d9688adc763fb77dcccbe34d605464ff60574309
[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/vmalloc.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 /* Just an arbitrary number so we can be sure this happened */
30 #define BACKREF_FOUND_SHARED 6
31
32 struct extent_inode_elem {
33         u64 inum;
34         u64 offset;
35         struct extent_inode_elem *next;
36 };
37
38 /*
39  * ref_root is used as the root of the ref tree that hold a collection
40  * of unique references.
41  */
42 struct ref_root {
43         struct rb_root rb_root;
44
45         /*
46          * The unique_refs represents the number of ref_nodes with a positive
47          * count stored in the tree. Even if a ref_node (the count is greater
48          * than one) is added, the unique_refs will only increase by one.
49          */
50         unsigned int unique_refs;
51 };
52
53 /* ref_node is used to store a unique reference to the ref tree. */
54 struct ref_node {
55         struct rb_node rb_node;
56
57         /* For NORMAL_REF, otherwise all these fields should be set to 0 */
58         u64 root_id;
59         u64 object_id;
60         u64 offset;
61
62         /* For SHARED_REF, otherwise parent field should be set to 0 */
63         u64 parent;
64
65         /* Ref to the ref_mod of btrfs_delayed_ref_node */
66         int ref_mod;
67 };
68
69 /* Dynamically allocate and initialize a ref_root */
70 static struct ref_root *ref_root_alloc(void)
71 {
72         struct ref_root *ref_tree;
73
74         ref_tree = kmalloc(sizeof(*ref_tree), GFP_NOFS);
75         if (!ref_tree)
76                 return NULL;
77
78         ref_tree->rb_root = RB_ROOT;
79         ref_tree->unique_refs = 0;
80
81         return ref_tree;
82 }
83
84 /* Free all nodes in the ref tree, and reinit ref_root */
85 static void ref_root_fini(struct ref_root *ref_tree)
86 {
87         struct ref_node *node;
88         struct rb_node *next;
89
90         while ((next = rb_first(&ref_tree->rb_root)) != NULL) {
91                 node = rb_entry(next, struct ref_node, rb_node);
92                 rb_erase(next, &ref_tree->rb_root);
93                 kfree(node);
94         }
95
96         ref_tree->rb_root = RB_ROOT;
97         ref_tree->unique_refs = 0;
98 }
99
100 static void ref_root_free(struct ref_root *ref_tree)
101 {
102         if (!ref_tree)
103                 return;
104
105         ref_root_fini(ref_tree);
106         kfree(ref_tree);
107 }
108
109 /*
110  * Compare ref_node with (root_id, object_id, offset, parent)
111  *
112  * The function compares two ref_node a and b. It returns an integer less
113  * than, equal to, or greater than zero , respectively, to be less than, to
114  * equal, or be greater than b.
115  */
116 static int ref_node_cmp(struct ref_node *a, struct ref_node *b)
117 {
118         if (a->root_id < b->root_id)
119                 return -1;
120         else if (a->root_id > b->root_id)
121                 return 1;
122
123         if (a->object_id < b->object_id)
124                 return -1;
125         else if (a->object_id > b->object_id)
126                 return 1;
127
128         if (a->offset < b->offset)
129                 return -1;
130         else if (a->offset > b->offset)
131                 return 1;
132
133         if (a->parent < b->parent)
134                 return -1;
135         else if (a->parent > b->parent)
136                 return 1;
137
138         return 0;
139 }
140
141 /*
142  * Search ref_node with (root_id, object_id, offset, parent) in the tree
143  *
144  * if found, the pointer of the ref_node will be returned;
145  * if not found, NULL will be returned and pos will point to the rb_node for
146  * insert, pos_parent will point to pos'parent for insert;
147 */
148 static struct ref_node *__ref_tree_search(struct ref_root *ref_tree,
149                                           struct rb_node ***pos,
150                                           struct rb_node **pos_parent,
151                                           u64 root_id, u64 object_id,
152                                           u64 offset, u64 parent)
153 {
154         struct ref_node *cur = NULL;
155         struct ref_node entry;
156         int ret;
157
158         entry.root_id = root_id;
159         entry.object_id = object_id;
160         entry.offset = offset;
161         entry.parent = parent;
162
163         *pos = &ref_tree->rb_root.rb_node;
164
165         while (**pos) {
166                 *pos_parent = **pos;
167                 cur = rb_entry(*pos_parent, struct ref_node, rb_node);
168
169                 ret = ref_node_cmp(cur, &entry);
170                 if (ret > 0)
171                         *pos = &(**pos)->rb_left;
172                 else if (ret < 0)
173                         *pos = &(**pos)->rb_right;
174                 else
175                         return cur;
176         }
177
178         return NULL;
179 }
180
181 /*
182  * Insert a ref_node to the ref tree
183  * @pos used for specifiy the position to insert
184  * @pos_parent for specifiy pos's parent
185  *
186  * success, return 0;
187  * ref_node already exists, return -EEXIST;
188 */
189 static int ref_tree_insert(struct ref_root *ref_tree, struct rb_node **pos,
190                            struct rb_node *pos_parent, struct ref_node *ins)
191 {
192         struct rb_node **p = NULL;
193         struct rb_node *parent = NULL;
194         struct ref_node *cur = NULL;
195
196         if (!pos) {
197                 cur = __ref_tree_search(ref_tree, &p, &parent, ins->root_id,
198                                         ins->object_id, ins->offset,
199                                         ins->parent);
200                 if (cur)
201                         return -EEXIST;
202         } else {
203                 p = pos;
204                 parent = pos_parent;
205         }
206
207         rb_link_node(&ins->rb_node, parent, p);
208         rb_insert_color(&ins->rb_node, &ref_tree->rb_root);
209
210         return 0;
211 }
212
213 /* Erase and free ref_node, caller should update ref_root->unique_refs */
214 static void ref_tree_remove(struct ref_root *ref_tree, struct ref_node *node)
215 {
216         rb_erase(&node->rb_node, &ref_tree->rb_root);
217         kfree(node);
218 }
219
220 /*
221  * Update ref_root->unique_refs
222  *
223  * Call __ref_tree_search
224  *      1. if ref_node doesn't exist, ref_tree_insert this node, and update
225  *      ref_root->unique_refs:
226  *              if ref_node->ref_mod > 0, ref_root->unique_refs++;
227  *              if ref_node->ref_mod < 0, do noting;
228  *
229  *      2. if ref_node is found, then get origin ref_node->ref_mod, and update
230  *      ref_node->ref_mod.
231  *              if ref_node->ref_mod is equal to 0,then call ref_tree_remove
232  *
233  *              according to origin_mod and new_mod, update ref_root->items
234  *              +----------------+--------------+-------------+
235  *              |                |new_count <= 0|new_count > 0|
236  *              +----------------+--------------+-------------+
237  *              |origin_count < 0|       0      |      1      |
238  *              +----------------+--------------+-------------+
239  *              |origin_count > 0|      -1      |      0      |
240  *              +----------------+--------------+-------------+
241  *
242  * In case of allocation failure, -ENOMEM is returned and the ref_tree stays
243  * unaltered.
244  * Success, return 0
245  */
246 static int ref_tree_add(struct ref_root *ref_tree, u64 root_id, u64 object_id,
247                         u64 offset, u64 parent, int count)
248 {
249         struct ref_node *node = NULL;
250         struct rb_node **pos = NULL;
251         struct rb_node *pos_parent = NULL;
252         int origin_count;
253         int ret;
254
255         if (!count)
256                 return 0;
257
258         node = __ref_tree_search(ref_tree, &pos, &pos_parent, root_id,
259                                  object_id, offset, parent);
260         if (node == NULL) {
261                 node = kmalloc(sizeof(*node), GFP_NOFS);
262                 if (!node)
263                         return -ENOMEM;
264
265                 node->root_id = root_id;
266                 node->object_id = object_id;
267                 node->offset = offset;
268                 node->parent = parent;
269                 node->ref_mod = count;
270
271                 ret = ref_tree_insert(ref_tree, pos, pos_parent, node);
272                 ASSERT(!ret);
273                 if (ret) {
274                         kfree(node);
275                         return ret;
276                 }
277
278                 ref_tree->unique_refs += node->ref_mod > 0 ? 1 : 0;
279
280                 return 0;
281         }
282
283         origin_count = node->ref_mod;
284         node->ref_mod += count;
285
286         if (node->ref_mod > 0)
287                 ref_tree->unique_refs += origin_count > 0 ? 0 : 1;
288         else if (node->ref_mod <= 0)
289                 ref_tree->unique_refs += origin_count > 0 ? -1 : 0;
290
291         if (!node->ref_mod)
292                 ref_tree_remove(ref_tree, node);
293
294         return 0;
295 }
296
297 static int check_extent_in_eb(struct btrfs_key *key, struct extent_buffer *eb,
298                                 struct btrfs_file_extent_item *fi,
299                                 u64 extent_item_pos,
300                                 struct extent_inode_elem **eie)
301 {
302         u64 offset = 0;
303         struct extent_inode_elem *e;
304
305         if (!btrfs_file_extent_compression(eb, fi) &&
306             !btrfs_file_extent_encryption(eb, fi) &&
307             !btrfs_file_extent_other_encoding(eb, fi)) {
308                 u64 data_offset;
309                 u64 data_len;
310
311                 data_offset = btrfs_file_extent_offset(eb, fi);
312                 data_len = btrfs_file_extent_num_bytes(eb, fi);
313
314                 if (extent_item_pos < data_offset ||
315                     extent_item_pos >= data_offset + data_len)
316                         return 1;
317                 offset = extent_item_pos - data_offset;
318         }
319
320         e = kmalloc(sizeof(*e), GFP_NOFS);
321         if (!e)
322                 return -ENOMEM;
323
324         e->next = *eie;
325         e->inum = key->objectid;
326         e->offset = key->offset + offset;
327         *eie = e;
328
329         return 0;
330 }
331
332 static void free_inode_elem_list(struct extent_inode_elem *eie)
333 {
334         struct extent_inode_elem *eie_next;
335
336         for (; eie; eie = eie_next) {
337                 eie_next = eie->next;
338                 kfree(eie);
339         }
340 }
341
342 static int find_extent_in_eb(struct extent_buffer *eb, u64 wanted_disk_byte,
343                                 u64 extent_item_pos,
344                                 struct extent_inode_elem **eie)
345 {
346         u64 disk_byte;
347         struct btrfs_key key;
348         struct btrfs_file_extent_item *fi;
349         int slot;
350         int nritems;
351         int extent_type;
352         int ret;
353
354         /*
355          * from the shared data ref, we only have the leaf but we need
356          * the key. thus, we must look into all items and see that we
357          * find one (some) with a reference to our extent item.
358          */
359         nritems = btrfs_header_nritems(eb);
360         for (slot = 0; slot < nritems; ++slot) {
361                 btrfs_item_key_to_cpu(eb, &key, slot);
362                 if (key.type != BTRFS_EXTENT_DATA_KEY)
363                         continue;
364                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
365                 extent_type = btrfs_file_extent_type(eb, fi);
366                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
367                         continue;
368                 /* don't skip BTRFS_FILE_EXTENT_PREALLOC, we can handle that */
369                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
370                 if (disk_byte != wanted_disk_byte)
371                         continue;
372
373                 ret = check_extent_in_eb(&key, eb, fi, extent_item_pos, eie);
374                 if (ret < 0)
375                         return ret;
376         }
377
378         return 0;
379 }
380
381 /*
382  * this structure records all encountered refs on the way up to the root
383  */
384 struct __prelim_ref {
385         struct list_head list;
386         u64 root_id;
387         struct btrfs_key key_for_search;
388         int level;
389         int count;
390         struct extent_inode_elem *inode_list;
391         u64 parent;
392         u64 wanted_disk_byte;
393 };
394
395 static struct kmem_cache *btrfs_prelim_ref_cache;
396
397 int __init btrfs_prelim_ref_init(void)
398 {
399         btrfs_prelim_ref_cache = kmem_cache_create("btrfs_prelim_ref",
400                                         sizeof(struct __prelim_ref),
401                                         0,
402                                         SLAB_MEM_SPREAD,
403                                         NULL);
404         if (!btrfs_prelim_ref_cache)
405                 return -ENOMEM;
406         return 0;
407 }
408
409 void btrfs_prelim_ref_exit(void)
410 {
411         kmem_cache_destroy(btrfs_prelim_ref_cache);
412 }
413
414 /*
415  * the rules for all callers of this function are:
416  * - obtaining the parent is the goal
417  * - if you add a key, you must know that it is a correct key
418  * - if you cannot add the parent or a correct key, then we will look into the
419  *   block later to set a correct key
420  *
421  * delayed refs
422  * ============
423  *        backref type | shared | indirect | shared | indirect
424  * information         |   tree |     tree |   data |     data
425  * --------------------+--------+----------+--------+----------
426  *      parent logical |    y   |     -    |    -   |     -
427  *      key to resolve |    -   |     y    |    y   |     y
428  *  tree block logical |    -   |     -    |    -   |     -
429  *  root for resolving |    y   |     y    |    y   |     y
430  *
431  * - column 1:       we've the parent -> done
432  * - column 2, 3, 4: we use the key to find the parent
433  *
434  * on disk refs (inline or keyed)
435  * ==============================
436  *        backref type | shared | indirect | shared | indirect
437  * information         |   tree |     tree |   data |     data
438  * --------------------+--------+----------+--------+----------
439  *      parent logical |    y   |     -    |    y   |     -
440  *      key to resolve |    -   |     -    |    -   |     y
441  *  tree block logical |    y   |     y    |    y   |     y
442  *  root for resolving |    -   |     y    |    y   |     y
443  *
444  * - column 1, 3: we've the parent -> done
445  * - column 2:    we take the first key from the block to find the parent
446  *                (see __add_missing_keys)
447  * - column 4:    we use the key to find the parent
448  *
449  * additional information that's available but not required to find the parent
450  * block might help in merging entries to gain some speed.
451  */
452
453 static int __add_prelim_ref(struct list_head *head, u64 root_id,
454                             struct btrfs_key *key, int level,
455                             u64 parent, u64 wanted_disk_byte, int count,
456                             gfp_t gfp_mask)
457 {
458         struct __prelim_ref *ref;
459
460         if (root_id == BTRFS_DATA_RELOC_TREE_OBJECTID)
461                 return 0;
462
463         ref = kmem_cache_alloc(btrfs_prelim_ref_cache, gfp_mask);
464         if (!ref)
465                 return -ENOMEM;
466
467         ref->root_id = root_id;
468         if (key) {
469                 ref->key_for_search = *key;
470                 /*
471                  * We can often find data backrefs with an offset that is too
472                  * large (>= LLONG_MAX, maximum allowed file offset) due to
473                  * underflows when subtracting a file's offset with the data
474                  * offset of its corresponding extent data item. This can
475                  * happen for example in the clone ioctl.
476                  * So if we detect such case we set the search key's offset to
477                  * zero to make sure we will find the matching file extent item
478                  * at add_all_parents(), otherwise we will miss it because the
479                  * offset taken form the backref is much larger then the offset
480                  * of the file extent item. This can make us scan a very large
481                  * number of file extent items, but at least it will not make
482                  * us miss any.
483                  * This is an ugly workaround for a behaviour that should have
484                  * never existed, but it does and a fix for the clone ioctl
485                  * would touch a lot of places, cause backwards incompatibility
486                  * and would not fix the problem for extents cloned with older
487                  * kernels.
488                  */
489                 if (ref->key_for_search.type == BTRFS_EXTENT_DATA_KEY &&
490                     ref->key_for_search.offset >= LLONG_MAX)
491                         ref->key_for_search.offset = 0;
492         } else {
493                 memset(&ref->key_for_search, 0, sizeof(ref->key_for_search));
494         }
495
496         ref->inode_list = NULL;
497         ref->level = level;
498         ref->count = count;
499         ref->parent = parent;
500         ref->wanted_disk_byte = wanted_disk_byte;
501         list_add_tail(&ref->list, head);
502
503         return 0;
504 }
505
506 static int add_all_parents(struct btrfs_root *root, struct btrfs_path *path,
507                            struct ulist *parents, struct __prelim_ref *ref,
508                            int level, u64 time_seq, const u64 *extent_item_pos,
509                            u64 total_refs)
510 {
511         int ret = 0;
512         int slot;
513         struct extent_buffer *eb;
514         struct btrfs_key key;
515         struct btrfs_key *key_for_search = &ref->key_for_search;
516         struct btrfs_file_extent_item *fi;
517         struct extent_inode_elem *eie = NULL, *old = NULL;
518         u64 disk_byte;
519         u64 wanted_disk_byte = ref->wanted_disk_byte;
520         u64 count = 0;
521
522         if (level != 0) {
523                 eb = path->nodes[level];
524                 ret = ulist_add(parents, eb->start, 0, GFP_NOFS);
525                 if (ret < 0)
526                         return ret;
527                 return 0;
528         }
529
530         /*
531          * We normally enter this function with the path already pointing to
532          * the first item to check. But sometimes, we may enter it with
533          * slot==nritems. In that case, go to the next leaf before we continue.
534          */
535         if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
536                 if (time_seq == (u64)-1)
537                         ret = btrfs_next_leaf(root, path);
538                 else
539                         ret = btrfs_next_old_leaf(root, path, time_seq);
540         }
541
542         while (!ret && count < total_refs) {
543                 eb = path->nodes[0];
544                 slot = path->slots[0];
545
546                 btrfs_item_key_to_cpu(eb, &key, slot);
547
548                 if (key.objectid != key_for_search->objectid ||
549                     key.type != BTRFS_EXTENT_DATA_KEY)
550                         break;
551
552                 fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
553                 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
554
555                 if (disk_byte == wanted_disk_byte) {
556                         eie = NULL;
557                         old = NULL;
558                         count++;
559                         if (extent_item_pos) {
560                                 ret = check_extent_in_eb(&key, eb, fi,
561                                                 *extent_item_pos,
562                                                 &eie);
563                                 if (ret < 0)
564                                         break;
565                         }
566                         if (ret > 0)
567                                 goto next;
568                         ret = ulist_add_merge_ptr(parents, eb->start,
569                                                   eie, (void **)&old, GFP_NOFS);
570                         if (ret < 0)
571                                 break;
572                         if (!ret && extent_item_pos) {
573                                 while (old->next)
574                                         old = old->next;
575                                 old->next = eie;
576                         }
577                         eie = NULL;
578                 }
579 next:
580                 if (time_seq == (u64)-1)
581                         ret = btrfs_next_item(root, path);
582                 else
583                         ret = btrfs_next_old_item(root, path, time_seq);
584         }
585
586         if (ret > 0)
587                 ret = 0;
588         else if (ret < 0)
589                 free_inode_elem_list(eie);
590         return ret;
591 }
592
593 /*
594  * resolve an indirect backref in the form (root_id, key, level)
595  * to a logical address
596  */
597 static int __resolve_indirect_ref(struct btrfs_fs_info *fs_info,
598                                   struct btrfs_path *path, u64 time_seq,
599                                   struct __prelim_ref *ref,
600                                   struct ulist *parents,
601                                   const u64 *extent_item_pos, u64 total_refs)
602 {
603         struct btrfs_root *root;
604         struct btrfs_key root_key;
605         struct extent_buffer *eb;
606         int ret = 0;
607         int root_level;
608         int level = ref->level;
609         int index;
610
611         root_key.objectid = ref->root_id;
612         root_key.type = BTRFS_ROOT_ITEM_KEY;
613         root_key.offset = (u64)-1;
614
615         index = srcu_read_lock(&fs_info->subvol_srcu);
616
617         root = btrfs_get_fs_root(fs_info, &root_key, false);
618         if (IS_ERR(root)) {
619                 srcu_read_unlock(&fs_info->subvol_srcu, index);
620                 ret = PTR_ERR(root);
621                 goto out;
622         }
623
624         if (btrfs_is_testing(fs_info)) {
625                 srcu_read_unlock(&fs_info->subvol_srcu, index);
626                 ret = -ENOENT;
627                 goto out;
628         }
629
630         if (path->search_commit_root)
631                 root_level = btrfs_header_level(root->commit_root);
632         else if (time_seq == (u64)-1)
633                 root_level = btrfs_header_level(root->node);
634         else
635                 root_level = btrfs_old_root_level(root, time_seq);
636
637         if (root_level + 1 == level) {
638                 srcu_read_unlock(&fs_info->subvol_srcu, index);
639                 goto out;
640         }
641
642         path->lowest_level = level;
643         if (time_seq == (u64)-1)
644                 ret = btrfs_search_slot(NULL, root, &ref->key_for_search, path,
645                                         0, 0);
646         else
647                 ret = btrfs_search_old_slot(root, &ref->key_for_search, path,
648                                             time_seq);
649
650         /* root node has been locked, we can release @subvol_srcu safely here */
651         srcu_read_unlock(&fs_info->subvol_srcu, index);
652
653         pr_debug("search slot in root %llu (level %d, ref count %d) returned "
654                  "%d for key (%llu %u %llu)\n",
655                  ref->root_id, level, ref->count, ret,
656                  ref->key_for_search.objectid, ref->key_for_search.type,
657                  ref->key_for_search.offset);
658         if (ret < 0)
659                 goto out;
660
661         eb = path->nodes[level];
662         while (!eb) {
663                 if (WARN_ON(!level)) {
664                         ret = 1;
665                         goto out;
666                 }
667                 level--;
668                 eb = path->nodes[level];
669         }
670
671         ret = add_all_parents(root, path, parents, ref, level, time_seq,
672                               extent_item_pos, total_refs);
673 out:
674         path->lowest_level = 0;
675         btrfs_release_path(path);
676         return ret;
677 }
678
679 /*
680  * resolve all indirect backrefs from the list
681  */
682 static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
683                                    struct btrfs_path *path, u64 time_seq,
684                                    struct list_head *head,
685                                    const u64 *extent_item_pos, u64 total_refs,
686                                    u64 root_objectid)
687 {
688         int err;
689         int ret = 0;
690         struct __prelim_ref *ref;
691         struct __prelim_ref *ref_safe;
692         struct __prelim_ref *new_ref;
693         struct ulist *parents;
694         struct ulist_node *node;
695         struct ulist_iterator uiter;
696
697         parents = ulist_alloc(GFP_NOFS);
698         if (!parents)
699                 return -ENOMEM;
700
701         /*
702          * _safe allows us to insert directly after the current item without
703          * iterating over the newly inserted items.
704          * we're also allowed to re-assign ref during iteration.
705          */
706         list_for_each_entry_safe(ref, ref_safe, head, list) {
707                 if (ref->parent)        /* already direct */
708                         continue;
709                 if (ref->count == 0)
710                         continue;
711                 if (root_objectid && ref->root_id != root_objectid) {
712                         ret = BACKREF_FOUND_SHARED;
713                         goto out;
714                 }
715                 err = __resolve_indirect_ref(fs_info, path, time_seq, ref,
716                                              parents, extent_item_pos,
717                                              total_refs);
718                 /*
719                  * we can only tolerate ENOENT,otherwise,we should catch error
720                  * and return directly.
721                  */
722                 if (err == -ENOENT) {
723                         continue;
724                 } else if (err) {
725                         ret = err;
726                         goto out;
727                 }
728
729                 /* we put the first parent into the ref at hand */
730                 ULIST_ITER_INIT(&uiter);
731                 node = ulist_next(parents, &uiter);
732                 ref->parent = node ? node->val : 0;
733                 ref->inode_list = node ?
734                         (struct extent_inode_elem *)(uintptr_t)node->aux : NULL;
735
736                 /* additional parents require new refs being added here */
737                 while ((node = ulist_next(parents, &uiter))) {
738                         new_ref = kmem_cache_alloc(btrfs_prelim_ref_cache,
739                                                    GFP_NOFS);
740                         if (!new_ref) {
741                                 ret = -ENOMEM;
742                                 goto out;
743                         }
744                         memcpy(new_ref, ref, sizeof(*ref));
745                         new_ref->parent = node->val;
746                         new_ref->inode_list = (struct extent_inode_elem *)
747                                                         (uintptr_t)node->aux;
748                         list_add(&new_ref->list, &ref->list);
749                 }
750                 ulist_reinit(parents);
751         }
752 out:
753         ulist_free(parents);
754         return ret;
755 }
756
757 static inline int ref_for_same_block(struct __prelim_ref *ref1,
758                                      struct __prelim_ref *ref2)
759 {
760         if (ref1->level != ref2->level)
761                 return 0;
762         if (ref1->root_id != ref2->root_id)
763                 return 0;
764         if (ref1->key_for_search.type != ref2->key_for_search.type)
765                 return 0;
766         if (ref1->key_for_search.objectid != ref2->key_for_search.objectid)
767                 return 0;
768         if (ref1->key_for_search.offset != ref2->key_for_search.offset)
769                 return 0;
770         if (ref1->parent != ref2->parent)
771                 return 0;
772
773         return 1;
774 }
775
776 /*
777  * read tree blocks and add keys where required.
778  */
779 static int __add_missing_keys(struct btrfs_fs_info *fs_info,
780                               struct list_head *head)
781 {
782         struct __prelim_ref *ref;
783         struct extent_buffer *eb;
784
785         list_for_each_entry(ref, head, list) {
786                 if (ref->parent)
787                         continue;
788                 if (ref->key_for_search.type)
789                         continue;
790                 BUG_ON(!ref->wanted_disk_byte);
791                 eb = read_tree_block(fs_info->tree_root, ref->wanted_disk_byte,
792                                      0);
793                 if (IS_ERR(eb)) {
794                         return PTR_ERR(eb);
795                 } else if (!extent_buffer_uptodate(eb)) {
796                         free_extent_buffer(eb);
797                         return -EIO;
798                 }
799                 btrfs_tree_read_lock(eb);
800                 if (btrfs_header_level(eb) == 0)
801                         btrfs_item_key_to_cpu(eb, &ref->key_for_search, 0);
802                 else
803                         btrfs_node_key_to_cpu(eb, &ref->key_for_search, 0);
804                 btrfs_tree_read_unlock(eb);
805                 free_extent_buffer(eb);
806         }
807         return 0;
808 }
809
810 /*
811  * merge backrefs and adjust counts accordingly
812  *
813  * mode = 1: merge identical keys, if key is set
814  *    FIXME: if we add more keys in __add_prelim_ref, we can merge more here.
815  *           additionally, we could even add a key range for the blocks we
816  *           looked into to merge even more (-> replace unresolved refs by those
817  *           having a parent).
818  * mode = 2: merge identical parents
819  */
820 static void __merge_refs(struct list_head *head, int mode)
821 {
822         struct __prelim_ref *pos1;
823
824         list_for_each_entry(pos1, head, list) {
825                 struct __prelim_ref *pos2 = pos1, *tmp;
826
827                 list_for_each_entry_safe_continue(pos2, tmp, head, list) {
828                         struct __prelim_ref *ref1 = pos1, *ref2 = pos2;
829                         struct extent_inode_elem *eie;
830
831                         if (!ref_for_same_block(ref1, ref2))
832                                 continue;
833                         if (mode == 1) {
834                                 if (!ref1->parent && ref2->parent)
835                                         swap(ref1, ref2);
836                         } else {
837                                 if (ref1->parent != ref2->parent)
838                                         continue;
839                         }
840
841                         eie = ref1->inode_list;
842                         while (eie && eie->next)
843                                 eie = eie->next;
844                         if (eie)
845                                 eie->next = ref2->inode_list;
846                         else
847                                 ref1->inode_list = ref2->inode_list;
848                         ref1->count += ref2->count;
849
850                         list_del(&ref2->list);
851                         kmem_cache_free(btrfs_prelim_ref_cache, ref2);
852                         cond_resched();
853                 }
854
855         }
856 }
857
858 /*
859  * add all currently queued delayed refs from this head whose seq nr is
860  * smaller or equal that seq to the list
861  */
862 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
863                               struct list_head *prefs, u64 *total_refs,
864                               u64 inum)
865 {
866         struct btrfs_delayed_ref_node *node;
867         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
868         struct btrfs_key key;
869         struct btrfs_key op_key = {0};
870         int sgn;
871         int ret = 0;
872
873         if (extent_op && extent_op->update_key)
874                 btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
875
876         spin_lock(&head->lock);
877         list_for_each_entry(node, &head->ref_list, list) {
878                 if (node->seq > seq)
879                         continue;
880
881                 switch (node->action) {
882                 case BTRFS_ADD_DELAYED_EXTENT:
883                 case BTRFS_UPDATE_DELAYED_HEAD:
884                         WARN_ON(1);
885                         continue;
886                 case BTRFS_ADD_DELAYED_REF:
887                         sgn = 1;
888                         break;
889                 case BTRFS_DROP_DELAYED_REF:
890                         sgn = -1;
891                         break;
892                 default:
893                         BUG_ON(1);
894                 }
895                 *total_refs += (node->ref_mod * sgn);
896                 switch (node->type) {
897                 case BTRFS_TREE_BLOCK_REF_KEY: {
898                         struct btrfs_delayed_tree_ref *ref;
899
900                         ref = btrfs_delayed_node_to_tree_ref(node);
901                         ret = __add_prelim_ref(prefs, ref->root, &op_key,
902                                                ref->level + 1, 0, node->bytenr,
903                                                node->ref_mod * sgn, GFP_ATOMIC);
904                         break;
905                 }
906                 case BTRFS_SHARED_BLOCK_REF_KEY: {
907                         struct btrfs_delayed_tree_ref *ref;
908
909                         ref = btrfs_delayed_node_to_tree_ref(node);
910                         ret = __add_prelim_ref(prefs, 0, NULL,
911                                                ref->level + 1, ref->parent,
912                                                node->bytenr,
913                                                node->ref_mod * sgn, GFP_ATOMIC);
914                         break;
915                 }
916                 case BTRFS_EXTENT_DATA_REF_KEY: {
917                         struct btrfs_delayed_data_ref *ref;
918                         ref = btrfs_delayed_node_to_data_ref(node);
919
920                         key.objectid = ref->objectid;
921                         key.type = BTRFS_EXTENT_DATA_KEY;
922                         key.offset = ref->offset;
923
924                         /*
925                          * Found a inum that doesn't match our known inum, we
926                          * know it's shared.
927                          */
928                         if (inum && ref->objectid != inum) {
929                                 ret = BACKREF_FOUND_SHARED;
930                                 break;
931                         }
932
933                         ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
934                                                node->bytenr,
935                                                node->ref_mod * sgn, GFP_ATOMIC);
936                         break;
937                 }
938                 case BTRFS_SHARED_DATA_REF_KEY: {
939                         struct btrfs_delayed_data_ref *ref;
940
941                         ref = btrfs_delayed_node_to_data_ref(node);
942                         ret = __add_prelim_ref(prefs, 0, NULL, 0,
943                                                ref->parent, node->bytenr,
944                                                node->ref_mod * sgn, GFP_ATOMIC);
945                         break;
946                 }
947                 default:
948                         WARN_ON(1);
949                 }
950                 if (ret)
951                         break;
952         }
953         spin_unlock(&head->lock);
954         return ret;
955 }
956
957 /*
958  * add all inline backrefs for bytenr to the list
959  */
960 static int __add_inline_refs(struct btrfs_fs_info *fs_info,
961                              struct btrfs_path *path, u64 bytenr,
962                              int *info_level, struct list_head *prefs,
963                              struct ref_root *ref_tree,
964                              u64 *total_refs, u64 inum)
965 {
966         int ret = 0;
967         int slot;
968         struct extent_buffer *leaf;
969         struct btrfs_key key;
970         struct btrfs_key found_key;
971         unsigned long ptr;
972         unsigned long end;
973         struct btrfs_extent_item *ei;
974         u64 flags;
975         u64 item_size;
976
977         /*
978          * enumerate all inline refs
979          */
980         leaf = path->nodes[0];
981         slot = path->slots[0];
982
983         item_size = btrfs_item_size_nr(leaf, slot);
984         BUG_ON(item_size < sizeof(*ei));
985
986         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
987         flags = btrfs_extent_flags(leaf, ei);
988         *total_refs += btrfs_extent_refs(leaf, ei);
989         btrfs_item_key_to_cpu(leaf, &found_key, slot);
990
991         ptr = (unsigned long)(ei + 1);
992         end = (unsigned long)ei + item_size;
993
994         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
995             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
996                 struct btrfs_tree_block_info *info;
997
998                 info = (struct btrfs_tree_block_info *)ptr;
999                 *info_level = btrfs_tree_block_level(leaf, info);
1000                 ptr += sizeof(struct btrfs_tree_block_info);
1001                 BUG_ON(ptr > end);
1002         } else if (found_key.type == BTRFS_METADATA_ITEM_KEY) {
1003                 *info_level = found_key.offset;
1004         } else {
1005                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1006         }
1007
1008         while (ptr < end) {
1009                 struct btrfs_extent_inline_ref *iref;
1010                 u64 offset;
1011                 int type;
1012
1013                 iref = (struct btrfs_extent_inline_ref *)ptr;
1014                 type = btrfs_extent_inline_ref_type(leaf, iref);
1015                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
1016
1017                 switch (type) {
1018                 case BTRFS_SHARED_BLOCK_REF_KEY:
1019                         ret = __add_prelim_ref(prefs, 0, NULL,
1020                                                 *info_level + 1, offset,
1021                                                 bytenr, 1, GFP_NOFS);
1022                         break;
1023                 case BTRFS_SHARED_DATA_REF_KEY: {
1024                         struct btrfs_shared_data_ref *sdref;
1025                         int count;
1026
1027                         sdref = (struct btrfs_shared_data_ref *)(iref + 1);
1028                         count = btrfs_shared_data_ref_count(leaf, sdref);
1029                         ret = __add_prelim_ref(prefs, 0, NULL, 0, offset,
1030                                                bytenr, count, GFP_NOFS);
1031                         if (ref_tree) {
1032                                 if (!ret)
1033                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1034                                                            bytenr, count);
1035                                 if (!ret && ref_tree->unique_refs > 1)
1036                                         ret = BACKREF_FOUND_SHARED;
1037                         }
1038                         break;
1039                 }
1040                 case BTRFS_TREE_BLOCK_REF_KEY:
1041                         ret = __add_prelim_ref(prefs, offset, NULL,
1042                                                *info_level + 1, 0,
1043                                                bytenr, 1, GFP_NOFS);
1044                         break;
1045                 case BTRFS_EXTENT_DATA_REF_KEY: {
1046                         struct btrfs_extent_data_ref *dref;
1047                         int count;
1048                         u64 root;
1049
1050                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1051                         count = btrfs_extent_data_ref_count(leaf, dref);
1052                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1053                                                                       dref);
1054                         key.type = BTRFS_EXTENT_DATA_KEY;
1055                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1056
1057                         if (inum && key.objectid != inum) {
1058                                 ret = BACKREF_FOUND_SHARED;
1059                                 break;
1060                         }
1061
1062                         root = btrfs_extent_data_ref_root(leaf, dref);
1063                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1064                                                bytenr, count, GFP_NOFS);
1065                         if (ref_tree) {
1066                                 if (!ret)
1067                                         ret = ref_tree_add(ref_tree, root,
1068                                                            key.objectid,
1069                                                            key.offset, 0,
1070                                                            count);
1071                                 if (!ret && ref_tree->unique_refs > 1)
1072                                         ret = BACKREF_FOUND_SHARED;
1073                         }
1074                         break;
1075                 }
1076                 default:
1077                         WARN_ON(1);
1078                 }
1079                 if (ret)
1080                         return ret;
1081                 ptr += btrfs_extent_inline_ref_size(type);
1082         }
1083
1084         return 0;
1085 }
1086
1087 /*
1088  * add all non-inline backrefs for bytenr to the list
1089  */
1090 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
1091                             struct btrfs_path *path, u64 bytenr,
1092                             int info_level, struct list_head *prefs,
1093                             struct ref_root *ref_tree, u64 inum)
1094 {
1095         struct btrfs_root *extent_root = fs_info->extent_root;
1096         int ret;
1097         int slot;
1098         struct extent_buffer *leaf;
1099         struct btrfs_key key;
1100
1101         while (1) {
1102                 ret = btrfs_next_item(extent_root, path);
1103                 if (ret < 0)
1104                         break;
1105                 if (ret) {
1106                         ret = 0;
1107                         break;
1108                 }
1109
1110                 slot = path->slots[0];
1111                 leaf = path->nodes[0];
1112                 btrfs_item_key_to_cpu(leaf, &key, slot);
1113
1114                 if (key.objectid != bytenr)
1115                         break;
1116                 if (key.type < BTRFS_TREE_BLOCK_REF_KEY)
1117                         continue;
1118                 if (key.type > BTRFS_SHARED_DATA_REF_KEY)
1119                         break;
1120
1121                 switch (key.type) {
1122                 case BTRFS_SHARED_BLOCK_REF_KEY:
1123                         ret = __add_prelim_ref(prefs, 0, NULL,
1124                                                 info_level + 1, key.offset,
1125                                                 bytenr, 1, GFP_NOFS);
1126                         break;
1127                 case BTRFS_SHARED_DATA_REF_KEY: {
1128                         struct btrfs_shared_data_ref *sdref;
1129                         int count;
1130
1131                         sdref = btrfs_item_ptr(leaf, slot,
1132                                               struct btrfs_shared_data_ref);
1133                         count = btrfs_shared_data_ref_count(leaf, sdref);
1134                         ret = __add_prelim_ref(prefs, 0, NULL, 0, key.offset,
1135                                                 bytenr, count, GFP_NOFS);
1136                         if (ref_tree) {
1137                                 if (!ret)
1138                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1139                                                            bytenr, count);
1140                                 if (!ret && ref_tree->unique_refs > 1)
1141                                         ret = BACKREF_FOUND_SHARED;
1142                         }
1143                         break;
1144                 }
1145                 case BTRFS_TREE_BLOCK_REF_KEY:
1146                         ret = __add_prelim_ref(prefs, key.offset, NULL,
1147                                                info_level + 1, 0,
1148                                                bytenr, 1, GFP_NOFS);
1149                         break;
1150                 case BTRFS_EXTENT_DATA_REF_KEY: {
1151                         struct btrfs_extent_data_ref *dref;
1152                         int count;
1153                         u64 root;
1154
1155                         dref = btrfs_item_ptr(leaf, slot,
1156                                               struct btrfs_extent_data_ref);
1157                         count = btrfs_extent_data_ref_count(leaf, dref);
1158                         key.objectid = btrfs_extent_data_ref_objectid(leaf,
1159                                                                       dref);
1160                         key.type = BTRFS_EXTENT_DATA_KEY;
1161                         key.offset = btrfs_extent_data_ref_offset(leaf, dref);
1162
1163                         if (inum && key.objectid != inum) {
1164                                 ret = BACKREF_FOUND_SHARED;
1165                                 break;
1166                         }
1167
1168                         root = btrfs_extent_data_ref_root(leaf, dref);
1169                         ret = __add_prelim_ref(prefs, root, &key, 0, 0,
1170                                                bytenr, count, GFP_NOFS);
1171                         if (ref_tree) {
1172                                 if (!ret)
1173                                         ret = ref_tree_add(ref_tree, root,
1174                                                            key.objectid,
1175                                                            key.offset, 0,
1176                                                            count);
1177                                 if (!ret && ref_tree->unique_refs > 1)
1178                                         ret = BACKREF_FOUND_SHARED;
1179                         }
1180                         break;
1181                 }
1182                 default:
1183                         WARN_ON(1);
1184                 }
1185                 if (ret)
1186                         return ret;
1187
1188         }
1189
1190         return ret;
1191 }
1192
1193 /*
1194  * this adds all existing backrefs (inline backrefs, backrefs and delayed
1195  * refs) for the given bytenr to the refs list, merges duplicates and resolves
1196  * indirect refs to their parent bytenr.
1197  * When roots are found, they're added to the roots list
1198  *
1199  * NOTE: This can return values > 0
1200  *
1201  * If time_seq is set to (u64)-1, it will not search delayed_refs, and behave
1202  * much like trans == NULL case, the difference only lies in it will not
1203  * commit root.
1204  * The special case is for qgroup to search roots in commit_transaction().
1205  *
1206  * If check_shared is set to 1, any extent has more than one ref item, will
1207  * be returned BACKREF_FOUND_SHARED immediately.
1208  *
1209  * FIXME some caching might speed things up
1210  */
1211 static int find_parent_nodes(struct btrfs_trans_handle *trans,
1212                              struct btrfs_fs_info *fs_info, u64 bytenr,
1213                              u64 time_seq, struct ulist *refs,
1214                              struct ulist *roots, const u64 *extent_item_pos,
1215                              u64 root_objectid, u64 inum, int check_shared)
1216 {
1217         struct btrfs_key key;
1218         struct btrfs_path *path;
1219         struct btrfs_delayed_ref_root *delayed_refs = NULL;
1220         struct btrfs_delayed_ref_head *head;
1221         int info_level = 0;
1222         int ret;
1223         struct list_head prefs_delayed;
1224         struct list_head prefs;
1225         struct __prelim_ref *ref;
1226         struct extent_inode_elem *eie = NULL;
1227         struct ref_root *ref_tree = NULL;
1228         u64 total_refs = 0;
1229
1230         INIT_LIST_HEAD(&prefs);
1231         INIT_LIST_HEAD(&prefs_delayed);
1232
1233         key.objectid = bytenr;
1234         key.offset = (u64)-1;
1235         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1236                 key.type = BTRFS_METADATA_ITEM_KEY;
1237         else
1238                 key.type = BTRFS_EXTENT_ITEM_KEY;
1239
1240         path = btrfs_alloc_path();
1241         if (!path)
1242                 return -ENOMEM;
1243         if (!trans) {
1244                 path->search_commit_root = 1;
1245                 path->skip_locking = 1;
1246         }
1247
1248         if (time_seq == (u64)-1)
1249                 path->skip_locking = 1;
1250
1251         /*
1252          * grab both a lock on the path and a lock on the delayed ref head.
1253          * We need both to get a consistent picture of how the refs look
1254          * at a specified point in time
1255          */
1256 again:
1257         head = NULL;
1258
1259         if (check_shared) {
1260                 if (!ref_tree) {
1261                         ref_tree = ref_root_alloc();
1262                         if (!ref_tree) {
1263                                 ret = -ENOMEM;
1264                                 goto out;
1265                         }
1266                 } else {
1267                         ref_root_fini(ref_tree);
1268                 }
1269         }
1270
1271         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
1272         if (ret < 0)
1273                 goto out;
1274         BUG_ON(ret == 0);
1275
1276 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
1277         if (trans && likely(trans->type != __TRANS_DUMMY) &&
1278             time_seq != (u64)-1) {
1279 #else
1280         if (trans && time_seq != (u64)-1) {
1281 #endif
1282                 /*
1283                  * look if there are updates for this ref queued and lock the
1284                  * head
1285                  */
1286                 delayed_refs = &trans->transaction->delayed_refs;
1287                 spin_lock(&delayed_refs->lock);
1288                 head = btrfs_find_delayed_ref_head(trans, bytenr);
1289                 if (head) {
1290                         if (!mutex_trylock(&head->mutex)) {
1291                                 atomic_inc(&head->node.refs);
1292                                 spin_unlock(&delayed_refs->lock);
1293
1294                                 btrfs_release_path(path);
1295
1296                                 /*
1297                                  * Mutex was contended, block until it's
1298                                  * released and try again
1299                                  */
1300                                 mutex_lock(&head->mutex);
1301                                 mutex_unlock(&head->mutex);
1302                                 btrfs_put_delayed_ref(&head->node);
1303                                 goto again;
1304                         }
1305                         spin_unlock(&delayed_refs->lock);
1306                         ret = __add_delayed_refs(head, time_seq,
1307                                                  &prefs_delayed, &total_refs,
1308                                                  inum);
1309                         mutex_unlock(&head->mutex);
1310                         if (ret)
1311                                 goto out;
1312                 } else {
1313                         spin_unlock(&delayed_refs->lock);
1314                 }
1315
1316                 if (check_shared && !list_empty(&prefs_delayed)) {
1317                         /*
1318                          * Add all delay_ref to the ref_tree and check if there
1319                          * are multiple ref items added.
1320                          */
1321                         list_for_each_entry(ref, &prefs_delayed, list) {
1322                                 if (ref->key_for_search.type) {
1323                                         ret = ref_tree_add(ref_tree,
1324                                                 ref->root_id,
1325                                                 ref->key_for_search.objectid,
1326                                                 ref->key_for_search.offset,
1327                                                 0, ref->count);
1328                                         if (ret)
1329                                                 goto out;
1330                                 } else {
1331                                         ret = ref_tree_add(ref_tree, 0, 0, 0,
1332                                                      ref->parent, ref->count);
1333                                         if (ret)
1334                                                 goto out;
1335                                 }
1336
1337                         }
1338
1339                         if (ref_tree->unique_refs > 1) {
1340                                 ret = BACKREF_FOUND_SHARED;
1341                                 goto out;
1342                         }
1343
1344                 }
1345         }
1346
1347         if (path->slots[0]) {
1348                 struct extent_buffer *leaf;
1349                 int slot;
1350
1351                 path->slots[0]--;
1352                 leaf = path->nodes[0];
1353                 slot = path->slots[0];
1354                 btrfs_item_key_to_cpu(leaf, &key, slot);
1355                 if (key.objectid == bytenr &&
1356                     (key.type == BTRFS_EXTENT_ITEM_KEY ||
1357                      key.type == BTRFS_METADATA_ITEM_KEY)) {
1358                         ret = __add_inline_refs(fs_info, path, bytenr,
1359                                                 &info_level, &prefs,
1360                                                 ref_tree, &total_refs,
1361                                                 inum);
1362                         if (ret)
1363                                 goto out;
1364                         ret = __add_keyed_refs(fs_info, path, bytenr,
1365                                                info_level, &prefs,
1366                                                ref_tree, inum);
1367                         if (ret)
1368                                 goto out;
1369                 }
1370         }
1371         btrfs_release_path(path);
1372
1373         list_splice_init(&prefs_delayed, &prefs);
1374
1375         ret = __add_missing_keys(fs_info, &prefs);
1376         if (ret)
1377                 goto out;
1378
1379         __merge_refs(&prefs, 1);
1380
1381         ret = __resolve_indirect_refs(fs_info, path, time_seq, &prefs,
1382                                       extent_item_pos, total_refs,
1383                                       root_objectid);
1384         if (ret)
1385                 goto out;
1386
1387         __merge_refs(&prefs, 2);
1388
1389         while (!list_empty(&prefs)) {
1390                 ref = list_first_entry(&prefs, struct __prelim_ref, list);
1391                 WARN_ON(ref->count < 0);
1392                 if (roots && ref->count && ref->root_id && ref->parent == 0) {
1393                         if (root_objectid && ref->root_id != root_objectid) {
1394                                 ret = BACKREF_FOUND_SHARED;
1395                                 goto out;
1396                         }
1397
1398                         /* no parent == root of tree */
1399                         ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
1400                         if (ret < 0)
1401                                 goto out;
1402                 }
1403                 if (ref->count && ref->parent) {
1404                         if (extent_item_pos && !ref->inode_list &&
1405                             ref->level == 0) {
1406                                 struct extent_buffer *eb;
1407
1408                                 eb = read_tree_block(fs_info->extent_root,
1409                                                            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->extent_root->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                 pr_debug("logical %llu is not within any extent\n", logical);
1839                 return -ENOENT;
1840         }
1841
1842         eb = path->nodes[0];
1843         item_size = btrfs_item_size_nr(eb, path->slots[0]);
1844         BUG_ON(item_size < sizeof(*ei));
1845
1846         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
1847         flags = btrfs_extent_flags(eb, ei);
1848
1849         pr_debug("logical %llu is at position %llu within the extent (%llu "
1850                  "EXTENT_ITEM %llu) flags %#llx size %u\n",
1851                  logical, logical - found_key->objectid, found_key->objectid,
1852                  found_key->offset, flags, item_size);
1853
1854         WARN_ON(!flags_ret);
1855         if (flags_ret) {
1856                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1857                         *flags_ret = BTRFS_EXTENT_FLAG_TREE_BLOCK;
1858                 else if (flags & BTRFS_EXTENT_FLAG_DATA)
1859                         *flags_ret = BTRFS_EXTENT_FLAG_DATA;
1860                 else
1861                         BUG_ON(1);
1862                 return 0;
1863         }
1864
1865         return -EIO;
1866 }
1867
1868 /*
1869  * helper function to iterate extent inline refs. ptr must point to a 0 value
1870  * for the first call and may be modified. it is used to track state.
1871  * if more refs exist, 0 is returned and the next call to
1872  * __get_extent_inline_ref must pass the modified ptr parameter to get the
1873  * next ref. after the last ref was processed, 1 is returned.
1874  * returns <0 on error
1875  */
1876 static int __get_extent_inline_ref(unsigned long *ptr, struct extent_buffer *eb,
1877                                    struct btrfs_key *key,
1878                                    struct btrfs_extent_item *ei, u32 item_size,
1879                                    struct btrfs_extent_inline_ref **out_eiref,
1880                                    int *out_type)
1881 {
1882         unsigned long end;
1883         u64 flags;
1884         struct btrfs_tree_block_info *info;
1885
1886         if (!*ptr) {
1887                 /* first call */
1888                 flags = btrfs_extent_flags(eb, ei);
1889                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1890                         if (key->type == BTRFS_METADATA_ITEM_KEY) {
1891                                 /* a skinny metadata extent */
1892                                 *out_eiref =
1893                                      (struct btrfs_extent_inline_ref *)(ei + 1);
1894                         } else {
1895                                 WARN_ON(key->type != BTRFS_EXTENT_ITEM_KEY);
1896                                 info = (struct btrfs_tree_block_info *)(ei + 1);
1897                                 *out_eiref =
1898                                    (struct btrfs_extent_inline_ref *)(info + 1);
1899                         }
1900                 } else {
1901                         *out_eiref = (struct btrfs_extent_inline_ref *)(ei + 1);
1902                 }
1903                 *ptr = (unsigned long)*out_eiref;
1904                 if ((unsigned long)(*ptr) >= (unsigned long)ei + item_size)
1905                         return -ENOENT;
1906         }
1907
1908         end = (unsigned long)ei + item_size;
1909         *out_eiref = (struct btrfs_extent_inline_ref *)(*ptr);
1910         *out_type = btrfs_extent_inline_ref_type(eb, *out_eiref);
1911
1912         *ptr += btrfs_extent_inline_ref_size(*out_type);
1913         WARN_ON(*ptr > end);
1914         if (*ptr == end)
1915                 return 1; /* last */
1916
1917         return 0;
1918 }
1919
1920 /*
1921  * reads the tree block backref for an extent. tree level and root are returned
1922  * through out_level and out_root. ptr must point to a 0 value for the first
1923  * call and may be modified (see __get_extent_inline_ref comment).
1924  * returns 0 if data was provided, 1 if there was no more data to provide or
1925  * <0 on error.
1926  */
1927 int tree_backref_for_extent(unsigned long *ptr, struct extent_buffer *eb,
1928                             struct btrfs_key *key, struct btrfs_extent_item *ei,
1929                             u32 item_size, u64 *out_root, u8 *out_level)
1930 {
1931         int ret;
1932         int type;
1933         struct btrfs_extent_inline_ref *eiref;
1934
1935         if (*ptr == (unsigned long)-1)
1936                 return 1;
1937
1938         while (1) {
1939                 ret = __get_extent_inline_ref(ptr, eb, key, ei, item_size,
1940                                               &eiref, &type);
1941                 if (ret < 0)
1942                         return ret;
1943
1944                 if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1945                     type == BTRFS_SHARED_BLOCK_REF_KEY)
1946                         break;
1947
1948                 if (ret == 1)
1949                         return 1;
1950         }
1951
1952         /* we can treat both ref types equally here */
1953         *out_root = btrfs_extent_inline_ref_offset(eb, eiref);
1954
1955         if (key->type == BTRFS_EXTENT_ITEM_KEY) {
1956                 struct btrfs_tree_block_info *info;
1957
1958                 info = (struct btrfs_tree_block_info *)(ei + 1);
1959                 *out_level = btrfs_tree_block_level(eb, info);
1960         } else {
1961                 ASSERT(key->type == BTRFS_METADATA_ITEM_KEY);
1962                 *out_level = (u8)key->offset;
1963         }
1964
1965         if (ret == 1)
1966                 *ptr = (unsigned long)-1;
1967
1968         return 0;
1969 }
1970
1971 static int iterate_leaf_refs(struct extent_inode_elem *inode_list,
1972                                 u64 root, u64 extent_item_objectid,
1973                                 iterate_extent_inodes_t *iterate, void *ctx)
1974 {
1975         struct extent_inode_elem *eie;
1976         int ret = 0;
1977
1978         for (eie = inode_list; eie; eie = eie->next) {
1979                 pr_debug("ref for %llu resolved, key (%llu EXTEND_DATA %llu), "
1980                          "root %llu\n", extent_item_objectid,
1981                          eie->inum, eie->offset, root);
1982                 ret = iterate(eie->inum, eie->offset, root, ctx);
1983                 if (ret) {
1984                         pr_debug("stopping iteration for %llu due to ret=%d\n",
1985                                  extent_item_objectid, ret);
1986                         break;
1987                 }
1988         }
1989
1990         return ret;
1991 }
1992
1993 /*
1994  * calls iterate() for every inode that references the extent identified by
1995  * the given parameters.
1996  * when the iterator function returns a non-zero value, iteration stops.
1997  */
1998 int iterate_extent_inodes(struct btrfs_fs_info *fs_info,
1999                                 u64 extent_item_objectid, u64 extent_item_pos,
2000                                 int search_commit_root,
2001                                 iterate_extent_inodes_t *iterate, void *ctx)
2002 {
2003         int ret;
2004         struct btrfs_trans_handle *trans = NULL;
2005         struct ulist *refs = NULL;
2006         struct ulist *roots = NULL;
2007         struct ulist_node *ref_node = NULL;
2008         struct ulist_node *root_node = NULL;
2009         struct seq_list tree_mod_seq_elem = SEQ_LIST_INIT(tree_mod_seq_elem);
2010         struct ulist_iterator ref_uiter;
2011         struct ulist_iterator root_uiter;
2012
2013         pr_debug("resolving all inodes for extent %llu\n",
2014                         extent_item_objectid);
2015
2016         if (!search_commit_root) {
2017                 trans = btrfs_join_transaction(fs_info->extent_root);
2018                 if (IS_ERR(trans))
2019                         return PTR_ERR(trans);
2020                 btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2021         } else {
2022                 down_read(&fs_info->commit_root_sem);
2023         }
2024
2025         ret = btrfs_find_all_leafs(trans, fs_info, extent_item_objectid,
2026                                    tree_mod_seq_elem.seq, &refs,
2027                                    &extent_item_pos);
2028         if (ret)
2029                 goto out;
2030
2031         ULIST_ITER_INIT(&ref_uiter);
2032         while (!ret && (ref_node = ulist_next(refs, &ref_uiter))) {
2033                 ret = __btrfs_find_all_roots(trans, fs_info, ref_node->val,
2034                                              tree_mod_seq_elem.seq, &roots);
2035                 if (ret)
2036                         break;
2037                 ULIST_ITER_INIT(&root_uiter);
2038                 while (!ret && (root_node = ulist_next(roots, &root_uiter))) {
2039                         pr_debug("root %llu references leaf %llu, data list "
2040                                  "%#llx\n", root_node->val, ref_node->val,
2041                                  ref_node->aux);
2042                         ret = iterate_leaf_refs((struct extent_inode_elem *)
2043                                                 (uintptr_t)ref_node->aux,
2044                                                 root_node->val,
2045                                                 extent_item_objectid,
2046                                                 iterate, ctx);
2047                 }
2048                 ulist_free(roots);
2049         }
2050
2051         free_leaf_list(refs);
2052 out:
2053         if (!search_commit_root) {
2054                 btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2055                 btrfs_end_transaction(trans, fs_info->extent_root);
2056         } else {
2057                 up_read(&fs_info->commit_root_sem);
2058         }
2059
2060         return ret;
2061 }
2062
2063 int iterate_inodes_from_logical(u64 logical, struct btrfs_fs_info *fs_info,
2064                                 struct btrfs_path *path,
2065                                 iterate_extent_inodes_t *iterate, void *ctx)
2066 {
2067         int ret;
2068         u64 extent_item_pos;
2069         u64 flags = 0;
2070         struct btrfs_key found_key;
2071         int search_commit_root = path->search_commit_root;
2072
2073         ret = extent_from_logical(fs_info, logical, path, &found_key, &flags);
2074         btrfs_release_path(path);
2075         if (ret < 0)
2076                 return ret;
2077         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2078                 return -EINVAL;
2079
2080         extent_item_pos = logical - found_key.objectid;
2081         ret = iterate_extent_inodes(fs_info, found_key.objectid,
2082                                         extent_item_pos, search_commit_root,
2083                                         iterate, ctx);
2084
2085         return ret;
2086 }
2087
2088 typedef int (iterate_irefs_t)(u64 parent, u32 name_len, unsigned long name_off,
2089                               struct extent_buffer *eb, void *ctx);
2090
2091 static int iterate_inode_refs(u64 inum, struct btrfs_root *fs_root,
2092                               struct btrfs_path *path,
2093                               iterate_irefs_t *iterate, void *ctx)
2094 {
2095         int ret = 0;
2096         int slot;
2097         u32 cur;
2098         u32 len;
2099         u32 name_len;
2100         u64 parent = 0;
2101         int found = 0;
2102         struct extent_buffer *eb;
2103         struct btrfs_item *item;
2104         struct btrfs_inode_ref *iref;
2105         struct btrfs_key found_key;
2106
2107         while (!ret) {
2108                 ret = btrfs_find_item(fs_root, path, inum,
2109                                 parent ? parent + 1 : 0, BTRFS_INODE_REF_KEY,
2110                                 &found_key);
2111
2112                 if (ret < 0)
2113                         break;
2114                 if (ret) {
2115                         ret = found ? 0 : -ENOENT;
2116                         break;
2117                 }
2118                 ++found;
2119
2120                 parent = found_key.offset;
2121                 slot = path->slots[0];
2122                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2123                 if (!eb) {
2124                         ret = -ENOMEM;
2125                         break;
2126                 }
2127                 extent_buffer_get(eb);
2128                 btrfs_tree_read_lock(eb);
2129                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2130                 btrfs_release_path(path);
2131
2132                 item = btrfs_item_nr(slot);
2133                 iref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
2134
2135                 for (cur = 0; cur < btrfs_item_size(eb, item); cur += len) {
2136                         name_len = btrfs_inode_ref_name_len(eb, iref);
2137                         /* path must be released before calling iterate()! */
2138                         pr_debug("following ref at offset %u for inode %llu in "
2139                                  "tree %llu\n", cur, found_key.objectid,
2140                                  fs_root->objectid);
2141                         ret = iterate(parent, name_len,
2142                                       (unsigned long)(iref + 1), eb, ctx);
2143                         if (ret)
2144                                 break;
2145                         len = sizeof(*iref) + name_len;
2146                         iref = (struct btrfs_inode_ref *)((char *)iref + len);
2147                 }
2148                 btrfs_tree_read_unlock_blocking(eb);
2149                 free_extent_buffer(eb);
2150         }
2151
2152         btrfs_release_path(path);
2153
2154         return ret;
2155 }
2156
2157 static int iterate_inode_extrefs(u64 inum, struct btrfs_root *fs_root,
2158                                  struct btrfs_path *path,
2159                                  iterate_irefs_t *iterate, void *ctx)
2160 {
2161         int ret;
2162         int slot;
2163         u64 offset = 0;
2164         u64 parent;
2165         int found = 0;
2166         struct extent_buffer *eb;
2167         struct btrfs_inode_extref *extref;
2168         u32 item_size;
2169         u32 cur_offset;
2170         unsigned long ptr;
2171
2172         while (1) {
2173                 ret = btrfs_find_one_extref(fs_root, inum, offset, path, &extref,
2174                                             &offset);
2175                 if (ret < 0)
2176                         break;
2177                 if (ret) {
2178                         ret = found ? 0 : -ENOENT;
2179                         break;
2180                 }
2181                 ++found;
2182
2183                 slot = path->slots[0];
2184                 eb = btrfs_clone_extent_buffer(path->nodes[0]);
2185                 if (!eb) {
2186                         ret = -ENOMEM;
2187                         break;
2188                 }
2189                 extent_buffer_get(eb);
2190
2191                 btrfs_tree_read_lock(eb);
2192                 btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
2193                 btrfs_release_path(path);
2194
2195                 item_size = btrfs_item_size_nr(eb, slot);
2196                 ptr = btrfs_item_ptr_offset(eb, slot);
2197                 cur_offset = 0;
2198
2199                 while (cur_offset < item_size) {
2200                         u32 name_len;
2201
2202                         extref = (struct btrfs_inode_extref *)(ptr + cur_offset);
2203                         parent = btrfs_inode_extref_parent(eb, extref);
2204                         name_len = btrfs_inode_extref_name_len(eb, extref);
2205                         ret = iterate(parent, name_len,
2206                                       (unsigned long)&extref->name, eb, ctx);
2207                         if (ret)
2208                                 break;
2209
2210                         cur_offset += btrfs_inode_extref_name_len(eb, extref);
2211                         cur_offset += sizeof(*extref);
2212                 }
2213                 btrfs_tree_read_unlock_blocking(eb);
2214                 free_extent_buffer(eb);
2215
2216                 offset++;
2217         }
2218
2219         btrfs_release_path(path);
2220
2221         return ret;
2222 }
2223
2224 static int iterate_irefs(u64 inum, struct btrfs_root *fs_root,
2225                          struct btrfs_path *path, iterate_irefs_t *iterate,
2226                          void *ctx)
2227 {
2228         int ret;
2229         int found_refs = 0;
2230
2231         ret = iterate_inode_refs(inum, fs_root, path, iterate, ctx);
2232         if (!ret)
2233                 ++found_refs;
2234         else if (ret != -ENOENT)
2235                 return ret;
2236
2237         ret = iterate_inode_extrefs(inum, fs_root, path, iterate, ctx);
2238         if (ret == -ENOENT && found_refs)
2239                 return 0;
2240
2241         return ret;
2242 }
2243
2244 /*
2245  * returns 0 if the path could be dumped (probably truncated)
2246  * returns <0 in case of an error
2247  */
2248 static int inode_to_path(u64 inum, u32 name_len, unsigned long name_off,
2249                          struct extent_buffer *eb, void *ctx)
2250 {
2251         struct inode_fs_paths *ipath = ctx;
2252         char *fspath;
2253         char *fspath_min;
2254         int i = ipath->fspath->elem_cnt;
2255         const int s_ptr = sizeof(char *);
2256         u32 bytes_left;
2257
2258         bytes_left = ipath->fspath->bytes_left > s_ptr ?
2259                                         ipath->fspath->bytes_left - s_ptr : 0;
2260
2261         fspath_min = (char *)ipath->fspath->val + (i + 1) * s_ptr;
2262         fspath = btrfs_ref_to_path(ipath->fs_root, ipath->btrfs_path, name_len,
2263                                    name_off, eb, inum, fspath_min, bytes_left);
2264         if (IS_ERR(fspath))
2265                 return PTR_ERR(fspath);
2266
2267         if (fspath > fspath_min) {
2268                 ipath->fspath->val[i] = (u64)(unsigned long)fspath;
2269                 ++ipath->fspath->elem_cnt;
2270                 ipath->fspath->bytes_left = fspath - fspath_min;
2271         } else {
2272                 ++ipath->fspath->elem_missed;
2273                 ipath->fspath->bytes_missing += fspath_min - fspath;
2274                 ipath->fspath->bytes_left = 0;
2275         }
2276
2277         return 0;
2278 }
2279
2280 /*
2281  * this dumps all file system paths to the inode into the ipath struct, provided
2282  * is has been created large enough. each path is zero-terminated and accessed
2283  * from ipath->fspath->val[i].
2284  * when it returns, there are ipath->fspath->elem_cnt number of paths available
2285  * in ipath->fspath->val[]. when the allocated space wasn't sufficient, the
2286  * number of missed paths is recorded in ipath->fspath->elem_missed, otherwise,
2287  * it's zero. ipath->fspath->bytes_missing holds the number of bytes that would
2288  * have been needed to return all paths.
2289  */
2290 int paths_from_inode(u64 inum, struct inode_fs_paths *ipath)
2291 {
2292         return iterate_irefs(inum, ipath->fs_root, ipath->btrfs_path,
2293                              inode_to_path, ipath);
2294 }
2295
2296 struct btrfs_data_container *init_data_container(u32 total_bytes)
2297 {
2298         struct btrfs_data_container *data;
2299         size_t alloc_bytes;
2300
2301         alloc_bytes = max_t(size_t, total_bytes, sizeof(*data));
2302         data = vmalloc(alloc_bytes);
2303         if (!data)
2304                 return ERR_PTR(-ENOMEM);
2305
2306         if (total_bytes >= sizeof(*data)) {
2307                 data->bytes_left = total_bytes - sizeof(*data);
2308                 data->bytes_missing = 0;
2309         } else {
2310                 data->bytes_missing = sizeof(*data) - total_bytes;
2311                 data->bytes_left = 0;
2312         }
2313
2314         data->elem_cnt = 0;
2315         data->elem_missed = 0;
2316
2317         return data;
2318 }
2319
2320 /*
2321  * allocates space to return multiple file system paths for an inode.
2322  * total_bytes to allocate are passed, note that space usable for actual path
2323  * information will be total_bytes - sizeof(struct inode_fs_paths).
2324  * the returned pointer must be freed with free_ipath() in the end.
2325  */
2326 struct inode_fs_paths *init_ipath(s32 total_bytes, struct btrfs_root *fs_root,
2327                                         struct btrfs_path *path)
2328 {
2329         struct inode_fs_paths *ifp;
2330         struct btrfs_data_container *fspath;
2331
2332         fspath = init_data_container(total_bytes);
2333         if (IS_ERR(fspath))
2334                 return (void *)fspath;
2335
2336         ifp = kmalloc(sizeof(*ifp), GFP_NOFS);
2337         if (!ifp) {
2338                 vfree(fspath);
2339                 return ERR_PTR(-ENOMEM);
2340         }
2341
2342         ifp->btrfs_path = path;
2343         ifp->fspath = fspath;
2344         ifp->fs_root = fs_root;
2345
2346         return ifp;
2347 }
2348
2349 void free_ipath(struct inode_fs_paths *ipath)
2350 {
2351         if (!ipath)
2352                 return;
2353         vfree(ipath->fspath);
2354         kfree(ipath);
2355 }