Merge branch 'for-4.14/block-postmerge' of git://git.kernel.dk/linux-block
[sfrench/cifs-2.6.git] / fs / btrfs / backref.c
1 /*
2  * Copyright (C) 2011 STRATO.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/mm.h>
20 #include <linux/rbtree.h>
21 #include "ctree.h"
22 #include "disk-io.h"
23 #include "backref.h"
24 #include "ulist.h"
25 #include "transaction.h"
26 #include "delayed-ref.h"
27 #include "locking.h"
28
29 enum merge_mode {
30         MERGE_IDENTICAL_KEYS = 1,
31         MERGE_IDENTICAL_PARENTS,
32 };
33
34 /* Just an arbitrary number so we can be sure this happened */
35 #define BACKREF_FOUND_SHARED 6
36
37 struct extent_inode_elem {
38         u64 inum;
39         u64 offset;
40         struct extent_inode_elem *next;
41 };
42
43 /*
44  * ref_root is used as the root of the ref tree that hold a collection
45  * of unique references.
46  */
47 struct ref_root {
48         struct rb_root rb_root;
49
50         /*
51          * The unique_refs represents the number of ref_nodes with a positive
52          * count stored in the tree. Even if a ref_node (the count is greater
53          * than one) is added, the unique_refs will only increase by one.
54          */
55         unsigned int unique_refs;
56 };
57
58 /* ref_node is used to store a unique reference to the ref tree. */
59 struct ref_node {
60         struct rb_node rb_node;
61
62         /* For NORMAL_REF, otherwise all these fields should be set to 0 */
63         u64 root_id;
64         u64 object_id;
65         u64 offset;
66
67         /* For SHARED_REF, otherwise parent field should be set to 0 */
68         u64 parent;
69
70         /* Ref to the ref_mod of btrfs_delayed_ref_node */
71         int ref_mod;
72 };
73
74 /* Dynamically allocate and initialize a ref_root */
75 static struct ref_root *ref_root_alloc(void)
76 {
77         struct ref_root *ref_tree;
78
79         ref_tree = kmalloc(sizeof(*ref_tree), GFP_NOFS);
80         if (!ref_tree)
81                 return NULL;
82
83         ref_tree->rb_root = RB_ROOT;
84         ref_tree->unique_refs = 0;
85
86         return ref_tree;
87 }
88
89 /* Free all nodes in the ref tree, and reinit ref_root */
90 static void ref_root_fini(struct ref_root *ref_tree)
91 {
92         struct ref_node *node;
93         struct rb_node *next;
94
95         while ((next = rb_first(&ref_tree->rb_root)) != NULL) {
96                 node = rb_entry(next, struct ref_node, rb_node);
97                 rb_erase(next, &ref_tree->rb_root);
98                 kfree(node);
99         }
100
101         ref_tree->rb_root = RB_ROOT;
102         ref_tree->unique_refs = 0;
103 }
104
105 static void ref_root_free(struct ref_root *ref_tree)
106 {
107         if (!ref_tree)
108                 return;
109
110         ref_root_fini(ref_tree);
111         kfree(ref_tree);
112 }
113
114 /*
115  * Compare ref_node with (root_id, object_id, offset, parent)
116  *
117  * The function compares two ref_node a and b. It returns an integer less
118  * than, equal to, or greater than zero , respectively, to be less than, to
119  * equal, or be greater than b.
120  */
121 static int ref_node_cmp(struct ref_node *a, struct ref_node *b)
122 {
123         if (a->root_id < b->root_id)
124                 return -1;
125         else if (a->root_id > b->root_id)
126                 return 1;
127
128         if (a->object_id < b->object_id)
129                 return -1;
130         else if (a->object_id > b->object_id)
131                 return 1;
132
133         if (a->offset < b->offset)
134                 return -1;
135         else if (a->offset > b->offset)
136                 return 1;
137
138         if (a->parent < b->parent)
139                 return -1;
140         else if (a->parent > b->parent)
141                 return 1;
142
143         return 0;
144 }
145
146 /*
147  * Search ref_node with (root_id, object_id, offset, parent) in the tree
148  *
149  * if found, the pointer of the ref_node will be returned;
150  * if not found, NULL will be returned and pos will point to the rb_node for
151  * insert, pos_parent will point to pos'parent for insert;
152 */
153 static struct ref_node *__ref_tree_search(struct ref_root *ref_tree,
154                                           struct rb_node ***pos,
155                                           struct rb_node **pos_parent,
156                                           u64 root_id, u64 object_id,
157                                           u64 offset, u64 parent)
158 {
159         struct ref_node *cur = NULL;
160         struct ref_node entry;
161         int ret;
162
163         entry.root_id = root_id;
164         entry.object_id = object_id;
165         entry.offset = offset;
166         entry.parent = parent;
167
168         *pos = &ref_tree->rb_root.rb_node;
169
170         while (**pos) {
171                 *pos_parent = **pos;
172                 cur = rb_entry(*pos_parent, struct ref_node, rb_node);
173
174                 ret = ref_node_cmp(cur, &entry);
175                 if (ret > 0)
176                         *pos = &(**pos)->rb_left;
177                 else if (ret < 0)
178                         *pos = &(**pos)->rb_right;
179                 else
180                         return cur;
181         }
182
183         return NULL;
184 }
185
186 /*
187  * Insert a ref_node to the ref tree
188  * @pos used for specifiy the position to insert
189  * @pos_parent for specifiy pos's parent
190  *
191  * success, return 0;
192  * ref_node already exists, return -EEXIST;
193 */
194 static int ref_tree_insert(struct ref_root *ref_tree, struct rb_node **pos,
195                            struct rb_node *pos_parent, struct ref_node *ins)
196 {
197         struct rb_node **p = NULL;
198         struct rb_node *parent = NULL;
199         struct ref_node *cur = NULL;
200
201         if (!pos) {
202                 cur = __ref_tree_search(ref_tree, &p, &parent, ins->root_id,
203                                         ins->object_id, ins->offset,
204                                         ins->parent);
205                 if (cur)
206                         return -EEXIST;
207         } else {
208                 p = pos;
209                 parent = pos_parent;
210         }
211
212         rb_link_node(&ins->rb_node, parent, p);
213         rb_insert_color(&ins->rb_node, &ref_tree->rb_root);
214
215         return 0;
216 }
217
218 /* Erase and free ref_node, caller should update ref_root->unique_refs */
219 static void ref_tree_remove(struct ref_root *ref_tree, struct ref_node *node)
220 {
221         rb_erase(&node->rb_node, &ref_tree->rb_root);
222         kfree(node);
223 }
224
225 /*
226  * Update ref_root->unique_refs
227  *
228  * Call __ref_tree_search
229  *      1. if ref_node doesn't exist, ref_tree_insert this node, and update
230  *      ref_root->unique_refs:
231  *              if ref_node->ref_mod > 0, ref_root->unique_refs++;
232  *              if ref_node->ref_mod < 0, do noting;
233  *
234  *      2. if ref_node is found, then get origin ref_node->ref_mod, and update
235  *      ref_node->ref_mod.
236  *              if ref_node->ref_mod is equal to 0,then call ref_tree_remove
237  *
238  *              according to origin_mod and new_mod, update ref_root->items
239  *              +----------------+--------------+-------------+
240  *              |                |new_count <= 0|new_count > 0|
241  *              +----------------+--------------+-------------+
242  *              |origin_count < 0|       0      |      1      |
243  *              +----------------+--------------+-------------+
244  *              |origin_count > 0|      -1      |      0      |
245  *              +----------------+--------------+-------------+
246  *
247  * In case of allocation failure, -ENOMEM is returned and the ref_tree stays
248  * unaltered.
249  * Success, return 0
250  */
251 static int ref_tree_add(struct ref_root *ref_tree, u64 root_id, u64 object_id,
252                         u64 offset, u64 parent, int count)
253 {
254         struct ref_node *node = NULL;
255         struct rb_node **pos = NULL;
256         struct rb_node *pos_parent = NULL;
257         int origin_count;
258         int ret;
259
260         if (!count)
261                 return 0;
262
263         node = __ref_tree_search(ref_tree, &pos, &pos_parent, root_id,
264                                  object_id, offset, parent);
265         if (node == NULL) {
266                 node = kmalloc(sizeof(*node), GFP_NOFS);
267                 if (!node)
268                         return -ENOMEM;
269
270                 node->root_id = root_id;
271                 node->object_id = object_id;
272                 node->offset = offset;
273                 node->parent = parent;
274                 node->ref_mod = count;
275
276                 ret = ref_tree_insert(ref_tree, pos, pos_parent, node);
277                 ASSERT(!ret);
278                 if (ret) {
279                         kfree(node);
280                         return ret;
281                 }
282
283                 ref_tree->unique_refs += node->ref_mod > 0 ? 1 : 0;
284
285                 return 0;
286         }
287
288         origin_count = node->ref_mod;
289         node->ref_mod += count;
290
291         if (node->ref_mod > 0)
292                 ref_tree->unique_refs += origin_count > 0 ? 0 : 1;
293         else if (node->ref_mod <= 0)
294                 ref_tree->unique_refs += origin_count > 0 ? -1 : 0;
295
296         if (!node->ref_mod)
297                 ref_tree_remove(ref_tree, node);
298
299         return 0;
300 }
301
302 static int check_extent_in_eb(struct btrfs_key *key, struct extent_buffer *eb,
303                                 struct btrfs_file_extent_item *fi,
304                                 u64 extent_item_pos,
305                                 struct extent_inode_elem **eie)
306 {
307         u64 offset = 0;
308         struct extent_inode_elem *e;
309
310         if (!btrfs_file_extent_compression(eb, fi) &&
311             !btrfs_file_extent_encryption(eb, fi) &&
312             !btrfs_file_extent_other_encoding(eb, fi)) {
313                 u64 data_offset;
314                 u64 data_len;
315
316                 data_offset = btrfs_file_extent_offset(eb, fi);
317                 data_len = btrfs_file_extent_num_bytes(eb, fi);
318
319                 if (extent_item_pos < data_offset ||
320                     extent_item_pos >= data_offset + data_len)
321                         return 1;
322                 offset = extent_item_pos - data_offset;
323         }
324
325         e = kmalloc(sizeof(*e), GFP_NOFS);
326         if (!e)
327                 return -ENOMEM;
328
329         e->next = *eie;
330         e->inum = key->objectid;
331         e->offset = key->offset + offset;
332         *eie = e;
333
334         return 0;
335 }
336
337 static void free_inode_elem_list(struct extent_inode_elem *eie)
338 {
339         struct extent_inode_elem *eie_next;
340
341         for (; eie; eie = eie_next) {
342                 eie_next = eie->next;
343                 kfree(eie);
344         }
345 }
346
347 static int find_extent_in_eb(struct extent_buffer *eb, u64 wanted_disk_byte,
348                                 u64 extent_item_pos,
349                                 struct extent_inode_elem **eie)
350 {
351         u64 disk_byte;
352         struct btrfs_key key;
353         struct btrfs_file_extent_item *fi;
354         int slot;
355         int nritems;
356         int extent_type;
357         int ret;
358
359         /*
360          * from the shared data ref, we only have the leaf but we need
361          * the key. thus, we must look into all items and see that we
362          * find one (some) with a reference to our extent item.
363          */
364         nritems = btrfs_header_nritems(eb);
365         for (slot = 0; slot < nritems; ++slot) {
366                 btrfs_item_key_to_cpu(eb, &key, slot);
367                 if (key.type != BTRFS_EXTENT_DATA_KEY)
368                         continue;
369                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
370                 extent_type = btrfs_file_extent_type(eb, fi);
371                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
372                         continue;
373                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
374                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
375                 if (disk_byte != wanted_disk_byte)
376                         continue;
377
378                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
379                 if (ret < 0)
380                         return ret;
381         }
382
383         return 0;
384 }
385
386 /*
387  * this structure records all encountered refs on the way up to the root
388  */
389 struct __prelim_ref {
390         struct list_head list;
391         u64 root_id;
392         struct btrfs_key key_for_search;
393         int level;
394         int count;
395         struct extent_inode_elem *inode_list;
396         u64 parent;
397         u64 wanted_disk_byte;
398 };
399
400 static struct kmem_cache *btrfs_prelim_ref_cache;
401
402 int __init btrfs_prelim_ref_init(void)
403 {
404         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
405                                         sizeof(struct __prelim_ref),
406                                         0,
407                                         SLAB_MEM_SPREAD,
408                                         NULL);
409         if (!btrfs_prelim_ref_cache)
410                 return -ENOMEM;
411         return 0;
412 }
413
414 void btrfs_prelim_ref_exit(void)
415 {
416         kmem_cache_destroy(btrfs_prelim_ref_cache);
417 }
418
419 /*
420  * the rules for all callers of this function are:
421  * - obtaining the parent is the goal
422  * - if you add a key, you must know that it is a correct key
423  * - if you cannot add the parent or a correct key, then we will look into the
424  *   block later to set a correct key
425  *
426  * delayed refs
427  * ============
428  *        backref type | shared | indirect | shared | indirect
429  * information         |   tree |     tree |   data |     data
430  * --------------------+--------+----------+--------+----------
431  *      parent logical |    y   |     -    |    -   |     -
432  *      key to resolve |    -   |     y    |    y   |     y
433  *  tree block logical |    -   |     -    |    -   |     -
434  *  root for resolving |    y   |     y    |    y   |     y
435  *
436  * - column 1:       we've the parent -> done
437  * - column 2, 3, 4: we use the key to find the parent
438  *
439  * on disk refs (inline or keyed)
440  * ==============================
441  *        backref type | shared | indirect | shared | indirect
442  * information         |   tree |     tree |   data |     data
443  * --------------------+--------+----------+--------+----------
444  *      parent logical |    y   |     -    |    y   |     -
445  *      key to resolve |    -   |     -    |    -   |     y
446  *  tree block logical |    y   |     y    |    y   |     y
447  *  root for resolving |    -   |     y    |    y   |     y
448  *
449  * - column 1, 3: we've the parent -> done
450  * - column 2:    we take the first key from the block to find the parent
451  *                (see __add_missing_keys)
452  * - column 4:    we use the key to find the parent
453  *
454  * additional information that's available but not required to find the parent
455  * block might help in merging entries to gain some speed.
456  */
457
458 static int __add_prelim_ref(struct list_head *head, u64 root_id,
459                             struct btrfs_key *key, int level,
460                             u64 parent, u64 wanted_disk_byte, int count,
461                             gfp_t gfp_mask)
462 {
463         struct __prelim_ref *ref;
464
465         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
466                 return 0;
467
468         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
469         if (!ref)
470                 return -ENOMEM;
471
472         ref->root_id = root_id;
473         if (key) {
474                 ref->key_for_search = *key;
475                 /*
476                  * We can often find data backrefs with an offset that is too
477                  * large (>= LLONG_MAX, maximum allowed file offset) due to
478                  * underflows when subtracting a file's offset with the data
479                  * offset of its corresponding extent data item. This can
480                  * happen for example in the clone ioctl.
481                  * So if we detect such case we set the search key's offset to
482                  * zero to make sure we will find the matching file extent item
483                  * at add_all_parents(), otherwise we will miss it because the
484                  * offset taken form the backref is much larger then the offset
485                  * of the file extent item. This can make us scan a very large
486                  * number of file extent items, but at least it will not make
487                  * us miss any.
488                  * This is an ugly workaround for a behaviour that should have
489                  * never existed, but it does and a fix for the clone ioctl
490                  * would touch a lot of places, cause backwards incompatibility
491                  * and would not fix the problem for extents cloned with older
492                  * kernels.
493                  */
494                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
495                     ref->key_for_search.offset >= LLONG_MAX)
496                         ref->key_for_search.offset = 0;
497         } else {
498                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
499         }
500
501         ref->inode_list = NULL;
502         ref->level = level;
503         ref->count = count;
504         ref->parent = parent;
505         ref->wanted_disk_byte = wanted_disk_byte;
506         list_add_tail(&ref->list, head);
507
508         return 0;
509 }
510
511 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
512                            struct ulist *parents, struct __prelim_ref *ref,
513                            int level, u64 time_seq, const u64 *extent_item_pos,
514                            u64 total_refs)
515 {
516         int ret = 0;
517         int slot;
518         struct extent_buffer *eb;
519         struct btrfs_key key;
520         struct btrfs_key *key_for_search = &ref->key_for_search;
521         struct btrfs_file_extent_item *fi;
522         struct extent_inode_elem *eie = NULL, *old = NULL;
523         u64 disk_byte;
524         u64 wanted_disk_byte = ref->wanted_disk_byte;
525         u64 count = 0;
526
527         if (level != 0) {
528                 eb = path->nodes[level];
529                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
530                 if (ret < 0)
531                         return ret;
532                 return 0;
533         }
534
535         /*
536          * We normally enter this function with the path already pointing to
537          * the first item to check. But sometimes, we may enter it with
538          * slot==nritems. In that case, go to the next leaf before we continue.
539          */
540         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
541                 if (time_seq == SEQ_LAST)
542                         ret = btrfs_next_leaf(root, path);
543                 else
544                         ret = btrfs_next_old_leaf(root, path, time_seq);
545         }
546
547         while (!ret && count < total_refs) {
548                 eb = path->nodes[0];
549                 slot = path->slots[0];
550
551                 btrfs_item_key_to_cpu(eb, &key, slot);
552
553                 if (key.objectid != key_for_search->objectid ||
554                     key.type != BTRFS_EXTENT_DATA_KEY)
555                         break;
556
557                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
558                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
559
560                 if (disk_byte == wanted_disk_byte) {
561                         eie = NULL;
562                         old = NULL;
563                         count++;
564                         if (extent_item_pos) {
565                                 ret = check_extent_in_eb(&key, eb, fi,
566                                                 *extent_item_pos,
567                                                 &eie);
568                                 if (ret < 0)
569                                         break;
570                         }
571                         if (ret > 0)
572                                 goto next;
573                         ret = ulist_add_merge_ptr(parents, eb->start,
574                                                   eie, (void **)&old, GFP_NOFS);
575                         if (ret < 0)
576                                 break;
577                         if (!ret && extent_item_pos) {
578                                 while (old->next)
579                                         old = old->next;
580                                 old->next = eie;
581                         }
582                         eie = NULL;
583                 }
584 next:
585                 if (time_seq == SEQ_LAST)
586                         ret = btrfs_next_item(root, path);
587                 else
588                         ret = btrfs_next_old_item(root, path, time_seq);
589         }
590
591         if (ret > 0)
592                 ret = 0;
593         else if (ret < 0)
594                 free_inode_elem_list(eie);
595         return ret;
596 }
597
598 /*
599  * resolve an indirect backref in the form (root_id, key, level)
600  * to a logical address
601  */
602 static int __resolve_indirect_ref(struct btrfs_fs_info *fs_info,
603                                   struct btrfs_path *path, u64 time_seq,
604                                   struct __prelim_ref *ref,
605                                   struct ulist *parents,
606                                   const u64 *extent_item_pos, u64 total_refs)
607 {
608         struct btrfs_root *root;
609         struct btrfs_key root_key;
610         struct extent_buffer *eb;
611         int ret = 0;
612         int root_level;
613         int level = ref->level;
614         int index;
615
616         root_key.objectid = ref->root_id;
617         root_key.type = BTRFS_ROOT_ITEM_KEY;
618         root_key.offset = (u64)-1;
619
620         index = srcu_read_lock(&fs_info->subvol_srcu);
621
622         root = btrfs_get_fs_root(fs_info, &root_key, false);
623         if (IS_ERR(root)) {
624                 srcu_read_unlock(&fs_info->subvol_srcu, index);
625                 ret = PTR_ERR(root);
626                 goto out;
627         }
628
629         if (btrfs_is_testing(fs_info)) {
630                 srcu_read_unlock(&fs_info->subvol_srcu, index);
631                 ret = -ENOENT;
632                 goto out;
633         }
634
635         if (path->search_commit_root)
636                 root_level = btrfs_header_level(root->commit_root);
637         else if (time_seq == SEQ_LAST)
638                 root_level = btrfs_header_level(root->node);
639         else
640                 root_level = btrfs_old_root_level(root, time_seq);
641
642         if (root_level + 1 == level) {
643                 srcu_read_unlock(&fs_info->subvol_srcu, index);
644                 goto out;
645         }
646
647         path->lowest_level = level;
648         if (time_seq == SEQ_LAST)
649                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
650                                         0, 0);
651         else
652                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
653                                             time_seq);
654
655         /* root node has been locked, we can release @subvol_srcu safely here */
656         srcu_read_unlock(&fs_info->subvol_srcu, index);
657
658         btrfs_debug(fs_info,
659                 "search slot in root %llu (level %d, ref count %d) returned %d for key (%llu %u %llu)",
660                  ref->root_id, level, ref->count, ret,
661                  ref->key_for_search.objectid, ref->key_for_search.type,
662                  ref->key_for_search.offset);
663         if (ret < 0)
664                 goto out;
665
666         eb = path->nodes[level];
667         while (!eb) {
668                 if (WARN_ON(!level)) {
669                         ret = 1;
670                         goto out;
671                 }
672                 level--;
673                 eb = path->nodes[level];
674         }
675
676         ret = add_all_parents(root, path, parents, ref, level, time_seq,
677                               extent_item_pos, total_refs);
678 out:
679         path->lowest_level = 0;
680         btrfs_release_path(path);
681         return ret;
682 }
683
684 /*
685  * resolve all indirect backrefs from the list
686  */
687 static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
688                                    struct btrfs_path *path, u64 time_seq,
689                                    struct list_head *head,
690                                    const u64 *extent_item_pos, u64 total_refs,
691                                    u64 root_objectid)
692 {
693         int err;
694         int ret = 0;
695         struct __prelim_ref *ref;
696         struct __prelim_ref *ref_safe;
697         struct __prelim_ref *new_ref;
698         struct ulist *parents;
699         struct ulist_node *node;
700         struct ulist_iterator uiter;
701
702         parents = ulist_alloc(GFP_NOFS);
703         if (!parents)
704                 return -ENOMEM;
705
706         /*
707          * _safe allows us to insert directly after the current item without
708          * iterating over the newly inserted items.
709          * we're also allowed to re-assign ref during iteration.
710          */
711         list_for_each_entry_safe(ref, ref_safe, head, list) {
712                 if (ref->parent)        /* already direct */
713                         continue;
714                 if (ref->count == 0)
715                         continue;
716                 if (root_objectid && ref->root_id != root_objectid) {
717                         ret = BACKREF_FOUND_SHARED;
718                         goto out;
719                 }
720                 err = __resolve_indirect_ref(fs_info, path, time_seq, ref,
721                                              parents, extent_item_pos,
722                                              total_refs);
723                 /*
724                  * we can only tolerate ENOENT,otherwise,we should catch error
725                  * and return directly.
726                  */
727                 if (err == -ENOENT) {
728                         continue;
729                 } else if (err) {
730                         ret = err;
731                         goto out;
732                 }
733
734                 /* we put the first parent into the ref at hand */
735                 ULIST_ITER_INIT(&uiter);
736                 node = ulist_next(parents, &uiter);
737                 ref->parent = node ? node->val : 0;
738                 ref->inode_list = node ?
739                         (struct extent_inode_elem *)(uintptr_t)node->aux : NULL;
740
741                 /* additional parents require new refs being added here */
742                 while ((node = ulist_next(parents, &uiter))) {
743                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
744                                                    GFP_NOFS);
745                         if (!new_ref) {
746                                 ret = -ENOMEM;
747                                 goto out;
748                         }
749                         memcpy(new_ref, ref, sizeof(*ref));
750                         new_ref->parent = node->val;
751                         new_ref->inode_list = (struct extent_inode_elem *)
752                                                         (uintptr_t)node->aux;
753                         list_add(&new_ref->list, &ref->list);
754                 }
755                 ulist_reinit(parents);
756         }
757 out:
758         ulist_free(parents);
759         return ret;
760 }
761
762 static inline int ref_for_same_block(struct __prelim_ref *ref1,
763                                      struct __prelim_ref *ref2)
764 {
765         if (ref1->level != ref2->level)
766                 return 0;
767         if (ref1->root_id != ref2->root_id)
768                 return 0;
769         if (ref1->key_for_search.type != ref2->key_for_search.type)
770                 return 0;
771         if (ref1->key_for_search.objectid != ref2->key_for_search.objectid)
772                 return 0;
773         if (ref1->key_for_search.offset != ref2->key_for_search.offset)
774                 return 0;
775         if (ref1->parent != ref2->parent)
776                 return 0;
777
778         return 1;
779 }
780
781 /*
782  * read tree blocks and add keys where required.
783  */
784 static int __add_missing_keys(struct btrfs_fs_info *fs_info,
785                               struct list_head *head)
786 {
787         struct __prelim_ref *ref;
788         struct extent_buffer *eb;
789
790         list_for_each_entry(ref, head, list) {
791                 if (ref->parent)
792                         continue;
793                 if (ref->key_for_search.type)
794                         continue;
795                 BUG_ON(!ref->wanted_disk_byte);
796                 eb = read_tree_block(fs_info, ref->wanted_disk_byte, 0);
797                 if (IS_ERR(eb)) {
798                         return PTR_ERR(eb);
799                 } else if (!extent_buffer_uptodate(eb)) {
800                         free_extent_buffer(eb);
801                         return -EIO;
802                 }
803                 btrfs_tree_read_lock(eb);
804                 if (btrfs_header_level(eb) == 0)
805                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
806                 else
807                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
808                 btrfs_tree_read_unlock(eb);
809                 free_extent_buffer(eb);
810         }
811         return 0;
812 }
813
814 /*
815  * merge backrefs and adjust counts accordingly
816  *
817  *    FIXME: For MERGE_IDENTICAL_KEYS, if we add more keys in __add_prelim_ref
818  *           then we can merge more here. Additionally, we could even add a key
819  *           range for the blocks we looked into to merge even more (-> replace
820  *           unresolved refs by those having a parent).
821  */
822 static void __merge_refs(struct list_head *head, enum merge_mode mode)
823 {
824         struct __prelim_ref *pos1;
825
826         list_for_each_entry(pos1, head, list) {
827                 struct __prelim_ref *pos2 = pos1, *tmp;
828
829                 list_for_each_entry_safe_continue(pos2, tmp, head, list) {
830                         struct __prelim_ref *ref1 = pos1, *ref2 = pos2;
831                         struct extent_inode_elem *eie;
832
833                         if (!ref_for_same_block(ref1, ref2))
834                                 continue;
835                         if (mode == MERGE_IDENTICAL_KEYS) {
836                                 if (!ref1->parent && ref2->parent)
837                                         swap(ref1, ref2);
838                         } else {
839                                 if (ref1->parent != ref2->parent)
840                                         continue;
841                         }
842
843                         eie = ref1->inode_list;
844                         while (eie && eie->next)
845                                 eie = eie->next;
846                         if (eie)
847                                 eie->next = ref2->inode_list;
848                         else
849                                 ref1->inode_list = ref2->inode_list;
850                         ref1->count += ref2->count;
851
852                         list_del(&ref2->list);
853                         kmem_cache_free(btrfs_prelim_ref_cache, ref2);
854                         cond_resched();
855                 }
856
857         }
858 }
859
860 /*
861  * add all currently queued delayed refs from this head whose seq nr is
862  * smaller or equal that seq to the list
863  */
864 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
865                               struct list_head *prefs, u64 *total_refs,
866                               u64 inum)
867 {
868         struct btrfs_delayed_ref_node *node;
869         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
870         struct btrfs_key key;
871         struct btrfs_key op_key = {0};
872         int sgn;
873         int ret = 0;
874
875         if (extent_op && extent_op->update_key)
876                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
877
878         spin_lock(&head->lock);
879         list_for_each_entry(node, &head->ref_list, list) {
880                 if (node->seq > seq)
881                         continue;
882
883                 switch (node->action) {
884                 case BTRFS_ADD_DELAYED_EXTENT:
885                 case BTRFS_UPDATE_DELAYED_HEAD:
886                         WARN_ON(1);
887                         continue;
888                 case BTRFS_ADD_DELAYED_REF:
889                         sgn = 1;
890                         break;
891                 case BTRFS_DROP_DELAYED_REF:
892                         sgn = -1;
893                         break;
894                 default:
895                         BUG_ON(1);
896                 }
897                 *total_refs += (node->ref_mod * sgn);
898                 switch (node->type) {
899                 case BTRFS_TREE_BLOCK_REF_KEY: {
900                         struct btrfs_delayed_tree_ref *ref;
901
902                         ref = btrfs_delayed_node_to_tree_ref(node);
903                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
904                                                ref->level + 1, 0, node->bytenr,
905                                                node->ref_mod * sgn, GFP_ATOMIC);
906                         break;
907                 }
908                 case BTRFS_SHARED_BLOCK_REF_KEY: {
909                         struct btrfs_delayed_tree_ref *ref;
910
911                         ref = btrfs_delayed_node_to_tree_ref(node);
912                         ret = __add_prelim_ref(prefs, 0, NULL,
913                                                ref->level + 1, ref->parent,
914                                                node->bytenr,
915                                                node->ref_mod * sgn, GFP_ATOMIC);
916                         break;
917                 }
918                 case BTRFS_EXTENT_DATA_REF_KEY: {
919                         struct btrfs_delayed_data_ref *ref;
920                         ref = btrfs_delayed_node_to_data_ref(node);
921
922                         key.objectid = ref->objectid;
923                         key.type = BTRFS_EXTENT_DATA_KEY;
924                         key.offset = ref->offset;
925
926                         /*
927                          * Found a inum that doesn't match our known inum, we
928                          * know it's shared.
929                          */
930                         if (inum && ref->objectid != inum) {
931                                 ret = BACKREF_FOUND_SHARED;
932                                 break;
933                         }
934
935                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
936                                                node->bytenr,
937                                                node->ref_mod * sgn, GFP_ATOMIC);
938                         break;
939                 }
940                 case BTRFS_SHARED_DATA_REF_KEY: {
941                         struct btrfs_delayed_data_ref *ref;
942
943                         ref = btrfs_delayed_node_to_data_ref(node);
944                         ret = __add_prelim_ref(prefs, 0, NULL, 0,
945                                                ref->parent, node->bytenr,
946                                                node->ref_mod * sgn, GFP_ATOMIC);
947                         break;
948                 }
949                 default:
950                         WARN_ON(1);
951                 }
952                 if (ret)
953                         break;
954         }
955         spin_unlock(&head->lock);
956         return ret;
957 }
958
959 /*
960  * add all inline backrefs for bytenr to the list
961  */
962 static int __add_inline_refs(struct btrfs_path *path, u64 bytenr,
963                              int *info_level, struct list_head *prefs,
964                              struct ref_root *ref_tree,
965                              u64 *total_refs, u64 inum)
966 {
967         int ret = 0;
968         int slot;
969         struct extent_buffer *leaf;
970         struct btrfs_key key;
971         struct btrfs_key found_key;
972         unsigned long ptr;
973         unsigned long end;
974         struct btrfs_extent_item *ei;
975         u64 flags;
976         u64 item_size;
977
978         /*
979          * enumerate all inline refs
980          */
981         leaf = path->nodes[0];
982         slot = path->slots[0];
983
984         item_size = btrfs_item_size_nr(leaf, slot);
985         BUG_ON(item_size < sizeof(*ei));
986
987         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
988         flags = btrfs_extent_flags(leaf, ei);
989         *total_refs += btrfs_extent_refs(leaf, ei);
990         btrfs_item_key_to_cpu(leaf, &found_key, slot);
991
992         ptr = (unsigned long)(ei + 1);
993         end = (unsigned long)ei + item_size;
994
995         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
996             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
997                 struct btrfs_tree_block_info *info;
998
999                 info = (struct btrfs_tree_block_info *)ptr;
1000                 *info_level = btrfs_tree_block_level(leaf, info);
1001                 ptr += sizeof(struct btrfs_tree_block_info);
1002                 BUG_ON(ptr > end);
1003         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1004                 *info_level = found_key.offset;
1005         } else {
1006                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1007         }
1008
1009         while (ptr < end) {
1010                 struct btrfs_extent_inline_ref *iref;
1011                 u64 offset;
1012                 int type;
1013
1014                 iref = (struct btrfs_extent_inline_ref *)ptr;
1015                 type = btrfs_extent_inline_ref_type(leaf, iref);
1016                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1017
1018                 switch (type) {
1019                 case BTRFS_SHARED_BLOCK_REF_KEY:
1020                         ret = __add_prelim_ref(prefs, 0, NULL,
1021                                                 *info_level + 1, offset,
1022                                                 bytenr, 1, GFP_NOFS);
1023                         break;
1024                 case BTRFS_SHARED_DATA_REF_KEY: {
1025                         struct btrfs_shared_data_ref *sdref;
1026                         int count;
1027
1028                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1029                         count = btrfs_shared_data_ref_count(leaf, sdref);
1030                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
1031                                                bytenr, count, GFP_NOFS);
1032                         if (ref_tree) {
1033                                 if (!ret)
1034                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1035                                                            bytenr, count);
1036                                 if (!ret && ref_tree->unique_refs > 1)
1037                                         ret = BACKREF_FOUND_SHARED;
1038                         }
1039                         break;
1040                 }
1041                 case BTRFS_TREE_BLOCK_REF_KEY:
1042                         ret = __add_prelim_ref(prefs, offset, NULL,
1043                                                *info_level + 1, 0,
1044                                                bytenr, 1, GFP_NOFS);
1045                         break;
1046                 case BTRFS_EXTENT_DATA_REF_KEY: {
1047                         struct btrfs_extent_data_ref *dref;
1048                         int count;
1049                         u64 root;
1050
1051                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1052                         count = btrfs_extent_data_ref_count(leaf, dref);
1053                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1054                                                                       dref);
1055                         key.type = BTRFS_EXTENT_DATA_KEY;
1056                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1057
1058                         if (inum && key.objectid != inum) {
1059                                 ret = BACKREF_FOUND_SHARED;
1060                                 break;
1061                         }
1062
1063                         root = btrfs_extent_data_ref_root(leaf, dref);
1064                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1065                                                bytenr, count, GFP_NOFS);
1066                         if (ref_tree) {
1067                                 if (!ret)
1068                                         ret = ref_tree_add(ref_tree, root,
1069                                                            key.objectid,
1070                                                            key.offset, 0,
1071                                                            count);
1072                                 if (!ret && ref_tree->unique_refs > 1)
1073                                         ret = BACKREF_FOUND_SHARED;
1074                         }
1075                         break;
1076                 }
1077                 default:
1078                         WARN_ON(1);
1079                 }
1080                 if (ret)
1081                         return ret;
1082                 ptr += btrfs_extent_inline_ref_size(type);
1083         }
1084
1085         return 0;
1086 }
1087
1088 /*
1089  * add all non-inline backrefs for bytenr to the list
1090  */
1091 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
1092                             struct btrfs_path *path, u64 bytenr,
1093                             int info_level, struct list_head *prefs,
1094                             struct ref_root *ref_tree, u64 inum)
1095 {
1096         struct btrfs_root *extent_root = fs_info->extent_root;
1097         int ret;
1098         int slot;
1099         struct extent_buffer *leaf;
1100         struct btrfs_key key;
1101
1102         while (1) {
1103                 ret = btrfs_next_item(extent_root, path);
1104                 if (ret < 0)
1105                         break;
1106                 if (ret) {
1107                         ret = 0;
1108                         break;
1109                 }
1110
1111                 slot = path->slots[0];
1112                 leaf = path->nodes[0];
1113                 btrfs_item_key_to_cpu(leaf, &key, slot);
1114
1115                 if (key.objectid != bytenr)
1116                         break;
1117                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1118                         continue;
1119                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1120                         break;
1121
1122                 switch (key.type) {
1123                 case BTRFS_SHARED_BLOCK_REF_KEY:
1124                         ret = __add_prelim_ref(prefs, 0, NULL,
1125                                                 info_level + 1, key.offset,
1126                                                 bytenr, 1, GFP_NOFS);
1127                         break;
1128                 case BTRFS_SHARED_DATA_REF_KEY: {
1129                         struct btrfs_shared_data_ref *sdref;
1130                         int count;
1131
1132                         sdref = btrfs_item_ptr(leaf, slot,
1133                                               struct btrfs_shared_data_ref);
1134                         count = btrfs_shared_data_ref_count(leaf, sdref);
1135                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
1136                                                 bytenr, count, GFP_NOFS);
1137                         if (ref_tree) {
1138                                 if (!ret)
1139                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1140                                                            bytenr, count);
1141                                 if (!ret && ref_tree->unique_refs > 1)
1142                                         ret = BACKREF_FOUND_SHARED;
1143                         }
1144                         break;
1145                 }
1146                 case BTRFS_TREE_BLOCK_REF_KEY:
1147                         ret = __add_prelim_ref(prefs, key.offset, NULL,
1148                                                info_level + 1, 0,
1149                                                bytenr, 1, GFP_NOFS);
1150                         break;
1151                 case BTRFS_EXTENT_DATA_REF_KEY: {
1152                         struct btrfs_extent_data_ref *dref;
1153                         int count;
1154                         u64 root;
1155
1156                         dref = btrfs_item_ptr(leaf, slot,
1157                                               struct btrfs_extent_data_ref);
1158                         count = btrfs_extent_data_ref_count(leaf, dref);
1159                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1160                                                                       dref);
1161                         key.type = BTRFS_EXTENT_DATA_KEY;
1162                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1163
1164                         if (inum && key.objectid != inum) {
1165                                 ret = BACKREF_FOUND_SHARED;
1166                                 break;
1167                         }
1168
1169                         root = btrfs_extent_data_ref_root(leaf, dref);
1170                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1171                                                bytenr, count, GFP_NOFS);
1172                         if (ref_tree) {
1173                                 if (!ret)
1174                                         ret = ref_tree_add(ref_tree, root,
1175                                                            key.objectid,
1176                                                            key.offset, 0,
1177                                                            count);
1178                                 if (!ret && ref_tree->unique_refs > 1)
1179                                         ret = BACKREF_FOUND_SHARED;
1180                         }
1181                         break;
1182                 }
1183                 default:
1184                         WARN_ON(1);
1185                 }
1186                 if (ret)
1187                         return ret;
1188
1189         }
1190
1191         return ret;
1192 }
1193
1194 /*
1195  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1196  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1197  * indirect refs to their parent bytenr.
1198  * When roots are found, they're added to the roots list
1199  *
1200  * NOTE: This can return values > 0
1201  *
1202  * If time_seq is set to SEQ_LAST, it will not search delayed_refs, and behave
1203  * much like trans == NULL case, the difference only lies in it will not
1204  * commit root.
1205  * The special case is for qgroup to search roots in commit_transaction().
1206  *
1207  * If check_shared is set to 1, any extent has more than one ref item, will
1208  * be returned BACKREF_FOUND_SHARED immediately.
1209  *
1210  * FIXME some caching might speed things up
1211  */
1212 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1213                              struct btrfs_fs_info *fs_info, u64 bytenr,
1214                              u64 time_seq, struct ulist *refs,
1215                              struct ulist *roots, const u64 *extent_item_pos,
1216                              u64 root_objectid, u64 inum, int check_shared)
1217 {
1218         struct btrfs_key key;
1219         struct btrfs_path *path;
1220         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1221         struct btrfs_delayed_ref_head *head;
1222         int info_level = 0;
1223         int ret;
1224         struct list_head prefs_delayed;
1225         struct list_head prefs;
1226         struct __prelim_ref *ref;
1227         struct extent_inode_elem *eie = NULL;
1228         struct ref_root *ref_tree = NULL;
1229         u64 total_refs = 0;
1230
1231         INIT_LIST_HEAD(&prefs);
1232         INIT_LIST_HEAD(&prefs_delayed);
1233
1234         key.objectid = bytenr;
1235         key.offset = (u64)-1;
1236         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1237                 key.type = BTRFS_METADATA_ITEM_KEY;
1238         else
1239                 key.type = BTRFS_EXTENT_ITEM_KEY;
1240
1241         path = btrfs_alloc_path();
1242         if (!path)
1243                 return -ENOMEM;
1244         if (!trans) {
1245                 path->search_commit_root = 1;
1246                 path->skip_locking = 1;
1247         }
1248
1249         if (time_seq == SEQ_LAST)
1250                 path->skip_locking = 1;
1251
1252         /*
1253          * grab both a lock on the path and a lock on the delayed ref head.
1254          * We need both to get a consistent picture of how the refs look
1255          * at a specified point in time
1256          */
1257 again:
1258         head = NULL;
1259
1260         if (check_shared) {
1261                 if (!ref_tree) {
1262                         ref_tree = ref_root_alloc();
1263                         if (!ref_tree) {
1264                                 ret = -ENOMEM;
1265                                 goto out;
1266                         }
1267                 } else {
1268                         ref_root_fini(ref_tree);
1269                 }
1270         }
1271
1272         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1273         if (ret < 0)
1274                 goto out;
1275         BUG_ON(ret == 0);
1276
1277 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1278         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1279             time_seq != SEQ_LAST) {
1280 #else
1281         if (trans && time_seq != SEQ_LAST) {
1282 #endif
1283                 /*
1284                  * look if there are updates for this ref queued and lock the
1285                  * head
1286                  */
1287                 delayed_refs = &trans->transaction->delayed_refs;
1288                 spin_lock(&delayed_refs->lock);
1289                 head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
1290                 if (head) {
1291                         if (!mutex_trylock(&head->mutex)) {
1292                                 refcount_inc(&head->node.refs);
1293                                 spin_unlock(&delayed_refs->lock);
1294
1295                                 btrfs_release_path(path);
1296
1297                                 /*
1298                                  * Mutex was contended, block until it's
1299                                  * released and try again
1300                                  */
1301                                 mutex_lock(&head->mutex);
1302                                 mutex_unlock(&head->mutex);
1303                                 btrfs_put_delayed_ref(&head->node);
1304                                 goto again;
1305                         }
1306                         spin_unlock(&delayed_refs->lock);
1307                         ret = __add_delayed_refs(head, time_seq,
1308                                                  &prefs_delayed, &total_refs,
1309                                                  inum);
1310                         mutex_unlock(&head->mutex);
1311                         if (ret)
1312                                 goto out;
1313                 } else {
1314                         spin_unlock(&delayed_refs->lock);
1315                 }
1316
1317                 if (check_shared && !list_empty(&prefs_delayed)) {
1318                         /*
1319                          * Add all delay_ref to the ref_tree and check if there
1320                          * are multiple ref items added.
1321                          */
1322                         list_for_each_entry(ref, &prefs_delayed, list) {
1323                                 if (ref->key_for_search.type) {
1324                                         ret = ref_tree_add(ref_tree,
1325                                                 ref->root_id,
1326                                                 ref->key_for_search.objectid,
1327                                                 ref->key_for_search.offset,
1328                                                 0, ref->count);
1329                                         if (ret)
1330                                                 goto out;
1331                                 } else {
1332                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1333                                                      ref->parent, ref->count);
1334                                         if (ret)
1335                                                 goto out;
1336                                 }
1337
1338                         }
1339
1340                         if (ref_tree->unique_refs > 1) {
1341                                 ret = BACKREF_FOUND_SHARED;
1342                                 goto out;
1343                         }
1344
1345                 }
1346         }
1347
1348         if (path->slots[0]) {
1349                 struct extent_buffer *leaf;
1350                 int slot;
1351
1352                 path->slots[0]--;
1353                 leaf = path->nodes[0];
1354                 slot = path->slots[0];
1355                 btrfs_item_key_to_cpu(leaf, &key, slot);
1356                 if (key.objectid == bytenr &&
1357                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1358                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1359                         ret = __add_inline_refs(path, bytenr,
1360                                                 &info_level, &prefs,
1361                                                 ref_tree, &total_refs,
1362                                                 inum);
1363                         if (ret)
1364                                 goto out;
1365                         ret = __add_keyed_refs(fs_info, path, bytenr,
1366                                                info_level, &prefs,
1367                                                ref_tree, inum);
1368                         if (ret)
1369                                 goto out;
1370                 }
1371         }
1372         btrfs_release_path(path);
1373
1374         list_splice_init(&prefs_delayed, &prefs);
1375
1376         ret = __add_missing_keys(fs_info, &prefs);
1377         if (ret)
1378                 goto out;
1379
1380         __merge_refs(&prefs, MERGE_IDENTICAL_KEYS);
1381
1382         ret = __resolve_indirect_refs(fs_info, path, time_seq, &prefs,
1383                                       extent_item_pos, total_refs,
1384                                       root_objectid);
1385         if (ret)
1386                 goto out;
1387
1388         __merge_refs(&prefs, MERGE_IDENTICAL_PARENTS);
1389
1390         while (!list_empty(&prefs)) {
1391                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1392                 WARN_ON(ref->count < 0);
1393                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1394                         if (root_objectid && ref->root_id != root_objectid) {
1395                                 ret = BACKREF_FOUND_SHARED;
1396                                 goto out;
1397                         }
1398
1399                         /* no parent == root of tree */
1400                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1401                         if (ret < 0)
1402                                 goto out;
1403                 }
1404                 if (ref->count && ref->parent) {
1405                         if (extent_item_pos && !ref->inode_list &&
1406                             ref->level == 0) {
1407                                 struct extent_buffer *eb;
1408
1409                                 eb = read_tree_block(fs_info, ref->parent, 0);
1410                                 if (IS_ERR(eb)) {
1411                                         ret = PTR_ERR(eb);
1412                                         goto out;
1413                                 } else if (!extent_buffer_uptodate(eb)) {
1414                                         free_extent_buffer(eb);
1415                                         ret = -EIO;
1416                                         goto out;
1417                                 }
1418                                 btrfs_tree_read_lock(eb);
1419                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1420                                 ret = find_extent_in_eb(eb, bytenr,
1421                                                         *extent_item_pos, &eie);
1422                                 btrfs_tree_read_unlock_blocking(eb);
1423                                 free_extent_buffer(eb);
1424                                 if (ret < 0)
1425                                         goto out;
1426                                 ref->inode_list = eie;
1427                         }
1428                         ret = ulist_add_merge_ptr(refs, ref->parent,
1429                                                   ref->inode_list,
1430                                                   (void **)&eie, GFP_NOFS);
1431                         if (ret < 0)
1432                                 goto out;
1433                         if (!ret && extent_item_pos) {
1434                                 /*
1435                                  * we've recorded that parent, so we must extend
1436                                  * its inode list here
1437                                  */
1438                                 BUG_ON(!eie);
1439                                 while (eie->next)
1440                                         eie = eie->next;
1441                                 eie->next = ref->inode_list;
1442                         }
1443                         eie = NULL;
1444                 }
1445                 list_del(&ref->list);
1446                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1447         }
1448
1449 out:
1450         btrfs_free_path(path);
1451         ref_root_free(ref_tree);
1452         while (!list_empty(&prefs)) {
1453                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1454                 list_del(&ref->list);
1455                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1456         }
1457         while (!list_empty(&prefs_delayed)) {
1458                 ref = list_first_entry(&prefs_delayed, struct __prelim_ref,
1459                                        list);
1460                 list_del(&ref->list);
1461                 kmem_cache_free(btrfs_prelim_ref_cache, ref);
1462         }
1463         if (ret < 0)
1464                 free_inode_elem_list(eie);
1465         return ret;
1466 }
1467
1468 static void free_leaf_list(struct ulist *blocks)
1469 {
1470         struct ulist_node *node = NULL;
1471         struct extent_inode_elem *eie;
1472         struct ulist_iterator uiter;
1473
1474         ULIST_ITER_INIT(&uiter);
1475         while ((node = ulist_next(blocks, &uiter))) {
1476                 if (!node->aux)
1477                         continue;
1478                 eie = (struct extent_inode_elem *)(uintptr_t)node->aux;
1479                 free_inode_elem_list(eie);
1480                 node->aux = 0;
1481         }
1482
1483         ulist_free(blocks);
1484 }
1485
1486 /*
1487  * Finds all leafs with a reference to the specified combination of bytenr and
1488  * offset. key_list_head will point to a list of corresponding keys (caller must
1489  * free each list element). The leafs will be stored in the leafs ulist, which
1490  * must be freed with ulist_free.
1491  *
1492  * returns 0 on success, <0 on error
1493  */
1494 static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
1495                                 struct btrfs_fs_info *fs_info, u64 bytenr,
1496                                 u64 time_seq, struct ulist **leafs,
1497                                 const u64 *extent_item_pos)
1498 {
1499         int ret;
1500
1501         *leafs = ulist_alloc(GFP_NOFS);
1502         if (!*leafs)
1503                 return -ENOMEM;
1504
1505         ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1506                                 *leafs, NULL, extent_item_pos, 0, 0, 0);
1507         if (ret < 0 && ret != -ENOENT) {
1508                 free_leaf_list(*leafs);
1509                 return ret;
1510         }
1511
1512         return 0;
1513 }
1514
1515 /*
1516  * walk all backrefs for a given extent to find all roots that reference this
1517  * extent. Walking a backref means finding all extents that reference this
1518  * extent and in turn walk the backrefs of those, too. Naturally this is a
1519  * recursive process, but here it is implemented in an iterative fashion: We
1520  * find all referencing extents for the extent in question and put them on a
1521  * list. In turn, we find all referencing extents for those, further appending
1522  * to the list. The way we iterate the list allows adding more elements after
1523  * the current while iterating. The process stops when we reach the end of the
1524  * list. Found roots are added to the roots list.
1525  *
1526  * returns 0 on success, < 0 on error.
1527  */
1528 static int __btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1529                                   struct btrfs_fs_info *fs_info, u64 bytenr,
1530                                   u64 time_seq, struct ulist **roots)
1531 {
1532         struct ulist *tmp;
1533         struct ulist_node *node = NULL;
1534         struct ulist_iterator uiter;
1535         int ret;
1536
1537         tmp = ulist_alloc(GFP_NOFS);
1538         if (!tmp)
1539                 return -ENOMEM;
1540         *roots = ulist_alloc(GFP_NOFS);
1541         if (!*roots) {
1542                 ulist_free(tmp);
1543                 return -ENOMEM;
1544         }
1545
1546         ULIST_ITER_INIT(&uiter);
1547         while (1) {
1548                 ret = find_parent_nodes(trans, fs_info, bytenr, time_seq,
1549                                         tmp, *roots, NULL, 0, 0, 0);
1550                 if (ret < 0 && ret != -ENOENT) {
1551                         ulist_free(tmp);
1552                         ulist_free(*roots);
1553                         return ret;
1554                 }
1555                 node = ulist_next(tmp, &uiter);
1556                 if (!node)
1557                         break;
1558                 bytenr = node->val;
1559                 cond_resched();
1560         }
1561
1562         ulist_free(tmp);
1563         return 0;
1564 }
1565
1566 int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
1567                          struct btrfs_fs_info *fs_info, u64 bytenr,
1568                          u64 time_seq, struct ulist **roots)
1569 {
1570         int ret;
1571
1572         if (!trans)
1573                 down_read(&fs_info->commit_root_sem);
1574         ret = __btrfs_find_all_roots(trans, fs_info, bytenr, time_seq, roots);
1575         if (!trans)
1576                 up_read(&fs_info->commit_root_sem);
1577         return ret;
1578 }
1579
1580 /**
1581  * btrfs_check_shared - tell us whether an extent is shared
1582  *
1583  * @trans: optional trans handle
1584  *
1585  * btrfs_check_shared uses the backref walking code but will short
1586  * circuit as soon as it finds a root or inode that doesn't match the
1587  * one passed in. This provides a significant performance benefit for
1588  * callers (such as fiemap) which want to know whether the extent is
1589  * shared but do not need a ref count.
1590  *
1591  * Return: 0 if extent is not shared, 1 if it is shared, < 0 on error.
1592  */
1593 int btrfs_check_shared(struct btrfs_trans_handle *trans,
1594                        struct btrfs_fs_info *fs_info, u64 root_objectid,
1595                        u64 inum, u64 bytenr)
1596 {
1597         struct ulist *tmp = NULL;
1598         struct ulist *roots = NULL;
1599         struct ulist_iterator uiter;
1600         struct ulist_node *node;
1601         struct seq_list elem = SEQ_LIST_INIT(elem);
1602         int ret = 0;
1603
1604         tmp = ulist_alloc(GFP_NOFS);
1605         roots = ulist_alloc(GFP_NOFS);
1606         if (!tmp || !roots) {
1607                 ulist_free(tmp);
1608                 ulist_free(roots);
1609                 return -ENOMEM;
1610         }
1611
1612         if (trans)
1613                 btrfs_get_tree_mod_seq(fs_info, &elem);
1614         else
1615                 down_read(&fs_info->commit_root_sem);
1616         ULIST_ITER_INIT(&uiter);
1617         while (1) {
1618                 ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
1619                                         roots, NULL, root_objectid, inum, 1);
1620                 if (ret == BACKREF_FOUND_SHARED) {
1621                         /* this is the only condition under which we return 1 */
1622                         ret = 1;
1623                         break;
1624                 }
1625                 if (ret < 0 && ret != -ENOENT)
1626                         break;
1627                 ret = 0;
1628                 node = ulist_next(tmp, &uiter);
1629                 if (!node)
1630                         break;
1631                 bytenr = node->val;
1632                 cond_resched();
1633         }
1634         if (trans)
1635                 btrfs_put_tree_mod_seq(fs_info, &elem);
1636         else
1637                 up_read(&fs_info->commit_root_sem);
1638         ulist_free(tmp);
1639         ulist_free(roots);
1640         return ret;
1641 }
1642
1643 int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
1644                           u64 start_off, struct btrfs_path *path,
1645                           struct btrfs_inode_extref **ret_extref,
1646                           u64 *found_off)
1647 {
1648         int ret, slot;
1649         struct btrfs_key key;
1650         struct btrfs_key found_key;
1651         struct btrfs_inode_extref *extref;
1652         struct extent_buffer *leaf;
1653         unsigned long ptr;
1654
1655         key.objectid = inode_objectid;
1656         key.type = BTRFS_INODE_EXTREF_KEY;
1657         key.offset = start_off;
1658
1659         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1660         if (ret < 0)
1661                 return ret;
1662
1663         while (1) {
1664                 leaf = path->nodes[0];
1665                 slot = path->slots[0];
1666                 if (slot >= btrfs_header_nritems(leaf)) {
1667                         /*
1668                          * If the item at offset is not found,
1669                          * btrfs_search_slot will point us to the slot
1670                          * where it should be inserted. In our case
1671                          * that will be the slot directly before the
1672                          * next INODE_REF_KEY_V2 item. In the case
1673                          * that we're pointing to the last slot in a
1674                          * leaf, we must move one leaf over.
1675                          */
1676                         ret = btrfs_next_leaf(root, path);
1677                         if (ret) {
1678                                 if (ret >= 1)
1679                                         ret = -ENOENT;
1680                                 break;
1681                         }
1682                         continue;
1683                 }
1684
1685                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1686
1687                 /*
1688                  * Check that we're still looking at an extended ref key for
1689                  * this particular objectid. If we have different
1690                  * objectid or type then there are no more to be found
1691                  * in the tree and we can exit.
1692                  */
1693                 ret = -ENOENT;
1694                 if (found_key.objectid != inode_objectid)
1695                         break;
1696                 if (found_key.type != BTRFS_INODE_EXTREF_KEY)
1697                         break;
1698
1699                 ret = 0;
1700                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1701                 extref = (struct btrfs_inode_extref *)ptr;
1702                 *ret_extref = extref;
1703                 if (found_off)
1704                         *found_off = found_key.offset;
1705                 break;
1706         }
1707
1708         return ret;
1709 }
1710
1711 /*
1712  * this iterates to turn a name (from iref/extref) into a full filesystem path.
1713  * Elements of the path are separated by '/' and the path is guaranteed to be
1714  * 0-terminated. the path is only given within the current file system.
1715  * Therefore, it never starts with a '/'. the caller is responsible to provide
1716  * "size" bytes in "dest". the dest buffer will be filled backwards. finally,
1717  * the start point of the resulting string is returned. this pointer is within
1718  * dest, normally.
1719  * in case the path buffer would overflow, the pointer is decremented further
1720  * as if output was written to the buffer, though no more output is actually
1721  * generated. that way, the caller can determine how much space would be
1722  * required for the path to fit into the buffer. in that case, the returned
1723  * value will be smaller than dest. callers must check this!
1724  */
1725 char *btrfs_ref_to_path(struct btrfs_root *fs_root, struct btrfs_path *path,
1726                         u32 name_len, unsigned long name_off,
1727                         struct extent_buffer *eb_in, u64 parent,
1728                         char *dest, u32 size)
1729 {
1730         int slot;
1731         u64 next_inum;
1732         int ret;
1733         s64 bytes_left = ((s64)size) - 1;
1734         struct extent_buffer *eb = eb_in;
1735         struct btrfs_key found_key;
1736         int leave_spinning = path->leave_spinning;
1737         struct btrfs_inode_ref *iref;
1738
1739         if (bytes_left >= 0)
1740                 dest[bytes_left] = '\0';
1741
1742         path->leave_spinning = 1;
1743         while (1) {
1744                 bytes_left -= name_len;
1745                 if (bytes_left >= 0)
1746                         read_extent_buffer(eb, dest + bytes_left,
1747                                            name_off, name_len);
1748                 if (eb != eb_in) {
1749                         if (!path->skip_locking)
1750                                 btrfs_tree_read_unlock_blocking(eb);
1751                         free_extent_buffer(eb);
1752                 }
1753                 ret = btrfs_find_item(fs_root, path, parent, 0,
1754                                 BTRFS_INODE_REF_KEY, &found_key);
1755                 if (ret > 0)
1756                         ret = -ENOENT;
1757                 if (ret)
1758                         break;
1759
1760                 next_inum = found_key.offset;
1761
1762                 /* regular exit ahead */
1763                 if (parent == next_inum)
1764                         break;
1765
1766                 slot = path->slots[0];
1767                 eb = path->nodes[0];
1768                 /* make sure we can use eb after releasing the path */
1769                 if (eb != eb_in) {
1770                         if (!path->skip_locking)
1771                                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1772                         path->nodes[0] = NULL;
1773                         path->locks[0] = 0;
1774                 }
1775                 btrfs_release_path(path);
1776                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
1777
1778                 name_len = btrfs_inode_ref_name_len(eb, iref);
1779                 name_off = (unsigned long)(iref + 1);
1780
1781                 parent = next_inum;
1782                 --bytes_left;
1783                 if (bytes_left >= 0)
1784                         dest[bytes_left] = '/';
1785         }
1786
1787         btrfs_release_path(path);
1788         path->leave_spinning = leave_spinning;
1789
1790         if (ret)
1791                 return ERR_PTR(ret);
1792
1793         return dest + bytes_left;
1794 }
1795
1796 /*
1797  * this makes the path point to (logical EXTENT_ITEM *)
1798  * returns BTRFS_EXTENT_FLAG_DATA for data, BTRFS_EXTENT_FLAG_TREE_BLOCK for
1799  * tree blocks and <0 on error.
1800  */
1801 int extent_from_logical(struct btrfs_fs_info *fs_info, u64 logical,
1802                         struct btrfs_path *path, struct btrfs_key *found_key,
1803                         u64 *flags_ret)
1804 {
1805         int ret;
1806         u64 flags;
1807         u64 size = 0;
1808         u32 item_size;
1809         struct extent_buffer *eb;
1810         struct btrfs_extent_item *ei;
1811         struct btrfs_key key;
1812
1813         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1814                 key.type = BTRFS_METADATA_ITEM_KEY;
1815         else
1816                 key.type = BTRFS_EXTENT_ITEM_KEY;
1817         key.objectid = logical;
1818         key.offset = (u64)-1;
1819
1820         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
1821         if (ret < 0)
1822                 return ret;
1823
1824         ret = btrfs_previous_extent_item(fs_info->extent_root, path, 0);
1825         if (ret) {
1826                 if (ret > 0)
1827                         ret = -ENOENT;
1828                 return ret;
1829         }
1830         btrfs_item_key_to_cpu(path->nodes[0], found_key, path->slots[0]);
1831         if (found_key->type == BTRFS_METADATA_ITEM_KEY)
1832                 size = fs_info->nodesize;
1833         else if (found_key->type == BTRFS_EXTENT_ITEM_KEY)
1834                 size = found_key->offset;
1835
1836         if (found_key->objectid > logical ||
1837             found_key->objectid + size <= logical) {
1838                 btrfs_debug(fs_info,
1839                         "logical %llu is not within any extent", logical);
1840                 return -ENOENT;
1841         }
1842
1843         eb = path->nodes[0];
1844         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1845         BUG_ON(item_size < sizeof(*ei));
1846
1847         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1848         flags = btrfs_extent_flags(eb, ei);
1849
1850         btrfs_debug(fs_info,
1851                 "logical %llu is at position %llu within the extent (%llu EXTENT_ITEM %llu) flags %#llx size %u",
1852                  logical, logical - found_key->objectid, found_key->objectid,
1853                  found_key->offset, flags, item_size);
1854
1855         WARN_ON(!flags_ret);
1856         if (flags_ret) {
1857                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1858                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1859                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1860                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1861                 else
1862                         BUG_ON(1);
1863                 return 0;
1864         }
1865
1866         return -EIO;
1867 }
1868
1869 /*
1870  * helper function to iterate extent inline refs. ptr must point to a 0 value
1871  * for the first call and may be modified. it is used to track state.
1872  * if more refs exist, 0 is returned and the next call to
1873  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1874  * next ref. after the last ref was processed, 1 is returned.
1875  * returns <0 on error
1876  */
1877 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1878                                    struct btrfs_key *key,
1879                                    struct btrfs_extent_item *ei, u32 item_size,
1880                                    struct btrfs_extent_inline_ref **out_eiref,
1881                                    int *out_type)
1882 {
1883         unsigned long end;
1884         u64 flags;
1885         struct btrfs_tree_block_info *info;
1886
1887         if (!*ptr) {
1888                 /* first call */
1889                 flags = btrfs_extent_flags(eb, ei);
1890                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1891                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1892                                 /* a skinny metadata extent */
1893                                 *out_eiref =
1894                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1895                         } else {
1896                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1897                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1898                                 *out_eiref =
1899                                    (struct btrfs_extent_inline_ref *)(info + 1);
1900                         }
1901                 } else {
1902                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1903                 }
1904                 *ptr = (unsigned long)*out_eiref;
1905                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1906                         return -ENOENT;
1907         }
1908
1909         end = (unsigned long)ei + item_size;
1910         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1911         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1912
1913         *ptr += btrfs_extent_inline_ref_size(*out_type);
1914         WARN_ON(*ptr > end);
1915         if (*ptr == end)
1916                 return 1; /* last */
1917
1918         return 0;
1919 }
1920
1921 /*
1922  * reads the tree block backref for an extent. tree level and root are returned
1923  * through out_level and out_root. ptr must point to a 0 value for the first
1924  * call and may be modified (see __get_extent_inline_ref comment).
1925  * returns 0 if data was provided, 1 if there was no more data to provide or
1926  * <0 on error.
1927  */
1928 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1929                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1930                             u32 item_size, u64 *out_root, u8 *out_level)
1931 {
1932         int ret;
1933         int type;
1934         struct btrfs_extent_inline_ref *eiref;
1935
1936         if (*ptr == (unsigned long)-1)
1937                 return 1;
1938
1939         while (1) {
1940                 ret = __get_extent_inline_ref(ptr, eb, key, ei, item_size,
1941                                               &eiref, &type);
1942                 if (ret < 0)
1943                         return ret;
1944
1945                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1946                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1947                         break;
1948
1949                 if (ret == 1)
1950                         return 1;
1951         }
1952
1953         /* we can treat both ref types equally here */
1954         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1955
1956         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1957                 struct btrfs_tree_block_info *info;
1958
1959                 info = (struct btrfs_tree_block_info *)(ei + 1);
1960                 *out_level = btrfs_tree_block_level(eb, info);
1961         } else {
1962                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1963                 *out_level = (u8)key->offset;
1964         }
1965
1966         if (ret == 1)
1967                 *ptr = (unsigned long)-1;
1968
1969         return 0;
1970 }
1971
1972 static int iterate_leaf_refs(struct btrfs_fs_info *fs_info,
1973                              struct extent_inode_elem *inode_list,
1974                              u64 root, u64 extent_item_objectid,
1975                              iterate_extent_inodes_t *iterate, void *ctx)
1976 {
1977         struct extent_inode_elem *eie;
1978         int ret = 0;
1979
1980         for (eie = inode_list; eie; eie = eie->next) {
1981                 btrfs_debug(fs_info,
1982                             "ref for %llu resolved, key (%llu EXTEND_DATA %llu), root %llu",
1983                             extent_item_objectid, eie->inum,
1984                             eie->offset, root);
1985                 ret = iterate(eie->inum, eie->offset, root, ctx);
1986                 if (ret) {
1987                         btrfs_debug(fs_info,
1988                                     "stopping iteration for %llu due to ret=%d",
1989                                     extent_item_objectid, ret);
1990                         break;
1991                 }
1992         }
1993
1994         return ret;
1995 }
1996
1997 /*
1998  * calls iterate() for every inode that references the extent identified by
1999  * the given parameters.
2000  * when the iterator function returns a non-zero value, iteration stops.
2001  */
2002 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
2003                                 u64 extent_item_objectid, u64 extent_item_pos,
2004                                 int search_commit_root,
2005                                 iterate_extent_inodes_t *iterate, void *ctx)
2006 {
2007         int ret;
2008         struct btrfs_trans_handle *trans = NULL;
2009         struct ulist *refs = NULL;
2010         struct ulist *roots = NULL;
2011         struct ulist_node *ref_node = NULL;
2012         struct ulist_node *root_node = NULL;
2013         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
2014         struct ulist_iterator ref_uiter;
2015         struct ulist_iterator root_uiter;
2016
2017         btrfs_debug(fs_info, "resolving all inodes for extent %llu",
2018                         extent_item_objectid);
2019
2020         if (!search_commit_root) {
2021                 trans = btrfs_join_transaction(fs_info->extent_root);
2022                 if (IS_ERR(trans))
2023                         return PTR_ERR(trans);
2024                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2025         } else {
2026                 down_read(&fs_info->commit_root_sem);
2027         }
2028
2029         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2030                                    tree_mod_seq_elem.seq, &refs,
2031                                    &extent_item_pos);
2032         if (ret)
2033                 goto out;
2034
2035         ULIST_ITER_INIT(&ref_uiter);
2036         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2037                 ret = __btrfs_find_all_roots(trans, fs_info, ref_node->val,
2038                                              tree_mod_seq_elem.seq, &roots);
2039                 if (ret)
2040                         break;
2041                 ULIST_ITER_INIT(&root_uiter);
2042                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2043                         btrfs_debug(fs_info,
2044                                     "root %llu references leaf %llu, data list %#llx",
2045                                     root_node->val, ref_node->val,
2046                                     ref_node->aux);
2047                         ret = iterate_leaf_refs(fs_info,
2048                                                 (struct extent_inode_elem *)
2049                                                 (uintptr_t)ref_node->aux,
2050                                                 root_node->val,
2051                                                 extent_item_objectid,
2052                                                 iterate, ctx);
2053                 }
2054                 ulist_free(roots);
2055         }
2056
2057         free_leaf_list(refs);
2058 out:
2059         if (!search_commit_root) {
2060                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2061                 btrfs_end_transaction(trans);
2062         } else {
2063                 up_read(&fs_info->commit_root_sem);
2064         }
2065
2066         return ret;
2067 }
2068
2069 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2070                                 struct btrfs_path *path,
2071                                 iterate_extent_inodes_t *iterate, void *ctx)
2072 {
2073         int ret;
2074         u64 extent_item_pos;
2075         u64 flags = 0;
2076         struct btrfs_key found_key;
2077         int search_commit_root = path->search_commit_root;
2078
2079         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2080         btrfs_release_path(path);
2081         if (ret < 0)
2082                 return ret;
2083         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2084                 return -EINVAL;
2085
2086         extent_item_pos = logical - found_key.objectid;
2087         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2088                                         extent_item_pos, search_commit_root,
2089                                         iterate, ctx);
2090
2091         return ret;
2092 }
2093
2094 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2095                               struct extent_buffer *eb, void *ctx);
2096
2097 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2098                               struct btrfs_path *path,
2099                               iterate_irefs_t *iterate, void *ctx)
2100 {
2101         int ret = 0;
2102         int slot;
2103         u32 cur;
2104         u32 len;
2105         u32 name_len;
2106         u64 parent = 0;
2107         int found = 0;
2108         struct extent_buffer *eb;
2109         struct btrfs_item *item;
2110         struct btrfs_inode_ref *iref;
2111         struct btrfs_key found_key;
2112
2113         while (!ret) {
2114                 ret = btrfs_find_item(fs_root, path, inum,
2115                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2116                                 &found_key);
2117
2118                 if (ret < 0)
2119                         break;
2120                 if (ret) {
2121                         ret = found ? 0 : -ENOENT;
2122                         break;
2123                 }
2124                 ++found;
2125
2126                 parent = found_key.offset;
2127                 slot = path->slots[0];
2128                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2129                 if (!eb) {
2130                         ret = -ENOMEM;
2131                         break;
2132                 }
2133                 extent_buffer_get(eb);
2134                 btrfs_tree_read_lock(eb);
2135                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2136                 btrfs_release_path(path);
2137
2138                 item = btrfs_item_nr(slot);
2139                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2140
2141                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2142                         name_len = btrfs_inode_ref_name_len(eb, iref);
2143                         /* path must be released before calling iterate()! */
2144                         btrfs_debug(fs_root->fs_info,
2145                                 "following ref at offset %u for inode %llu in tree %llu",
2146                                 cur, found_key.objectid, fs_root->objectid);
2147                         ret = iterate(parent, name_len,
2148                                       (unsigned long)(iref + 1), eb, ctx);
2149                         if (ret)
2150                                 break;
2151                         len = sizeof(*iref) + name_len;
2152                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2153                 }
2154                 btrfs_tree_read_unlock_blocking(eb);
2155                 free_extent_buffer(eb);
2156         }
2157
2158         btrfs_release_path(path);
2159
2160         return ret;
2161 }
2162
2163 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2164                                  struct btrfs_path *path,
2165                                  iterate_irefs_t *iterate, void *ctx)
2166 {
2167         int ret;
2168         int slot;
2169         u64 offset = 0;
2170         u64 parent;
2171         int found = 0;
2172         struct extent_buffer *eb;
2173         struct btrfs_inode_extref *extref;
2174         u32 item_size;
2175         u32 cur_offset;
2176         unsigned long ptr;
2177
2178         while (1) {
2179                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2180                                             &offset);
2181                 if (ret < 0)
2182                         break;
2183                 if (ret) {
2184                         ret = found ? 0 : -ENOENT;
2185                         break;
2186                 }
2187                 ++found;
2188
2189                 slot = path->slots[0];
2190                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2191                 if (!eb) {
2192                         ret = -ENOMEM;
2193                         break;
2194                 }
2195                 extent_buffer_get(eb);
2196
2197                 btrfs_tree_read_lock(eb);
2198                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2199                 btrfs_release_path(path);
2200
2201                 item_size = btrfs_item_size_nr(eb, slot);
2202                 ptr = btrfs_item_ptr_offset(eb, slot);
2203                 cur_offset = 0;
2204
2205                 while (cur_offset < item_size) {
2206                         u32 name_len;
2207
2208                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2209                         parent = btrfs_inode_extref_parent(eb, extref);
2210                         name_len = btrfs_inode_extref_name_len(eb, extref);
2211                         ret = iterate(parent, name_len,
2212                                       (unsigned long)&extref->name, eb, ctx);
2213                         if (ret)
2214                                 break;
2215
2216                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2217                         cur_offset += sizeof(*extref);
2218                 }
2219                 btrfs_tree_read_unlock_blocking(eb);
2220                 free_extent_buffer(eb);
2221
2222                 offset++;
2223         }
2224
2225         btrfs_release_path(path);
2226
2227         return ret;
2228 }
2229
2230 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2231                          struct btrfs_path *path, iterate_irefs_t *iterate,
2232                          void *ctx)
2233 {
2234         int ret;
2235         int found_refs = 0;
2236
2237         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2238         if (!ret)
2239                 ++found_refs;
2240         else if (ret != -ENOENT)
2241                 return ret;
2242
2243         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2244         if (ret == -ENOENT && found_refs)
2245                 return 0;
2246
2247         return ret;
2248 }
2249
2250 /*
2251  * returns 0 if the path could be dumped (probably truncated)
2252  * returns <0 in case of an error
2253  */
2254 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2255                          struct extent_buffer *eb, void *ctx)
2256 {
2257         struct inode_fs_paths *ipath = ctx;
2258         char *fspath;
2259         char *fspath_min;
2260         int i = ipath->fspath->elem_cnt;
2261         const int s_ptr = sizeof(char *);
2262         u32 bytes_left;
2263
2264         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2265                                         ipath->fspath->bytes_left - s_ptr : 0;
2266
2267         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2268         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2269                                    name_off, eb, inum, fspath_min, bytes_left);
2270         if (IS_ERR(fspath))
2271                 return PTR_ERR(fspath);
2272
2273         if (fspath > fspath_min) {
2274                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2275                 ++ipath->fspath->elem_cnt;
2276                 ipath->fspath->bytes_left = fspath - fspath_min;
2277         } else {
2278                 ++ipath->fspath->elem_missed;
2279                 ipath->fspath->bytes_missing += fspath_min - fspath;
2280                 ipath->fspath->bytes_left = 0;
2281         }
2282
2283         return 0;
2284 }
2285
2286 /*
2287  * this dumps all file system paths to the inode into the ipath struct, provided
2288  * is has been created large enough. each path is zero-terminated and accessed
2289  * from ipath->fspath->val[i].
2290  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2291  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2292  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2293  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2294  * have been needed to return all paths.
2295  */
2296 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2297 {
2298         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2299                              inode_to_path, ipath);
2300 }
2301
2302 struct btrfs_data_container *init_data_container(u32 total_bytes)
2303 {
2304         struct btrfs_data_container *data;
2305         size_t alloc_bytes;
2306
2307         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2308         data = kvmalloc(alloc_bytes, GFP_KERNEL);
2309         if (!data)
2310                 return ERR_PTR(-ENOMEM);
2311
2312         if (total_bytes >= sizeof(*data)) {
2313                 data->bytes_left = total_bytes - sizeof(*data);
2314                 data->bytes_missing = 0;
2315         } else {
2316                 data->bytes_missing = sizeof(*data) - total_bytes;
2317                 data->bytes_left = 0;
2318         }
2319
2320         data->elem_cnt = 0;
2321         data->elem_missed = 0;
2322
2323         return data;
2324 }
2325
2326 /*
2327  * allocates space to return multiple file system paths for an inode.
2328  * total_bytes to allocate are passed, note that space usable for actual path
2329  * information will be total_bytes - sizeof(struct inode_fs_paths).
2330  * the returned pointer must be freed with free_ipath() in the end.
2331  */
2332 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2333                                         struct btrfs_path *path)
2334 {
2335         struct inode_fs_paths *ifp;
2336         struct btrfs_data_container *fspath;
2337
2338         fspath = init_data_container(total_bytes);
2339         if (IS_ERR(fspath))
2340                 return (void *)fspath;
2341
2342         ifp = kmalloc(sizeof(*ifp), GFP_KERNEL);
2343         if (!ifp) {
2344                 kvfree(fspath);
2345                 return ERR_PTR(-ENOMEM);
2346         }
2347
2348         ifp->btrfs_path = path;
2349         ifp->fspath = fspath;
2350         ifp->fs_root = fs_root;
2351
2352         return ifp;
2353 }
2354
2355 void free_ipath(struct inode_fs_paths *ipath)
2356 {
2357         if (!ipath)
2358                 return;
2359         kvfree(ipath->fspath);
2360         kfree(ipath);
2361 }