136454dbb4afdde1330f07eae441392be18c0973
[sfrench/cifs-2.6.git] / fs / btrfs / backref.c
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
17 /* Just an arbitrary number so we can be sure this happened */
18 #define BACKREF_FOUND_SHARED 6
19
20 struct extent_inode_elem {
21         u64 inum;
22         u64 offset;
23         struct extent_inode_elem *next;
24 };
25
26 static int check_extent_in_eb(const struct btrfs_key *key,
27                               const struct extent_buffer *eb,
28                               const struct btrfs_file_extent_item *fi,
29                               u64 extent_item_pos,
30                               struct extent_inode_elem **eie,
31                               bool ignore_offset)
32 {
33         u64 offset = 0;
34         struct extent_inode_elem *e;
35
36         if (!ignore_offset &&
37             !btrfs_file_extent_compression(eb, fi) &&
38             !btrfs_file_extent_encryption(eb, fi) &&
39             !btrfs_file_extent_other_encoding(eb, fi)) {
40                 u64 data_offset;
41                 u64 data_len;
42
43                 data_offset = btrfs_file_extent_offset(eb, fi);
44                 data_len = btrfs_file_extent_num_bytes(eb, fi);
45
46                 if (extent_item_pos < data_offset ||
47                     extent_item_pos >= data_offset + data_len)
48                         return 1;
49                 offset = extent_item_pos - data_offset;
50         }
51
52         e = kmalloc(sizeof(*e), GFP_NOFS);
53         if (!e)
54                 return -ENOMEM;
55
56         e->next = *eie;
57         e->inum = key->objectid;
58         e->offset = key->offset + offset;
59         *eie = e;
60
61         return 0;
62 }
63
64 static void free_inode_elem_list(struct extent_inode_elem *eie)
65 {
66         struct extent_inode_elem *eie_next;
67
68         for (; eie; eie = eie_next) {
69                 eie_next = eie->next;
70                 kfree(eie);
71         }
72 }
73
74 static int find_extent_in_eb(const struct extent_buffer *eb,
75                              u64 wanted_disk_byte, u64 extent_item_pos,
76                              struct extent_inode_elem **eie,
77                              bool ignore_offset)
78 {
79         u64 disk_byte;
80         struct btrfs_key key;
81         struct btrfs_file_extent_item *fi;
82         int slot;
83         int nritems;
84         int extent_type;
85         int ret;
86
87         /*
88          * from the shared data ref, we only have the leaf but we need
89          * the key. thus, we must look into all items and see that we
90          * find one (some) with a reference to our extent item.
91          */
92         nritems = btrfs_header_nritems(eb);
93         for (slot = 0; slot < nritems; ++slot) {
94                 btrfs_item_key_to_cpu(eb, &key, slot);
95                 if (key.type != BTRFS_EXTENT_DATA_KEY)
96                         continue;
97                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
98                 extent_type = btrfs_file_extent_type(eb, fi);
99                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
100                         continue;
101                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
102                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
103                 if (disk_byte != wanted_disk_byte)
104                         continue;
105
106                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie, ignore_offset);
107                 if (ret < 0)
108                         return ret;
109         }
110
111         return 0;
112 }
113
114 struct preftree {
115         struct rb_root_cached root;
116         unsigned int count;
117 };
118
119 #define PREFTREE_INIT   { .root = RB_ROOT_CACHED, .count = 0 }
120
121 struct preftrees {
122         struct preftree direct;    /* BTRFS_SHARED_[DATA|BLOCK]_REF_KEY */
123         struct preftree indirect;  /* BTRFS_[TREE_BLOCK|EXTENT_DATA]_REF_KEY */
124         struct preftree indirect_missing_keys;
125 };
126
127 /*
128  * Checks for a shared extent during backref search.
129  *
130  * The share_count tracks prelim_refs (direct and indirect) having a
131  * ref->count >0:
132  *  - incremented when a ref->count transitions to >0
133  *  - decremented when a ref->count transitions to <1
134  */
135 struct share_check {
136         u64 root_objectid;
137         u64 inum;
138         int share_count;
139 };
140
141 static inline int extent_is_shared(struct share_check *sc)
142 {
143         return (sc && sc->share_count > 1) ? BACKREF_FOUND_SHARED : 0;
144 }
145
146 static struct kmem_cache *btrfs_prelim_ref_cache;
147
148 int __init btrfs_prelim_ref_init(void)
149 {
150         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
151                                         sizeof(struct prelim_ref),
152                                         0,
153                                         SLAB_MEM_SPREAD,
154                                         NULL);
155         if (!btrfs_prelim_ref_cache)
156                 return -ENOMEM;
157         return 0;
158 }
159
160 void __cold btrfs_prelim_ref_exit(void)
161 {
162         kmem_cache_destroy(btrfs_prelim_ref_cache);
163 }
164
165 static void free_pref(struct prelim_ref *ref)
166 {
167         kmem_cache_free(btrfs_prelim_ref_cache, ref);
168 }
169
170 /*
171  * Return 0 when both refs are for the same block (and can be merged).
172  * A -1 return indicates ref1 is a 'lower' block than ref2, while 1
173  * indicates a 'higher' block.
174  */
175 static int prelim_ref_compare(struct prelim_ref *ref1,
176                               struct prelim_ref *ref2)
177 {
178         if (ref1->level < ref2->level)
179                 return -1;
180         if (ref1->level > ref2->level)
181                 return 1;
182         if (ref1->root_id < ref2->root_id)
183                 return -1;
184         if (ref1->root_id > ref2->root_id)
185                 return 1;
186         if (ref1->key_for_search.type < ref2->key_for_search.type)
187                 return -1;
188         if (ref1->key_for_search.type > ref2->key_for_search.type)
189                 return 1;
190         if (ref1->key_for_search.objectid < ref2->key_for_search.objectid)
191                 return -1;
192         if (ref1->key_for_search.objectid > ref2->key_for_search.objectid)
193                 return 1;
194         if (ref1->key_for_search.offset < ref2->key_for_search.offset)
195                 return -1;
196         if (ref1->key_for_search.offset > ref2->key_for_search.offset)
197                 return 1;
198         if (ref1->parent < ref2->parent)
199                 return -1;
200         if (ref1->parent > ref2->parent)
201                 return 1;
202
203         return 0;
204 }
205
206 static void update_share_count(struct share_check *sc, int oldcount,
207                                int newcount)
208 {
209         if ((!sc) || (oldcount == 0 && newcount < 1))
210                 return;
211
212         if (oldcount > 0 && newcount < 1)
213                 sc->share_count--;
214         else if (oldcount < 1 && newcount > 0)
215                 sc->share_count++;
216 }
217
218 /*
219  * Add @newref to the @root rbtree, merging identical refs.
220  *
221  * Callers should assume that newref has been freed after calling.
222  */
223 static void prelim_ref_insert(const struct btrfs_fs_info *fs_info,
224                               struct preftree *preftree,
225                               struct prelim_ref *newref,
226                               struct share_check *sc)
227 {
228         struct rb_root_cached *root;
229         struct rb_node **p;
230         struct rb_node *parent = NULL;
231         struct prelim_ref *ref;
232         int result;
233         bool leftmost = true;
234
235         root = &preftree->root;
236         p = &root->rb_root.rb_node;
237
238         while (*p) {
239                 parent = *p;
240                 ref = rb_entry(parent, struct prelim_ref, rbnode);
241                 result = prelim_ref_compare(ref, newref);
242                 if (result < 0) {
243                         p = &(*p)->rb_left;
244                 } else if (result > 0) {
245                         p = &(*p)->rb_right;
246                         leftmost = false;
247                 } else {
248                         /* Identical refs, merge them and free @newref */
249                         struct extent_inode_elem *eie = ref->inode_list;
250
251                         while (eie && eie->next)
252                                 eie = eie->next;
253
254                         if (!eie)
255                                 ref->inode_list = newref->inode_list;
256                         else
257                                 eie->next = newref->inode_list;
258                         trace_btrfs_prelim_ref_merge(fs_info, ref, newref,
259                                                      preftree->count);
260                         /*
261                          * A delayed ref can have newref->count < 0.
262                          * The ref->count is updated to follow any
263                          * BTRFS_[ADD|DROP]_DELAYED_REF actions.
264                          */
265                         update_share_count(sc, ref->count,
266                                            ref->count + newref->count);
267                         ref->count += newref->count;
268                         free_pref(newref);
269                         return;
270                 }
271         }
272
273         update_share_count(sc, 0, newref->count);
274         preftree->count++;
275         trace_btrfs_prelim_ref_insert(fs_info, newref, NULL, preftree->count);
276         rb_link_node(&newref->rbnode, parent, p);
277         rb_insert_color_cached(&newref->rbnode, root, leftmost);
278 }
279
280 /*
281  * Release the entire tree.  We don't care about internal consistency so
282  * just free everything and then reset the tree root.
283  */
284 static void prelim_release(struct preftree *preftree)
285 {
286         struct prelim_ref *ref, *next_ref;
287
288         rbtree_postorder_for_each_entry_safe(ref, next_ref,
289                                              &preftree->root.rb_root, rbnode)
290                 free_pref(ref);
291
292         preftree->root = RB_ROOT_CACHED;
293         preftree->count = 0;
294 }
295
296 /*
297  * the rules for all callers of this function are:
298  * - obtaining the parent is the goal
299  * - if you add a key, you must know that it is a correct key
300  * - if you cannot add the parent or a correct key, then we will look into the
301  *   block later to set a correct key
302  *
303  * delayed refs
304  * ============
305  *        backref type | shared | indirect | shared | indirect
306  * information         |   tree |     tree |   data |     data
307  * --------------------+--------+----------+--------+----------
308  *      parent logical |    y   |     -    |    -   |     -
309  *      key to resolve |    -   |     y    |    y   |     y
310  *  tree block logical |    -   |     -    |    -   |     -
311  *  root for resolving |    y   |     y    |    y   |     y
312  *
313  * - column 1:       we've the parent -> done
314  * - column 2, 3, 4: we use the key to find the parent
315  *
316  * on disk refs (inline or keyed)
317  * ==============================
318  *        backref type | shared | indirect | shared | indirect
319  * information         |   tree |     tree |   data |     data
320  * --------------------+--------+----------+--------+----------
321  *      parent logical |    y   |     -    |    y   |     -
322  *      key to resolve |    -   |     -    |    -   |     y
323  *  tree block logical |    y   |     y    |    y   |     y
324  *  root for resolving |    -   |     y    |    y   |     y
325  *
326  * - column 1, 3: we've the parent -> done
327  * - column 2:    we take the first key from the block to find the parent
328  *                (see add_missing_keys)
329  * - column 4:    we use the key to find the parent
330  *
331  * additional information that's available but not required to find the parent
332  * block might help in merging entries to gain some speed.
333  */
334 static int add_prelim_ref(const struct btrfs_fs_info *fs_info,
335                           struct preftree *preftree, u64 root_id,
336                           const struct btrfs_key *key, int level, u64 parent,
337                           u64 wanted_disk_byte, int count,
338                           struct share_check *sc, gfp_t gfp_mask)
339 {
340         struct prelim_ref *ref;
341
342         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
343                 return 0;
344
345         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
346         if (!ref)
347                 return -ENOMEM;
348
349         ref->root_id = root_id;
350         if (key) {
351                 ref->key_for_search = *key;
352                 /*
353                  * We can often find data backrefs with an offset that is too
354                  * large (>= LLONG_MAX, maximum allowed file offset) due to
355                  * underflows when subtracting a file's offset with the data
356                  * offset of its corresponding extent data item. This can
357                  * happen for example in the clone ioctl.
358                  * So if we detect such case we set the search key's offset to
359                  * zero to make sure we will find the matching file extent item
360                  * at add_all_parents(), otherwise we will miss it because the
361                  * offset taken form the backref is much larger then the offset
362                  * of the file extent item. This can make us scan a very large
363                  * number of file extent items, but at least it will not make
364                  * us miss any.
365                  * This is an ugly workaround for a behaviour that should have
366                  * never existed, but it does and a fix for the clone ioctl
367                  * would touch a lot of places, cause backwards incompatibility
368                  * and would not fix the problem for extents cloned with older
369                  * kernels.
370                  */
371                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
372                     ref->key_for_search.offset >= LLONG_MAX)
373                         ref->key_for_search.offset = 0;
374         } else {
375                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
376         }
377
378         ref->inode_list = NULL;
379         ref->level = level;
380         ref->count = count;
381         ref->parent = parent;
382         ref->wanted_disk_byte = wanted_disk_byte;
383         prelim_ref_insert(fs_info, preftree, ref, sc);
384         return extent_is_shared(sc);
385 }
386
387 /* direct refs use root == 0, key == NULL */
388 static int add_direct_ref(const struct btrfs_fs_info *fs_info,
389                           struct preftrees *preftrees, int level, u64 parent,
390                           u64 wanted_disk_byte, int count,
391                           struct share_check *sc, gfp_t gfp_mask)
392 {
393         return add_prelim_ref(fs_info, &preftrees->direct, 0, NULL, level,
394                               parent, wanted_disk_byte, count, sc, gfp_mask);
395 }
396
397 /* indirect refs use parent == 0 */
398 static int add_indirect_ref(const struct btrfs_fs_info *fs_info,
399                             struct preftrees *preftrees, u64 root_id,
400                             const struct btrfs_key *key, int level,
401                             u64 wanted_disk_byte, int count,
402                             struct share_check *sc, gfp_t gfp_mask)
403 {
404         struct preftree *tree = &preftrees->indirect;
405
406         if (!key)
407                 tree = &preftrees->indirect_missing_keys;
408         return add_prelim_ref(fs_info, tree, root_id, key, level, 0,
409                               wanted_disk_byte, count, sc, gfp_mask);
410 }
411
412 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
413                            struct ulist *parents, struct prelim_ref *ref,
414                            int level, u64 time_seq, const u64 *extent_item_pos,
415                            u64 total_refs, bool ignore_offset)
416 {
417         int ret = 0;
418         int slot;
419         struct extent_buffer *eb;
420         struct btrfs_key key;
421         struct btrfs_key *key_for_search = &ref->key_for_search;
422         struct btrfs_file_extent_item *fi;
423         struct extent_inode_elem *eie = NULL, *old = NULL;
424         u64 disk_byte;
425         u64 wanted_disk_byte = ref->wanted_disk_byte;
426         u64 count = 0;
427
428         if (level != 0) {
429                 eb = path->nodes[level];
430                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
431                 if (ret < 0)
432                         return ret;
433                 return 0;
434         }
435
436         /*
437          * We normally enter this function with the path already pointing to
438          * the first item to check. But sometimes, we may enter it with
439          * slot==nritems. In that case, go to the next leaf before we continue.
440          */
441         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
442                 if (time_seq == SEQ_LAST)
443                         ret = btrfs_next_leaf(root, path);
444                 else
445                         ret = btrfs_next_old_leaf(root, path, time_seq);
446         }
447
448         while (!ret && count < total_refs) {
449                 eb = path->nodes[0];
450                 slot = path->slots[0];
451
452                 btrfs_item_key_to_cpu(eb, &key, slot);
453
454                 if (key.objectid != key_for_search->objectid ||
455                     key.type != BTRFS_EXTENT_DATA_KEY)
456                         break;
457
458                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
459                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
460
461                 if (disk_byte == wanted_disk_byte) {
462                         eie = NULL;
463                         old = NULL;
464                         count++;
465                         if (extent_item_pos) {
466                                 ret = check_extent_in_eb(&key, eb, fi,
467                                                 *extent_item_pos,
468                                                 &eie, ignore_offset);
469                                 if (ret < 0)
470                                         break;
471                         }
472                         if (ret > 0)
473                                 goto next;
474                         ret = ulist_add_merge_ptr(parents, eb->start,
475                                                   eie, (void **)&old, GFP_NOFS);
476                         if (ret < 0)
477                                 break;
478                         if (!ret && extent_item_pos) {
479                                 while (old->next)
480                                         old = old->next;
481                                 old->next = eie;
482                         }
483                         eie = NULL;
484                 }
485 next:
486                 if (time_seq == SEQ_LAST)
487                         ret = btrfs_next_item(root, path);
488                 else
489                         ret = btrfs_next_old_item(root, path, time_seq);
490         }
491
492         if (ret > 0)
493                 ret = 0;
494         else if (ret < 0)
495                 free_inode_elem_list(eie);
496         return ret;
497 }
498
499 /*
500  * resolve an indirect backref in the form (root_id, key, level)
501  * to a logical address
502  */
503 static int resolve_indirect_ref(struct btrfs_fs_info *fs_info,
504                                 struct btrfs_path *path, u64 time_seq,
505                                 struct prelim_ref *ref, struct ulist *parents,
506                                 const u64 *extent_item_pos, u64 total_refs,
507                                 bool ignore_offset)
508 {
509         struct btrfs_root *root;
510         struct btrfs_key root_key;
511         struct extent_buffer *eb;
512         int ret = 0;
513         int root_level;
514         int level = ref->level;
515         int index;
516
517         root_key.objectid = ref->root_id;
518         root_key.type = BTRFS_ROOT_ITEM_KEY;
519         root_key.offset = (u64)-1;
520
521         index = srcu_read_lock(&fs_info->subvol_srcu);
522
523         root = btrfs_get_fs_root(fs_info, &root_key, false);
524         if (IS_ERR(root)) {
525                 srcu_read_unlock(&fs_info->subvol_srcu, index);
526                 ret = PTR_ERR(root);
527                 goto out;
528         }
529
530         if (btrfs_is_testing(fs_info)) {
531                 srcu_read_unlock(&fs_info->subvol_srcu, index);
532                 ret = -ENOENT;
533                 goto out;
534         }
535
536         if (path->search_commit_root)
537                 root_level = btrfs_header_level(root->commit_root);
538         else if (time_seq == SEQ_LAST)
539                 root_level = btrfs_header_level(root->node);
540         else
541                 root_level = btrfs_old_root_level(root, time_seq);
542
543         if (root_level + 1 == level) {
544                 srcu_read_unlock(&fs_info->subvol_srcu, index);
545                 goto out;
546         }
547
548         path->lowest_level = level;
549         if (time_seq == SEQ_LAST)
550                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
551                                         0, 0);
552         else
553                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
554                                             time_seq);
555
556         /* root node has been locked, we can release @subvol_srcu safely here */
557         srcu_read_unlock(&fs_info->subvol_srcu, index);
558
559         btrfs_debug(fs_info,
560                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
561                  ref->root_id, level, ref->count, ret,
562                  ref->key_for_search.objectid, ref->key_for_search.type,
563                  ref->key_for_search.offset);
564         if (ret < 0)
565                 goto out;
566
567         eb = path->nodes[level];
568         while (!eb) {
569                 if (WARN_ON(!level)) {
570                         ret = 1;
571                         goto out;
572                 }
573                 level--;
574                 eb = path->nodes[level];
575         }
576
577         ret = add_all_parents(root, path, parents, ref, level, time_seq,
578                               extent_item_pos, total_refs, ignore_offset);
579 out:
580         path->lowest_level = 0;
581         btrfs_release_path(path);
582         return ret;
583 }
584
585 static struct extent_inode_elem *
586 unode_aux_to_inode_list(struct ulist_node *node)
587 {
588         if (!node)
589                 return NULL;
590         return (struct extent_inode_elem *)(uintptr_t)node->aux;
591 }
592
593 /*
594  * We maintain three separate rbtrees: one for direct refs, one for
595  * indirect refs which have a key, and one for indirect refs which do not
596  * have a key. Each tree does merge on insertion.
597  *
598  * Once all of the references are located, we iterate over the tree of
599  * indirect refs with missing keys. An appropriate key is located and
600  * the ref is moved onto the tree for indirect refs. After all missing
601  * keys are thus located, we iterate over the indirect ref tree, resolve
602  * each reference, and then insert the resolved reference onto the
603  * direct tree (merging there too).
604  *
605  * New backrefs (i.e., for parent nodes) are added to the appropriate
606  * rbtree as they are encountered. The new backrefs are subsequently
607  * resolved as above.
608  */
609 static int resolve_indirect_refs(struct btrfs_fs_info *fs_info,
610                                  struct btrfs_path *path, u64 time_seq,
611                                  struct preftrees *preftrees,
612                                  const u64 *extent_item_pos, u64 total_refs,
613                                  struct share_check *sc, bool ignore_offset)
614 {
615         int err;
616         int ret = 0;
617         struct ulist *parents;
618         struct ulist_node *node;
619         struct ulist_iterator uiter;
620         struct rb_node *rnode;
621
622         parents = ulist_alloc(GFP_NOFS);
623         if (!parents)
624                 return -ENOMEM;
625
626         /*
627          * We could trade memory usage for performance here by iterating
628          * the tree, allocating new refs for each insertion, and then
629          * freeing the entire indirect tree when we're done.  In some test
630          * cases, the tree can grow quite large (~200k objects).
631          */
632         while ((rnode = rb_first_cached(&preftrees->indirect.root))) {
633                 struct prelim_ref *ref;
634
635                 ref = rb_entry(rnode, struct prelim_ref, rbnode);
636                 if (WARN(ref->parent,
637                          "BUG: direct ref found in indirect tree")) {
638                         ret = -EINVAL;
639                         goto out;
640                 }
641
642                 rb_erase_cached(&ref->rbnode, &preftrees->indirect.root);
643                 preftrees->indirect.count--;
644
645                 if (ref->count == 0) {
646                         free_pref(ref);
647                         continue;
648                 }
649
650                 if (sc && sc->root_objectid &&
651                     ref->root_id != sc->root_objectid) {
652                         free_pref(ref);
653                         ret = BACKREF_FOUND_SHARED;
654                         goto out;
655                 }
656                 err = resolve_indirect_ref(fs_info, path, time_seq, ref,
657                                            parents, extent_item_pos,
658                                            total_refs, ignore_offset);
659                 /*
660                  * we can only tolerate ENOENT,otherwise,we should catch error
661                  * and return directly.
662                  */
663                 if (err == -ENOENT) {
664                         prelim_ref_insert(fs_info, &preftrees->direct, ref,
665                                           NULL);
666                         continue;
667                 } else if (err) {
668                         free_pref(ref);
669                         ret = err;
670                         goto out;
671                 }
672
673                 /* we put the first parent into the ref at hand */
674                 ULIST_ITER_INIT(&uiter);
675                 node = ulist_next(parents, &uiter);
676                 ref->parent = node ? node->val : 0;
677                 ref->inode_list = unode_aux_to_inode_list(node);
678
679                 /* Add a prelim_ref(s) for any other parent(s). */
680                 while ((node = ulist_next(parents, &uiter))) {
681                         struct prelim_ref *new_ref;
682
683                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
684                                                    GFP_NOFS);
685                         if (!new_ref) {
686                                 free_pref(ref);
687                                 ret = -ENOMEM;
688                                 goto out;
689                         }
690                         memcpy(new_ref, ref, sizeof(*ref));
691                         new_ref->parent = node->val;
692                         new_ref->inode_list = unode_aux_to_inode_list(node);
693                         prelim_ref_insert(fs_info, &preftrees->direct,
694                                           new_ref, NULL);
695                 }
696
697                 /*
698                  * Now it's a direct ref, put it in the direct tree. We must
699                  * do this last because the ref could be merged/freed here.
700                  */
701                 prelim_ref_insert(fs_info, &preftrees->direct, ref, NULL);
702
703                 ulist_reinit(parents);
704                 cond_resched();
705         }
706 out:
707         ulist_free(parents);
708         return ret;
709 }
710
711 /*
712  * read tree blocks and add keys where required.
713  */
714 static int add_missing_keys(struct btrfs_fs_info *fs_info,
715                             struct preftrees *preftrees)
716 {
717         struct prelim_ref *ref;
718         struct extent_buffer *eb;
719         struct preftree *tree = &preftrees->indirect_missing_keys;
720         struct rb_node *node;
721
722         while ((node = rb_first_cached(&tree->root))) {
723                 ref = rb_entry(node, struct prelim_ref, rbnode);
724                 rb_erase_cached(node, &tree->root);
725
726                 BUG_ON(ref->parent);    /* should not be a direct ref */
727                 BUG_ON(ref->key_for_search.type);
728                 BUG_ON(!ref->wanted_disk_byte);
729
730                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0,
731                                      ref->level - 1, NULL);
732                 if (IS_ERR(eb)) {
733                         free_pref(ref);
734                         return PTR_ERR(eb);
735                 } else if (!extent_buffer_uptodate(eb)) {
736                         free_pref(ref);
737                         free_extent_buffer(eb);
738                         return -EIO;
739                 }
740                 btrfs_tree_read_lock(eb);
741                 if (btrfs_header_level(eb) == 0)
742                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
743                 else
744                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
745                 btrfs_tree_read_unlock(eb);
746                 free_extent_buffer(eb);
747                 prelim_ref_insert(fs_info, &preftrees->indirect, ref, NULL);
748                 cond_resched();
749         }
750         return 0;
751 }
752
753 /*
754  * add all currently queued delayed refs from this head whose seq nr is
755  * smaller or equal that seq to the list
756  */
757 static int add_delayed_refs(const struct btrfs_fs_info *fs_info,
758                             struct btrfs_delayed_ref_head *head, u64 seq,
759                             struct preftrees *preftrees, u64 *total_refs,
760                             struct share_check *sc)
761 {
762         struct btrfs_delayed_ref_node *node;
763         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
764         struct btrfs_key key;
765         struct btrfs_key tmp_op_key;
766         struct rb_node *n;
767         int count;
768         int ret = 0;
769
770         if (extent_op && extent_op->update_key)
771                 btrfs_disk_key_to_cpu(&tmp_op_key, &extent_op->key);
772
773         spin_lock(&head->lock);
774         for (n = rb_first_cached(&head->ref_tree); n; n = rb_next(n)) {
775                 node = rb_entry(n, struct btrfs_delayed_ref_node,
776                                 ref_node);
777                 if (node->seq > seq)
778                         continue;
779
780                 switch (node->action) {
781                 case BTRFS_ADD_DELAYED_EXTENT:
782                 case BTRFS_UPDATE_DELAYED_HEAD:
783                         WARN_ON(1);
784                         continue;
785                 case BTRFS_ADD_DELAYED_REF:
786                         count = node->ref_mod;
787                         break;
788                 case BTRFS_DROP_DELAYED_REF:
789                         count = node->ref_mod * -1;
790                         break;
791                 default:
792                         BUG_ON(1);
793                 }
794                 *total_refs += count;
795                 switch (node->type) {
796                 case BTRFS_TREE_BLOCK_REF_KEY: {
797                         /* NORMAL INDIRECT METADATA backref */
798                         struct btrfs_delayed_tree_ref *ref;
799
800                         ref = btrfs_delayed_node_to_tree_ref(node);
801                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
802                                                &tmp_op_key, ref->level + 1,
803                                                node->bytenr, count, sc,
804                                                GFP_ATOMIC);
805                         break;
806                 }
807                 case BTRFS_SHARED_BLOCK_REF_KEY: {
808                         /* SHARED DIRECT METADATA backref */
809                         struct btrfs_delayed_tree_ref *ref;
810
811                         ref = btrfs_delayed_node_to_tree_ref(node);
812
813                         ret = add_direct_ref(fs_info, preftrees, ref->level + 1,
814                                              ref->parent, node->bytenr, count,
815                                              sc, GFP_ATOMIC);
816                         break;
817                 }
818                 case BTRFS_EXTENT_DATA_REF_KEY: {
819                         /* NORMAL INDIRECT DATA backref */
820                         struct btrfs_delayed_data_ref *ref;
821                         ref = btrfs_delayed_node_to_data_ref(node);
822
823                         key.objectid = ref->objectid;
824                         key.type = BTRFS_EXTENT_DATA_KEY;
825                         key.offset = ref->offset;
826
827                         /*
828                          * Found a inum that doesn't match our known inum, we
829                          * know it's shared.
830                          */
831                         if (sc && sc->inum && ref->objectid != sc->inum) {
832                                 ret = BACKREF_FOUND_SHARED;
833                                 goto out;
834                         }
835
836                         ret = add_indirect_ref(fs_info, preftrees, ref->root,
837                                                &key, 0, node->bytenr, count, sc,
838                                                GFP_ATOMIC);
839                         break;
840                 }
841                 case BTRFS_SHARED_DATA_REF_KEY: {
842                         /* SHARED DIRECT FULL backref */
843                         struct btrfs_delayed_data_ref *ref;
844
845                         ref = btrfs_delayed_node_to_data_ref(node);
846
847                         ret = add_direct_ref(fs_info, preftrees, 0, ref->parent,
848                                              node->bytenr, count, sc,
849                                              GFP_ATOMIC);
850                         break;
851                 }
852                 default:
853                         WARN_ON(1);
854                 }
855                 /*
856                  * We must ignore BACKREF_FOUND_SHARED until all delayed
857                  * refs have been checked.
858                  */
859                 if (ret && (ret != BACKREF_FOUND_SHARED))
860                         break;
861         }
862         if (!ret)
863                 ret = extent_is_shared(sc);
864 out:
865         spin_unlock(&head->lock);
866         return ret;
867 }
868
869 /*
870  * add all inline backrefs for bytenr to the list
871  *
872  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
873  */
874 static int add_inline_refs(const struct btrfs_fs_info *fs_info,
875                            struct btrfs_path *path, u64 bytenr,
876                            int *info_level, struct preftrees *preftrees,
877                            u64 *total_refs, struct share_check *sc)
878 {
879         int ret = 0;
880         int slot;
881         struct extent_buffer *leaf;
882         struct btrfs_key key;
883         struct btrfs_key found_key;
884         unsigned long ptr;
885         unsigned long end;
886         struct btrfs_extent_item *ei;
887         u64 flags;
888         u64 item_size;
889
890         /*
891          * enumerate all inline refs
892          */
893         leaf = path->nodes[0];
894         slot = path->slots[0];
895
896         item_size = btrfs_item_size_nr(leaf, slot);
897         BUG_ON(item_size < sizeof(*ei));
898
899         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
900         flags = btrfs_extent_flags(leaf, ei);
901         *total_refs += btrfs_extent_refs(leaf, ei);
902         btrfs_item_key_to_cpu(leaf, &found_key, slot);
903
904         ptr = (unsigned long)(ei + 1);
905         end = (unsigned long)ei + item_size;
906
907         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
908             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
909                 struct btrfs_tree_block_info *info;
910
911                 info = (struct btrfs_tree_block_info *)ptr;
912                 *info_level = btrfs_tree_block_level(leaf, info);
913                 ptr += sizeof(struct btrfs_tree_block_info);
914                 BUG_ON(ptr > end);
915         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
916                 *info_level = found_key.offset;
917         } else {
918                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
919         }
920
921         while (ptr < end) {
922                 struct btrfs_extent_inline_ref *iref;
923                 u64 offset;
924                 int type;
925
926                 iref = (struct btrfs_extent_inline_ref *)ptr;
927                 type = btrfs_get_extent_inline_ref_type(leaf, iref,
928                                                         BTRFS_REF_TYPE_ANY);
929                 if (type == BTRFS_REF_TYPE_INVALID)
930                         return -EUCLEAN;
931
932                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
933
934                 switch (type) {
935                 case BTRFS_SHARED_BLOCK_REF_KEY:
936                         ret = add_direct_ref(fs_info, preftrees,
937                                              *info_level + 1, offset,
938                                              bytenr, 1, NULL, GFP_NOFS);
939                         break;
940                 case BTRFS_SHARED_DATA_REF_KEY: {
941                         struct btrfs_shared_data_ref *sdref;
942                         int count;
943
944                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
945                         count = btrfs_shared_data_ref_count(leaf, sdref);
946
947                         ret = add_direct_ref(fs_info, preftrees, 0, offset,
948                                              bytenr, count, sc, GFP_NOFS);
949                         break;
950                 }
951                 case BTRFS_TREE_BLOCK_REF_KEY:
952                         ret = add_indirect_ref(fs_info, preftrees, offset,
953                                                NULL, *info_level + 1,
954                                                bytenr, 1, NULL, GFP_NOFS);
955                         break;
956                 case BTRFS_EXTENT_DATA_REF_KEY: {
957                         struct btrfs_extent_data_ref *dref;
958                         int count;
959                         u64 root;
960
961                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
962                         count = btrfs_extent_data_ref_count(leaf, dref);
963                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
964                                                                       dref);
965                         key.type = BTRFS_EXTENT_DATA_KEY;
966                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
967
968                         if (sc && sc->inum && key.objectid != sc->inum) {
969                                 ret = BACKREF_FOUND_SHARED;
970                                 break;
971                         }
972
973                         root = btrfs_extent_data_ref_root(leaf, dref);
974
975                         ret = add_indirect_ref(fs_info, preftrees, root,
976                                                &key, 0, bytenr, count,
977                                                sc, GFP_NOFS);
978                         break;
979                 }
980                 default:
981                         WARN_ON(1);
982                 }
983                 if (ret)
984                         return ret;
985                 ptr += btrfs_extent_inline_ref_size(type);
986         }
987
988         return 0;
989 }
990
991 /*
992  * add all non-inline backrefs for bytenr to the list
993  *
994  * Returns 0 on success, <0 on error, or BACKREF_FOUND_SHARED.
995  */
996 static int add_keyed_refs(struct btrfs_fs_info *fs_info,
997                           struct btrfs_path *path, u64 bytenr,
998                           int info_level, struct preftrees *preftrees,
999                           struct share_check *sc)
1000 {
1001         struct btrfs_root *extent_root = fs_info->extent_root;
1002         int ret;
1003         int slot;
1004         struct extent_buffer *leaf;
1005         struct btrfs_key key;
1006
1007         while (1) {
1008                 ret = btrfs_next_item(extent_root, path);
1009                 if (ret < 0)
1010                         break;
1011                 if (ret) {
1012                         ret = 0;
1013                         break;
1014                 }
1015
1016                 slot = path->slots[0];
1017                 leaf = path->nodes[0];
1018                 btrfs_item_key_to_cpu(leaf, &key, slot);
1019
1020                 if (key.objectid != bytenr)
1021                         break;
1022                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1023                         continue;
1024                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1025                         break;
1026
1027                 switch (key.type) {
1028                 case BTRFS_SHARED_BLOCK_REF_KEY:
1029                         /* SHARED DIRECT METADATA backref */
1030                         ret = add_direct_ref(fs_info, preftrees,
1031                                              info_level + 1, key.offset,
1032                                              bytenr, 1, NULL, GFP_NOFS);
1033                         break;
1034                 case BTRFS_SHARED_DATA_REF_KEY: {
1035                         /* SHARED DIRECT FULL backref */
1036                         struct btrfs_shared_data_ref *sdref;
1037                         int count;
1038
1039                         sdref = btrfs_item_ptr(leaf, slot,
1040                                               struct btrfs_shared_data_ref);
1041                         count = btrfs_shared_data_ref_count(leaf, sdref);
1042                         ret = add_direct_ref(fs_info, preftrees, 0,
1043                                              key.offset, bytenr, count,
1044                                              sc, GFP_NOFS);
1045                         break;
1046                 }
1047                 case BTRFS_TREE_BLOCK_REF_KEY:
1048                         /* NORMAL INDIRECT METADATA backref */
1049                         ret = add_indirect_ref(fs_info, preftrees, key.offset,
1050                                                NULL, info_level + 1, bytenr,
1051                                                1, NULL, GFP_NOFS);
1052                         break;
1053                 case BTRFS_EXTENT_DATA_REF_KEY: {
1054                         /* NORMAL INDIRECT DATA backref */
1055                         struct btrfs_extent_data_ref *dref;
1056                         int count;
1057                         u64 root;
1058
1059                         dref = btrfs_item_ptr(leaf, slot,
1060                                               struct btrfs_extent_data_ref);
1061                         count = btrfs_extent_data_ref_count(leaf, dref);
1062                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1063                                                                       dref);
1064                         key.type = BTRFS_EXTENT_DATA_KEY;
1065                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1066
1067                         if (sc && sc->inum && key.objectid != sc->inum) {
1068                                 ret = BACKREF_FOUND_SHARED;
1069                                 break;
1070                         }
1071
1072                         root = btrfs_extent_data_ref_root(leaf, dref);
1073                         ret = add_indirect_ref(fs_info, preftrees, root,
1074                                                &key, 0, bytenr, count,
1075                                                sc, GFP_NOFS);
1076                         break;
1077                 }
1078                 default:
1079                         WARN_ON(1);
1080                 }
1081                 if (ret)
1082                         return ret;
1083
1084         }
1085
1086         return ret;
1087 }
1088
1089 /*
1090  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1091  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1092  * indirect refs to their parent bytenr.
1093  * When roots are found, they're added to the roots list
1094  *
1095  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1096  * much like trans == NULL case, the difference only lies in it will not
1097  * commit root.
1098  * The special case is for qgroup to search roots in commit_transaction().
1099  *
1100  * @sc - if !NULL, then immediately return BACKREF_FOUND_SHARED when a
1101  * shared extent is detected.
1102  *
1103  * Otherwise this returns 0 for success and <0 for an error.
1104  *
1105  * If ignore_offset is set to false, only extent refs whose offsets match
1106  * extent_item_pos are returned.  If true, every extent ref is returned
1107  * and extent_item_pos is ignored.
1108  *
1109  * FIXME some caching might speed things up
1110  */
1111 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1112                              struct btrfs_fs_info *fs_info, u64 bytenr,
1113                              u64 time_seq, struct ulist *refs,
1114                              struct ulist *roots, const u64 *extent_item_pos,
1115                              struct share_check *sc, bool ignore_offset)
1116 {
1117         struct btrfs_key key;
1118         struct btrfs_path *path;
1119         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1120         struct btrfs_delayed_ref_head *head;
1121         int info_level = 0;
1122         int ret;
1123         struct prelim_ref *ref;
1124         struct rb_node *node;
1125         struct extent_inode_elem *eie = NULL;
1126         /* total of both direct AND indirect refs! */
1127         u64 total_refs = 0;
1128         struct preftrees preftrees = {
1129                 .direct = PREFTREE_INIT,
1130                 .indirect = PREFTREE_INIT,
1131                 .indirect_missing_keys = PREFTREE_INIT
1132         };
1133
1134         key.objectid = bytenr;
1135         key.offset = (u64)-1;
1136         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1137                 key.type = BTRFS_METADATA_ITEM_KEY;
1138         else
1139                 key.type = BTRFS_EXTENT_ITEM_KEY;
1140
1141         path = btrfs_alloc_path();
1142         if (!path)
1143                 return -ENOMEM;
1144         if (!trans) {
1145                 path->search_commit_root = 1;
1146                 path->skip_locking = 1;
1147         }
1148
1149         if (time_seq == SEQ_LAST)
1150                 path->skip_locking = 1;
1151
1152         /*
1153          * grab both a lock on the path and a lock on the delayed ref head.
1154          * We need both to get a consistent picture of how the refs look
1155          * at a specified point in time
1156          */
1157 again:
1158         head = NULL;
1159
1160         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1161         if (ret < 0)
1162                 goto out;
1163         BUG_ON(ret == 0);
1164
1165 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1166         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1167             time_seq != SEQ_LAST) {
1168 #else
1169         if (trans && time_seq != SEQ_LAST) {
1170 #endif
1171                 /*
1172                  * look if there are updates for this ref queued and lock the
1173                  * head
1174                  */
1175                 delayed_refs = &trans->transaction->delayed_refs;
1176                 spin_lock(&delayed_refs->lock);
1177                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1178                 if (head) {
1179                         if (!mutex_trylock(&head->mutex)) {
1180                                 refcount_inc(&head->refs);
1181                                 spin_unlock(&delayed_refs->lock);
1182
1183                                 btrfs_release_path(path);
1184
1185                                 /*
1186                                  * Mutex was contended, block until it's
1187                                  * released and try again
1188                                  */
1189                                 mutex_lock(&head->mutex);
1190                                 mutex_unlock(&head->mutex);
1191                                 btrfs_put_delayed_ref_head(head);
1192                                 goto again;
1193                         }
1194                         spin_unlock(&delayed_refs->lock);
1195                         ret = add_delayed_refs(fs_info, head, time_seq,
1196                                                &preftrees, &total_refs, sc);
1197                         mutex_unlock(&head->mutex);
1198                         if (ret)
1199                                 goto out;
1200                 } else {
1201                         spin_unlock(&delayed_refs->lock);
1202                 }
1203         }
1204
1205         if (path->slots[0]) {
1206                 struct extent_buffer *leaf;
1207                 int slot;
1208
1209                 path->slots[0]--;
1210                 leaf = path->nodes[0];
1211                 slot = path->slots[0];
1212                 btrfs_item_key_to_cpu(leaf, &key, slot);
1213                 if (key.objectid == bytenr &&
1214                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1215                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1216                         ret = add_inline_refs(fs_info, path, bytenr,
1217                                               &info_level, &preftrees,
1218                                               &total_refs, sc);
1219                         if (ret)
1220                                 goto out;
1221                         ret = add_keyed_refs(fs_info, path, bytenr, info_level,
1222                                              &preftrees, sc);
1223                         if (ret)
1224                                 goto out;
1225                 }
1226         }
1227
1228         btrfs_release_path(path);
1229
1230         ret = add_missing_keys(fs_info, &preftrees);
1231         if (ret)
1232                 goto out;
1233
1234         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect_missing_keys.root.rb_root));
1235
1236         ret = resolve_indirect_refs(fs_info, path, time_seq, &preftrees,
1237                                     extent_item_pos, total_refs, sc, ignore_offset);
1238         if (ret)
1239                 goto out;
1240
1241         WARN_ON(!RB_EMPTY_ROOT(&preftrees.indirect.root.rb_root));
1242
1243         /*
1244          * This walks the tree of merged and resolved refs. Tree blocks are
1245          * read in as needed. Unique entries are added to the ulist, and
1246          * the list of found roots is updated.
1247          *
1248          * We release the entire tree in one go before returning.
1249          */
1250         node = rb_first_cached(&preftrees.direct.root);
1251         while (node) {
1252                 ref = rb_entry(node, struct prelim_ref, rbnode);
1253                 node = rb_next(&ref->rbnode);
1254                 /*
1255                  * ref->count < 0 can happen here if there are delayed
1256                  * refs with a node->action of BTRFS_DROP_DELAYED_REF.
1257                  * prelim_ref_insert() relies on this when merging
1258                  * identical refs to keep the overall count correct.
1259                  * prelim_ref_insert() will merge only those refs
1260                  * which compare identically.  Any refs having
1261                  * e.g. different offsets would not be merged,
1262                  * and would retain their original ref->count < 0.
1263                  */
1264                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1265                         if (sc && sc->root_objectid &&
1266                             ref->root_id != sc->root_objectid) {
1267                                 ret = BACKREF_FOUND_SHARED;
1268                                 goto out;
1269                         }
1270
1271                         /* no parent == root of tree */
1272                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1273                         if (ret < 0)
1274                                 goto out;
1275                 }
1276                 if (ref->count && ref->parent) {
1277                         if (extent_item_pos && !ref->inode_list &&
1278                             ref->level == 0) {
1279                                 struct extent_buffer *eb;
1280
1281                                 eb = read_tree_block(fs_info, ref->parent, 0,
1282                                                      ref->level, NULL);
1283                                 if (IS_ERR(eb)) {
1284                                         ret = PTR_ERR(eb);
1285                                         goto out;
1286                                 } else if (!extent_buffer_uptodate(eb)) {
1287                                         free_extent_buffer(eb);
1288                                         ret = -EIO;
1289                                         goto out;
1290                                 }
1291                                 btrfs_tree_read_lock(eb);
1292                                 btrfs_set_lock_blocking_read(eb);
1293                                 ret = find_extent_in_eb(eb, bytenr,
1294                                                         *extent_item_pos, &eie, ignore_offset);
1295                                 btrfs_tree_read_unlock_blocking(eb);
1296                                 free_extent_buffer(eb);
1297                                 if (ret < 0)
1298                                         goto out;
1299                                 ref->inode_list = eie;
1300                         }
1301                         ret = ulist_add_merge_ptr(refs, ref->parent,
1302                                                   ref->inode_list,
1303                                                   (void **)&eie, GFP_NOFS);
1304                         if (ret < 0)
1305                                 goto out;
1306                         if (!ret && extent_item_pos) {
1307                                 /*
1308                                  * we've recorded that parent, so we must extend
1309                                  * its inode list here
1310                                  */
1311                                 BUG_ON(!eie);
1312                                 while (eie->next)
1313                                         eie = eie->next;
1314                                 eie->next = ref->inode_list;
1315                         }
1316                         eie = NULL;
1317                 }
1318                 cond_resched();
1319         }
1320
1321 out:
1322         btrfs_free_path(path);
1323
1324         prelim_release(&preftrees.direct);
1325         prelim_release(&preftrees.indirect);
1326         prelim_release(&preftrees.indirect_missing_keys);
1327
1328         if (ret < 0)
1329                 free_inode_elem_list(eie);
1330         return ret;
1331 }
1332
1333 static void free_leaf_list(struct ulist *blocks)
1334 {
1335         struct ulist_node *node = NULL;
1336         struct extent_inode_elem *eie;
1337         struct ulist_iterator uiter;
1338
1339         ULIST_ITER_INIT(&uiter);
1340         while ((node = ulist_next(blocks, &uiter))) {
1341                 if (!node->aux)
1342                         continue;
1343                 eie = unode_aux_to_inode_list(node);
1344                 free_inode_elem_list(eie);
1345                 node->aux = 0;
1346         }
1347
1348         ulist_free(blocks);
1349 }
1350
1351 /*
1352  * Finds all leafs with a reference to the specified combination of bytenr and
1353  * offset. key_list_head will point to a list of corresponding keys (caller must
1354  * free each list element). The leafs will be stored in the leafs ulist, which
1355  * must be freed with ulist_free.
1356  *
1357  * returns 0 on success, <0 on error
1358  */
1359 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1360                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1361                                 u64 time_seq, struct ulist **leafs,
1362                                 const u64 *extent_item_pos, bool ignore_offset)
1363 {
1364         int ret;
1365
1366         *leafs = ulist_alloc(GFP_NOFS);
1367         if (!*leafs)
1368                 return -ENOMEM;
1369
1370         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1371                                 *leafs, NULL, extent_item_pos, NULL, ignore_offset);
1372         if (ret < 0 && ret != -ENOENT) {
1373                 free_leaf_list(*leafs);
1374                 return ret;
1375         }
1376
1377         return 0;
1378 }
1379
1380 /*
1381  * walk all backrefs for a given extent to find all roots that reference this
1382  * extent. Walking a backref means finding all extents that reference this
1383  * extent and in turn walk the backrefs of those, too. Naturally this is a
1384  * recursive process, but here it is implemented in an iterative fashion: We
1385  * find all referencing extents for the extent in question and put them on a
1386  * list. In turn, we find all referencing extents for those, further appending
1387  * to the list. The way we iterate the list allows adding more elements after
1388  * the current while iterating. The process stops when we reach the end of the
1389  * list. Found roots are added to the roots list.
1390  *
1391  * returns 0 on success, < 0 on error.
1392  */
1393 static int btrfs_find_all_roots_safe(struct btrfs_trans_handle *trans,
1394                                      struct btrfs_fs_info *fs_info, u64 bytenr,
1395                                      u64 time_seq, struct ulist **roots,
1396                                      bool ignore_offset)
1397 {
1398         struct ulist *tmp;
1399         struct ulist_node *node = NULL;
1400         struct ulist_iterator uiter;
1401         int ret;
1402
1403         tmp = ulist_alloc(GFP_NOFS);
1404         if (!tmp)
1405                 return -ENOMEM;
1406         *roots = ulist_alloc(GFP_NOFS);
1407         if (!*roots) {
1408                 ulist_free(tmp);
1409                 return -ENOMEM;
1410         }
1411
1412         ULIST_ITER_INIT(&uiter);
1413         while (1) {
1414                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1415                                         tmp, *roots, NULL, NULL, ignore_offset);
1416                 if (ret < 0 && ret != -ENOENT) {
1417                         ulist_free(tmp);
1418                         ulist_free(*roots);
1419                         return ret;
1420                 }
1421                 node = ulist_next(tmp, &uiter);
1422                 if (!node)
1423                         break;
1424                 bytenr = node->val;
1425                 cond_resched();
1426         }
1427
1428         ulist_free(tmp);
1429         return 0;
1430 }
1431
1432 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1433                          struct btrfs_fs_info *fs_info, u64 bytenr,
1434                          u64 time_seq, struct ulist **roots,
1435                          bool ignore_offset)
1436 {
1437         int ret;
1438
1439         if (!trans)
1440                 down_read(&fs_info->commit_root_sem);
1441         ret = btrfs_find_all_roots_safe(trans, fs_info, bytenr,
1442                                         time_seq, roots, ignore_offset);
1443         if (!trans)
1444                 up_read(&fs_info->commit_root_sem);
1445         return ret;
1446 }
1447
1448 /**
1449  * btrfs_check_shared - tell us whether an extent is shared
1450  *
1451  * btrfs_check_shared uses the backref walking code but will short
1452  * circuit as soon as it finds a root or inode that doesn't match the
1453  * one passed in. This provides a significant performance benefit for
1454  * callers (such as fiemap) which want to know whether the extent is
1455  * shared but do not need a ref count.
1456  *
1457  * This attempts to allocate a transaction in order to account for
1458  * delayed refs, but continues on even when the alloc fails.
1459  *
1460  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1461  */
1462 int btrfs_check_shared(struct btrfs_root *root, u64 inum, u64 bytenr)
1463 {
1464         struct btrfs_fs_info *fs_info = root->fs_info;
1465         struct btrfs_trans_handle *trans;
1466         struct ulist *tmp = NULL;
1467         struct ulist *roots = NULL;
1468         struct ulist_iterator uiter;
1469         struct ulist_node *node;
1470         struct seq_list elem = SEQ_LIST_INIT(elem);
1471         int ret = 0;
1472         struct share_check shared = {
1473                 .root_objectid = root->root_key.objectid,
1474                 .inum = inum,
1475                 .share_count = 0,
1476         };
1477
1478         tmp = ulist_alloc(GFP_NOFS);
1479         roots = ulist_alloc(GFP_NOFS);
1480         if (!tmp || !roots) {
1481                 ulist_free(tmp);
1482                 ulist_free(roots);
1483                 return -ENOMEM;
1484         }
1485
1486         trans = btrfs_join_transaction(root);
1487         if (IS_ERR(trans)) {
1488                 trans = NULL;
1489                 down_read(&fs_info->commit_root_sem);
1490         } else {
1491                 btrfs_get_tree_mod_seq(fs_info, &elem);
1492         }
1493
1494         ULIST_ITER_INIT(&uiter);
1495         while (1) {
1496                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1497                                         roots, NULL, &shared, false);
1498                 if (ret == BACKREF_FOUND_SHARED) {
1499                         /* this is the only condition under which we return 1 */
1500                         ret = 1;
1501                         break;
1502                 }
1503                 if (ret < 0 && ret != -ENOENT)
1504                         break;
1505                 ret = 0;
1506                 node = ulist_next(tmp, &uiter);
1507                 if (!node)
1508                         break;
1509                 bytenr = node->val;
1510                 shared.share_count = 0;
1511                 cond_resched();
1512         }
1513
1514         if (trans) {
1515                 btrfs_put_tree_mod_seq(fs_info, &elem);
1516                 btrfs_end_transaction(trans);
1517         } else {
1518                 up_read(&fs_info->commit_root_sem);
1519         }
1520         ulist_free(tmp);
1521         ulist_free(roots);
1522         return ret;
1523 }
1524
1525 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1526                           u64 start_off, struct btrfs_path *path,
1527                           struct btrfs_inode_extref **ret_extref,
1528                           u64 *found_off)
1529 {
1530         int ret, slot;
1531         struct btrfs_key key;
1532         struct btrfs_key found_key;
1533         struct btrfs_inode_extref *extref;
1534         const struct extent_buffer *leaf;
1535         unsigned long ptr;
1536
1537         key.objectid = inode_objectid;
1538         key.type = BTRFS_INODE_EXTREF_KEY;
1539         key.offset = start_off;
1540
1541         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1542         if (ret < 0)
1543                 return ret;
1544
1545         while (1) {
1546                 leaf = path->nodes[0];
1547                 slot = path->slots[0];
1548                 if (slot >= btrfs_header_nritems(leaf)) {
1549                         /*
1550                          * If the item at offset is not found,
1551                          * btrfs_search_slot will point us to the slot
1552                          * where it should be inserted. In our case
1553                          * that will be the slot directly before the
1554                          * next INODE_REF_KEY_V2 item. In the case
1555                          * that we're pointing to the last slot in a
1556                          * leaf, we must move one leaf over.
1557                          */
1558                         ret = btrfs_next_leaf(root, path);
1559                         if (ret) {
1560                                 if (ret >= 1)
1561                                         ret = -ENOENT;
1562                                 break;
1563                         }
1564                         continue;
1565                 }
1566
1567                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1568
1569                 /*
1570                  * Check that we're still looking at an extended ref key for
1571                  * this particular objectid. If we have different
1572                  * objectid or type then there are no more to be found
1573                  * in the tree and we can exit.
1574                  */
1575                 ret = -ENOENT;
1576                 if (found_key.objectid != inode_objectid)
1577                         break;
1578                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1579                         break;
1580
1581                 ret = 0;
1582                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1583                 extref = (struct btrfs_inode_extref *)ptr;
1584                 *ret_extref = extref;
1585                 if (found_off)
1586                         *found_off = found_key.offset;
1587                 break;
1588         }
1589
1590         return ret;
1591 }
1592
1593 /*
1594  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1595  * Elements of the path are separated by '/' and the path is guaranteed to be
1596  * 0-terminated. the path is only given within the current file system.
1597  * Therefore, it never starts with a '/'. the caller is responsible to provide
1598  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1599  * the start point of the resulting string is returned. this pointer is within
1600  * dest, normally.
1601  * in case the path buffer would overflow, the pointer is decremented further
1602  * as if output was written to the buffer, though no more output is actually
1603  * generated. that way, the caller can determine how much space would be
1604  * required for the path to fit into the buffer. in that case, the returned
1605  * value will be smaller than dest. callers must check this!
1606  */
1607 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1608                         u32 name_len, unsigned long name_off,
1609                         struct extent_buffer *eb_in, u64 parent,
1610                         char *dest, u32 size)
1611 {
1612         int slot;
1613         u64 next_inum;
1614         int ret;
1615         s64 bytes_left = ((s64)size) - 1;
1616         struct extent_buffer *eb = eb_in;
1617         struct btrfs_key found_key;
1618         int leave_spinning = path->leave_spinning;
1619         struct btrfs_inode_ref *iref;
1620
1621         if (bytes_left >= 0)
1622                 dest[bytes_left] = '\0';
1623
1624         path->leave_spinning = 1;
1625         while (1) {
1626                 bytes_left -= name_len;
1627                 if (bytes_left >= 0)
1628                         read_extent_buffer(eb, dest + bytes_left,
1629                                            name_off, name_len);
1630                 if (eb != eb_in) {
1631                         if (!path->skip_locking)
1632                                 btrfs_tree_read_unlock_blocking(eb);
1633                         free_extent_buffer(eb);
1634                 }
1635                 ret = btrfs_find_item(fs_root, path, parent, 0,
1636                                 BTRFS_INODE_REF_KEY, &found_key);
1637                 if (ret > 0)
1638                         ret = -ENOENT;
1639                 if (ret)
1640                         break;
1641
1642                 next_inum = found_key.offset;
1643
1644                 /* regular exit ahead */
1645                 if (parent == next_inum)
1646                         break;
1647
1648                 slot = path->slots[0];
1649                 eb = path->nodes[0];
1650                 /* make sure we can use eb after releasing the path */
1651                 if (eb != eb_in) {
1652                         if (!path->skip_locking)
1653                                 btrfs_set_lock_blocking_read(eb);
1654                         path->nodes[0] = NULL;
1655                         path->locks[0] = 0;
1656                 }
1657                 btrfs_release_path(path);
1658                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1659
1660                 name_len = btrfs_inode_ref_name_len(eb, iref);
1661                 name_off = (unsigned long)(iref + 1);
1662
1663                 parent = next_inum;
1664                 --bytes_left;
1665                 if (bytes_left >= 0)
1666                         dest[bytes_left] = '/';
1667         }
1668
1669         btrfs_release_path(path);
1670         path->leave_spinning = leave_spinning;
1671
1672         if (ret)
1673                 return ERR_PTR(ret);
1674
1675         return dest + bytes_left;
1676 }
1677
1678 /*
1679  * this makes the path point to (logical EXTENT_ITEM *)
1680  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1681  * tree blocks and <0 on error.
1682  */
1683 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1684                         struct btrfs_path *path, struct btrfs_key *found_key,
1685                         u64 *flags_ret)
1686 {
1687         int ret;
1688         u64 flags;
1689         u64 size = 0;
1690         u32 item_size;
1691         const struct extent_buffer *eb;
1692         struct btrfs_extent_item *ei;
1693         struct btrfs_key key;
1694
1695         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1696                 key.type = BTRFS_METADATA_ITEM_KEY;
1697         else
1698                 key.type = BTRFS_EXTENT_ITEM_KEY;
1699         key.objectid = logical;
1700         key.offset = (u64)-1;
1701
1702         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1703         if (ret < 0)
1704                 return ret;
1705
1706         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1707         if (ret) {
1708                 if (ret > 0)
1709                         ret = -ENOENT;
1710                 return ret;
1711         }
1712         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1713         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1714                 size = fs_info->nodesize;
1715         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1716                 size = found_key->offset;
1717
1718         if (found_key->objectid > logical ||
1719             found_key->objectid + size <= logical) {
1720                 btrfs_debug(fs_info,
1721                         "logical %llu is not within any extent", logical);
1722                 return -ENOENT;
1723         }
1724
1725         eb = path->nodes[0];
1726         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1727         BUG_ON(item_size < sizeof(*ei));
1728
1729         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1730         flags = btrfs_extent_flags(eb, ei);
1731
1732         btrfs_debug(fs_info,
1733                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1734                  logical, logical - found_key->objectid, found_key->objectid,
1735                  found_key->offset, flags, item_size);
1736
1737         WARN_ON(!flags_ret);
1738         if (flags_ret) {
1739                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1740                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1741                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1742                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1743                 else
1744                         BUG_ON(1);
1745                 return 0;
1746         }
1747
1748         return -EIO;
1749 }
1750
1751 /*
1752  * helper function to iterate extent inline refs. ptr must point to a 0 value
1753  * for the first call and may be modified. it is used to track state.
1754  * if more refs exist, 0 is returned and the next call to
1755  * get_extent_inline_ref must pass the modified ptr parameter to get the
1756  * next ref. after the last ref was processed, 1 is returned.
1757  * returns <0 on error
1758  */
1759 static int get_extent_inline_ref(unsigned long *ptr,
1760                                  const struct extent_buffer *eb,
1761                                  const struct btrfs_key *key,
1762                                  const struct btrfs_extent_item *ei,
1763                                  u32 item_size,
1764                                  struct btrfs_extent_inline_ref **out_eiref,
1765                                  int *out_type)
1766 {
1767         unsigned long end;
1768         u64 flags;
1769         struct btrfs_tree_block_info *info;
1770
1771         if (!*ptr) {
1772                 /* first call */
1773                 flags = btrfs_extent_flags(eb, ei);
1774                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1775                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1776                                 /* a skinny metadata extent */
1777                                 *out_eiref =
1778                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1779                         } else {
1780                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1781                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1782                                 *out_eiref =
1783                                    (struct btrfs_extent_inline_ref *)(info + 1);
1784                         }
1785                 } else {
1786                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1787                 }
1788                 *ptr = (unsigned long)*out_eiref;
1789                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1790                         return -ENOENT;
1791         }
1792
1793         end = (unsigned long)ei + item_size;
1794         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1795         *out_type = btrfs_get_extent_inline_ref_type(eb, *out_eiref,
1796                                                      BTRFS_REF_TYPE_ANY);
1797         if (*out_type == BTRFS_REF_TYPE_INVALID)
1798                 return -EUCLEAN;
1799
1800         *ptr += btrfs_extent_inline_ref_size(*out_type);
1801         WARN_ON(*ptr > end);
1802         if (*ptr == end)
1803                 return 1; /* last */
1804
1805         return 0;
1806 }
1807
1808 /*
1809  * reads the tree block backref for an extent. tree level and root are returned
1810  * through out_level and out_root. ptr must point to a 0 value for the first
1811  * call and may be modified (see get_extent_inline_ref comment).
1812  * returns 0 if data was provided, 1 if there was no more data to provide or
1813  * <0 on error.
1814  */
1815 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1816                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1817                             u32 item_size, u64 *out_root, u8 *out_level)
1818 {
1819         int ret;
1820         int type;
1821         struct btrfs_extent_inline_ref *eiref;
1822
1823         if (*ptr == (unsigned long)-1)
1824                 return 1;
1825
1826         while (1) {
1827                 ret = get_extent_inline_ref(ptr, eb, key, ei, item_size,
1828                                               &eiref, &type);
1829                 if (ret < 0)
1830                         return ret;
1831
1832                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1833                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1834                         break;
1835
1836                 if (ret == 1)
1837                         return 1;
1838         }
1839
1840         /* we can treat both ref types equally here */
1841         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1842
1843         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1844                 struct btrfs_tree_block_info *info;
1845
1846                 info = (struct btrfs_tree_block_info *)(ei + 1);
1847                 *out_level = btrfs_tree_block_level(eb, info);
1848         } else {
1849                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1850                 *out_level = (u8)key->offset;
1851         }
1852
1853         if (ret == 1)
1854                 *ptr = (unsigned long)-1;
1855
1856         return 0;
1857 }
1858
1859 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1860                              struct extent_inode_elem *inode_list,
1861                              u64 root, u64 extent_item_objectid,
1862                              iterate_extent_inodes_t *iterate, void *ctx)
1863 {
1864         struct extent_inode_elem *eie;
1865         int ret = 0;
1866
1867         for (eie = inode_list; eie; eie = eie->next) {
1868                 btrfs_debug(fs_info,
1869                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1870                             extent_item_objectid, eie->inum,
1871                             eie->offset, root);
1872                 ret = iterate(eie->inum, eie->offset, root, ctx);
1873                 if (ret) {
1874                         btrfs_debug(fs_info,
1875                                     "stopping iteration for %llu due to ret=%d",
1876                                     extent_item_objectid, ret);
1877                         break;
1878                 }
1879         }
1880
1881         return ret;
1882 }
1883
1884 /*
1885  * calls iterate() for every inode that references the extent identified by
1886  * the given parameters.
1887  * when the iterator function returns a non-zero value, iteration stops.
1888  */
1889 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1890                                 u64 extent_item_objectid, u64 extent_item_pos,
1891                                 int search_commit_root,
1892                                 iterate_extent_inodes_t *iterate, void *ctx,
1893                                 bool ignore_offset)
1894 {
1895         int ret;
1896         struct btrfs_trans_handle *trans = NULL;
1897         struct ulist *refs = NULL;
1898         struct ulist *roots = NULL;
1899         struct ulist_node *ref_node = NULL;
1900         struct ulist_node *root_node = NULL;
1901         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
1902         struct ulist_iterator ref_uiter;
1903         struct ulist_iterator root_uiter;
1904
1905         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
1906                         extent_item_objectid);
1907
1908         if (!search_commit_root) {
1909                 trans = btrfs_join_transaction(fs_info->extent_root);
1910                 if (IS_ERR(trans))
1911                         return PTR_ERR(trans);
1912                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1913         } else {
1914                 down_read(&fs_info->commit_root_sem);
1915         }
1916
1917         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
1918                                    tree_mod_seq_elem.seq, &refs,
1919                                    &extent_item_pos, ignore_offset);
1920         if (ret)
1921                 goto out;
1922
1923         ULIST_ITER_INIT(&ref_uiter);
1924         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
1925                 ret = btrfs_find_all_roots_safe(trans, fs_info, ref_node->val,
1926                                                 tree_mod_seq_elem.seq, &roots,
1927                                                 ignore_offset);
1928                 if (ret)
1929                         break;
1930                 ULIST_ITER_INIT(&root_uiter);
1931                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
1932                         btrfs_debug(fs_info,
1933                                     "root %llu references leaf %llu, data list %#llx",
1934                                     root_node->val, ref_node->val,
1935                                     ref_node->aux);
1936                         ret = iterate_leaf_refs(fs_info,
1937                                                 (struct extent_inode_elem *)
1938                                                 (uintptr_t)ref_node->aux,
1939                                                 root_node->val,
1940                                                 extent_item_objectid,
1941                                                 iterate, ctx);
1942                 }
1943                 ulist_free(roots);
1944         }
1945
1946         free_leaf_list(refs);
1947 out:
1948         if (!search_commit_root) {
1949                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
1950                 btrfs_end_transaction(trans);
1951         } else {
1952                 up_read(&fs_info->commit_root_sem);
1953         }
1954
1955         return ret;
1956 }
1957
1958 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
1959                                 struct btrfs_path *path,
1960                                 iterate_extent_inodes_t *iterate, void *ctx,
1961                                 bool ignore_offset)
1962 {
1963         int ret;
1964         u64 extent_item_pos;
1965         u64 flags = 0;
1966         struct btrfs_key found_key;
1967         int search_commit_root = path->search_commit_root;
1968
1969         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
1970         btrfs_release_path(path);
1971         if (ret < 0)
1972                 return ret;
1973         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1974                 return -EINVAL;
1975
1976         extent_item_pos = logical - found_key.objectid;
1977         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1978                                         extent_item_pos, search_commit_root,
1979                                         iterate, ctx, ignore_offset);
1980
1981         return ret;
1982 }
1983
1984 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
1985                               struct extent_buffer *eb, void *ctx);
1986
1987 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
1988                               struct btrfs_path *path,
1989                               iterate_irefs_t *iterate, void *ctx)
1990 {
1991         int ret = 0;
1992         int slot;
1993         u32 cur;
1994         u32 len;
1995         u32 name_len;
1996         u64 parent = 0;
1997         int found = 0;
1998         struct extent_buffer *eb;
1999         struct btrfs_item *item;
2000         struct btrfs_inode_ref *iref;
2001         struct btrfs_key found_key;
2002
2003         while (!ret) {
2004                 ret = btrfs_find_item(fs_root, path, inum,
2005                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2006                                 &found_key);
2007
2008                 if (ret < 0)
2009                         break;
2010                 if (ret) {
2011                         ret = found ? 0 : -ENOENT;
2012                         break;
2013                 }
2014                 ++found;
2015
2016                 parent = found_key.offset;
2017                 slot = path->slots[0];
2018                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2019                 if (!eb) {
2020                         ret = -ENOMEM;
2021                         break;
2022                 }
2023                 btrfs_release_path(path);
2024
2025                 item = btrfs_item_nr(slot);
2026                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2027
2028                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2029                         name_len = btrfs_inode_ref_name_len(eb, iref);
2030                         /* path must be released before calling iterate()! */
2031                         btrfs_debug(fs_root->fs_info,
2032                                 "following ref at offset %u for inode %llu in tree %llu",
2033                                 cur, found_key.objectid,
2034                                 fs_root->root_key.objectid);
2035                         ret = iterate(parent, name_len,
2036                                       (unsigned long)(iref + 1), eb, ctx);
2037                         if (ret)
2038                                 break;
2039                         len = sizeof(*iref) + name_len;
2040                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2041                 }
2042                 free_extent_buffer(eb);
2043         }
2044
2045         btrfs_release_path(path);
2046
2047         return ret;
2048 }
2049
2050 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2051                                  struct btrfs_path *path,
2052                                  iterate_irefs_t *iterate, void *ctx)
2053 {
2054         int ret;
2055         int slot;
2056         u64 offset = 0;
2057         u64 parent;
2058         int found = 0;
2059         struct extent_buffer *eb;
2060         struct btrfs_inode_extref *extref;
2061         u32 item_size;
2062         u32 cur_offset;
2063         unsigned long ptr;
2064
2065         while (1) {
2066                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2067                                             &offset);
2068                 if (ret < 0)
2069                         break;
2070                 if (ret) {
2071                         ret = found ? 0 : -ENOENT;
2072                         break;
2073                 }
2074                 ++found;
2075
2076                 slot = path->slots[0];
2077                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2078                 if (!eb) {
2079                         ret = -ENOMEM;
2080                         break;
2081                 }
2082                 btrfs_release_path(path);
2083
2084                 item_size = btrfs_item_size_nr(eb, slot);
2085                 ptr = btrfs_item_ptr_offset(eb, slot);
2086                 cur_offset = 0;
2087
2088                 while (cur_offset < item_size) {
2089                         u32 name_len;
2090
2091                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2092                         parent = btrfs_inode_extref_parent(eb, extref);
2093                         name_len = btrfs_inode_extref_name_len(eb, extref);
2094                         ret = iterate(parent, name_len,
2095                                       (unsigned long)&extref->name, eb, ctx);
2096                         if (ret)
2097                                 break;
2098
2099                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2100                         cur_offset += sizeof(*extref);
2101                 }
2102                 free_extent_buffer(eb);
2103
2104                 offset++;
2105         }
2106
2107         btrfs_release_path(path);
2108
2109         return ret;
2110 }
2111
2112 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2113                          struct btrfs_path *path, iterate_irefs_t *iterate,
2114                          void *ctx)
2115 {
2116         int ret;
2117         int found_refs = 0;
2118
2119         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2120         if (!ret)
2121                 ++found_refs;
2122         else if (ret != -ENOENT)
2123                 return ret;
2124
2125         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2126         if (ret == -ENOENT && found_refs)
2127                 return 0;
2128
2129         return ret;
2130 }
2131
2132 /*
2133  * returns 0 if the path could be dumped (probably truncated)
2134  * returns <0 in case of an error
2135  */
2136 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2137                          struct extent_buffer *eb, void *ctx)
2138 {
2139         struct inode_fs_paths *ipath = ctx;
2140         char *fspath;
2141         char *fspath_min;
2142         int i = ipath->fspath->elem_cnt;
2143         const int s_ptr = sizeof(char *);
2144         u32 bytes_left;
2145
2146         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2147                                         ipath->fspath->bytes_left - s_ptr : 0;
2148
2149         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2150         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2151                                    name_off, eb, inum, fspath_min, bytes_left);
2152         if (IS_ERR(fspath))
2153                 return PTR_ERR(fspath);
2154
2155         if (fspath > fspath_min) {
2156                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2157                 ++ipath->fspath->elem_cnt;
2158                 ipath->fspath->bytes_left = fspath - fspath_min;
2159         } else {
2160                 ++ipath->fspath->elem_missed;
2161                 ipath->fspath->bytes_missing += fspath_min - fspath;
2162                 ipath->fspath->bytes_left = 0;
2163         }
2164
2165         return 0;
2166 }
2167
2168 /*
2169  * this dumps all file system paths to the inode into the ipath struct, provided
2170  * is has been created large enough. each path is zero-terminated and accessed
2171  * from ipath->fspath->val[i].
2172  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2173  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2174  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2175  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2176  * have been needed to return all paths.
2177  */
2178 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2179 {
2180         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2181                              inode_to_path, ipath);
2182 }
2183
2184 struct btrfs_data_container *init_data_container(u32 total_bytes)
2185 {
2186         struct btrfs_data_container *data;
2187         size_t alloc_bytes;
2188
2189         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2190         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2191         if (!data)
2192                 return ERR_PTR(-ENOMEM);
2193
2194         if (total_bytes >= sizeof(*data)) {
2195                 data->bytes_left = total_bytes - sizeof(*data);
2196                 data->bytes_missing = 0;
2197         } else {
2198                 data->bytes_missing = sizeof(*data) - total_bytes;
2199                 data->bytes_left = 0;
2200         }
2201
2202         data->elem_cnt = 0;
2203         data->elem_missed = 0;
2204
2205         return data;
2206 }
2207
2208 /*
2209  * allocates space to return multiple file system paths for an inode.
2210  * total_bytes to allocate are passed, note that space usable for actual path
2211  * information will be total_bytes - sizeof(struct inode_fs_paths).
2212  * the returned pointer must be freed with free_ipath() in the end.
2213  */
2214 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2215                                         struct btrfs_path *path)
2216 {
2217         struct inode_fs_paths *ifp;
2218         struct btrfs_data_container *fspath;
2219
2220         fspath = init_data_container(total_bytes);
2221         if (IS_ERR(fspath))
2222                 return ERR_CAST(fspath);
2223
2224         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2225         if (!ifp) {
2226                 kvfree(fspath);
2227                 return ERR_PTR(-ENOMEM);
2228         }
2229
2230         ifp->btrfs_path = path;
2231         ifp->fspath = fspath;
2232         ifp->fs_root = fs_root;
2233
2234         return ifp;
2235 }
2236
2237 void free_ipath(struct inode_fs_paths *ipath)
2238 {
2239         if (!ipath)
2240                 return;
2241         kvfree(ipath->fspath);
2242         kfree(ipath);
2243 }