_missing_keys);
1679
1680 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP || ret < 0)
1681 free_inode_elem_list(eie);
1682 return ret;
1683}
1684
1685/*
1686 * Finds all leaves with a reference to the specified combination of
1687 * @ctx->bytenr and @ctx->extent_item_pos. The bytenr of the found leaves are
1688 * added to the ulist at @ctx->refs, and that ulist is allocated by this
1689 * function. The caller should free the ulist with free_leaf_list() if
1690 * @ctx->ignore_extent_item_pos is false, otherwise a fimple ulist_free() is
1691 * enough.
1692 *
1693 * Returns 0 on success and < 0 on error. On error @ctx->refs is not allocated.
1694 */
1695int btrfs_find_all_leafs(struct btrfs_backref_walk_ctx *ctx)
1696{
1697 int ret;
1698
1699 ASSERT(ctx->refs == NULL);
1700
1701 ctx->refs = ulist_alloc(GFP_NOFS);
1702 if (!ctx->refs)
1703 return -ENOMEM;
1704
1705 ret = find_parent_nodes(ctx, NULL);
1706 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP ||
1707 (ret < 0 && ret != -ENOENT)) {
1708 free_leaf_list(ctx->refs);
1709 ctx->refs = NULL;
1710 return ret;
1711 }
1712
1713 return 0;
1714}
1715
1716/*
1717 * Walk all backrefs for a given extent to find all roots that reference this
1718 * extent. Walking a backref means finding all extents that reference this
1719 * extent and in turn walk the backrefs of those, too. Naturally this is a
1720 * recursive process, but here it is implemented in an iterative fashion: We
1721 * find all referencing extents for the extent in question and put them on a
1722 * list. In turn, we find all referencing extents for those, further appending
1723 * to the list. The way we iterate the list allows adding more elements after
1724 * the current while iterating. The process stops when we reach the end of the
1725 * list.
1726 *
1727 * Found roots are added to @ctx->roots, which is allocated by this function if
1728 * it points to NULL, in which case the caller is responsible for freeing it
1729 * after it's not needed anymore.
1730 * This function requires @ctx->refs to be NULL, as it uses it for allocating a
1731 * ulist to do temporary work, and frees it before returning.
1732 *
1733 * Returns 0 on success, < 0 on error.
1734 */
1735static int btrfs_find_all_roots_safe(struct btrfs_backref_walk_ctx *ctx)
1736{
1737 const u64 orig_bytenr = ctx->bytenr;
1738 const bool orig_ignore_extent_item_pos = ctx->ignore_extent_item_pos;
1739 bool roots_ulist_allocated = false;
1740 struct ulist_iterator uiter;
1741 int ret = 0;
1742
1743 ASSERT(ctx->refs == NULL);
1744
1745 ctx->refs = ulist_alloc(GFP_NOFS);
1746 if (!ctx->refs)
1747 return -ENOMEM;
1748
1749 if (!ctx->roots) {
1750 ctx->roots = ulist_alloc(GFP_NOFS);
1751 if (!ctx->roots) {
1752 ulist_free(ctx->refs);
1753 ctx->refs = NULL;
1754 return -ENOMEM;
1755 }
1756 roots_ulist_allocated = true;
1757 }
1758
1759 ctx->ignore_extent_item_pos = true;
1760
1761 ULIST_ITER_INIT(&uiter);
1762 while (1) {
1763 struct ulist_node *node;
1764
1765 ret = find_parent_nodes(ctx, NULL);
1766 if (ret < 0 && ret != -ENOENT) {
1767 if (roots_ulist_allocated) {
1768 ulist_free(ctx->roots);
1769 ctx->roots = NULL;
1770 }
1771 break;
1772 }
1773 ret = 0;
1774 node = ulist_next(ctx->refs, &uiter);
1775 if (!node)
1776 break;
1777 ctx->bytenr = node->val;
1778 cond_resched();
1779 }
1780
1781 ulist_free(ctx->refs);
1782 ctx->refs = NULL;
1783 ctx->bytenr = orig_bytenr;
1784 ctx->ignore_extent_item_pos = orig_ignore_extent_item_pos;
1785
1786 return ret;
1787}
1788
1789int btrfs_find_all_roots(struct btrfs_backref_walk_ctx *ctx,
1790 bool skip_commit_root_sem)
1791{
1792 int ret;
1793
1794 if (!ctx->trans && !skip_commit_root_sem)
1795 down_read(&ctx->fs_info->commit_root_sem);
1796 ret = btrfs_find_all_roots_safe(ctx);
1797 if (!ctx->trans && !skip_commit_root_sem)
1798 up_read(&ctx->fs_info->commit_root_sem);
1799 return ret;
1800}
1801
1802struct btrfs_backref_share_check_ctx *btrfs_alloc_backref_share_check_ctx(void)
1803{
1804 struct btrfs_backref_share_check_ctx *ctx;
1805
1806 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1807 if (!ctx)
1808 return NULL;
1809
1810 ulist_init(&ctx->refs);
1811
1812 return ctx;
1813}
1814
1815void btrfs_free_backref_share_ctx(struct btrfs_backref_share_check_ctx *ctx)
1816{
1817 if (!ctx)
1818 return;
1819
1820 ulist_release(&ctx->refs);
1821 kfree(ctx);
1822}
1823
1824/*
1825 * Check if a data extent is shared or not.
1826 *
1827 * @inode: The inode whose extent we are checking.
1828 * @bytenr: Logical bytenr of the extent we are checking.
1829 * @extent_gen: Generation of the extent (file extent item) or 0 if it is
1830 * not known.
1831 * @ctx: A backref sharedness check context.
1832 *
1833 * btrfs_is_data_extent_shared uses the backref walking code but will short
1834 * circuit as soon as it finds a root or inode that doesn't match the
1835 * one passed in. This provides a significant performance benefit for
1836 * callers (such as fiemap) which want to know whether the extent is
1837 * shared but do not need a ref count.
1838 *
1839 * This attempts to attach to the running transaction in order to account for
1840 * delayed refs, but continues on even when no running transaction exists.
1841 *
1842 * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1843 */
1844int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
1845 u64 extent_gen,
1846 struct btrfs_backref_share_check_ctx *ctx)
1847{
1848 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
1849 struct btrfs_root *root = inode->root;
1850 struct btrfs_fs_info *fs_info = root->fs_info;
1851 struct btrfs_trans_handle *trans;
1852 struct ulist_iterator uiter;
1853 struct ulist_node *node;
1854 struct btrfs_seq_list elem = BTRFS_SEQ_LIST_INIT(elem);
1855 int ret = 0;
1856 struct share_check shared = {
1857 .ctx = ctx,
1858 .root = root,
1859 .inum = btrfs_ino(inode),
1860 .data_bytenr = bytenr,
1861 .data_extent_gen = extent_gen,
1862 .share_count = 0,
1863 .self_ref_count = 0,
1864 .have_delayed_delete_refs = false,
1865 };
1866 int level;
1867
1868 for (int i = 0; i < BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE; i++) {
1869 if (ctx->prev_extents_cache[i].bytenr == bytenr)
1870 return ctx->prev_extents_cache[i].is_shared;
1871 }
1872
1873 ulist_init(&ctx->refs);
1874
1875 trans = btrfs_join_transaction_nostart(root);
1876 if (IS_ERR(trans)) {
1877 if (PTR_ERR(trans) != -ENOENT && PTR_ERR(trans) != -EROFS) {
1878 ret = PTR_ERR(trans);
1879 goto out;
1880 }
1881 trans = NULL;
1882 down_read(&fs_info->commit_root_sem);
1883 } else {
1884 btrfs_get_tree_mod_seq(fs_info, &elem);
1885 walk_ctx.time_seq = elem.seq;
1886 }
1887
1888 walk_ctx.ignore_extent_item_pos = true;
1889 walk_ctx.trans = trans;
1890 walk_ctx.fs_info = fs_info;
1891 walk_ctx.refs = &ctx->refs;
1892
1893 /* -1 means we are in the bytenr of the data extent. */
1894 level = -1;
1895 ULIST_ITER_INIT(&uiter);
1896 ctx->use_path_cache = true;
1897 while (1) {
1898 bool is_shared;
1899 bool cached;
1900
1901 walk_ctx.bytenr = bytenr;
1902 ret = find_parent_nodes(&walk_ctx, &shared);
1903 if (ret == BACKREF_FOUND_SHARED ||
1904 ret == BACKREF_FOUND_NOT_SHARED) {
1905 /* If shared must return 1, otherwise return 0. */
1906 ret = (ret == BACKREF_FOUND_SHARED) ? 1 : 0;
1907 if (level >= 0)
1908 store_backref_shared_cache(ctx, root, bytenr,
1909 level, ret == 1);
1910 break;
1911 }
1912 if (ret < 0 && ret != -ENOENT)
1913 break;
1914 ret = 0;
1915
1916 /*
1917 * If our data extent was not directly shared (without multiple
1918 * reference items), than it might have a single reference item
1919 * with a count > 1 for the same offset, which means there are 2
1920 * (or more) file extent items that point to the data extent -
1921 * this happens when a file extent item needs to be split and
1922 * then one item gets moved to another leaf due to a b+tree leaf
1923 * split when inserting some item. In this case the file extent
1924 * items may be located in different leaves and therefore some
1925 * of the leaves may be referenced through shared subtrees while
1926 * others are not. Since our extent buffer cache only works for
1927 * a single path (by far the most common case and simpler to
1928 * deal with), we can not use it if we have multiple leaves
1929 * (which implies multiple paths).
1930 */
1931 if (level == -1 && ctx->refs.nnodes > 1)
1932 ctx->use_path_cache = false;
1933
1934 if (level >= 0)
1935 store_backref_shared_cache(ctx, root, bytenr,
1936 level, false);
1937 node = ulist_next(&ctx->refs, &uiter);
1938 if (!node)
1939 break;
1940 bytenr = node->val;
1941 level++;
1942 cached = lookup_backref_shared_cache(ctx, root, bytenr, level,
1943 &is_shared);
1944 if (cached) {
1945 ret = (is_shared ? 1 : 0);
1946 break;
1947 }
1948 shared.share_count = 0;
1949 shared.have_delayed_delete_refs = false;
1950 cond_resched();
1951 }
1952
1953 /*
1954 * Cache the sharedness result for the data extent if we know our inode
1955 * has more than 1 file extent item that refers to the data extent.
1956 */
1957 if (ret >= 0 && shared.self_ref_count > 1) {
1958 int slot = ctx->prev_extents_cache_slot;
1959
1960 ctx->prev_extents_cache[slot].bytenr = shared.data_bytenr;
1961 ctx->prev_extents_cache[slot].is_shared = (ret == 1);
1962
1963 slot = (slot + 1) % BTRFS_BACKREF_CTX_PREV_EXTENTS_SIZE;
1964 ctx->prev_extents_cache_slot = slot;
1965 }
1966
1967 if (trans) {
1968 btrfs_put_tree_mod_seq(fs_info, &elem);
1969 btrfs_end_transaction(trans);
1970 } else {
1971 up_read(&fs_info->commit_root_sem);
1972 }
1973out:
1974 ulist_release(&ctx->refs);
1975 ctx->prev_leaf_bytenr = ctx->curr_leaf_bytenr;
1976
1977 return ret;
1978}
1979
1980int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1981 u64 start_off, struct btrfs_path *path,
1982 struct btrfs_inode_extref **ret_extref,
1983 u64 *found_off)
1984{
1985 int ret, slot;
1986 struct btrfs_key key;
1987 struct btrfs_key found_key;
1988 struct btrfs_inode_extref *extref;
1989 const struct extent_buffer *leaf;
1990 unsigned long ptr;
1991
1992 key.objectid = inode_objectid;
1993 key.type = BTRFS_INODE_EXTREF_KEY;
1994 key.offset = start_off;
1995
1996 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1997 if (ret < 0)
1998 return ret;
1999
2000 while (1) {
2001 leaf = path->nodes[0];
2002 slot = path->slots[0];
2003 if (slot >= btrfs_header_nritems(leaf)) {
2004 /*
2005 * If the item at offset is not found,
2006 * btrfs_search_slot will point us to the slot
2007 * where it should be inserted. In our case
2008 * that will be the slot directly before the
2009 * next INODE_REF_KEY_V2 item. In the case
2010 * that we're pointing to the last slot in a
2011 * leaf, we must move one leaf over.
2012 */
2013 ret = btrfs_next_leaf(root, path);
2014 if (ret) {
2015 if (ret >= 1)
2016 ret = -ENOENT;
2017 break;
2018 }
2019 continue;
2020 }
2021
2022 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2023
2024 /*
2025 * Check that we're still looking at an extended ref key for
2026 * this particular objectid. If we have different
2027 * objectid or type then there are no more to be found
2028 * in the tree and we can exit.
2029 */
2030 ret = -ENOENT;
2031 if (found_key.objectid != inode_objectid)
2032 break;
2033 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
2034 break;
2035
2036 ret = 0;
2037 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2038 extref = (struct btrfs_inode_extref *)ptr;
2039 *ret_extref = extref;
2040 if (found_off)
2041 *found_off = found_key.offset;
2042 break;
2043 }
2044
2045 return ret;
2046}
2047
2048/*
2049 * this iterates to turn a name (from iref/extref) into a full filesystem path.
2050 * Elements of the path are separated by '/' and the path is guaranteed to be
2051 * 0-terminated. the path is only given within the current file system.
2052 * Therefore, it never starts with a '/'. the caller is responsible to provide
2053 * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
2054 * the start point of the resulting string is returned. this pointer is within
2055 * dest, normally.
2056 * in case the path buffer would overflow, the pointer is decremented further
2057 * as if output was written to the buffer, though no more output is actually
2058 * generated. that way, the caller can determine how much space would be
2059 * required for the path to fit into the buffer. in that case, the returned
2060 * value will be smaller than dest. callers must check this!
2061 */
2062char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
2063 u32 name_len, unsigned long name_off,
2064 struct extent_buffer *eb_in, u64 parent,
2065 char *dest, u32 size)
2066{
2067 int slot;
2068 u64 next_inum;
2069 int ret;
2070 s64 bytes_left = ((s64)size) - 1;
2071 struct extent_buffer *eb = eb_in;
2072 struct btrfs_key found_key;
2073 struct btrfs_inode_ref *iref;
2074
2075 if (bytes_left >= 0)
2076 dest[bytes_left] = '\0';
2077
2078 while (1) {
2079 bytes_left -= name_len;
2080 if (bytes_left >= 0)
2081 read_extent_buffer(eb, dest + bytes_left,
2082 name_off, name_len);
2083 if (eb != eb_in) {
2084 if (!path->skip_locking)
2085 btrfs_tree_read_unlock(eb);
2086 free_extent_buffer(eb);
2087 }
2088 ret = btrfs_find_item(fs_root, path, parent, 0,
2089 BTRFS_INODE_REF_KEY, &found_key);
2090 if (ret > 0)
2091 ret = -ENOENT;
2092 if (ret)
2093 break;
2094
2095 next_inum = found_key.offset;
2096
2097 /* regular exit ahead */
2098 if (parent == next_inum)
2099 break;
2100
2101 slot = path->slots[0];
2102 eb = path->nodes[0];
2103 /* make sure we can use eb after releasing the path */
2104 if (eb != eb_in) {
2105 path->nodes[0] = NULL;
2106 path->locks[0] = 0;
2107 }
2108 btrfs_release_path(path);
2109 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2110
2111 name_len = btrfs_inode_ref_name_len(eb, iref);
2112 name_off = (unsigned long)(iref + 1);
2113
2114 parent = next_inum;
2115 --bytes_left;
2116 if (bytes_left >= 0)
2117 dest[bytes_left] = '/';
2118 }
2119
2120 btrfs_release_path(path);
2121
2122 if (ret)
2123 return ERR_PTR(ret);
2124
2125 return dest + bytes_left;
2126}
2127
2128/*
2129 * this makes the path point to (logical EXTENT_ITEM *)
2130 * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
2131 * tree blocks and <0 on error.
2132 */
2133int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
2134 struct btrfs_path *path, struct btrfs_key *found_key,
2135 u64 *flags_ret)
2136{
2137 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
2138 int ret;
2139 u64 flags;
2140 u64 size = 0;
2141 u32 item_size;
2142 const struct extent_buffer *eb;
2143 struct btrfs_extent_item *ei;
2144 struct btrfs_key key;
2145
2146 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2147 key.type = BTRFS_METADATA_ITEM_KEY;
2148 else
2149 key.type = BTRFS_EXTENT_ITEM_KEY;
2150 key.objectid = logical;
2151 key.offset = (u64)-1;
2152
2153 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2154 if (ret < 0)
2155 return ret;
2156
2157 ret = btrfs_previous_extent_item(extent_root, path, 0);
2158 if (ret) {
2159 if (ret > 0)
2160 ret = -ENOENT;
2161 return ret;
2162 }
2163 btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
2164 if (found_key->type == BTRFS_METADATA_ITEM_KEY)
2165 size = fs_info->nodesize;
2166 else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
2167 size = found_key->offset;
2168
2169 if (found_key->objectid > logical ||
2170 found_key->objectid + size <= logical) {
2171 btrfs_debug(fs_info,
2172 "logical %llu is not within any extent", logical);
2173 return -ENOENT;
2174 }
2175
2176 eb = path->nodes[0];
2177 item_size = btrfs_item_size(eb, path->slots[0]);
2178 BUG_ON(item_size < sizeof(*ei));
2179
2180 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
2181 flags = btrfs_extent_flags(eb, ei);
2182
2183 btrfs_debug(fs_info,
2184 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
2185 logical, logical - found_key->objectid, found_key->objectid,
2186 found_key->offset, flags, item_size);
2187
2188 WARN_ON(!flags_ret);
2189 if (flags_ret) {
2190 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2191 *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
2192 else if (flags & BTRFS_EXTENT_FLAG_DATA)
2193 *flags_ret = BTRFS_EXTENT_FLAG_DATA;
2194 else
2195 BUG();
2196 return 0;
2197 }
2198
2199 return -EIO;
2200}
2201
2202/*
2203 * helper function to iterate extent inline refs. ptr must point to a 0 value
2204 * for the first call and may be modified. it is used to track state.
2205 * if more refs exist, 0 is returned and the next call to
2206 * get_extent_inline_ref must pass the modified ptr parameter to get the
2207 * next ref. after the last ref was processed, 1 is returned.
2208 * returns <0 on error
2209 */
2210static int get_extent_inline_ref(unsigned long *ptr,
2211 const struct extent_buffer *eb,
2212 const struct btrfs_key *key,
2213 const struct btrfs_extent_item *ei,
2214 u32 item_size,
2215 struct btrfs_extent_inline_ref **out_eiref,
2216 int *out_type)
2217{
2218 unsigned long end;
2219 u64 flags;
2220 struct btrfs_tree_block_info *info;
2221
2222 if (!*ptr) {
2223 /* first call */
2224 flags = btrfs_extent_flags(eb, ei);
2225 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2226 if (key->type == BTRFS_METADATA_ITEM_KEY) {
2227 /* a skinny metadata extent */
2228 *out_eiref =
2229 (struct btrfs_extent_inline_ref *)(ei + 1);
2230 } else {
2231 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
2232 info = (struct btrfs_tree_block_info *)(ei + 1);
2233 *out_eiref =
2234 (struct btrfs_extent_inline_ref *)(info + 1);
2235 }
2236 } else {
2237 *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
2238 }
2239 *ptr = (unsigned long)*out_eiref;
2240 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
2241 return -ENOENT;
2242 }
2243
2244 end = (unsigned long)ei + item_size;
2245 *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
2246 *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
2247 BTRFS_REF_TYPE_ANY);
2248 if (*out_type == BTRFS_REF_TYPE_INVALID)
2249 return -EUCLEAN;
2250
2251 *ptr += btrfs_extent_inline_ref_size(*out_type);
2252 WARN_ON(*ptr > end);
2253 if (*ptr == end)
2254 return 1; /* last */
2255
2256 return 0;
2257}
2258
2259/*
2260 * reads the tree block backref for an extent. tree level and root are returned
2261 * through out_level and out_root. ptr must point to a 0 value for the first
2262 * call and may be modified (see get_extent_inline_ref comment).
2263 * returns 0 if data was provided, 1 if there was no more data to provide or
2264 * <0 on error.
2265 */
2266int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
2267 struct btrfs_key *key, struct btrfs_extent_item *ei,
2268 u32 item_size, u64 *out_root, u8 *out_level)
2269{
2270 int ret;
2271 int type;
2272 struct btrfs_extent_inline_ref *eiref;
2273
2274 if (*ptr == (unsigned long)-1)
2275 return 1;
2276
2277 while (1) {
2278 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
2279 &eiref, &type);
2280 if (ret < 0)
2281 return ret;
2282
2283 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
2284 type == BTRFS_SHARED_BLOCK_REF_KEY)
2285 break;
2286
2287 if (ret == 1)
2288 return 1;
2289 }
2290
2291 /* we can treat both ref types equally here */
2292 *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
2293
2294 if (key->type == BTRFS_EXTENT_ITEM_KEY) {
2295 struct btrfs_tree_block_info *info;
2296
2297 info = (struct btrfs_tree_block_info *)(ei + 1);
2298 *out_level = btrfs_tree_block_level(eb, info);
2299 } else {
2300 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
2301 *out_level = (u8)key->offset;
2302 }
2303
2304 if (ret == 1)
2305 *ptr = (unsigned long)-1;
2306
2307 return 0;
2308}
2309
2310static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
2311 struct extent_inode_elem *inode_list,
2312 u64 root, u64 extent_item_objectid,
2313 iterate_extent_inodes_t *iterate, void *ctx)
2314{
2315 struct extent_inode_elem *eie;
2316 int ret = 0;
2317
2318 for (eie = inode_list; eie; eie = eie->next) {
2319 btrfs_debug(fs_info,
2320 "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
2321 extent_item_objectid, eie->inum,
2322 eie->offset, root);
2323 ret = iterate(eie->inum, eie->offset, eie->num_bytes, root, ctx);
2324 if (ret) {
2325 btrfs_debug(fs_info,
2326 "stopping iteration for %llu due to ret=%d",
2327 extent_item_objectid, ret);
2328 break;
2329 }
2330 }
2331
2332 return ret;
2333}
2334
2335/*
2336 * calls iterate() for every inode that references the extent identified by
2337 * the given parameters.
2338 * when the iterator function returns a non-zero value, iteration stops.
2339 */
2340int iterate_extent_inodes(struct btrfs_backref_walk_ctx *ctx,
2341 bool search_commit_root,
2342 iterate_extent_inodes_t *iterate, void *user_ctx)
2343{
2344 int ret;
2345 struct ulist *refs;
2346 struct ulist_node *ref_node;
2347 struct btrfs_seq_list seq_elem = BTRFS_SEQ_LIST_INIT(seq_elem);
2348 struct ulist_iterator ref_uiter;
2349
2350 btrfs_debug(ctx->fs_info, "resolving all inodes for extent %llu",
2351 ctx->bytenr);
2352
2353 ASSERT(ctx->trans == NULL);
2354 ASSERT(ctx->roots == NULL);
2355
2356 if (!search_commit_root) {
2357 struct btrfs_trans_handle *trans;
2358
2359 trans = btrfs_attach_transaction(ctx->fs_info->tree_root);
2360 if (IS_ERR(trans)) {
2361 if (PTR_ERR(trans) != -ENOENT &&
2362 PTR_ERR(trans) != -EROFS)
2363 return PTR_ERR(trans);
2364 trans = NULL;
2365 }
2366 ctx->trans = trans;
2367 }
2368
2369 if (ctx->trans) {
2370 btrfs_get_tree_mod_seq(ctx->fs_info, &seq_elem);
2371 ctx->time_seq = seq_elem.seq;
2372 } else {
2373 down_read(&ctx->fs_info->commit_root_sem);
2374 }
2375
2376 ret = btrfs_find_all_leafs(ctx);
2377 if (ret)
2378 goto out;
2379 refs = ctx->refs;
2380 ctx->refs = NULL;
2381
2382 ULIST_ITER_INIT(&ref_uiter);
2383 while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2384 const u64 leaf_bytenr = ref_node->val;
2385 struct ulist_node *root_node;
2386 struct ulist_iterator root_uiter;
2387 struct extent_inode_elem *inode_list;
2388
2389 inode_list = (struct extent_inode_elem *)(uintptr_t)ref_node->aux;
2390
2391 if (ctx->cache_lookup) {
2392 const u64 *root_ids;
2393 int root_count;
2394 bool cached;
2395
2396 cached = ctx->cache_lookup(leaf_bytenr, ctx->user_ctx,
2397 &root_ids, &root_count);
2398 if (cached) {
2399 for (int i = 0; i < root_count; i++) {
2400 ret = iterate_leaf_refs(ctx->fs_info,
2401 inode_list,
2402 root_ids[i],
2403 leaf_bytenr,
2404 iterate,
2405 user_ctx);
2406 if (ret)
2407 break;
2408 }
2409 continue;
2410 }
2411 }
2412
2413 if (!ctx->roots) {
2414 ctx->roots = ulist_alloc(GFP_NOFS);
2415 if (!ctx->roots) {
2416 ret = -ENOMEM;
2417 break;
2418 }
2419 }
2420
2421 ctx->bytenr = leaf_bytenr;
2422 ret = btrfs_find_all_roots_safe(ctx);
2423 if (ret)
2424 break;
2425
2426 if (ctx->cache_store)
2427 ctx->cache_store(leaf_bytenr, ctx->roots, ctx->user_ctx);
2428
2429 ULIST_ITER_INIT(&root_uiter);
2430 while (!ret && (root_node = ulist_next(ctx->roots, &root_uiter))) {
2431 btrfs_debug(ctx->fs_info,
2432 "root %llu references leaf %llu, data list %#llx",
2433 root_node->val, ref_node->val,
2434 ref_node->aux);
2435 ret = iterate_leaf_refs(ctx->fs_info, inode_list,
2436 root_node->val, ctx->bytenr,
2437 iterate, user_ctx);
2438 }
2439 ulist_reinit(ctx->roots);
2440 }
2441
2442 free_leaf_list(refs);
2443out:
2444 if (ctx->trans) {
2445 btrfs_put_tree_mod_seq(ctx->fs_info, &seq_elem);
2446 btrfs_end_transaction(ctx->trans);
2447 ctx->trans = NULL;
2448 } else {
2449 up_read(&ctx->fs_info->commit_root_sem);
2450 }
2451
2452 ulist_free(ctx->roots);
2453 ctx->roots = NULL;
2454
2455 if (ret == BTRFS_ITERATE_EXTENT_INODES_STOP)
2456 ret = 0;
2457
2458 return ret;
2459}
2460
2461static int build_ino_list(u64 inum, u64 offset, u64 num_bytes, u64 root, void *ctx)
2462{
2463 struct btrfs_data_container *inodes = ctx;
2464 const size_t c = 3 * sizeof(u64);
2465
2466 if (inodes->bytes_left >= c) {
2467 inodes->bytes_left -= c;
2468 inodes->val[inodes->elem_cnt] = inum;
2469 inodes->val[inodes->elem_cnt + 1] = offset;
2470 inodes->val[inodes->elem_cnt + 2] = root;
2471 inodes->elem_cnt += 3;
2472 } else {
2473 inodes->bytes_missing += c - inodes->bytes_left;
2474 inodes->bytes_left = 0;
2475 inodes->elem_missed += 3;
2476 }
2477
2478 return 0;
2479}
2480
2481int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2482 struct btrfs_path *path,
2483 void *ctx, bool ignore_offset)
2484{
2485 struct btrfs_backref_walk_ctx walk_ctx = { 0 };
2486 int ret;
2487 u64 flags = 0;
2488 struct btrfs_key found_key;
2489 int search_commit_root = path->search_commit_root;
2490
2491 ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2492 btrfs_release_path(path);
2493 if (ret < 0)
2494 return ret;
2495 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2496 return -EINVAL;
2497
2498 walk_ctx.bytenr = found_key.objectid;
2499 if (ignore_offset)
2500 walk_ctx.ignore_extent_item_pos = true;
2501 else
2502 walk_ctx.extent_item_pos = logical - found_key.objectid;
2503 walk_ctx.fs_info = fs_info;
2504
2505 return iterate_extent_inodes(&walk_ctx, search_commit_root,
2506 build_ino_list, ctx);
2507}
2508
2509static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2510 struct extent_buffer *eb, struct inode_fs_paths *ipath);
2511
2512static int iterate_inode_refs(u64 inum, struct inode_fs_paths *ipath)
2513{
2514 int ret = 0;
2515 int slot;
2516 u32 cur;
2517 u32 len;
2518 u32 name_len;
2519 u64 parent = 0;
2520 int found = 0;
2521 struct btrfs_root *fs_root = ipath->fs_root;
2522 struct btrfs_path *path = ipath->btrfs_path;
2523 struct extent_buffer *eb;
2524 struct btrfs_inode_ref *iref;
2525 struct btrfs_key found_key;
2526
2527 while (!ret) {
2528 ret = btrfs_find_item(fs_root, path, inum,
2529 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2530 &found_key);
2531
2532 if (ret < 0)
2533 break;
2534 if (ret) {
2535 ret = found ? 0 : -ENOENT;
2536 break;
2537 }
2538 ++found;
2539
2540 parent = found_key.offset;
2541 slot = path->slots[0];
2542 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2543 if (!eb) {
2544 ret = -ENOMEM;
2545 break;
2546 }
2547 btrfs_release_path(path);
2548
2549 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2550
2551 for (cur = 0; cur < btrfs_item_size(eb, slot); cur += len) {
2552 name_len = btrfs_inode_ref_name_len(eb, iref);
2553 /* path must be released before calling iterate()! */
2554 btrfs_debug(fs_root->fs_info,
2555 "following ref at offset %u for inode %llu in tree %llu",
2556 cur, found_key.objectid,
2557 fs_root->root_key.objectid);
2558 ret = inode_to_path(parent, name_len,
2559 (unsigned long)(iref + 1), eb, ipath);
2560 if (ret)
2561 break;
2562 len = sizeof(*iref) + name_len;
2563 iref = (struct btrfs_inode_ref *)((char *)iref + len);
2564 }
2565 free_extent_buffer(eb);
2566 }
2567
2568 btrfs_release_path(path);
2569
2570 return ret;
2571}
2572
2573static int iterate_inode_extrefs(u64 inum, struct inode_fs_paths *ipath)
2574{
2575 int ret;
2576 int slot;
2577 u64 offset = 0;
2578 u64 parent;
2579 int found = 0;
2580 struct btrfs_root *fs_root = ipath->fs_root;
2581 struct btrfs_path *path = ipath->btrfs_path;
2582 struct extent_buffer *eb;
2583 struct btrfs_inode_extref *extref;
2584 u32 item_size;
2585 u32 cur_offset;
2586 unsigned long ptr;
2587
2588 while (1) {
2589 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2590 &offset);
2591 if (ret < 0)
2592 break;
2593 if (ret) {
2594 ret = found ? 0 : -ENOENT;
2595 break;
2596 }
2597 ++found;
2598
2599 slot = path->slots[0];
2600 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2601 if (!eb) {
2602 ret = -ENOMEM;
2603 break;
2604 }
2605 btrfs_release_path(path);
2606
2607 item_size = btrfs_item_size(eb, slot);
2608 ptr = btrfs_item_ptr_offset(eb, slot);
2609 cur_offset = 0;
2610
2611 while (cur_offset < item_size) {
2612 u32 name_len;
2613
2614 extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2615 parent = btrfs_inode_extref_parent(eb, extref);
2616 name_len = btrfs_inode_extref_name_len(eb, extref);
2617 ret = inode_to_path(parent, name_len,
2618 (unsigned long)&extref->name, eb, ipath);
2619 if (ret)
2620 break;
2621
2622 cur_offset += btrfs_inode_extref_name_len(eb, extref);
2623 cur_offset += sizeof(*extref);
2624 }
2625 free_extent_buffer(eb);
2626
2627 offset++;
2628 }
2629
2630 btrfs_release_path(path);
2631
2632 return ret;
2633}
2634
2635/*
2636 * returns 0 if the path could be dumped (probably truncated)
2637 * returns <0 in case of an error
2638 */
2639static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2640 struct extent_buffer *eb, struct inode_fs_paths *ipath)
2641{
2642 char *fspath;
2643 char *fspath_min;
2644 int i = ipath->fspath->elem_cnt;
2645 const int s_ptr = sizeof(char *);
2646 u32 bytes_left;
2647
2648 bytes_left = ipath->fspath->bytes_left > s_ptr ?
2649 ipath->fspath->bytes_left - s_ptr : 0;
2650
2651 fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2652 fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2653 name_off, eb, inum, fspath_min, bytes_left);
2654 if (IS_ERR(fspath))
2655 return PTR_ERR(fspath);
2656
2657 if (fspath > fspath_min) {
2658 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2659 ++ipath->fspath->elem_cnt;
2660 ipath->fspath->bytes_left = fspath - fspath_min;
2661 } else {
2662 ++ipath->fspath->elem_missed;
2663 ipath->fspath->bytes_missing += fspath_min - fspath;
2664 ipath->fspath->bytes_left = 0;
2665 }
2666
2667 return 0;
2668}
2669
2670/*
2671 * this dumps all file system paths to the inode into the ipath struct, provided
2672 * is has been created large enough. each path is zero-terminated and accessed
2673 * from ipath->fspath->val[i].
2674 * when it returns, there are ipath->fspath->elem_cnt number of paths available
2675 * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2676 * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2677 * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2678 * have been needed to return all paths.
2679 */
2680int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2681{
2682 int ret;
2683 int found_refs = 0;
2684
2685 ret = iterate_inode_refs(inum, ipath);
2686 if (!ret)
2687 ++found_refs;
2688 else if (ret != -ENOENT)
2689 return ret;
2690
2691 ret = iterate_inode_extrefs(inum, ipath);
2692 if (ret == -ENOENT && found_refs)
2693 return 0;
2694
2695 return ret;
2696}
2697
2698struct btrfs_data_container *init_data_container(u32 total_bytes)
2699{
2700 struct btrfs_data_container *data;
2701 size_t alloc_bytes;
2702
2703 alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2704 data = kvmalloc(alloc_bytes, GFP_KERNEL);
2705 if (!data)
2706 return ERR_PTR(-ENOMEM);
2707
2708 if (total_bytes >= sizeof(*data)) {
2709 data->bytes_left = total_bytes - sizeof(*data);
2710 data->bytes_missing = 0;
2711 } else {
2712 data->bytes_missing = sizeof(*data) - total_bytes;
2713 data->bytes_left = 0;
2714 }
2715
2716 data->elem_cnt = 0;
2717 data->elem_missed = 0;
2718
2719 return data;
2720}
2721
2722/*
2723 * allocates space to return multiple file system paths for an inode.
2724 * total_bytes to allocate are passed, note that space usable for actual path
2725 * information will be total_bytes - sizeof(struct inode_fs_paths).
2726 * the returned pointer must be freed with free_ipath() in the end.
2727 */
2728struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2729 struct btrfs_path *path)
2730{
2731 struct inode_fs_paths *ifp;
2732 struct btrfs_data_container *fspath;
2733
2734 fspath = init_data_container(total_bytes);
2735 if (IS_ERR(fspath))
2736 return ERR_CAST(fspath);
2737
2738 ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2739 if (!ifp) {
2740 kvfree(fspath);
2741 return ERR_PTR(-ENOMEM);
2742 }
2743
2744 ifp->btrfs_path = path;
2745 ifp->fspath = fspath;
2746 ifp->fs_root = fs_root;
2747
2748 return ifp;
2749}
2750
2751void free_ipath(struct inode_fs_paths *ipath)
2752{
2753 if (!ipath)
2754 return;
2755 kvfree(ipath->fspath);
2756 kfree(ipath);
2757}
2758
2759struct btrfs_backref_iter *btrfs_backref_iter_alloc(struct btrfs_fs_info *fs_info)
2760{
2761 struct btrfs_backref_iter *ret;
2762
2763 ret = kzalloc(sizeof(*ret), GFP_NOFS);
2764 if (!ret)
2765 return NULL;
2766
2767 ret->path = btrfs_alloc_path();
2768 if (!ret->path) {
2769 kfree(ret);
2770 return NULL;
2771 }
2772
2773 /* Current backref iterator only supports iteration in commit root */
2774 ret->path->search_commit_root = 1;
2775 ret->path->skip_locking = 1;
2776 ret->fs_info = fs_info;
2777
2778 return ret;
2779}
2780
2781int btrfs_backref_iter_start(struct btrfs_backref_iter *iter, u64 bytenr)
2782{
2783 struct btrfs_fs_info *fs_info = iter->fs_info;
2784 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bytenr);
2785 struct btrfs_path *path = iter->path;
2786 struct btrfs_extent_item *ei;
2787 struct btrfs_key key;
2788 int ret;
2789
2790 key.objectid = bytenr;
2791 key.type = BTRFS_METADATA_ITEM_KEY;
2792 key.offset = (u64)-1;
2793 iter->bytenr = bytenr;
2794
2795 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2796 if (ret < 0)
2797 return ret;
2798 if (ret == 0) {
2799 ret = -EUCLEAN;
2800 goto release;
2801 }
2802 if (path->slots[0] == 0) {
2803 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
2804 ret = -EUCLEAN;
2805 goto release;
2806 }
2807 path->slots[0]--;
2808
2809 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2810 if ((key.type != BTRFS_EXTENT_ITEM_KEY &&
2811 key.type != BTRFS_METADATA_ITEM_KEY) || key.objectid != bytenr) {
2812 ret = -ENOENT;
2813 goto release;
2814 }
2815 memcpy(&iter->cur_key, &key, sizeof(key));
2816 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2817 path->slots[0]);
2818 iter->end_ptr = (u32)(iter->item_ptr +
2819 btrfs_item_size(path->nodes[0], path->slots[0]));
2820 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
2821 struct btrfs_extent_item);
2822
2823 /*
2824 * Only support iteration on tree backref yet.
2825 *
2826 * This is an extra precaution for non skinny-metadata, where
2827 * EXTENT_ITEM is also used for tree blocks, that we can only use
2828 * extent flags to determine if it's a tree block.
2829 */
2830 if (btrfs_extent_flags(path->nodes[0], ei) & BTRFS_EXTENT_FLAG_DATA) {
2831 ret = -ENOTSUPP;
2832 goto release;
2833 }
2834 iter->cur_ptr = (u32)(iter->item_ptr + sizeof(*ei));
2835
2836 /* If there is no inline backref, go search for keyed backref */
2837 if (iter->cur_ptr >= iter->end_ptr) {
2838 ret = btrfs_next_item(extent_root, path);
2839
2840 /* No inline nor keyed ref */
2841 if (ret > 0) {
2842 ret = -ENOENT;
2843 goto release;
2844 }
2845 if (ret < 0)
2846 goto release;
2847
2848 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key,
2849 path->slots[0]);
2850 if (iter->cur_key.objectid != bytenr ||
2851 (iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2852 iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY)) {
2853 ret = -ENOENT;
2854 goto release;
2855 }
2856 iter->cur_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2857 path->slots[0]);
2858 iter->item_ptr = iter->cur_ptr;
2859 iter->end_ptr = (u32)(iter->item_ptr + btrfs_item_size(
2860 path->nodes[0], path->slots[0]));
2861 }
2862
2863 return 0;
2864release:
2865 btrfs_backref_iter_release(iter);
2866 return ret;
2867}
2868
2869/*
2870 * Go to the next backref item of current bytenr, can be either inlined or
2871 * keyed.
2872 *
2873 * Caller needs to check whether it's inline ref or not by iter->cur_key.
2874 *
2875 * Return 0 if we get next backref without problem.
2876 * Return >0 if there is no extra backref for this bytenr.
2877 * Return <0 if there is something wrong happened.
2878 */
2879int btrfs_backref_iter_next(struct btrfs_backref_iter *iter)
2880{
2881 struct extent_buffer *eb = btrfs_backref_get_eb(iter);
2882 struct btrfs_root *extent_root;
2883 struct btrfs_path *path = iter->path;
2884 struct btrfs_extent_inline_ref *iref;
2885 int ret;
2886 u32 size;
2887
2888 if (btrfs_backref_iter_is_inline_ref(iter)) {
2889 /* We're still inside the inline refs */
2890 ASSERT(iter->cur_ptr < iter->end_ptr);
2891
2892 if (btrfs_backref_has_tree_block_info(iter)) {
2893 /* First tree block info */
2894 size = sizeof(struct btrfs_tree_block_info);
2895 } else {
2896 /* Use inline ref type to determine the size */
2897 int type;
2898
2899 iref = (struct btrfs_extent_inline_ref *)
2900 ((unsigned long)iter->cur_ptr);
2901 type = btrfs_extent_inline_ref_type(eb, iref);
2902
2903 size = btrfs_extent_inline_ref_size(type);
2904 }
2905 iter->cur_ptr += size;
2906 if (iter->cur_ptr < iter->end_ptr)
2907 return 0;
2908
2909 /* All inline items iterated, fall through */
2910 }
2911
2912 /* We're at keyed items, there is no inline item, go to the next one */
2913 extent_root = btrfs_extent_root(iter->fs_info, iter->bytenr);
2914 ret = btrfs_next_item(extent_root, iter->path);
2915 if (ret)
2916 return ret;
2917
2918 btrfs_item_key_to_cpu(path->nodes[0], &iter->cur_key, path->slots[0]);
2919 if (iter->cur_key.objectid != iter->bytenr ||
2920 (iter->cur_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2921 iter->cur_key.type != BTRFS_SHARED_BLOCK_REF_KEY))
2922 return 1;
2923 iter->item_ptr = (u32)btrfs_item_ptr_offset(path->nodes[0],
2924 path->slots[0]);
2925 iter->cur_ptr = iter->item_ptr;
2926 iter->end_ptr = iter->item_ptr + (u32)btrfs_item_size(path->nodes[0],
2927 path->slots[0]);
2928 return 0;
2929}
2930
2931void btrfs_backref_init_cache(struct btrfs_fs_info *fs_info,
2932 struct btrfs_backref_cache *cache, int is_reloc)
2933{
2934 int i;
2935
2936 cache->rb_root = RB_ROOT;
2937 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
2938 INIT_LIST_HEAD(&cache->pending[i]);
2939 INIT_LIST_HEAD(&cache->changed);
2940 INIT_LIST_HEAD(&cache->detached);
2941 INIT_LIST_HEAD(&cache->leaves);
2942 INIT_LIST_HEAD(&cache->pending_edge);
2943 INIT_LIST_HEAD(&cache->useless_node);
2944 cache->fs_info = fs_info;
2945 cache->is_reloc = is_reloc;
2946}
2947
2948struct btrfs_backref_node *btrfs_backref_alloc_node(
2949 struct btrfs_backref_cache *cache, u64 bytenr, int level)
2950{
2951 struct btrfs_backref_node *node;
2952
2953 ASSERT(level >= 0 && level < BTRFS_MAX_LEVEL);
2954 node = kzalloc(sizeof(*node), GFP_NOFS);
2955 if (!node)
2956 return node;
2957
2958 INIT_LIST_HEAD(&node->list);
2959 INIT_LIST_HEAD(&node->upper);
2960 INIT_LIST_HEAD(&node->lower);
2961 RB_CLEAR_NODE(&node->rb_node);
2962 cache->nr_nodes++;
2963 node->level = level;
2964 node->bytenr = bytenr;
2965
2966 return node;
2967}
2968
2969struct btrfs_backref_edge *btrfs_backref_alloc_edge(
2970 struct btrfs_backref_cache *cache)
2971{
2972 struct btrfs_backref_edge *edge;
2973
2974 edge = kzalloc(sizeof(*edge), GFP_NOFS);
2975 if (edge)
2976 cache->nr_edges++;
2977 return edge;
2978}
2979
2980/*
2981 * Drop the backref node from cache, also cleaning up all its
2982 * upper edges and any uncached nodes in the path.
2983 *
2984 * This cleanup happens bottom up, thus the node should either
2985 * be the lowest node in the cache or a detached node.
2986 */
2987void btrfs_backref_cleanup_node(struct btrfs_backref_cache *cache,
2988 struct btrfs_backref_node *node)
2989{
2990 struct btrfs_backref_node *upper;
2991 struct btrfs_backref_edge *edge;
2992
2993 if (!node)
2994 return;
2995
2996 BUG_ON(!node->lowest && !node->detached);
2997 while (!list_empty(&node->upper)) {
2998 edge = list_entry(node->upper.next, struct btrfs_backref_edge,
2999 list[LOWER]);
3000 upper = edge->node[UPPER];
3001 list_del(&edge->list[LOWER]);
3002 list_del(&edge->list[UPPER]);
3003 btrfs_backref_free_edge(cache, edge);
3004
3005 /*
3006 * Add the node to leaf node list if no other child block
3007 * cached.
3008 */
3009 if (list_empty(&upper->lower)) {
3010 list_add_tail(&upper->lower, &cache->leaves);
3011 upper->lowest = 1;
3012 }
3013 }
3014
3015 btrfs_backref_drop_node(cache, node);
3016}
3017
3018/*
3019 * Release all nodes/edges from current cache
3020 */
3021void btrfs_backref_release_cache(struct btrfs_backref_cache *cache)
3022{
3023 struct btrfs_backref_node *node;
3024 int i;
3025
3026 while (!list_empty(&cache->detached)) {
3027 node = list_entry(cache->detached.next,
3028 struct btrfs_backref_node, list);
3029 btrfs_backref_cleanup_node(cache, node);
3030 }
3031
3032 while (!list_empty(&cache->leaves)) {
3033 node = list_entry(cache->leaves.next,
3034 struct btrfs_backref_node, lower);
3035 btrfs_backref_cleanup_node(cache, node);
3036 }
3037
3038 cache->last_trans = 0;
3039
3040 for (i = 0; i < BTRFS_MAX_LEVEL; i++)
3041 ASSERT(list_empty(&cache->pending[i]));
3042 ASSERT(list_empty(&cache->pending_edge));
3043 ASSERT(list_empty(&cache->useless_node));
3044 ASSERT(list_empty(&cache->changed));
3045 ASSERT(list_empty(&cache->detached));
3046 ASSERT(RB_EMPTY_ROOT(&cache->rb_root));
3047 ASSERT(!cache->nr_nodes);
3048 ASSERT(!cache->nr_edges);
3049}
3050
3051/*
3052 * Handle direct tree backref
3053 *
3054 * Direct tree backref means, the backref item shows its parent bytenr
3055 * directly. This is for SHARED_BLOCK_REF backref (keyed or inlined).
3056 *
3057 * @ref_key: The converted backref key.
3058 * For keyed backref, it's the item key.
3059 * For inlined backref, objectid is the bytenr,
3060 * type is btrfs_inline_ref_type, offset is
3061 * btrfs_inline_ref_offset.
3062 */
3063static int handle_direct_tree_backref(struct btrfs_backref_cache *cache,
3064 struct btrfs_key *ref_key,
3065 struct btrfs_backref_node *cur)
3066{
3067 struct btrfs_backref_edge *edge;
3068 struct btrfs_backref_node *upper;
3069 struct rb_node *rb_node;
3070
3071 ASSERT(ref_key->type == BTRFS_SHARED_BLOCK_REF_KEY);
3072
3073 /* Only reloc root uses backref pointing to itself */
3074 if (ref_key->objectid == ref_key->offset) {
3075 struct btrfs_root *root;
3076
3077 cur->is_reloc_root = 1;
3078 /* Only reloc backref cache cares about a specific root */
3079 if (cache->is_reloc) {
3080 root = find_reloc_root(cache->fs_info, cur->bytenr);
3081 if (!root)
3082 return -ENOENT;
3083 cur->root = root;
3084 } else {
3085 /*
3086 * For generic purpose backref cache, reloc root node
3087 * is useless.
3088 */
3089 list_add(&cur->list, &cache->useless_node);
3090 }
3091 return 0;
3092 }
3093
3094 edge = btrfs_backref_alloc_edge(cache);
3095 if (!edge)
3096 return -ENOMEM;
3097
3098 rb_node = rb_simple_search(&cache->rb_root, ref_key->offset);
3099 if (!rb_node) {
3100 /* Parent node not yet cached */
3101 upper = btrfs_backref_alloc_node(cache, ref_key->offset,
3102 cur->level + 1);
3103 if (!upper) {
3104 btrfs_backref_free_edge(cache, edge);
3105 return -ENOMEM;
3106 }
3107
3108 /*
3109 * Backrefs for the upper level block isn't cached, add the
3110 * block to pending list
3111 */
3112 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3113 } else {
3114 /* Parent node already cached */
3115 upper = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
3116 ASSERT(upper->checked);
3117 INIT_LIST_HEAD(&edge->list[UPPER]);
3118 }
3119 btrfs_backref_link_edge(edge, cur, upper, LINK_LOWER);
3120 return 0;
3121}
3122
3123/*
3124 * Handle indirect tree backref
3125 *
3126 * Indirect tree backref means, we only know which tree the node belongs to.
3127 * We still need to do a tree search to find out the parents. This is for
3128 * TREE_BLOCK_REF backref (keyed or inlined).
3129 *
3130 * @ref_key: The same as @ref_key in handle_direct_tree_backref()
3131 * @tree_key: The first key of this tree block.
3132 * @path: A clean (released) path, to avoid allocating path every time
3133 * the function get called.
3134 */
3135static int handle_indirect_tree_backref(struct btrfs_backref_cache *cache,
3136 struct btrfs_path *path,
3137 struct btrfs_key *ref_key,
3138 struct btrfs_key *tree_key,
3139 struct btrfs_backref_node *cur)
3140{
3141 struct btrfs_fs_info *fs_info = cache->fs_info;
3142 struct btrfs_backref_node *upper;
3143 struct btrfs_backref_node *lower;
3144 struct btrfs_backref_edge *edge;
3145 struct extent_buffer *eb;
3146 struct btrfs_root *root;
3147 struct rb_node *rb_node;
3148 int level;
3149 bool need_check = true;
3150 int ret;
3151
3152 root = btrfs_get_fs_root(fs_info, ref_key->offset, false);
3153 if (IS_ERR(root))
3154 return PTR_ERR(root);
3155 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3156 cur->cowonly = 1;
3157
3158 if (btrfs_root_level(&root->root_item) == cur->level) {
3159 /* Tree root */
3160 ASSERT(btrfs_root_bytenr(&root->root_item) == cur->bytenr);
3161 /*
3162 * For reloc backref cache, we may ignore reloc root. But for
3163 * general purpose backref cache, we can't rely on
3164 * btrfs_should_ignore_reloc_root() as it may conflict with
3165 * current running relocation and lead to missing root.
3166 *
3167 * For general purpose backref cache, reloc root detection is
3168 * completely relying on direct backref (key->offset is parent
3169 * bytenr), thus only do such check for reloc cache.
3170 */
3171 if (btrfs_should_ignore_reloc_root(root) && cache->is_reloc) {
3172 btrfs_put_root(root);
3173 list_add(&cur->list, &cache->useless_node);
3174 } else {
3175 cur->root = root;
3176 }
3177 return 0;
3178 }
3179
3180 level = cur->level + 1;
3181
3182 /* Search the tree to find parent blocks referring to the block */
3183 path->search_commit_root = 1;
3184 path->skip_locking = 1;
3185 path->lowest_level = level;
3186 ret = btrfs_search_slot(NULL, root, tree_key, path, 0, 0);
3187 path->lowest_level = 0;
3188 if (ret < 0) {
3189 btrfs_put_root(root);
3190 return ret;
3191 }
3192 if (ret > 0 && path->slots[level] > 0)
3193 path->slots[level]--;
3194
3195 eb = path->nodes[level];
3196 if (btrfs_node_blockptr(eb, path->slots[level]) != cur->bytenr) {
3197 btrfs_err(fs_info,
3198"couldn't find block (%llu) (level %d) in tree (%llu) with key (%llu %u %llu)",
3199 cur->bytenr, level - 1, root->root_key.objectid,
3200 tree_key->objectid, tree_key->type, tree_key->offset);
3201 btrfs_put_root(root);
3202 ret = -ENOENT;
3203 goto out;
3204 }
3205 lower = cur;
3206
3207 /* Add all nodes and edges in the path */
3208 for (; level < BTRFS_MAX_LEVEL; level++) {
3209 if (!path->nodes[level]) {
3210 ASSERT(btrfs_root_bytenr(&root->root_item) ==
3211 lower->bytenr);
3212 /* Same as previous should_ignore_reloc_root() call */
3213 if (btrfs_should_ignore_reloc_root(root) &&
3214 cache->is_reloc) {
3215 btrfs_put_root(root);
3216 list_add(&lower->list, &cache->useless_node);
3217 } else {
3218 lower->root = root;
3219 }
3220 break;
3221 }
3222
3223 edge = btrfs_backref_alloc_edge(cache);
3224 if (!edge) {
3225 btrfs_put_root(root);
3226 ret = -ENOMEM;
3227 goto out;
3228 }
3229
3230 eb = path->nodes[level];
3231 rb_node = rb_simple_search(&cache->rb_root, eb->start);
3232 if (!rb_node) {
3233 upper = btrfs_backref_alloc_node(cache, eb->start,
3234 lower->level + 1);
3235 if (!upper) {
3236 btrfs_put_root(root);
3237 btrfs_backref_free_edge(cache, edge);
3238 ret = -ENOMEM;
3239 goto out;
3240 }
3241 upper->owner = btrfs_header_owner(eb);
3242 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
3243 upper->cowonly = 1;
3244
3245 /*
3246 * If we know the block isn't shared we can avoid
3247 * checking its backrefs.
3248 */
3249 if (btrfs_block_can_be_shared(root, eb))
3250 upper->checked = 0;
3251 else
3252 upper->checked = 1;
3253
3254 /*
3255 * Add the block to pending list if we need to check its
3256 * backrefs, we only do this once while walking up a
3257 * tree as we will catch anything else later on.
3258 */
3259 if (!upper->checked && need_check) {
3260 need_check = false;
3261 list_add_tail(&edge->list[UPPER],
3262 &cache->pending_edge);
3263 } else {
3264 if (upper->checked)
3265 need_check = true;
3266 INIT_LIST_HEAD(&edge->list[UPPER]);
3267 }
3268 } else {
3269 upper = rb_entry(rb_node, struct btrfs_backref_node,
3270 rb_node);
3271 ASSERT(upper->checked);
3272 INIT_LIST_HEAD(&edge->list[UPPER]);
3273 if (!upper->owner)
3274 upper->owner = btrfs_header_owner(eb);
3275 }
3276 btrfs_backref_link_edge(edge, lower, upper, LINK_LOWER);
3277
3278 if (rb_node) {
3279 btrfs_put_root(root);
3280 break;
3281 }
3282 lower = upper;
3283 upper = NULL;
3284 }
3285out:
3286 btrfs_release_path(path);
3287 return ret;
3288}
3289
3290/*
3291 * Add backref node @cur into @cache.
3292 *
3293 * NOTE: Even if the function returned 0, @cur is not yet cached as its upper
3294 * links aren't yet bi-directional. Needs to finish such links.
3295 * Use btrfs_backref_finish_upper_links() to finish such linkage.
3296 *
3297 * @path: Released path for indirect tree backref lookup
3298 * @iter: Released backref iter for extent tree search
3299 * @node_key: The first key of the tree block
3300 */
3301int btrfs_backref_add_tree_node(struct btrfs_backref_cache *cache,
3302 struct btrfs_path *path,
3303 struct btrfs_backref_iter *iter,
3304 struct btrfs_key *node_key,
3305 struct btrfs_backref_node *cur)
3306{
3307 struct btrfs_fs_info *fs_info = cache->fs_info;
3308 struct btrfs_backref_edge *edge;
3309 struct btrfs_backref_node *exist;
3310 int ret;
3311
3312 ret = btrfs_backref_iter_start(iter, cur->bytenr);
3313 if (ret < 0)
3314 return ret;
3315 /*
3316 * We skip the first btrfs_tree_block_info, as we don't use the key
3317 * stored in it, but fetch it from the tree block
3318 */
3319 if (btrfs_backref_has_tree_block_info(iter)) {
3320 ret = btrfs_backref_iter_next(iter);
3321 if (ret < 0)
3322 goto out;
3323 /* No extra backref? This means the tree block is corrupted */
3324 if (ret > 0) {
3325 ret = -EUCLEAN;
3326 goto out;
3327 }
3328 }
3329 WARN_ON(cur->checked);
3330 if (!list_empty(&cur->upper)) {
3331 /*
3332 * The backref was added previously when processing backref of
3333 * type BTRFS_TREE_BLOCK_REF_KEY
3334 */
3335 ASSERT(list_is_singular(&cur->upper));
3336 edge = list_entry(cur->upper.next, struct btrfs_backref_edge,
3337 list[LOWER]);
3338 ASSERT(list_empty(&edge->list[UPPER]));
3339 exist = edge->node[UPPER];
3340 /*
3341 * Add the upper level block to pending list if we need check
3342 * its backrefs
3343 */
3344 if (!exist->checked)
3345 list_add_tail(&edge->list[UPPER], &cache->pending_edge);
3346 } else {
3347 exist = NULL;
3348 }
3349
3350 for (; ret == 0; ret = btrfs_backref_iter_next(iter)) {
3351 struct extent_buffer *eb;
3352 struct btrfs_key key;
3353 int type;
3354
3355 cond_resched();
3356 eb = btrfs_backref_get_eb(iter);
3357
3358 key.objectid = iter->bytenr;
3359 if (btrfs_backref_iter_is_inline_ref(iter)) {
3360 struct btrfs_extent_inline_ref *iref;
3361
3362 /* Update key for inline backref */
3363 iref = (struct btrfs_extent_inline_ref *)
3364 ((unsigned long)iter->cur_ptr);
3365 type = btrfs_get_extent_inline_ref_type(eb, iref,
3366 BTRFS_REF_TYPE_BLOCK);
3367 if (type == BTRFS_REF_TYPE_INVALID) {
3368 ret = -EUCLEAN;
3369 goto out;
3370 }
3371 key.type = type;
3372 key.offset = btrfs_extent_inline_ref_offset(eb, iref);
3373 } else {
3374 key.type = iter->cur_key.type;
3375 key.offset = iter->cur_key.offset;
3376 }
3377
3378 /*
3379 * Parent node found and matches current inline ref, no need to
3380 * rebuild this node for this inline ref
3381 */
3382 if (exist &&
3383 ((key.type == BTRFS_TREE_BLOCK_REF_KEY &&
3384 exist->owner == key.offset) ||
3385 (key.type == BTRFS_SHARED_BLOCK_REF_KEY &&
3386 exist->bytenr == key.offset))) {
3387 exist = NULL;
3388 continue;
3389 }
3390
3391 /* SHARED_BLOCK_REF means key.offset is the parent bytenr */
3392 if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3393 ret = handle_direct_tree_backref(cache, &key, cur);
3394 if (ret < 0)
3395 goto out;
3396 continue;
3397 } else if (unlikely(key.type == BTRFS_EXTENT_REF_V0_KEY)) {
3398 ret = -EINVAL;
3399 btrfs_print_v0_err(fs_info);
3400 btrfs_handle_fs_error(fs_info, ret, NULL);
3401 goto out;
3402 } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) {
3403 continue;
3404 }
3405
3406 /*
3407 * key.type == BTRFS_TREE_BLOCK_REF_KEY, inline ref offset
3408 * means the root objectid. We need to search the tree to get
3409 * its parent bytenr.
3410 */
3411 ret = handle_indirect_tree_backref(cache, path, &key, node_key,
3412 cur);
3413 if (ret < 0)
3414 goto out;
3415 }
3416 ret = 0;
3417 cur->checked = 1;
3418 WARN_ON(exist);
3419out:
3420 btrfs_backref_iter_release(iter);
3421 return ret;
3422}
3423
3424/*
3425 * Finish the upwards linkage created by btrfs_backref_add_tree_node()
3426 */
3427int btrfs_backref_finish_upper_links(struct btrfs_backref_cache *cache,
3428 struct btrfs_backref_node *start)
3429{
3430 struct list_head *useless_node = &cache->useless_node;
3431 struct btrfs_backref_edge *edge;
3432 struct rb_node *rb_node;
3433 LIST_HEAD(pending_edge);
3434
3435 ASSERT(start->checked);
3436
3437 /* Insert this node to cache if it's not COW-only */
3438 if (!start->cowonly) {
3439 rb_node = rb_simple_insert(&cache->rb_root, start->bytenr,
3440 &start->rb_node);
3441 if (rb_node)
3442 btrfs_backref_panic(cache->fs_info, start->bytenr,
3443 -EEXIST);
3444 list_add_tail(&start->lower, &cache->leaves);
3445 }
3446
3447 /*
3448 * Use breadth first search to iterate all related edges.
3449 *
3450 * The starting points are all the edges of this node
3451 */
3452 list_for_each_entry(edge, &start->upper, list[LOWER])
3453 list_add_tail(&edge->list[UPPER], &pending_edge);
3454
3455 while (!list_empty(&pending_edge)) {
3456 struct btrfs_backref_node *upper;
3457 struct btrfs_backref_node *lower;
3458
3459 edge = list_first_entry(&pending_edge,
3460 struct btrfs_backref_edge, list[UPPER]);
3461 list_del_init(&edge->list[UPPER]);
3462 upper = edge->node[UPPER];
3463 lower = edge->node[LOWER];
3464
3465 /* Parent is detached, no need to keep any edges */
3466 if (upper->detached) {
3467 list_del(&edge->list[LOWER]);
3468 btrfs_backref_free_edge(cache, edge);
3469
3470 /* Lower node is orphan, queue for cleanup */
3471 if (list_empty(&lower->upper))
3472 list_add(&lower->list, useless_node);
3473 continue;
3474 }
3475
3476 /*
3477 * All new nodes added in current build_backref_tree() haven't
3478 * been linked to the cache rb tree.
3479 * So if we have upper->rb_node populated, this means a cache
3480 * hit. We only need to link the edge, as @upper and all its
3481 * parents have already been linked.
3482 */
3483 if (!RB_EMPTY_NODE(&upper->rb_node)) {
3484 if (upper->lowest) {
3485 list_del_init(&upper->lower);
3486 upper->lowest = 0;
3487 }
3488
3489 list_add_tail(&edge->list[UPPER], &upper->lower);
3490 continue;
3491 }
3492
3493 /* Sanity check, we shouldn't have any unchecked nodes */
3494 if (!upper->checked) {
3495 ASSERT(0);
3496 return -EUCLEAN;
3497 }
3498
3499 /* Sanity check, COW-only node has non-COW-only parent */
3500 if (start->cowonly != upper->cowonly) {
3501 ASSERT(0);
3502 return -EUCLEAN;
3503 }
3504
3505 /* Only cache non-COW-only (subvolume trees) tree blocks */
3506 if (!upper->cowonly) {
3507 rb_node = rb_simple_insert(&cache->rb_root, upper->bytenr,
3508 &upper->rb_node);
3509 if (rb_node) {
3510 btrfs_backref_panic(cache->fs_info,
3511 upper->bytenr, -EEXIST);
3512 return -EUCLEAN;
3513 }
3514 }
3515
3516 list_add_tail(&edge->list[UPPER], &upper->lower);
3517
3518 /*
3519 * Also queue all the parent edges of this uncached node
3520 * to finish the upper linkage
3521 */
3522 list_for_each_entry(edge, &upper->upper, list[LOWER])
3523 list_add_tail(&edge->list[UPPER], &pending_edge);
3524 }
3525 return 0;
3526}
3527
3528void btrfs_backref_error_cleanup(struct btrfs_backref_cache *cache,
3529 struct btrfs_backref_node *node)
3530{
3531 struct btrfs_backref_node *lower;
3532 struct btrfs_backref_node *upper;
3533 struct btrfs_backref_edge *edge;
3534
3535 while (!list_empty(&cache->useless_node)) {
3536 lower = list_first_entry(&cache->useless_node,
3537 struct btrfs_backref_node, list);
3538 list_del_init(&lower->list);
3539 }
3540 while (!list_empty(&cache->pending_edge)) {
3541 edge = list_first_entry(&cache->pending_edge,
3542 struct btrfs_backref_edge, list[UPPER]);
3543 list_del(&edge->list[UPPER]);
3544 list_del(&edge->list[LOWER]);
3545 lower = edge->node[LOWER];
3546 upper = edge->node[UPPER];
3547 btrfs_backref_free_edge(cache, edge);
3548
3549 /*
3550 * Lower is no longer linked to any upper backref nodes and
3551 * isn't in the cache, we can free it ourselves.
3552 */
3553 if (list_empty(&lower->upper) &&
3554 RB_EMPTY_NODE(&lower->rb_node))
3555 list_add(&lower->list, &cache->useless_node);
3556
3557 if (!RB_EMPTY_NODE(&upper->rb_node))
3558 continue;
3559
3560 /* Add this guy's upper edges to the list to process */
3561 list_for_each_entry(edge, &upper->upper, list[LOWER])
3562 list_add_tail(&edge->list[UPPER],
3563 &cache->pending_edge);
3564 if (list_empty(&upper->upper))
3565 list_add(&upper->list, &cache->useless_node);
3566 }
3567
3568 while (!list_empty(&cache->useless_node)) {
3569 lower = list_first_entry(&cache->useless_node,
3570 struct btrfs_backref_node, list);
3571 list_del_init(&lower->list);
3572 if (lower == node)
3573 node = NULL;
3574 btrfs_backref_drop_node(cache, lower);
3575 }
3576
3577 btrfs_backref_cleanup_node(cache, node);
3578 ASSERT(list_empty(&cache->useless_node) &&
3579 list_empty(&cache->pending_edge));
3580}