GNU Linux-libre 5.10.215-gnu1
[releases.git] / fs / btrfs / relocation.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2009 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/pagemap.h>
8 #include <linux/writeback.h>
9 #include <linux/blkdev.h>
10 #include <linux/rbtree.h>
11 #include <linux/slab.h>
12 #include <linux/error-injection.h>
13 #include "ctree.h"
14 #include "disk-io.h"
15 #include "transaction.h"
16 #include "volumes.h"
17 #include "locking.h"
18 #include "btrfs_inode.h"
19 #include "async-thread.h"
20 #include "free-space-cache.h"
21 #include "inode-map.h"
22 #include "qgroup.h"
23 #include "print-tree.h"
24 #include "delalloc-space.h"
25 #include "block-group.h"
26 #include "backref.h"
27 #include "misc.h"
28
29 /*
30  * Relocation overview
31  *
32  * [What does relocation do]
33  *
34  * The objective of relocation is to relocate all extents of the target block
35  * group to other block groups.
36  * This is utilized by resize (shrink only), profile converting, compacting
37  * space, or balance routine to spread chunks over devices.
38  *
39  *              Before          |               After
40  * ------------------------------------------------------------------
41  *  BG A: 10 data extents       | BG A: deleted
42  *  BG B:  2 data extents       | BG B: 10 data extents (2 old + 8 relocated)
43  *  BG C:  1 extents            | BG C:  3 data extents (1 old + 2 relocated)
44  *
45  * [How does relocation work]
46  *
47  * 1.   Mark the target block group read-only
48  *      New extents won't be allocated from the target block group.
49  *
50  * 2.1  Record each extent in the target block group
51  *      To build a proper map of extents to be relocated.
52  *
53  * 2.2  Build data reloc tree and reloc trees
54  *      Data reloc tree will contain an inode, recording all newly relocated
55  *      data extents.
56  *      There will be only one data reloc tree for one data block group.
57  *
58  *      Reloc tree will be a special snapshot of its source tree, containing
59  *      relocated tree blocks.
60  *      Each tree referring to a tree block in target block group will get its
61  *      reloc tree built.
62  *
63  * 2.3  Swap source tree with its corresponding reloc tree
64  *      Each involved tree only refers to new extents after swap.
65  *
66  * 3.   Cleanup reloc trees and data reloc tree.
67  *      As old extents in the target block group are still referenced by reloc
68  *      trees, we need to clean them up before really freeing the target block
69  *      group.
70  *
71  * The main complexity is in steps 2.2 and 2.3.
72  *
73  * The entry point of relocation is relocate_block_group() function.
74  */
75
76 #define RELOCATION_RESERVED_NODES       256
77 /*
78  * map address of tree root to tree
79  */
80 struct mapping_node {
81         struct {
82                 struct rb_node rb_node;
83                 u64 bytenr;
84         }; /* Use rb_simle_node for search/insert */
85         void *data;
86 };
87
88 struct mapping_tree {
89         struct rb_root rb_root;
90         spinlock_t lock;
91 };
92
93 /*
94  * present a tree block to process
95  */
96 struct tree_block {
97         struct {
98                 struct rb_node rb_node;
99                 u64 bytenr;
100         }; /* Use rb_simple_node for search/insert */
101         struct btrfs_key key;
102         unsigned int level:8;
103         unsigned int key_ready:1;
104 };
105
106 #define MAX_EXTENTS 128
107
108 struct file_extent_cluster {
109         u64 start;
110         u64 end;
111         u64 boundary[MAX_EXTENTS];
112         unsigned int nr;
113 };
114
115 struct reloc_control {
116         /* block group to relocate */
117         struct btrfs_block_group *block_group;
118         /* extent tree */
119         struct btrfs_root *extent_root;
120         /* inode for moving data */
121         struct inode *data_inode;
122
123         struct btrfs_block_rsv *block_rsv;
124
125         struct btrfs_backref_cache backref_cache;
126
127         struct file_extent_cluster cluster;
128         /* tree blocks have been processed */
129         struct extent_io_tree processed_blocks;
130         /* map start of tree root to corresponding reloc tree */
131         struct mapping_tree reloc_root_tree;
132         /* list of reloc trees */
133         struct list_head reloc_roots;
134         /* list of subvolume trees that get relocated */
135         struct list_head dirty_subvol_roots;
136         /* size of metadata reservation for merging reloc trees */
137         u64 merging_rsv_size;
138         /* size of relocated tree nodes */
139         u64 nodes_relocated;
140         /* reserved size for block group relocation*/
141         u64 reserved_bytes;
142
143         u64 search_start;
144         u64 extents_found;
145
146         unsigned int stage:8;
147         unsigned int create_reloc_tree:1;
148         unsigned int merge_reloc_tree:1;
149         unsigned int found_file_extent:1;
150 };
151
152 /* stages of data relocation */
153 #define MOVE_DATA_EXTENTS       0
154 #define UPDATE_DATA_PTRS        1
155
156 static void mark_block_processed(struct reloc_control *rc,
157                                  struct btrfs_backref_node *node)
158 {
159         u32 blocksize;
160
161         if (node->level == 0 ||
162             in_range(node->bytenr, rc->block_group->start,
163                      rc->block_group->length)) {
164                 blocksize = rc->extent_root->fs_info->nodesize;
165                 set_extent_bits(&rc->processed_blocks, node->bytenr,
166                                 node->bytenr + blocksize - 1, EXTENT_DIRTY);
167         }
168         node->processed = 1;
169 }
170
171
172 static void mapping_tree_init(struct mapping_tree *tree)
173 {
174         tree->rb_root = RB_ROOT;
175         spin_lock_init(&tree->lock);
176 }
177
178 /*
179  * walk up backref nodes until reach node presents tree root
180  */
181 static struct btrfs_backref_node *walk_up_backref(
182                 struct btrfs_backref_node *node,
183                 struct btrfs_backref_edge *edges[], int *index)
184 {
185         struct btrfs_backref_edge *edge;
186         int idx = *index;
187
188         while (!list_empty(&node->upper)) {
189                 edge = list_entry(node->upper.next,
190                                   struct btrfs_backref_edge, list[LOWER]);
191                 edges[idx++] = edge;
192                 node = edge->node[UPPER];
193         }
194         BUG_ON(node->detached);
195         *index = idx;
196         return node;
197 }
198
199 /*
200  * walk down backref nodes to find start of next reference path
201  */
202 static struct btrfs_backref_node *walk_down_backref(
203                 struct btrfs_backref_edge *edges[], int *index)
204 {
205         struct btrfs_backref_edge *edge;
206         struct btrfs_backref_node *lower;
207         int idx = *index;
208
209         while (idx > 0) {
210                 edge = edges[idx - 1];
211                 lower = edge->node[LOWER];
212                 if (list_is_last(&edge->list[LOWER], &lower->upper)) {
213                         idx--;
214                         continue;
215                 }
216                 edge = list_entry(edge->list[LOWER].next,
217                                   struct btrfs_backref_edge, list[LOWER]);
218                 edges[idx - 1] = edge;
219                 *index = idx;
220                 return edge->node[UPPER];
221         }
222         *index = 0;
223         return NULL;
224 }
225
226 static void update_backref_node(struct btrfs_backref_cache *cache,
227                                 struct btrfs_backref_node *node, u64 bytenr)
228 {
229         struct rb_node *rb_node;
230         rb_erase(&node->rb_node, &cache->rb_root);
231         node->bytenr = bytenr;
232         rb_node = rb_simple_insert(&cache->rb_root, node->bytenr, &node->rb_node);
233         if (rb_node)
234                 btrfs_backref_panic(cache->fs_info, bytenr, -EEXIST);
235 }
236
237 /*
238  * update backref cache after a transaction commit
239  */
240 static int update_backref_cache(struct btrfs_trans_handle *trans,
241                                 struct btrfs_backref_cache *cache)
242 {
243         struct btrfs_backref_node *node;
244         int level = 0;
245
246         if (cache->last_trans == 0) {
247                 cache->last_trans = trans->transid;
248                 return 0;
249         }
250
251         if (cache->last_trans == trans->transid)
252                 return 0;
253
254         /*
255          * detached nodes are used to avoid unnecessary backref
256          * lookup. transaction commit changes the extent tree.
257          * so the detached nodes are no longer useful.
258          */
259         while (!list_empty(&cache->detached)) {
260                 node = list_entry(cache->detached.next,
261                                   struct btrfs_backref_node, list);
262                 btrfs_backref_cleanup_node(cache, node);
263         }
264
265         while (!list_empty(&cache->changed)) {
266                 node = list_entry(cache->changed.next,
267                                   struct btrfs_backref_node, list);
268                 list_del_init(&node->list);
269                 BUG_ON(node->pending);
270                 update_backref_node(cache, node, node->new_bytenr);
271         }
272
273         /*
274          * some nodes can be left in the pending list if there were
275          * errors during processing the pending nodes.
276          */
277         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
278                 list_for_each_entry(node, &cache->pending[level], list) {
279                         BUG_ON(!node->pending);
280                         if (node->bytenr == node->new_bytenr)
281                                 continue;
282                         update_backref_node(cache, node, node->new_bytenr);
283                 }
284         }
285
286         cache->last_trans = 0;
287         return 1;
288 }
289
290 static bool reloc_root_is_dead(struct btrfs_root *root)
291 {
292         /*
293          * Pair with set_bit/clear_bit in clean_dirty_subvols and
294          * btrfs_update_reloc_root. We need to see the updated bit before
295          * trying to access reloc_root
296          */
297         smp_rmb();
298         if (test_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state))
299                 return true;
300         return false;
301 }
302
303 /*
304  * Check if this subvolume tree has valid reloc tree.
305  *
306  * Reloc tree after swap is considered dead, thus not considered as valid.
307  * This is enough for most callers, as they don't distinguish dead reloc root
308  * from no reloc root.  But btrfs_should_ignore_reloc_root() below is a
309  * special case.
310  */
311 static bool have_reloc_root(struct btrfs_root *root)
312 {
313         if (reloc_root_is_dead(root))
314                 return false;
315         if (!root->reloc_root)
316                 return false;
317         return true;
318 }
319
320 int btrfs_should_ignore_reloc_root(struct btrfs_root *root)
321 {
322         struct btrfs_root *reloc_root;
323
324         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
325                 return 0;
326
327         /* This root has been merged with its reloc tree, we can ignore it */
328         if (reloc_root_is_dead(root))
329                 return 1;
330
331         reloc_root = root->reloc_root;
332         if (!reloc_root)
333                 return 0;
334
335         if (btrfs_header_generation(reloc_root->commit_root) ==
336             root->fs_info->running_transaction->transid)
337                 return 0;
338         /*
339          * if there is reloc tree and it was created in previous
340          * transaction backref lookup can find the reloc tree,
341          * so backref node for the fs tree root is useless for
342          * relocation.
343          */
344         return 1;
345 }
346
347 /*
348  * find reloc tree by address of tree root
349  */
350 struct btrfs_root *find_reloc_root(struct btrfs_fs_info *fs_info, u64 bytenr)
351 {
352         struct reloc_control *rc = fs_info->reloc_ctl;
353         struct rb_node *rb_node;
354         struct mapping_node *node;
355         struct btrfs_root *root = NULL;
356
357         ASSERT(rc);
358         spin_lock(&rc->reloc_root_tree.lock);
359         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root, bytenr);
360         if (rb_node) {
361                 node = rb_entry(rb_node, struct mapping_node, rb_node);
362                 root = (struct btrfs_root *)node->data;
363         }
364         spin_unlock(&rc->reloc_root_tree.lock);
365         return btrfs_grab_root(root);
366 }
367
368 /*
369  * For useless nodes, do two major clean ups:
370  *
371  * - Cleanup the children edges and nodes
372  *   If child node is also orphan (no parent) during cleanup, then the child
373  *   node will also be cleaned up.
374  *
375  * - Freeing up leaves (level 0), keeps nodes detached
376  *   For nodes, the node is still cached as "detached"
377  *
378  * Return false if @node is not in the @useless_nodes list.
379  * Return true if @node is in the @useless_nodes list.
380  */
381 static bool handle_useless_nodes(struct reloc_control *rc,
382                                  struct btrfs_backref_node *node)
383 {
384         struct btrfs_backref_cache *cache = &rc->backref_cache;
385         struct list_head *useless_node = &cache->useless_node;
386         bool ret = false;
387
388         while (!list_empty(useless_node)) {
389                 struct btrfs_backref_node *cur;
390
391                 cur = list_first_entry(useless_node, struct btrfs_backref_node,
392                                  list);
393                 list_del_init(&cur->list);
394
395                 /* Only tree root nodes can be added to @useless_nodes */
396                 ASSERT(list_empty(&cur->upper));
397
398                 if (cur == node)
399                         ret = true;
400
401                 /* The node is the lowest node */
402                 if (cur->lowest) {
403                         list_del_init(&cur->lower);
404                         cur->lowest = 0;
405                 }
406
407                 /* Cleanup the lower edges */
408                 while (!list_empty(&cur->lower)) {
409                         struct btrfs_backref_edge *edge;
410                         struct btrfs_backref_node *lower;
411
412                         edge = list_entry(cur->lower.next,
413                                         struct btrfs_backref_edge, list[UPPER]);
414                         list_del(&edge->list[UPPER]);
415                         list_del(&edge->list[LOWER]);
416                         lower = edge->node[LOWER];
417                         btrfs_backref_free_edge(cache, edge);
418
419                         /* Child node is also orphan, queue for cleanup */
420                         if (list_empty(&lower->upper))
421                                 list_add(&lower->list, useless_node);
422                 }
423                 /* Mark this block processed for relocation */
424                 mark_block_processed(rc, cur);
425
426                 /*
427                  * Backref nodes for tree leaves are deleted from the cache.
428                  * Backref nodes for upper level tree blocks are left in the
429                  * cache to avoid unnecessary backref lookup.
430                  */
431                 if (cur->level > 0) {
432                         list_add(&cur->list, &cache->detached);
433                         cur->detached = 1;
434                 } else {
435                         rb_erase(&cur->rb_node, &cache->rb_root);
436                         btrfs_backref_free_node(cache, cur);
437                 }
438         }
439         return ret;
440 }
441
442 /*
443  * Build backref tree for a given tree block. Root of the backref tree
444  * corresponds the tree block, leaves of the backref tree correspond roots of
445  * b-trees that reference the tree block.
446  *
447  * The basic idea of this function is check backrefs of a given block to find
448  * upper level blocks that reference the block, and then check backrefs of
449  * these upper level blocks recursively. The recursion stops when tree root is
450  * reached or backrefs for the block is cached.
451  *
452  * NOTE: if we find that backrefs for a block are cached, we know backrefs for
453  * all upper level blocks that directly/indirectly reference the block are also
454  * cached.
455  */
456 static noinline_for_stack struct btrfs_backref_node *build_backref_tree(
457                         struct reloc_control *rc, struct btrfs_key *node_key,
458                         int level, u64 bytenr)
459 {
460         struct btrfs_backref_iter *iter;
461         struct btrfs_backref_cache *cache = &rc->backref_cache;
462         /* For searching parent of TREE_BLOCK_REF */
463         struct btrfs_path *path;
464         struct btrfs_backref_node *cur;
465         struct btrfs_backref_node *node = NULL;
466         struct btrfs_backref_edge *edge;
467         int ret;
468         int err = 0;
469
470         iter = btrfs_backref_iter_alloc(rc->extent_root->fs_info, GFP_NOFS);
471         if (!iter)
472                 return ERR_PTR(-ENOMEM);
473         path = btrfs_alloc_path();
474         if (!path) {
475                 err = -ENOMEM;
476                 goto out;
477         }
478
479         node = btrfs_backref_alloc_node(cache, bytenr, level);
480         if (!node) {
481                 err = -ENOMEM;
482                 goto out;
483         }
484
485         node->lowest = 1;
486         cur = node;
487
488         /* Breadth-first search to build backref cache */
489         do {
490                 ret = btrfs_backref_add_tree_node(cache, path, iter, node_key,
491                                                   cur);
492                 if (ret < 0) {
493                         err = ret;
494                         goto out;
495                 }
496                 edge = list_first_entry_or_null(&cache->pending_edge,
497                                 struct btrfs_backref_edge, list[UPPER]);
498                 /*
499                  * The pending list isn't empty, take the first block to
500                  * process
501                  */
502                 if (edge) {
503                         list_del_init(&edge->list[UPPER]);
504                         cur = edge->node[UPPER];
505                 }
506         } while (edge);
507
508         /* Finish the upper linkage of newly added edges/nodes */
509         ret = btrfs_backref_finish_upper_links(cache, node);
510         if (ret < 0) {
511                 err = ret;
512                 goto out;
513         }
514
515         if (handle_useless_nodes(rc, node))
516                 node = NULL;
517 out:
518         btrfs_backref_iter_free(iter);
519         btrfs_free_path(path);
520         if (err) {
521                 btrfs_backref_error_cleanup(cache, node);
522                 return ERR_PTR(err);
523         }
524         ASSERT(!node || !node->detached);
525         ASSERT(list_empty(&cache->useless_node) &&
526                list_empty(&cache->pending_edge));
527         return node;
528 }
529
530 /*
531  * helper to add backref node for the newly created snapshot.
532  * the backref node is created by cloning backref node that
533  * corresponds to root of source tree
534  */
535 static int clone_backref_node(struct btrfs_trans_handle *trans,
536                               struct reloc_control *rc,
537                               struct btrfs_root *src,
538                               struct btrfs_root *dest)
539 {
540         struct btrfs_root *reloc_root = src->reloc_root;
541         struct btrfs_backref_cache *cache = &rc->backref_cache;
542         struct btrfs_backref_node *node = NULL;
543         struct btrfs_backref_node *new_node;
544         struct btrfs_backref_edge *edge;
545         struct btrfs_backref_edge *new_edge;
546         struct rb_node *rb_node;
547
548         if (cache->last_trans > 0)
549                 update_backref_cache(trans, cache);
550
551         rb_node = rb_simple_search(&cache->rb_root, src->commit_root->start);
552         if (rb_node) {
553                 node = rb_entry(rb_node, struct btrfs_backref_node, rb_node);
554                 if (node->detached)
555                         node = NULL;
556                 else
557                         BUG_ON(node->new_bytenr != reloc_root->node->start);
558         }
559
560         if (!node) {
561                 rb_node = rb_simple_search(&cache->rb_root,
562                                            reloc_root->commit_root->start);
563                 if (rb_node) {
564                         node = rb_entry(rb_node, struct btrfs_backref_node,
565                                         rb_node);
566                         BUG_ON(node->detached);
567                 }
568         }
569
570         if (!node)
571                 return 0;
572
573         new_node = btrfs_backref_alloc_node(cache, dest->node->start,
574                                             node->level);
575         if (!new_node)
576                 return -ENOMEM;
577
578         new_node->lowest = node->lowest;
579         new_node->checked = 1;
580         new_node->root = btrfs_grab_root(dest);
581         ASSERT(new_node->root);
582
583         if (!node->lowest) {
584                 list_for_each_entry(edge, &node->lower, list[UPPER]) {
585                         new_edge = btrfs_backref_alloc_edge(cache);
586                         if (!new_edge)
587                                 goto fail;
588
589                         btrfs_backref_link_edge(new_edge, edge->node[LOWER],
590                                                 new_node, LINK_UPPER);
591                 }
592         } else {
593                 list_add_tail(&new_node->lower, &cache->leaves);
594         }
595
596         rb_node = rb_simple_insert(&cache->rb_root, new_node->bytenr,
597                                    &new_node->rb_node);
598         if (rb_node)
599                 btrfs_backref_panic(trans->fs_info, new_node->bytenr, -EEXIST);
600
601         if (!new_node->lowest) {
602                 list_for_each_entry(new_edge, &new_node->lower, list[UPPER]) {
603                         list_add_tail(&new_edge->list[LOWER],
604                                       &new_edge->node[LOWER]->upper);
605                 }
606         }
607         return 0;
608 fail:
609         while (!list_empty(&new_node->lower)) {
610                 new_edge = list_entry(new_node->lower.next,
611                                       struct btrfs_backref_edge, list[UPPER]);
612                 list_del(&new_edge->list[UPPER]);
613                 btrfs_backref_free_edge(cache, new_edge);
614         }
615         btrfs_backref_free_node(cache, new_node);
616         return -ENOMEM;
617 }
618
619 /*
620  * helper to add 'address of tree root -> reloc tree' mapping
621  */
622 static int __must_check __add_reloc_root(struct btrfs_root *root)
623 {
624         struct btrfs_fs_info *fs_info = root->fs_info;
625         struct rb_node *rb_node;
626         struct mapping_node *node;
627         struct reloc_control *rc = fs_info->reloc_ctl;
628
629         node = kmalloc(sizeof(*node), GFP_NOFS);
630         if (!node)
631                 return -ENOMEM;
632
633         node->bytenr = root->commit_root->start;
634         node->data = root;
635
636         spin_lock(&rc->reloc_root_tree.lock);
637         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
638                                    node->bytenr, &node->rb_node);
639         spin_unlock(&rc->reloc_root_tree.lock);
640         if (rb_node) {
641                 btrfs_panic(fs_info, -EEXIST,
642                             "Duplicate root found for start=%llu while inserting into relocation tree",
643                             node->bytenr);
644         }
645
646         list_add_tail(&root->root_list, &rc->reloc_roots);
647         return 0;
648 }
649
650 /*
651  * helper to delete the 'address of tree root -> reloc tree'
652  * mapping
653  */
654 static void __del_reloc_root(struct btrfs_root *root)
655 {
656         struct btrfs_fs_info *fs_info = root->fs_info;
657         struct rb_node *rb_node;
658         struct mapping_node *node = NULL;
659         struct reloc_control *rc = fs_info->reloc_ctl;
660         bool put_ref = false;
661
662         if (rc && root->node) {
663                 spin_lock(&rc->reloc_root_tree.lock);
664                 rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
665                                            root->commit_root->start);
666                 if (rb_node) {
667                         node = rb_entry(rb_node, struct mapping_node, rb_node);
668                         rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
669                         RB_CLEAR_NODE(&node->rb_node);
670                 }
671                 spin_unlock(&rc->reloc_root_tree.lock);
672                 ASSERT(!node || (struct btrfs_root *)node->data == root);
673         }
674
675         /*
676          * We only put the reloc root here if it's on the list.  There's a lot
677          * of places where the pattern is to splice the rc->reloc_roots, process
678          * the reloc roots, and then add the reloc root back onto
679          * rc->reloc_roots.  If we call __del_reloc_root while it's off of the
680          * list we don't want the reference being dropped, because the guy
681          * messing with the list is in charge of the reference.
682          */
683         spin_lock(&fs_info->trans_lock);
684         if (!list_empty(&root->root_list)) {
685                 put_ref = true;
686                 list_del_init(&root->root_list);
687         }
688         spin_unlock(&fs_info->trans_lock);
689         if (put_ref)
690                 btrfs_put_root(root);
691         kfree(node);
692 }
693
694 /*
695  * helper to update the 'address of tree root -> reloc tree'
696  * mapping
697  */
698 static int __update_reloc_root(struct btrfs_root *root)
699 {
700         struct btrfs_fs_info *fs_info = root->fs_info;
701         struct rb_node *rb_node;
702         struct mapping_node *node = NULL;
703         struct reloc_control *rc = fs_info->reloc_ctl;
704
705         spin_lock(&rc->reloc_root_tree.lock);
706         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
707                                    root->commit_root->start);
708         if (rb_node) {
709                 node = rb_entry(rb_node, struct mapping_node, rb_node);
710                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
711         }
712         spin_unlock(&rc->reloc_root_tree.lock);
713
714         if (!node)
715                 return 0;
716         BUG_ON((struct btrfs_root *)node->data != root);
717
718         spin_lock(&rc->reloc_root_tree.lock);
719         node->bytenr = root->node->start;
720         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
721                                    node->bytenr, &node->rb_node);
722         spin_unlock(&rc->reloc_root_tree.lock);
723         if (rb_node)
724                 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
725         return 0;
726 }
727
728 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
729                                         struct btrfs_root *root, u64 objectid)
730 {
731         struct btrfs_fs_info *fs_info = root->fs_info;
732         struct btrfs_root *reloc_root;
733         struct extent_buffer *eb;
734         struct btrfs_root_item *root_item;
735         struct btrfs_key root_key;
736         int ret = 0;
737         bool must_abort = false;
738
739         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
740         if (!root_item)
741                 return ERR_PTR(-ENOMEM);
742
743         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
744         root_key.type = BTRFS_ROOT_ITEM_KEY;
745         root_key.offset = objectid;
746
747         if (root->root_key.objectid == objectid) {
748                 u64 commit_root_gen;
749
750                 /* called by btrfs_init_reloc_root */
751                 ret = btrfs_copy_root(trans, root, root->commit_root, &eb,
752                                       BTRFS_TREE_RELOC_OBJECTID);
753                 if (ret)
754                         goto fail;
755
756                 /*
757                  * Set the last_snapshot field to the generation of the commit
758                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
759                  * correctly (returns true) when the relocation root is created
760                  * either inside the critical section of a transaction commit
761                  * (through transaction.c:qgroup_account_snapshot()) and when
762                  * it's created before the transaction commit is started.
763                  */
764                 commit_root_gen = btrfs_header_generation(root->commit_root);
765                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
766         } else {
767                 /*
768                  * called by btrfs_reloc_post_snapshot_hook.
769                  * the source tree is a reloc tree, all tree blocks
770                  * modified after it was created have RELOC flag
771                  * set in their headers. so it's OK to not update
772                  * the 'last_snapshot'.
773                  */
774                 ret = btrfs_copy_root(trans, root, root->node, &eb,
775                                       BTRFS_TREE_RELOC_OBJECTID);
776                 if (ret)
777                         goto fail;
778         }
779
780         /*
781          * We have changed references at this point, we must abort the
782          * transaction if anything fails.
783          */
784         must_abort = true;
785
786         memcpy(root_item, &root->root_item, sizeof(*root_item));
787         btrfs_set_root_bytenr(root_item, eb->start);
788         btrfs_set_root_level(root_item, btrfs_header_level(eb));
789         btrfs_set_root_generation(root_item, trans->transid);
790
791         if (root->root_key.objectid == objectid) {
792                 btrfs_set_root_refs(root_item, 0);
793                 memset(&root_item->drop_progress, 0,
794                        sizeof(struct btrfs_disk_key));
795                 root_item->drop_level = 0;
796         }
797
798         btrfs_tree_unlock(eb);
799         free_extent_buffer(eb);
800
801         ret = btrfs_insert_root(trans, fs_info->tree_root,
802                                 &root_key, root_item);
803         if (ret)
804                 goto fail;
805
806         kfree(root_item);
807
808         reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
809         if (IS_ERR(reloc_root)) {
810                 ret = PTR_ERR(reloc_root);
811                 goto abort;
812         }
813         set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
814         reloc_root->last_trans = trans->transid;
815         return reloc_root;
816 fail:
817         kfree(root_item);
818 abort:
819         if (must_abort)
820                 btrfs_abort_transaction(trans, ret);
821         return ERR_PTR(ret);
822 }
823
824 /*
825  * create reloc tree for a given fs tree. reloc tree is just a
826  * snapshot of the fs tree with special root objectid.
827  *
828  * The reloc_root comes out of here with two references, one for
829  * root->reloc_root, and another for being on the rc->reloc_roots list.
830  */
831 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
832                           struct btrfs_root *root)
833 {
834         struct btrfs_fs_info *fs_info = root->fs_info;
835         struct btrfs_root *reloc_root;
836         struct reloc_control *rc = fs_info->reloc_ctl;
837         struct btrfs_block_rsv *rsv;
838         int clear_rsv = 0;
839         int ret;
840
841         if (!rc)
842                 return 0;
843
844         /*
845          * The subvolume has reloc tree but the swap is finished, no need to
846          * create/update the dead reloc tree
847          */
848         if (reloc_root_is_dead(root))
849                 return 0;
850
851         /*
852          * This is subtle but important.  We do not do
853          * record_root_in_transaction for reloc roots, instead we record their
854          * corresponding fs root, and then here we update the last trans for the
855          * reloc root.  This means that we have to do this for the entire life
856          * of the reloc root, regardless of which stage of the relocation we are
857          * in.
858          */
859         if (root->reloc_root) {
860                 reloc_root = root->reloc_root;
861                 reloc_root->last_trans = trans->transid;
862                 return 0;
863         }
864
865         /*
866          * We are merging reloc roots, we do not need new reloc trees.  Also
867          * reloc trees never need their own reloc tree.
868          */
869         if (!rc->create_reloc_tree ||
870             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
871                 return 0;
872
873         if (!trans->reloc_reserved) {
874                 rsv = trans->block_rsv;
875                 trans->block_rsv = rc->block_rsv;
876                 clear_rsv = 1;
877         }
878         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
879         if (clear_rsv)
880                 trans->block_rsv = rsv;
881
882         ret = __add_reloc_root(reloc_root);
883         BUG_ON(ret < 0);
884         root->reloc_root = btrfs_grab_root(reloc_root);
885         return 0;
886 }
887
888 /*
889  * update root item of reloc tree
890  */
891 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
892                             struct btrfs_root *root)
893 {
894         struct btrfs_fs_info *fs_info = root->fs_info;
895         struct btrfs_root *reloc_root;
896         struct btrfs_root_item *root_item;
897         int ret;
898
899         if (!have_reloc_root(root))
900                 return 0;
901
902         reloc_root = root->reloc_root;
903         root_item = &reloc_root->root_item;
904
905         /*
906          * We are probably ok here, but __del_reloc_root() will drop its ref of
907          * the root.  We have the ref for root->reloc_root, but just in case
908          * hold it while we update the reloc root.
909          */
910         btrfs_grab_root(reloc_root);
911
912         /* root->reloc_root will stay until current relocation finished */
913         if (fs_info->reloc_ctl->merge_reloc_tree &&
914             btrfs_root_refs(root_item) == 0) {
915                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
916                 /*
917                  * Mark the tree as dead before we change reloc_root so
918                  * have_reloc_root will not touch it from now on.
919                  */
920                 smp_wmb();
921                 __del_reloc_root(reloc_root);
922         }
923
924         if (reloc_root->commit_root != reloc_root->node) {
925                 __update_reloc_root(reloc_root);
926                 btrfs_set_root_node(root_item, reloc_root->node);
927                 free_extent_buffer(reloc_root->commit_root);
928                 reloc_root->commit_root = btrfs_root_node(reloc_root);
929         }
930
931         ret = btrfs_update_root(trans, fs_info->tree_root,
932                                 &reloc_root->root_key, root_item);
933         btrfs_put_root(reloc_root);
934         return ret;
935 }
936
937 /*
938  * helper to find first cached inode with inode number >= objectid
939  * in a subvolume
940  */
941 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
942 {
943         struct rb_node *node;
944         struct rb_node *prev;
945         struct btrfs_inode *entry;
946         struct inode *inode;
947
948         spin_lock(&root->inode_lock);
949 again:
950         node = root->inode_tree.rb_node;
951         prev = NULL;
952         while (node) {
953                 prev = node;
954                 entry = rb_entry(node, struct btrfs_inode, rb_node);
955
956                 if (objectid < btrfs_ino(entry))
957                         node = node->rb_left;
958                 else if (objectid > btrfs_ino(entry))
959                         node = node->rb_right;
960                 else
961                         break;
962         }
963         if (!node) {
964                 while (prev) {
965                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
966                         if (objectid <= btrfs_ino(entry)) {
967                                 node = prev;
968                                 break;
969                         }
970                         prev = rb_next(prev);
971                 }
972         }
973         while (node) {
974                 entry = rb_entry(node, struct btrfs_inode, rb_node);
975                 inode = igrab(&entry->vfs_inode);
976                 if (inode) {
977                         spin_unlock(&root->inode_lock);
978                         return inode;
979                 }
980
981                 objectid = btrfs_ino(entry) + 1;
982                 if (cond_resched_lock(&root->inode_lock))
983                         goto again;
984
985                 node = rb_next(node);
986         }
987         spin_unlock(&root->inode_lock);
988         return NULL;
989 }
990
991 /*
992  * get new location of data
993  */
994 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
995                             u64 bytenr, u64 num_bytes)
996 {
997         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
998         struct btrfs_path *path;
999         struct btrfs_file_extent_item *fi;
1000         struct extent_buffer *leaf;
1001         int ret;
1002
1003         path = btrfs_alloc_path();
1004         if (!path)
1005                 return -ENOMEM;
1006
1007         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
1008         ret = btrfs_lookup_file_extent(NULL, root, path,
1009                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
1010         if (ret < 0)
1011                 goto out;
1012         if (ret > 0) {
1013                 ret = -ENOENT;
1014                 goto out;
1015         }
1016
1017         leaf = path->nodes[0];
1018         fi = btrfs_item_ptr(leaf, path->slots[0],
1019                             struct btrfs_file_extent_item);
1020
1021         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1022                btrfs_file_extent_compression(leaf, fi) ||
1023                btrfs_file_extent_encryption(leaf, fi) ||
1024                btrfs_file_extent_other_encoding(leaf, fi));
1025
1026         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1027                 ret = -EINVAL;
1028                 goto out;
1029         }
1030
1031         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1032         ret = 0;
1033 out:
1034         btrfs_free_path(path);
1035         return ret;
1036 }
1037
1038 /*
1039  * update file extent items in the tree leaf to point to
1040  * the new locations.
1041  */
1042 static noinline_for_stack
1043 int replace_file_extents(struct btrfs_trans_handle *trans,
1044                          struct reloc_control *rc,
1045                          struct btrfs_root *root,
1046                          struct extent_buffer *leaf)
1047 {
1048         struct btrfs_fs_info *fs_info = root->fs_info;
1049         struct btrfs_key key;
1050         struct btrfs_file_extent_item *fi;
1051         struct inode *inode = NULL;
1052         u64 parent;
1053         u64 bytenr;
1054         u64 new_bytenr = 0;
1055         u64 num_bytes;
1056         u64 end;
1057         u32 nritems;
1058         u32 i;
1059         int ret = 0;
1060         int first = 1;
1061         int dirty = 0;
1062
1063         if (rc->stage != UPDATE_DATA_PTRS)
1064                 return 0;
1065
1066         /* reloc trees always use full backref */
1067         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1068                 parent = leaf->start;
1069         else
1070                 parent = 0;
1071
1072         nritems = btrfs_header_nritems(leaf);
1073         for (i = 0; i < nritems; i++) {
1074                 struct btrfs_ref ref = { 0 };
1075
1076                 cond_resched();
1077                 btrfs_item_key_to_cpu(leaf, &key, i);
1078                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1079                         continue;
1080                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1081                 if (btrfs_file_extent_type(leaf, fi) ==
1082                     BTRFS_FILE_EXTENT_INLINE)
1083                         continue;
1084                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1085                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1086                 if (bytenr == 0)
1087                         continue;
1088                 if (!in_range(bytenr, rc->block_group->start,
1089                               rc->block_group->length))
1090                         continue;
1091
1092                 /*
1093                  * if we are modifying block in fs tree, wait for readpage
1094                  * to complete and drop the extent cache
1095                  */
1096                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1097                         if (first) {
1098                                 inode = find_next_inode(root, key.objectid);
1099                                 first = 0;
1100                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1101                                 btrfs_add_delayed_iput(inode);
1102                                 inode = find_next_inode(root, key.objectid);
1103                         }
1104                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1105                                 end = key.offset +
1106                                       btrfs_file_extent_num_bytes(leaf, fi);
1107                                 WARN_ON(!IS_ALIGNED(key.offset,
1108                                                     fs_info->sectorsize));
1109                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1110                                 end--;
1111                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1112                                                       key.offset, end);
1113                                 if (!ret)
1114                                         continue;
1115
1116                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1117                                                 key.offset,     end, 1);
1118                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1119                                               key.offset, end);
1120                         }
1121                 }
1122
1123                 ret = get_new_location(rc->data_inode, &new_bytenr,
1124                                        bytenr, num_bytes);
1125                 if (ret) {
1126                         /*
1127                          * Don't have to abort since we've not changed anything
1128                          * in the file extent yet.
1129                          */
1130                         break;
1131                 }
1132
1133                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1134                 dirty = 1;
1135
1136                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1137                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1138                                        num_bytes, parent);
1139                 ref.real_root = root->root_key.objectid;
1140                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1141                                     key.objectid, key.offset);
1142                 ret = btrfs_inc_extent_ref(trans, &ref);
1143                 if (ret) {
1144                         btrfs_abort_transaction(trans, ret);
1145                         break;
1146                 }
1147
1148                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1149                                        num_bytes, parent);
1150                 ref.real_root = root->root_key.objectid;
1151                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1152                                     key.objectid, key.offset);
1153                 ret = btrfs_free_extent(trans, &ref);
1154                 if (ret) {
1155                         btrfs_abort_transaction(trans, ret);
1156                         break;
1157                 }
1158         }
1159         if (dirty)
1160                 btrfs_mark_buffer_dirty(leaf);
1161         if (inode)
1162                 btrfs_add_delayed_iput(inode);
1163         return ret;
1164 }
1165
1166 static noinline_for_stack
1167 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1168                      struct btrfs_path *path, int level)
1169 {
1170         struct btrfs_disk_key key1;
1171         struct btrfs_disk_key key2;
1172         btrfs_node_key(eb, &key1, slot);
1173         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1174         return memcmp(&key1, &key2, sizeof(key1));
1175 }
1176
1177 /*
1178  * try to replace tree blocks in fs tree with the new blocks
1179  * in reloc tree. tree blocks haven't been modified since the
1180  * reloc tree was create can be replaced.
1181  *
1182  * if a block was replaced, level of the block + 1 is returned.
1183  * if no block got replaced, 0 is returned. if there are other
1184  * errors, a negative error number is returned.
1185  */
1186 static noinline_for_stack
1187 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1188                  struct btrfs_root *dest, struct btrfs_root *src,
1189                  struct btrfs_path *path, struct btrfs_key *next_key,
1190                  int lowest_level, int max_level)
1191 {
1192         struct btrfs_fs_info *fs_info = dest->fs_info;
1193         struct extent_buffer *eb;
1194         struct extent_buffer *parent;
1195         struct btrfs_ref ref = { 0 };
1196         struct btrfs_key key;
1197         u64 old_bytenr;
1198         u64 new_bytenr;
1199         u64 old_ptr_gen;
1200         u64 new_ptr_gen;
1201         u64 last_snapshot;
1202         u32 blocksize;
1203         int cow = 0;
1204         int level;
1205         int ret;
1206         int slot;
1207
1208         ASSERT(src->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1209         ASSERT(dest->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1210
1211         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1212 again:
1213         slot = path->slots[lowest_level];
1214         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1215
1216         eb = btrfs_lock_root_node(dest);
1217         btrfs_set_lock_blocking_write(eb);
1218         level = btrfs_header_level(eb);
1219
1220         if (level < lowest_level) {
1221                 btrfs_tree_unlock(eb);
1222                 free_extent_buffer(eb);
1223                 return 0;
1224         }
1225
1226         if (cow) {
1227                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1228                                       BTRFS_NESTING_COW);
1229                 BUG_ON(ret);
1230         }
1231         btrfs_set_lock_blocking_write(eb);
1232
1233         if (next_key) {
1234                 next_key->objectid = (u64)-1;
1235                 next_key->type = (u8)-1;
1236                 next_key->offset = (u64)-1;
1237         }
1238
1239         parent = eb;
1240         while (1) {
1241                 struct btrfs_key first_key;
1242
1243                 level = btrfs_header_level(parent);
1244                 ASSERT(level >= lowest_level);
1245
1246                 ret = btrfs_bin_search(parent, &key, &slot);
1247                 if (ret < 0)
1248                         break;
1249                 if (ret && slot > 0)
1250                         slot--;
1251
1252                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1253                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1254
1255                 old_bytenr = btrfs_node_blockptr(parent, slot);
1256                 blocksize = fs_info->nodesize;
1257                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1258                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1259
1260                 if (level <= max_level) {
1261                         eb = path->nodes[level];
1262                         new_bytenr = btrfs_node_blockptr(eb,
1263                                                         path->slots[level]);
1264                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1265                                                         path->slots[level]);
1266                 } else {
1267                         new_bytenr = 0;
1268                         new_ptr_gen = 0;
1269                 }
1270
1271                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1272                         ret = level;
1273                         break;
1274                 }
1275
1276                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1277                     memcmp_node_keys(parent, slot, path, level)) {
1278                         if (level <= lowest_level) {
1279                                 ret = 0;
1280                                 break;
1281                         }
1282
1283                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1284                                              level - 1, &first_key);
1285                         if (IS_ERR(eb)) {
1286                                 ret = PTR_ERR(eb);
1287                                 break;
1288                         } else if (!extent_buffer_uptodate(eb)) {
1289                                 ret = -EIO;
1290                                 free_extent_buffer(eb);
1291                                 break;
1292                         }
1293                         btrfs_tree_lock(eb);
1294                         if (cow) {
1295                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1296                                                       slot, &eb,
1297                                                       BTRFS_NESTING_COW);
1298                                 BUG_ON(ret);
1299                         }
1300                         btrfs_set_lock_blocking_write(eb);
1301
1302                         btrfs_tree_unlock(parent);
1303                         free_extent_buffer(parent);
1304
1305                         parent = eb;
1306                         continue;
1307                 }
1308
1309                 if (!cow) {
1310                         btrfs_tree_unlock(parent);
1311                         free_extent_buffer(parent);
1312                         cow = 1;
1313                         goto again;
1314                 }
1315
1316                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1317                                       path->slots[level]);
1318                 btrfs_release_path(path);
1319
1320                 path->lowest_level = level;
1321                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1322                 path->lowest_level = 0;
1323                 BUG_ON(ret);
1324
1325                 /*
1326                  * Info qgroup to trace both subtrees.
1327                  *
1328                  * We must trace both trees.
1329                  * 1) Tree reloc subtree
1330                  *    If not traced, we will leak data numbers
1331                  * 2) Fs subtree
1332                  *    If not traced, we will double count old data
1333                  *
1334                  * We don't scan the subtree right now, but only record
1335                  * the swapped tree blocks.
1336                  * The real subtree rescan is delayed until we have new
1337                  * CoW on the subtree root node before transaction commit.
1338                  */
1339                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1340                                 rc->block_group, parent, slot,
1341                                 path->nodes[level], path->slots[level],
1342                                 last_snapshot);
1343                 if (ret < 0)
1344                         break;
1345                 /*
1346                  * swap blocks in fs tree and reloc tree.
1347                  */
1348                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1349                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1350                 btrfs_mark_buffer_dirty(parent);
1351
1352                 btrfs_set_node_blockptr(path->nodes[level],
1353                                         path->slots[level], old_bytenr);
1354                 btrfs_set_node_ptr_generation(path->nodes[level],
1355                                               path->slots[level], old_ptr_gen);
1356                 btrfs_mark_buffer_dirty(path->nodes[level]);
1357
1358                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1359                                        blocksize, path->nodes[level]->start);
1360                 ref.skip_qgroup = true;
1361                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1362                 ret = btrfs_inc_extent_ref(trans, &ref);
1363                 BUG_ON(ret);
1364                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1365                                        blocksize, 0);
1366                 ref.skip_qgroup = true;
1367                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1368                 ret = btrfs_inc_extent_ref(trans, &ref);
1369                 BUG_ON(ret);
1370
1371                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1372                                        blocksize, path->nodes[level]->start);
1373                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1374                 ref.skip_qgroup = true;
1375                 ret = btrfs_free_extent(trans, &ref);
1376                 BUG_ON(ret);
1377
1378                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1379                                        blocksize, 0);
1380                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1381                 ref.skip_qgroup = true;
1382                 ret = btrfs_free_extent(trans, &ref);
1383                 BUG_ON(ret);
1384
1385                 btrfs_unlock_up_safe(path, 0);
1386
1387                 ret = level;
1388                 break;
1389         }
1390         btrfs_tree_unlock(parent);
1391         free_extent_buffer(parent);
1392         return ret;
1393 }
1394
1395 /*
1396  * helper to find next relocated block in reloc tree
1397  */
1398 static noinline_for_stack
1399 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1400                        int *level)
1401 {
1402         struct extent_buffer *eb;
1403         int i;
1404         u64 last_snapshot;
1405         u32 nritems;
1406
1407         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1408
1409         for (i = 0; i < *level; i++) {
1410                 free_extent_buffer(path->nodes[i]);
1411                 path->nodes[i] = NULL;
1412         }
1413
1414         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1415                 eb = path->nodes[i];
1416                 nritems = btrfs_header_nritems(eb);
1417                 while (path->slots[i] + 1 < nritems) {
1418                         path->slots[i]++;
1419                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1420                             last_snapshot)
1421                                 continue;
1422
1423                         *level = i;
1424                         return 0;
1425                 }
1426                 free_extent_buffer(path->nodes[i]);
1427                 path->nodes[i] = NULL;
1428         }
1429         return 1;
1430 }
1431
1432 /*
1433  * walk down reloc tree to find relocated block of lowest level
1434  */
1435 static noinline_for_stack
1436 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1437                          int *level)
1438 {
1439         struct btrfs_fs_info *fs_info = root->fs_info;
1440         struct extent_buffer *eb = NULL;
1441         int i;
1442         u64 bytenr;
1443         u64 ptr_gen = 0;
1444         u64 last_snapshot;
1445         u32 nritems;
1446
1447         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1448
1449         for (i = *level; i > 0; i--) {
1450                 struct btrfs_key first_key;
1451
1452                 eb = path->nodes[i];
1453                 nritems = btrfs_header_nritems(eb);
1454                 while (path->slots[i] < nritems) {
1455                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1456                         if (ptr_gen > last_snapshot)
1457                                 break;
1458                         path->slots[i]++;
1459                 }
1460                 if (path->slots[i] >= nritems) {
1461                         if (i == *level)
1462                                 break;
1463                         *level = i + 1;
1464                         return 0;
1465                 }
1466                 if (i == 1) {
1467                         *level = i;
1468                         return 0;
1469                 }
1470
1471                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1472                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
1473                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
1474                                      &first_key);
1475                 if (IS_ERR(eb)) {
1476                         return PTR_ERR(eb);
1477                 } else if (!extent_buffer_uptodate(eb)) {
1478                         free_extent_buffer(eb);
1479                         return -EIO;
1480                 }
1481                 BUG_ON(btrfs_header_level(eb) != i - 1);
1482                 path->nodes[i - 1] = eb;
1483                 path->slots[i - 1] = 0;
1484         }
1485         return 1;
1486 }
1487
1488 /*
1489  * invalidate extent cache for file extents whose key in range of
1490  * [min_key, max_key)
1491  */
1492 static int invalidate_extent_cache(struct btrfs_root *root,
1493                                    struct btrfs_key *min_key,
1494                                    struct btrfs_key *max_key)
1495 {
1496         struct btrfs_fs_info *fs_info = root->fs_info;
1497         struct inode *inode = NULL;
1498         u64 objectid;
1499         u64 start, end;
1500         u64 ino;
1501
1502         objectid = min_key->objectid;
1503         while (1) {
1504                 cond_resched();
1505                 iput(inode);
1506
1507                 if (objectid > max_key->objectid)
1508                         break;
1509
1510                 inode = find_next_inode(root, objectid);
1511                 if (!inode)
1512                         break;
1513                 ino = btrfs_ino(BTRFS_I(inode));
1514
1515                 if (ino > max_key->objectid) {
1516                         iput(inode);
1517                         break;
1518                 }
1519
1520                 objectid = ino + 1;
1521                 if (!S_ISREG(inode->i_mode))
1522                         continue;
1523
1524                 if (unlikely(min_key->objectid == ino)) {
1525                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1526                                 continue;
1527                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1528                                 start = 0;
1529                         else {
1530                                 start = min_key->offset;
1531                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1532                         }
1533                 } else {
1534                         start = 0;
1535                 }
1536
1537                 if (unlikely(max_key->objectid == ino)) {
1538                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1539                                 continue;
1540                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1541                                 end = (u64)-1;
1542                         } else {
1543                                 if (max_key->offset == 0)
1544                                         continue;
1545                                 end = max_key->offset;
1546                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1547                                 end--;
1548                         }
1549                 } else {
1550                         end = (u64)-1;
1551                 }
1552
1553                 /* the lock_extent waits for readpage to complete */
1554                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1555                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
1556                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1557         }
1558         return 0;
1559 }
1560
1561 static int find_next_key(struct btrfs_path *path, int level,
1562                          struct btrfs_key *key)
1563
1564 {
1565         while (level < BTRFS_MAX_LEVEL) {
1566                 if (!path->nodes[level])
1567                         break;
1568                 if (path->slots[level] + 1 <
1569                     btrfs_header_nritems(path->nodes[level])) {
1570                         btrfs_node_key_to_cpu(path->nodes[level], key,
1571                                               path->slots[level] + 1);
1572                         return 0;
1573                 }
1574                 level++;
1575         }
1576         return 1;
1577 }
1578
1579 /*
1580  * Insert current subvolume into reloc_control::dirty_subvol_roots
1581  */
1582 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
1583                                 struct reloc_control *rc,
1584                                 struct btrfs_root *root)
1585 {
1586         struct btrfs_root *reloc_root = root->reloc_root;
1587         struct btrfs_root_item *reloc_root_item;
1588
1589         /* @root must be a subvolume tree root with a valid reloc tree */
1590         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1591         ASSERT(reloc_root);
1592
1593         reloc_root_item = &reloc_root->root_item;
1594         memset(&reloc_root_item->drop_progress, 0,
1595                 sizeof(reloc_root_item->drop_progress));
1596         reloc_root_item->drop_level = 0;
1597         btrfs_set_root_refs(reloc_root_item, 0);
1598         btrfs_update_reloc_root(trans, root);
1599
1600         if (list_empty(&root->reloc_dirty_list)) {
1601                 btrfs_grab_root(root);
1602                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1603         }
1604 }
1605
1606 static int clean_dirty_subvols(struct reloc_control *rc)
1607 {
1608         struct btrfs_root *root;
1609         struct btrfs_root *next;
1610         int ret = 0;
1611         int ret2;
1612
1613         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1614                                  reloc_dirty_list) {
1615                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1616                         /* Merged subvolume, cleanup its reloc root */
1617                         struct btrfs_root *reloc_root = root->reloc_root;
1618
1619                         list_del_init(&root->reloc_dirty_list);
1620                         root->reloc_root = NULL;
1621                         /*
1622                          * Need barrier to ensure clear_bit() only happens after
1623                          * root->reloc_root = NULL. Pairs with have_reloc_root.
1624                          */
1625                         smp_wmb();
1626                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1627                         if (reloc_root) {
1628                                 /*
1629                                  * btrfs_drop_snapshot drops our ref we hold for
1630                                  * ->reloc_root.  If it fails however we must
1631                                  * drop the ref ourselves.
1632                                  */
1633                                 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1634                                 if (ret2 < 0) {
1635                                         btrfs_put_root(reloc_root);
1636                                         if (!ret)
1637                                                 ret = ret2;
1638                                 }
1639                         }
1640                         btrfs_put_root(root);
1641                 } else {
1642                         /* Orphan reloc tree, just clean it up */
1643                         ret2 = btrfs_drop_snapshot(root, 0, 1);
1644                         if (ret2 < 0) {
1645                                 btrfs_put_root(root);
1646                                 if (!ret)
1647                                         ret = ret2;
1648                         }
1649                 }
1650         }
1651         return ret;
1652 }
1653
1654 /*
1655  * merge the relocated tree blocks in reloc tree with corresponding
1656  * fs tree.
1657  */
1658 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1659                                                struct btrfs_root *root)
1660 {
1661         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1662         struct btrfs_key key;
1663         struct btrfs_key next_key;
1664         struct btrfs_trans_handle *trans = NULL;
1665         struct btrfs_root *reloc_root;
1666         struct btrfs_root_item *root_item;
1667         struct btrfs_path *path;
1668         struct extent_buffer *leaf;
1669         int reserve_level;
1670         int level;
1671         int max_level;
1672         int replaced = 0;
1673         int ret;
1674         int err = 0;
1675         u32 min_reserved;
1676
1677         path = btrfs_alloc_path();
1678         if (!path)
1679                 return -ENOMEM;
1680         path->reada = READA_FORWARD;
1681
1682         reloc_root = root->reloc_root;
1683         root_item = &reloc_root->root_item;
1684
1685         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1686                 level = btrfs_root_level(root_item);
1687                 atomic_inc(&reloc_root->node->refs);
1688                 path->nodes[level] = reloc_root->node;
1689                 path->slots[level] = 0;
1690         } else {
1691                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1692
1693                 level = root_item->drop_level;
1694                 BUG_ON(level == 0);
1695                 path->lowest_level = level;
1696                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1697                 path->lowest_level = 0;
1698                 if (ret < 0) {
1699                         btrfs_free_path(path);
1700                         return ret;
1701                 }
1702
1703                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1704                                       path->slots[level]);
1705                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1706
1707                 btrfs_unlock_up_safe(path, 0);
1708         }
1709
1710         /*
1711          * In merge_reloc_root(), we modify the upper level pointer to swap the
1712          * tree blocks between reloc tree and subvolume tree.  Thus for tree
1713          * block COW, we COW at most from level 1 to root level for each tree.
1714          *
1715          * Thus the needed metadata size is at most root_level * nodesize,
1716          * and * 2 since we have two trees to COW.
1717          */
1718         reserve_level = max_t(int, 1, btrfs_root_level(root_item));
1719         min_reserved = fs_info->nodesize * reserve_level * 2;
1720         memset(&next_key, 0, sizeof(next_key));
1721
1722         while (1) {
1723                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
1724                                              BTRFS_RESERVE_FLUSH_LIMIT);
1725                 if (ret) {
1726                         err = ret;
1727                         goto out;
1728                 }
1729                 trans = btrfs_start_transaction(root, 0);
1730                 if (IS_ERR(trans)) {
1731                         err = PTR_ERR(trans);
1732                         trans = NULL;
1733                         goto out;
1734                 }
1735
1736                 /*
1737                  * At this point we no longer have a reloc_control, so we can't
1738                  * depend on btrfs_init_reloc_root to update our last_trans.
1739                  *
1740                  * But that's ok, we started the trans handle on our
1741                  * corresponding fs_root, which means it's been added to the
1742                  * dirty list.  At commit time we'll still call
1743                  * btrfs_update_reloc_root() and update our root item
1744                  * appropriately.
1745                  */
1746                 reloc_root->last_trans = trans->transid;
1747                 trans->block_rsv = rc->block_rsv;
1748
1749                 replaced = 0;
1750                 max_level = level;
1751
1752                 ret = walk_down_reloc_tree(reloc_root, path, &level);
1753                 if (ret < 0) {
1754                         err = ret;
1755                         goto out;
1756                 }
1757                 if (ret > 0)
1758                         break;
1759
1760                 if (!find_next_key(path, level, &key) &&
1761                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1762                         ret = 0;
1763                 } else {
1764                         ret = replace_path(trans, rc, root, reloc_root, path,
1765                                            &next_key, level, max_level);
1766                 }
1767                 if (ret < 0) {
1768                         err = ret;
1769                         goto out;
1770                 }
1771
1772                 if (ret > 0) {
1773                         level = ret;
1774                         btrfs_node_key_to_cpu(path->nodes[level], &key,
1775                                               path->slots[level]);
1776                         replaced = 1;
1777                 }
1778
1779                 ret = walk_up_reloc_tree(reloc_root, path, &level);
1780                 if (ret > 0)
1781                         break;
1782
1783                 BUG_ON(level == 0);
1784                 /*
1785                  * save the merging progress in the drop_progress.
1786                  * this is OK since root refs == 1 in this case.
1787                  */
1788                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1789                                path->slots[level]);
1790                 root_item->drop_level = level;
1791
1792                 btrfs_end_transaction_throttle(trans);
1793                 trans = NULL;
1794
1795                 btrfs_btree_balance_dirty(fs_info);
1796
1797                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1798                         invalidate_extent_cache(root, &key, &next_key);
1799         }
1800
1801         /*
1802          * handle the case only one block in the fs tree need to be
1803          * relocated and the block is tree root.
1804          */
1805         leaf = btrfs_lock_root_node(root);
1806         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1807                               BTRFS_NESTING_COW);
1808         btrfs_tree_unlock(leaf);
1809         free_extent_buffer(leaf);
1810         if (ret < 0)
1811                 err = ret;
1812 out:
1813         btrfs_free_path(path);
1814
1815         if (err == 0)
1816                 insert_dirty_subvol(trans, rc, root);
1817
1818         if (trans)
1819                 btrfs_end_transaction_throttle(trans);
1820
1821         btrfs_btree_balance_dirty(fs_info);
1822
1823         if (replaced && rc->stage == UPDATE_DATA_PTRS)
1824                 invalidate_extent_cache(root, &key, &next_key);
1825
1826         return err;
1827 }
1828
1829 static noinline_for_stack
1830 int prepare_to_merge(struct reloc_control *rc, int err)
1831 {
1832         struct btrfs_root *root = rc->extent_root;
1833         struct btrfs_fs_info *fs_info = root->fs_info;
1834         struct btrfs_root *reloc_root;
1835         struct btrfs_trans_handle *trans;
1836         LIST_HEAD(reloc_roots);
1837         u64 num_bytes = 0;
1838         int ret;
1839
1840         mutex_lock(&fs_info->reloc_mutex);
1841         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1842         rc->merging_rsv_size += rc->nodes_relocated * 2;
1843         mutex_unlock(&fs_info->reloc_mutex);
1844
1845 again:
1846         if (!err) {
1847                 num_bytes = rc->merging_rsv_size;
1848                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
1849                                           BTRFS_RESERVE_FLUSH_ALL);
1850                 if (ret)
1851                         err = ret;
1852         }
1853
1854         trans = btrfs_join_transaction(rc->extent_root);
1855         if (IS_ERR(trans)) {
1856                 if (!err)
1857                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1858                                                 num_bytes, NULL);
1859                 return PTR_ERR(trans);
1860         }
1861
1862         if (!err) {
1863                 if (num_bytes != rc->merging_rsv_size) {
1864                         btrfs_end_transaction(trans);
1865                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1866                                                 num_bytes, NULL);
1867                         goto again;
1868                 }
1869         }
1870
1871         rc->merge_reloc_tree = 1;
1872
1873         while (!list_empty(&rc->reloc_roots)) {
1874                 reloc_root = list_entry(rc->reloc_roots.next,
1875                                         struct btrfs_root, root_list);
1876                 list_del_init(&reloc_root->root_list);
1877
1878                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1879                                 false);
1880                 BUG_ON(IS_ERR(root));
1881                 BUG_ON(root->reloc_root != reloc_root);
1882
1883                 /*
1884                  * set reference count to 1, so btrfs_recover_relocation
1885                  * knows it should resumes merging
1886                  */
1887                 if (!err)
1888                         btrfs_set_root_refs(&reloc_root->root_item, 1);
1889                 btrfs_update_reloc_root(trans, root);
1890
1891                 list_add(&reloc_root->root_list, &reloc_roots);
1892                 btrfs_put_root(root);
1893         }
1894
1895         list_splice(&reloc_roots, &rc->reloc_roots);
1896
1897         if (!err)
1898                 err = btrfs_commit_transaction(trans);
1899         else
1900                 btrfs_end_transaction(trans);
1901         return err;
1902 }
1903
1904 static noinline_for_stack
1905 void free_reloc_roots(struct list_head *list)
1906 {
1907         struct btrfs_root *reloc_root, *tmp;
1908
1909         list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1910                 __del_reloc_root(reloc_root);
1911 }
1912
1913 static noinline_for_stack
1914 void merge_reloc_roots(struct reloc_control *rc)
1915 {
1916         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1917         struct btrfs_root *root;
1918         struct btrfs_root *reloc_root;
1919         LIST_HEAD(reloc_roots);
1920         int found = 0;
1921         int ret = 0;
1922 again:
1923         root = rc->extent_root;
1924
1925         /*
1926          * this serializes us with btrfs_record_root_in_transaction,
1927          * we have to make sure nobody is in the middle of
1928          * adding their roots to the list while we are
1929          * doing this splice
1930          */
1931         mutex_lock(&fs_info->reloc_mutex);
1932         list_splice_init(&rc->reloc_roots, &reloc_roots);
1933         mutex_unlock(&fs_info->reloc_mutex);
1934
1935         while (!list_empty(&reloc_roots)) {
1936                 found = 1;
1937                 reloc_root = list_entry(reloc_roots.next,
1938                                         struct btrfs_root, root_list);
1939
1940                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1941                                          false);
1942                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1943                         BUG_ON(IS_ERR(root));
1944                         BUG_ON(root->reloc_root != reloc_root);
1945                         ret = merge_reloc_root(rc, root);
1946                         btrfs_put_root(root);
1947                         if (ret) {
1948                                 if (list_empty(&reloc_root->root_list))
1949                                         list_add_tail(&reloc_root->root_list,
1950                                                       &reloc_roots);
1951                                 goto out;
1952                         }
1953                 } else {
1954                         if (!IS_ERR(root)) {
1955                                 if (root->reloc_root == reloc_root) {
1956                                         root->reloc_root = NULL;
1957                                         btrfs_put_root(reloc_root);
1958                                 }
1959                                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
1960                                           &root->state);
1961                                 btrfs_put_root(root);
1962                         }
1963
1964                         list_del_init(&reloc_root->root_list);
1965                         /* Don't forget to queue this reloc root for cleanup */
1966                         list_add_tail(&reloc_root->reloc_dirty_list,
1967                                       &rc->dirty_subvol_roots);
1968                 }
1969         }
1970
1971         if (found) {
1972                 found = 0;
1973                 goto again;
1974         }
1975 out:
1976         if (ret) {
1977                 btrfs_handle_fs_error(fs_info, ret, NULL);
1978                 free_reloc_roots(&reloc_roots);
1979
1980                 /* new reloc root may be added */
1981                 mutex_lock(&fs_info->reloc_mutex);
1982                 list_splice_init(&rc->reloc_roots, &reloc_roots);
1983                 mutex_unlock(&fs_info->reloc_mutex);
1984                 free_reloc_roots(&reloc_roots);
1985         }
1986
1987         /*
1988          * We used to have
1989          *
1990          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1991          *
1992          * here, but it's wrong.  If we fail to start the transaction in
1993          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
1994          * have actually been removed from the reloc_root_tree rb tree.  This is
1995          * fine because we're bailing here, and we hold a reference on the root
1996          * for the list that holds it, so these roots will be cleaned up when we
1997          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
1998          * will be cleaned up on unmount.
1999          *
2000          * The remaining nodes will be cleaned up by free_reloc_control.
2001          */
2002 }
2003
2004 static void free_block_list(struct rb_root *blocks)
2005 {
2006         struct tree_block *block;
2007         struct rb_node *rb_node;
2008         while ((rb_node = rb_first(blocks))) {
2009                 block = rb_entry(rb_node, struct tree_block, rb_node);
2010                 rb_erase(rb_node, blocks);
2011                 kfree(block);
2012         }
2013 }
2014
2015 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
2016                                       struct btrfs_root *reloc_root)
2017 {
2018         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
2019         struct btrfs_root *root;
2020         int ret;
2021
2022         if (reloc_root->last_trans == trans->transid)
2023                 return 0;
2024
2025         root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2026         BUG_ON(IS_ERR(root));
2027         BUG_ON(root->reloc_root != reloc_root);
2028         ret = btrfs_record_root_in_trans(trans, root);
2029         btrfs_put_root(root);
2030
2031         return ret;
2032 }
2033
2034 static noinline_for_stack
2035 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2036                                      struct reloc_control *rc,
2037                                      struct btrfs_backref_node *node,
2038                                      struct btrfs_backref_edge *edges[])
2039 {
2040         struct btrfs_backref_node *next;
2041         struct btrfs_root *root;
2042         int index = 0;
2043
2044         next = node;
2045         while (1) {
2046                 cond_resched();
2047                 next = walk_up_backref(next, edges, &index);
2048                 root = next->root;
2049                 BUG_ON(!root);
2050                 BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state));
2051
2052                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2053                         record_reloc_root_in_trans(trans, root);
2054                         break;
2055                 }
2056
2057                 btrfs_record_root_in_trans(trans, root);
2058                 root = root->reloc_root;
2059
2060                 if (next->new_bytenr != root->node->start) {
2061                         BUG_ON(next->new_bytenr);
2062                         BUG_ON(!list_empty(&next->list));
2063                         next->new_bytenr = root->node->start;
2064                         btrfs_put_root(next->root);
2065                         next->root = btrfs_grab_root(root);
2066                         ASSERT(next->root);
2067                         list_add_tail(&next->list,
2068                                       &rc->backref_cache.changed);
2069                         mark_block_processed(rc, next);
2070                         break;
2071                 }
2072
2073                 WARN_ON(1);
2074                 root = NULL;
2075                 next = walk_down_backref(edges, &index);
2076                 if (!next || next->level <= node->level)
2077                         break;
2078         }
2079         if (!root)
2080                 return NULL;
2081
2082         next = node;
2083         /* setup backref node path for btrfs_reloc_cow_block */
2084         while (1) {
2085                 rc->backref_cache.path[next->level] = next;
2086                 if (--index < 0)
2087                         break;
2088                 next = edges[index]->node[UPPER];
2089         }
2090         return root;
2091 }
2092
2093 /*
2094  * Select a tree root for relocation.
2095  *
2096  * Return NULL if the block is not shareable. We should use do_relocation() in
2097  * this case.
2098  *
2099  * Return a tree root pointer if the block is shareable.
2100  * Return -ENOENT if the block is root of reloc tree.
2101  */
2102 static noinline_for_stack
2103 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2104 {
2105         struct btrfs_backref_node *next;
2106         struct btrfs_root *root;
2107         struct btrfs_root *fs_root = NULL;
2108         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2109         int index = 0;
2110
2111         next = node;
2112         while (1) {
2113                 cond_resched();
2114                 next = walk_up_backref(next, edges, &index);
2115                 root = next->root;
2116                 BUG_ON(!root);
2117
2118                 /* No other choice for non-shareable tree */
2119                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2120                         return root;
2121
2122                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2123                         fs_root = root;
2124
2125                 if (next != node)
2126                         return NULL;
2127
2128                 next = walk_down_backref(edges, &index);
2129                 if (!next || next->level <= node->level)
2130                         break;
2131         }
2132
2133         if (!fs_root)
2134                 return ERR_PTR(-ENOENT);
2135         return fs_root;
2136 }
2137
2138 static noinline_for_stack
2139 u64 calcu_metadata_size(struct reloc_control *rc,
2140                         struct btrfs_backref_node *node, int reserve)
2141 {
2142         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2143         struct btrfs_backref_node *next = node;
2144         struct btrfs_backref_edge *edge;
2145         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2146         u64 num_bytes = 0;
2147         int index = 0;
2148
2149         BUG_ON(reserve && node->processed);
2150
2151         while (next) {
2152                 cond_resched();
2153                 while (1) {
2154                         if (next->processed && (reserve || next != node))
2155                                 break;
2156
2157                         num_bytes += fs_info->nodesize;
2158
2159                         if (list_empty(&next->upper))
2160                                 break;
2161
2162                         edge = list_entry(next->upper.next,
2163                                         struct btrfs_backref_edge, list[LOWER]);
2164                         edges[index++] = edge;
2165                         next = edge->node[UPPER];
2166                 }
2167                 next = walk_down_backref(edges, &index);
2168         }
2169         return num_bytes;
2170 }
2171
2172 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2173                                   struct reloc_control *rc,
2174                                   struct btrfs_backref_node *node)
2175 {
2176         struct btrfs_root *root = rc->extent_root;
2177         struct btrfs_fs_info *fs_info = root->fs_info;
2178         u64 num_bytes;
2179         int ret;
2180         u64 tmp;
2181
2182         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2183
2184         trans->block_rsv = rc->block_rsv;
2185         rc->reserved_bytes += num_bytes;
2186
2187         /*
2188          * We are under a transaction here so we can only do limited flushing.
2189          * If we get an enospc just kick back -EAGAIN so we know to drop the
2190          * transaction and try to refill when we can flush all the things.
2191          */
2192         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2193                                 BTRFS_RESERVE_FLUSH_LIMIT);
2194         if (ret) {
2195                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2196                 while (tmp <= rc->reserved_bytes)
2197                         tmp <<= 1;
2198                 /*
2199                  * only one thread can access block_rsv at this point,
2200                  * so we don't need hold lock to protect block_rsv.
2201                  * we expand more reservation size here to allow enough
2202                  * space for relocation and we will return earlier in
2203                  * enospc case.
2204                  */
2205                 rc->block_rsv->size = tmp + fs_info->nodesize *
2206                                       RELOCATION_RESERVED_NODES;
2207                 return -EAGAIN;
2208         }
2209
2210         return 0;
2211 }
2212
2213 /*
2214  * relocate a block tree, and then update pointers in upper level
2215  * blocks that reference the block to point to the new location.
2216  *
2217  * if called by link_to_upper, the block has already been relocated.
2218  * in that case this function just updates pointers.
2219  */
2220 static int do_relocation(struct btrfs_trans_handle *trans,
2221                          struct reloc_control *rc,
2222                          struct btrfs_backref_node *node,
2223                          struct btrfs_key *key,
2224                          struct btrfs_path *path, int lowest)
2225 {
2226         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2227         struct btrfs_backref_node *upper;
2228         struct btrfs_backref_edge *edge;
2229         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2230         struct btrfs_root *root;
2231         struct extent_buffer *eb;
2232         u32 blocksize;
2233         u64 bytenr;
2234         u64 generation;
2235         int slot;
2236         int ret;
2237         int err = 0;
2238
2239         BUG_ON(lowest && node->eb);
2240
2241         path->lowest_level = node->level + 1;
2242         rc->backref_cache.path[node->level] = node;
2243         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2244                 struct btrfs_key first_key;
2245                 struct btrfs_ref ref = { 0 };
2246
2247                 cond_resched();
2248
2249                 upper = edge->node[UPPER];
2250                 root = select_reloc_root(trans, rc, upper, edges);
2251                 BUG_ON(!root);
2252
2253                 if (upper->eb && !upper->locked) {
2254                         if (!lowest) {
2255                                 ret = btrfs_bin_search(upper->eb, key, &slot);
2256                                 if (ret < 0) {
2257                                         err = ret;
2258                                         goto next;
2259                                 }
2260                                 BUG_ON(ret);
2261                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2262                                 if (node->eb->start == bytenr)
2263                                         goto next;
2264                         }
2265                         btrfs_backref_drop_node_buffer(upper);
2266                 }
2267
2268                 if (!upper->eb) {
2269                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2270                         if (ret) {
2271                                 if (ret < 0)
2272                                         err = ret;
2273                                 else
2274                                         err = -ENOENT;
2275
2276                                 btrfs_release_path(path);
2277                                 break;
2278                         }
2279
2280                         if (!upper->eb) {
2281                                 upper->eb = path->nodes[upper->level];
2282                                 path->nodes[upper->level] = NULL;
2283                         } else {
2284                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2285                         }
2286
2287                         upper->locked = 1;
2288                         path->locks[upper->level] = 0;
2289
2290                         slot = path->slots[upper->level];
2291                         btrfs_release_path(path);
2292                 } else {
2293                         ret = btrfs_bin_search(upper->eb, key, &slot);
2294                         if (ret < 0) {
2295                                 err = ret;
2296                                 goto next;
2297                         }
2298                         BUG_ON(ret);
2299                 }
2300
2301                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2302                 if (lowest) {
2303                         if (bytenr != node->bytenr) {
2304                                 btrfs_err(root->fs_info,
2305                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2306                                           bytenr, node->bytenr, slot,
2307                                           upper->eb->start);
2308                                 err = -EIO;
2309                                 goto next;
2310                         }
2311                 } else {
2312                         if (node->eb->start == bytenr)
2313                                 goto next;
2314                 }
2315
2316                 blocksize = root->fs_info->nodesize;
2317                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2318                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2319                 eb = read_tree_block(fs_info, bytenr, generation,
2320                                      upper->level - 1, &first_key);
2321                 if (IS_ERR(eb)) {
2322                         err = PTR_ERR(eb);
2323                         goto next;
2324                 } else if (!extent_buffer_uptodate(eb)) {
2325                         free_extent_buffer(eb);
2326                         err = -EIO;
2327                         goto next;
2328                 }
2329                 btrfs_tree_lock(eb);
2330                 btrfs_set_lock_blocking_write(eb);
2331
2332                 if (!node->eb) {
2333                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2334                                               slot, &eb, BTRFS_NESTING_COW);
2335                         btrfs_tree_unlock(eb);
2336                         free_extent_buffer(eb);
2337                         if (ret < 0) {
2338                                 err = ret;
2339                                 goto next;
2340                         }
2341                         BUG_ON(node->eb != eb);
2342                 } else {
2343                         btrfs_set_node_blockptr(upper->eb, slot,
2344                                                 node->eb->start);
2345                         btrfs_set_node_ptr_generation(upper->eb, slot,
2346                                                       trans->transid);
2347                         btrfs_mark_buffer_dirty(upper->eb);
2348
2349                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2350                                                node->eb->start, blocksize,
2351                                                upper->eb->start);
2352                         ref.real_root = root->root_key.objectid;
2353                         btrfs_init_tree_ref(&ref, node->level,
2354                                             btrfs_header_owner(upper->eb));
2355                         ret = btrfs_inc_extent_ref(trans, &ref);
2356                         BUG_ON(ret);
2357
2358                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2359                         BUG_ON(ret);
2360                 }
2361 next:
2362                 if (!upper->pending)
2363                         btrfs_backref_drop_node_buffer(upper);
2364                 else
2365                         btrfs_backref_unlock_node_buffer(upper);
2366                 if (err)
2367                         break;
2368         }
2369
2370         if (!err && node->pending) {
2371                 btrfs_backref_drop_node_buffer(node);
2372                 list_move_tail(&node->list, &rc->backref_cache.changed);
2373                 node->pending = 0;
2374         }
2375
2376         path->lowest_level = 0;
2377         BUG_ON(err == -ENOSPC);
2378         return err;
2379 }
2380
2381 static int link_to_upper(struct btrfs_trans_handle *trans,
2382                          struct reloc_control *rc,
2383                          struct btrfs_backref_node *node,
2384                          struct btrfs_path *path)
2385 {
2386         struct btrfs_key key;
2387
2388         btrfs_node_key_to_cpu(node->eb, &key, 0);
2389         return do_relocation(trans, rc, node, &key, path, 0);
2390 }
2391
2392 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2393                                 struct reloc_control *rc,
2394                                 struct btrfs_path *path, int err)
2395 {
2396         LIST_HEAD(list);
2397         struct btrfs_backref_cache *cache = &rc->backref_cache;
2398         struct btrfs_backref_node *node;
2399         int level;
2400         int ret;
2401
2402         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2403                 while (!list_empty(&cache->pending[level])) {
2404                         node = list_entry(cache->pending[level].next,
2405                                           struct btrfs_backref_node, list);
2406                         list_move_tail(&node->list, &list);
2407                         BUG_ON(!node->pending);
2408
2409                         if (!err) {
2410                                 ret = link_to_upper(trans, rc, node, path);
2411                                 if (ret < 0)
2412                                         err = ret;
2413                         }
2414                 }
2415                 list_splice_init(&list, &cache->pending[level]);
2416         }
2417         return err;
2418 }
2419
2420 /*
2421  * mark a block and all blocks directly/indirectly reference the block
2422  * as processed.
2423  */
2424 static void update_processed_blocks(struct reloc_control *rc,
2425                                     struct btrfs_backref_node *node)
2426 {
2427         struct btrfs_backref_node *next = node;
2428         struct btrfs_backref_edge *edge;
2429         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2430         int index = 0;
2431
2432         while (next) {
2433                 cond_resched();
2434                 while (1) {
2435                         if (next->processed)
2436                                 break;
2437
2438                         mark_block_processed(rc, next);
2439
2440                         if (list_empty(&next->upper))
2441                                 break;
2442
2443                         edge = list_entry(next->upper.next,
2444                                         struct btrfs_backref_edge, list[LOWER]);
2445                         edges[index++] = edge;
2446                         next = edge->node[UPPER];
2447                 }
2448                 next = walk_down_backref(edges, &index);
2449         }
2450 }
2451
2452 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2453 {
2454         u32 blocksize = rc->extent_root->fs_info->nodesize;
2455
2456         if (test_range_bit(&rc->processed_blocks, bytenr,
2457                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2458                 return 1;
2459         return 0;
2460 }
2461
2462 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2463                               struct tree_block *block)
2464 {
2465         struct extent_buffer *eb;
2466
2467         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2468                              block->level, NULL);
2469         if (IS_ERR(eb)) {
2470                 return PTR_ERR(eb);
2471         } else if (!extent_buffer_uptodate(eb)) {
2472                 free_extent_buffer(eb);
2473                 return -EIO;
2474         }
2475         if (block->level == 0)
2476                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2477         else
2478                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2479         free_extent_buffer(eb);
2480         block->key_ready = 1;
2481         return 0;
2482 }
2483
2484 /*
2485  * helper function to relocate a tree block
2486  */
2487 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2488                                 struct reloc_control *rc,
2489                                 struct btrfs_backref_node *node,
2490                                 struct btrfs_key *key,
2491                                 struct btrfs_path *path)
2492 {
2493         struct btrfs_root *root;
2494         int ret = 0;
2495
2496         if (!node)
2497                 return 0;
2498
2499         /*
2500          * If we fail here we want to drop our backref_node because we are going
2501          * to start over and regenerate the tree for it.
2502          */
2503         ret = reserve_metadata_space(trans, rc, node);
2504         if (ret)
2505                 goto out;
2506
2507         BUG_ON(node->processed);
2508         root = select_one_root(node);
2509         if (root == ERR_PTR(-ENOENT)) {
2510                 update_processed_blocks(rc, node);
2511                 goto out;
2512         }
2513
2514         if (root) {
2515                 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2516                         BUG_ON(node->new_bytenr);
2517                         BUG_ON(!list_empty(&node->list));
2518                         btrfs_record_root_in_trans(trans, root);
2519                         root = root->reloc_root;
2520                         node->new_bytenr = root->node->start;
2521                         btrfs_put_root(node->root);
2522                         node->root = btrfs_grab_root(root);
2523                         ASSERT(node->root);
2524                         list_add_tail(&node->list, &rc->backref_cache.changed);
2525                 } else {
2526                         path->lowest_level = node->level;
2527                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2528                         btrfs_release_path(path);
2529                         if (ret > 0)
2530                                 ret = 0;
2531                 }
2532                 if (!ret)
2533                         update_processed_blocks(rc, node);
2534         } else {
2535                 ret = do_relocation(trans, rc, node, key, path, 1);
2536         }
2537 out:
2538         if (ret || node->level == 0 || node->cowonly)
2539                 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2540         return ret;
2541 }
2542
2543 /*
2544  * relocate a list of blocks
2545  */
2546 static noinline_for_stack
2547 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2548                          struct reloc_control *rc, struct rb_root *blocks)
2549 {
2550         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2551         struct btrfs_backref_node *node;
2552         struct btrfs_path *path;
2553         struct tree_block *block;
2554         struct tree_block *next;
2555         int ret;
2556         int err = 0;
2557
2558         path = btrfs_alloc_path();
2559         if (!path) {
2560                 err = -ENOMEM;
2561                 goto out_free_blocks;
2562         }
2563
2564         /* Kick in readahead for tree blocks with missing keys */
2565         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2566                 if (!block->key_ready)
2567                         readahead_tree_block(fs_info, block->bytenr);
2568         }
2569
2570         /* Get first keys */
2571         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2572                 if (!block->key_ready) {
2573                         err = get_tree_block_key(fs_info, block);
2574                         if (err)
2575                                 goto out_free_path;
2576                 }
2577         }
2578
2579         /* Do tree relocation */
2580         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2581                 node = build_backref_tree(rc, &block->key,
2582                                           block->level, block->bytenr);
2583                 if (IS_ERR(node)) {
2584                         err = PTR_ERR(node);
2585                         goto out;
2586                 }
2587
2588                 ret = relocate_tree_block(trans, rc, node, &block->key,
2589                                           path);
2590                 if (ret < 0) {
2591                         err = ret;
2592                         break;
2593                 }
2594         }
2595 out:
2596         err = finish_pending_nodes(trans, rc, path, err);
2597
2598 out_free_path:
2599         btrfs_free_path(path);
2600 out_free_blocks:
2601         free_block_list(blocks);
2602         return err;
2603 }
2604
2605 static noinline_for_stack int prealloc_file_extent_cluster(
2606                                 struct btrfs_inode *inode,
2607                                 struct file_extent_cluster *cluster)
2608 {
2609         u64 alloc_hint = 0;
2610         u64 start;
2611         u64 end;
2612         u64 offset = inode->index_cnt;
2613         u64 num_bytes;
2614         int nr;
2615         int ret = 0;
2616         u64 prealloc_start = cluster->start - offset;
2617         u64 prealloc_end = cluster->end - offset;
2618         u64 cur_offset = prealloc_start;
2619
2620         BUG_ON(cluster->start != cluster->boundary[0]);
2621         ret = btrfs_alloc_data_chunk_ondemand(inode,
2622                                               prealloc_end + 1 - prealloc_start);
2623         if (ret)
2624                 return ret;
2625
2626         inode_lock(&inode->vfs_inode);
2627         for (nr = 0; nr < cluster->nr; nr++) {
2628                 start = cluster->boundary[nr] - offset;
2629                 if (nr + 1 < cluster->nr)
2630                         end = cluster->boundary[nr + 1] - 1 - offset;
2631                 else
2632                         end = cluster->end - offset;
2633
2634                 lock_extent(&inode->io_tree, start, end);
2635                 num_bytes = end + 1 - start;
2636                 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2637                                                 num_bytes, num_bytes,
2638                                                 end + 1, &alloc_hint);
2639                 cur_offset = end + 1;
2640                 unlock_extent(&inode->io_tree, start, end);
2641                 if (ret)
2642                         break;
2643         }
2644         inode_unlock(&inode->vfs_inode);
2645
2646         if (cur_offset < prealloc_end)
2647                 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2648                                                prealloc_end + 1 - cur_offset);
2649         return ret;
2650 }
2651
2652 static noinline_for_stack
2653 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2654                          u64 block_start)
2655 {
2656         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2657         struct extent_map *em;
2658         int ret = 0;
2659
2660         em = alloc_extent_map();
2661         if (!em)
2662                 return -ENOMEM;
2663
2664         em->start = start;
2665         em->len = end + 1 - start;
2666         em->block_len = em->len;
2667         em->block_start = block_start;
2668         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2669
2670         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2671         while (1) {
2672                 write_lock(&em_tree->lock);
2673                 ret = add_extent_mapping(em_tree, em, 0);
2674                 write_unlock(&em_tree->lock);
2675                 if (ret != -EEXIST) {
2676                         free_extent_map(em);
2677                         break;
2678                 }
2679                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2680         }
2681         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2682         return ret;
2683 }
2684
2685 /*
2686  * Allow error injection to test balance cancellation
2687  */
2688 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2689 {
2690         return atomic_read(&fs_info->balance_cancel_req) ||
2691                 fatal_signal_pending(current);
2692 }
2693 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2694
2695 static int relocate_file_extent_cluster(struct inode *inode,
2696                                         struct file_extent_cluster *cluster)
2697 {
2698         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2699         u64 page_start;
2700         u64 page_end;
2701         u64 offset = BTRFS_I(inode)->index_cnt;
2702         unsigned long index;
2703         unsigned long last_index;
2704         struct page *page;
2705         struct file_ra_state *ra;
2706         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2707         int nr = 0;
2708         int ret = 0;
2709
2710         if (!cluster->nr)
2711                 return 0;
2712
2713         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2714         if (!ra)
2715                 return -ENOMEM;
2716
2717         ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
2718         if (ret)
2719                 goto out;
2720
2721         file_ra_state_init(ra, inode->i_mapping);
2722
2723         ret = setup_extent_mapping(inode, cluster->start - offset,
2724                                    cluster->end - offset, cluster->start);
2725         if (ret)
2726                 goto out;
2727
2728         index = (cluster->start - offset) >> PAGE_SHIFT;
2729         last_index = (cluster->end - offset) >> PAGE_SHIFT;
2730         while (index <= last_index) {
2731                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2732                                 PAGE_SIZE);
2733                 if (ret)
2734                         goto out;
2735
2736                 page = find_lock_page(inode->i_mapping, index);
2737                 if (!page) {
2738                         page_cache_sync_readahead(inode->i_mapping,
2739                                                   ra, NULL, index,
2740                                                   last_index + 1 - index);
2741                         page = find_or_create_page(inode->i_mapping, index,
2742                                                    mask);
2743                         if (!page) {
2744                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2745                                                         PAGE_SIZE, true);
2746                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2747                                                         PAGE_SIZE);
2748                                 ret = -ENOMEM;
2749                                 goto out;
2750                         }
2751                 }
2752
2753                 if (PageReadahead(page)) {
2754                         page_cache_async_readahead(inode->i_mapping,
2755                                                    ra, NULL, page, index,
2756                                                    last_index + 1 - index);
2757                 }
2758
2759                 if (!PageUptodate(page)) {
2760                         btrfs_readpage(NULL, page);
2761                         lock_page(page);
2762                         if (!PageUptodate(page)) {
2763                                 unlock_page(page);
2764                                 put_page(page);
2765                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2766                                                         PAGE_SIZE, true);
2767                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2768                                                                PAGE_SIZE);
2769                                 ret = -EIO;
2770                                 goto out;
2771                         }
2772                 }
2773
2774                 page_start = page_offset(page);
2775                 page_end = page_start + PAGE_SIZE - 1;
2776
2777                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
2778
2779                 set_page_extent_mapped(page);
2780
2781                 if (nr < cluster->nr &&
2782                     page_start + offset == cluster->boundary[nr]) {
2783                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2784                                         page_start, page_end,
2785                                         EXTENT_BOUNDARY);
2786                         nr++;
2787                 }
2788
2789                 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start,
2790                                                 page_end, 0, NULL);
2791                 if (ret) {
2792                         unlock_page(page);
2793                         put_page(page);
2794                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
2795                                                          PAGE_SIZE, true);
2796                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2797                                                        PAGE_SIZE);
2798
2799                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
2800                                           page_start, page_end,
2801                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
2802                         goto out;
2803
2804                 }
2805                 set_page_dirty(page);
2806
2807                 unlock_extent(&BTRFS_I(inode)->io_tree,
2808                               page_start, page_end);
2809                 unlock_page(page);
2810                 put_page(page);
2811
2812                 index++;
2813                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2814                 balance_dirty_pages_ratelimited(inode->i_mapping);
2815                 btrfs_throttle(fs_info);
2816                 if (btrfs_should_cancel_balance(fs_info)) {
2817                         ret = -ECANCELED;
2818                         goto out;
2819                 }
2820         }
2821         WARN_ON(nr != cluster->nr);
2822 out:
2823         kfree(ra);
2824         return ret;
2825 }
2826
2827 static noinline_for_stack
2828 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
2829                          struct file_extent_cluster *cluster)
2830 {
2831         int ret;
2832
2833         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
2834                 ret = relocate_file_extent_cluster(inode, cluster);
2835                 if (ret)
2836                         return ret;
2837                 cluster->nr = 0;
2838         }
2839
2840         if (!cluster->nr)
2841                 cluster->start = extent_key->objectid;
2842         else
2843                 BUG_ON(cluster->nr >= MAX_EXTENTS);
2844         cluster->end = extent_key->objectid + extent_key->offset - 1;
2845         cluster->boundary[cluster->nr] = extent_key->objectid;
2846         cluster->nr++;
2847
2848         if (cluster->nr >= MAX_EXTENTS) {
2849                 ret = relocate_file_extent_cluster(inode, cluster);
2850                 if (ret)
2851                         return ret;
2852                 cluster->nr = 0;
2853         }
2854         return 0;
2855 }
2856
2857 /*
2858  * helper to add a tree block to the list.
2859  * the major work is getting the generation and level of the block
2860  */
2861 static int add_tree_block(struct reloc_control *rc,
2862                           struct btrfs_key *extent_key,
2863                           struct btrfs_path *path,
2864                           struct rb_root *blocks)
2865 {
2866         struct extent_buffer *eb;
2867         struct btrfs_extent_item *ei;
2868         struct btrfs_tree_block_info *bi;
2869         struct tree_block *block;
2870         struct rb_node *rb_node;
2871         u32 item_size;
2872         int level = -1;
2873         u64 generation;
2874
2875         eb =  path->nodes[0];
2876         item_size = btrfs_item_size_nr(eb, path->slots[0]);
2877
2878         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
2879             item_size >= sizeof(*ei) + sizeof(*bi)) {
2880                 ei = btrfs_item_ptr(eb, path->slots[0],
2881                                 struct btrfs_extent_item);
2882                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
2883                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2884                         level = btrfs_tree_block_level(eb, bi);
2885                 } else {
2886                         level = (int)extent_key->offset;
2887                 }
2888                 generation = btrfs_extent_generation(eb, ei);
2889         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
2890                 btrfs_print_v0_err(eb->fs_info);
2891                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
2892                 return -EINVAL;
2893         } else {
2894                 BUG();
2895         }
2896
2897         btrfs_release_path(path);
2898
2899         BUG_ON(level == -1);
2900
2901         block = kmalloc(sizeof(*block), GFP_NOFS);
2902         if (!block)
2903                 return -ENOMEM;
2904
2905         block->bytenr = extent_key->objectid;
2906         block->key.objectid = rc->extent_root->fs_info->nodesize;
2907         block->key.offset = generation;
2908         block->level = level;
2909         block->key_ready = 0;
2910
2911         rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
2912         if (rb_node)
2913                 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
2914                                     -EEXIST);
2915
2916         return 0;
2917 }
2918
2919 /*
2920  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
2921  */
2922 static int __add_tree_block(struct reloc_control *rc,
2923                             u64 bytenr, u32 blocksize,
2924                             struct rb_root *blocks)
2925 {
2926         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2927         struct btrfs_path *path;
2928         struct btrfs_key key;
2929         int ret;
2930         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2931
2932         if (tree_block_processed(bytenr, rc))
2933                 return 0;
2934
2935         if (rb_simple_search(blocks, bytenr))
2936                 return 0;
2937
2938         path = btrfs_alloc_path();
2939         if (!path)
2940                 return -ENOMEM;
2941 again:
2942         key.objectid = bytenr;
2943         if (skinny) {
2944                 key.type = BTRFS_METADATA_ITEM_KEY;
2945                 key.offset = (u64)-1;
2946         } else {
2947                 key.type = BTRFS_EXTENT_ITEM_KEY;
2948                 key.offset = blocksize;
2949         }
2950
2951         path->search_commit_root = 1;
2952         path->skip_locking = 1;
2953         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
2954         if (ret < 0)
2955                 goto out;
2956
2957         if (ret > 0 && skinny) {
2958                 if (path->slots[0]) {
2959                         path->slots[0]--;
2960                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2961                                               path->slots[0]);
2962                         if (key.objectid == bytenr &&
2963                             (key.type == BTRFS_METADATA_ITEM_KEY ||
2964                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
2965                               key.offset == blocksize)))
2966                                 ret = 0;
2967                 }
2968
2969                 if (ret) {
2970                         skinny = false;
2971                         btrfs_release_path(path);
2972                         goto again;
2973                 }
2974         }
2975         if (ret) {
2976                 ASSERT(ret == 1);
2977                 btrfs_print_leaf(path->nodes[0]);
2978                 btrfs_err(fs_info,
2979              "tree block extent item (%llu) is not found in extent tree",
2980                      bytenr);
2981                 WARN_ON(1);
2982                 ret = -EINVAL;
2983                 goto out;
2984         }
2985
2986         ret = add_tree_block(rc, &key, path, blocks);
2987 out:
2988         btrfs_free_path(path);
2989         return ret;
2990 }
2991
2992 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
2993                                     struct btrfs_block_group *block_group,
2994                                     struct inode *inode,
2995                                     u64 ino)
2996 {
2997         struct btrfs_root *root = fs_info->tree_root;
2998         struct btrfs_trans_handle *trans;
2999         int ret = 0;
3000
3001         if (inode)
3002                 goto truncate;
3003
3004         inode = btrfs_iget(fs_info->sb, ino, root);
3005         if (IS_ERR(inode))
3006                 return -ENOENT;
3007
3008 truncate:
3009         ret = btrfs_check_trunc_cache_free_space(fs_info,
3010                                                  &fs_info->global_block_rsv);
3011         if (ret)
3012                 goto out;
3013
3014         trans = btrfs_join_transaction(root);
3015         if (IS_ERR(trans)) {
3016                 ret = PTR_ERR(trans);
3017                 goto out;
3018         }
3019
3020         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3021
3022         btrfs_end_transaction(trans);
3023         btrfs_btree_balance_dirty(fs_info);
3024 out:
3025         iput(inode);
3026         return ret;
3027 }
3028
3029 /*
3030  * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3031  * cache inode, to avoid free space cache data extent blocking data relocation.
3032  */
3033 static int delete_v1_space_cache(struct extent_buffer *leaf,
3034                                  struct btrfs_block_group *block_group,
3035                                  u64 data_bytenr)
3036 {
3037         u64 space_cache_ino;
3038         struct btrfs_file_extent_item *ei;
3039         struct btrfs_key key;
3040         bool found = false;
3041         int i;
3042         int ret;
3043
3044         if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3045                 return 0;
3046
3047         for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3048                 u8 type;
3049
3050                 btrfs_item_key_to_cpu(leaf, &key, i);
3051                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3052                         continue;
3053                 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3054                 type = btrfs_file_extent_type(leaf, ei);
3055
3056                 if ((type == BTRFS_FILE_EXTENT_REG ||
3057                      type == BTRFS_FILE_EXTENT_PREALLOC) &&
3058                     btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3059                         found = true;
3060                         space_cache_ino = key.objectid;
3061                         break;
3062                 }
3063         }
3064         if (!found)
3065                 return -ENOENT;
3066         ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3067                                         space_cache_ino);
3068         return ret;
3069 }
3070
3071 /*
3072  * helper to find all tree blocks that reference a given data extent
3073  */
3074 static noinline_for_stack
3075 int add_data_references(struct reloc_control *rc,
3076                         struct btrfs_key *extent_key,
3077                         struct btrfs_path *path,
3078                         struct rb_root *blocks)
3079 {
3080         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3081         struct ulist *leaves = NULL;
3082         struct ulist_iterator leaf_uiter;
3083         struct ulist_node *ref_node = NULL;
3084         const u32 blocksize = fs_info->nodesize;
3085         int ret = 0;
3086
3087         btrfs_release_path(path);
3088         ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3089                                    0, &leaves, NULL, true);
3090         if (ret < 0)
3091                 return ret;
3092
3093         ULIST_ITER_INIT(&leaf_uiter);
3094         while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3095                 struct extent_buffer *eb;
3096
3097                 eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL);
3098                 if (IS_ERR(eb)) {
3099                         ret = PTR_ERR(eb);
3100                         break;
3101                 }
3102                 ret = delete_v1_space_cache(eb, rc->block_group,
3103                                             extent_key->objectid);
3104                 free_extent_buffer(eb);
3105                 if (ret < 0)
3106                         break;
3107                 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3108                 if (ret < 0)
3109                         break;
3110         }
3111         if (ret < 0)
3112                 free_block_list(blocks);
3113         ulist_free(leaves);
3114         return ret;
3115 }
3116
3117 /*
3118  * helper to find next unprocessed extent
3119  */
3120 static noinline_for_stack
3121 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3122                      struct btrfs_key *extent_key)
3123 {
3124         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3125         struct btrfs_key key;
3126         struct extent_buffer *leaf;
3127         u64 start, end, last;
3128         int ret;
3129
3130         last = rc->block_group->start + rc->block_group->length;
3131         while (1) {
3132                 cond_resched();
3133                 if (rc->search_start >= last) {
3134                         ret = 1;
3135                         break;
3136                 }
3137
3138                 key.objectid = rc->search_start;
3139                 key.type = BTRFS_EXTENT_ITEM_KEY;
3140                 key.offset = 0;
3141
3142                 path->search_commit_root = 1;
3143                 path->skip_locking = 1;
3144                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3145                                         0, 0);
3146                 if (ret < 0)
3147                         break;
3148 next:
3149                 leaf = path->nodes[0];
3150                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3151                         ret = btrfs_next_leaf(rc->extent_root, path);
3152                         if (ret != 0)
3153                                 break;
3154                         leaf = path->nodes[0];
3155                 }
3156
3157                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3158                 if (key.objectid >= last) {
3159                         ret = 1;
3160                         break;
3161                 }
3162
3163                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3164                     key.type != BTRFS_METADATA_ITEM_KEY) {
3165                         path->slots[0]++;
3166                         goto next;
3167                 }
3168
3169                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3170                     key.objectid + key.offset <= rc->search_start) {
3171                         path->slots[0]++;
3172                         goto next;
3173                 }
3174
3175                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3176                     key.objectid + fs_info->nodesize <=
3177                     rc->search_start) {
3178                         path->slots[0]++;
3179                         goto next;
3180                 }
3181
3182                 ret = find_first_extent_bit(&rc->processed_blocks,
3183                                             key.objectid, &start, &end,
3184                                             EXTENT_DIRTY, NULL);
3185
3186                 if (ret == 0 && start <= key.objectid) {
3187                         btrfs_release_path(path);
3188                         rc->search_start = end + 1;
3189                 } else {
3190                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3191                                 rc->search_start = key.objectid + key.offset;
3192                         else
3193                                 rc->search_start = key.objectid +
3194                                         fs_info->nodesize;
3195                         memcpy(extent_key, &key, sizeof(key));
3196                         return 0;
3197                 }
3198         }
3199         btrfs_release_path(path);
3200         return ret;
3201 }
3202
3203 static void set_reloc_control(struct reloc_control *rc)
3204 {
3205         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3206
3207         mutex_lock(&fs_info->reloc_mutex);
3208         fs_info->reloc_ctl = rc;
3209         mutex_unlock(&fs_info->reloc_mutex);
3210 }
3211
3212 static void unset_reloc_control(struct reloc_control *rc)
3213 {
3214         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3215
3216         mutex_lock(&fs_info->reloc_mutex);
3217         fs_info->reloc_ctl = NULL;
3218         mutex_unlock(&fs_info->reloc_mutex);
3219 }
3220
3221 static int check_extent_flags(u64 flags)
3222 {
3223         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3224             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3225                 return 1;
3226         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3227             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3228                 return 1;
3229         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3230             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3231                 return 1;
3232         return 0;
3233 }
3234
3235 static noinline_for_stack
3236 int prepare_to_relocate(struct reloc_control *rc)
3237 {
3238         struct btrfs_trans_handle *trans;
3239         int ret;
3240
3241         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3242                                               BTRFS_BLOCK_RSV_TEMP);
3243         if (!rc->block_rsv)
3244                 return -ENOMEM;
3245
3246         memset(&rc->cluster, 0, sizeof(rc->cluster));
3247         rc->search_start = rc->block_group->start;
3248         rc->extents_found = 0;
3249         rc->nodes_relocated = 0;
3250         rc->merging_rsv_size = 0;
3251         rc->reserved_bytes = 0;
3252         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3253                               RELOCATION_RESERVED_NODES;
3254         ret = btrfs_block_rsv_refill(rc->extent_root,
3255                                      rc->block_rsv, rc->block_rsv->size,
3256                                      BTRFS_RESERVE_FLUSH_ALL);
3257         if (ret)
3258                 return ret;
3259
3260         rc->create_reloc_tree = 1;
3261         set_reloc_control(rc);
3262
3263         trans = btrfs_join_transaction(rc->extent_root);
3264         if (IS_ERR(trans)) {
3265                 unset_reloc_control(rc);
3266                 /*
3267                  * extent tree is not a ref_cow tree and has no reloc_root to
3268                  * cleanup.  And callers are responsible to free the above
3269                  * block rsv.
3270                  */
3271                 return PTR_ERR(trans);
3272         }
3273
3274         ret = btrfs_commit_transaction(trans);
3275         if (ret)
3276                 unset_reloc_control(rc);
3277
3278         return ret;
3279 }
3280
3281 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3282 {
3283         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3284         struct rb_root blocks = RB_ROOT;
3285         struct btrfs_key key;
3286         struct btrfs_trans_handle *trans = NULL;
3287         struct btrfs_path *path;
3288         struct btrfs_extent_item *ei;
3289         u64 flags;
3290         u32 item_size;
3291         int ret;
3292         int err = 0;
3293         int progress = 0;
3294
3295         path = btrfs_alloc_path();
3296         if (!path)
3297                 return -ENOMEM;
3298         path->reada = READA_FORWARD;
3299
3300         ret = prepare_to_relocate(rc);
3301         if (ret) {
3302                 err = ret;
3303                 goto out_free;
3304         }
3305
3306         while (1) {
3307                 rc->reserved_bytes = 0;
3308                 ret = btrfs_block_rsv_refill(rc->extent_root,
3309                                         rc->block_rsv, rc->block_rsv->size,
3310                                         BTRFS_RESERVE_FLUSH_ALL);
3311                 if (ret) {
3312                         err = ret;
3313                         break;
3314                 }
3315                 progress++;
3316                 trans = btrfs_start_transaction(rc->extent_root, 0);
3317                 if (IS_ERR(trans)) {
3318                         err = PTR_ERR(trans);
3319                         trans = NULL;
3320                         break;
3321                 }
3322 restart:
3323                 if (update_backref_cache(trans, &rc->backref_cache)) {
3324                         btrfs_end_transaction(trans);
3325                         trans = NULL;
3326                         continue;
3327                 }
3328
3329                 ret = find_next_extent(rc, path, &key);
3330                 if (ret < 0)
3331                         err = ret;
3332                 if (ret != 0)
3333                         break;
3334
3335                 rc->extents_found++;
3336
3337                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3338                                     struct btrfs_extent_item);
3339                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3340                 if (item_size >= sizeof(*ei)) {
3341                         flags = btrfs_extent_flags(path->nodes[0], ei);
3342                         ret = check_extent_flags(flags);
3343                         BUG_ON(ret);
3344                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3345                         err = -EINVAL;
3346                         btrfs_print_v0_err(trans->fs_info);
3347                         btrfs_abort_transaction(trans, err);
3348                         break;
3349                 } else {
3350                         BUG();
3351                 }
3352
3353                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3354                         ret = add_tree_block(rc, &key, path, &blocks);
3355                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3356                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3357                         ret = add_data_references(rc, &key, path, &blocks);
3358                 } else {
3359                         btrfs_release_path(path);
3360                         ret = 0;
3361                 }
3362                 if (ret < 0) {
3363                         err = ret;
3364                         break;
3365                 }
3366
3367                 if (!RB_EMPTY_ROOT(&blocks)) {
3368                         ret = relocate_tree_blocks(trans, rc, &blocks);
3369                         if (ret < 0) {
3370                                 if (ret != -EAGAIN) {
3371                                         err = ret;
3372                                         break;
3373                                 }
3374                                 rc->extents_found--;
3375                                 rc->search_start = key.objectid;
3376                         }
3377                 }
3378
3379                 btrfs_end_transaction_throttle(trans);
3380                 btrfs_btree_balance_dirty(fs_info);
3381                 trans = NULL;
3382
3383                 if (rc->stage == MOVE_DATA_EXTENTS &&
3384                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3385                         rc->found_file_extent = 1;
3386                         ret = relocate_data_extent(rc->data_inode,
3387                                                    &key, &rc->cluster);
3388                         if (ret < 0) {
3389                                 err = ret;
3390                                 break;
3391                         }
3392                 }
3393                 if (btrfs_should_cancel_balance(fs_info)) {
3394                         err = -ECANCELED;
3395                         break;
3396                 }
3397         }
3398         if (trans && progress && err == -ENOSPC) {
3399                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3400                 if (ret == 1) {
3401                         err = 0;
3402                         progress = 0;
3403                         goto restart;
3404                 }
3405         }
3406
3407         btrfs_release_path(path);
3408         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3409
3410         if (trans) {
3411                 btrfs_end_transaction_throttle(trans);
3412                 btrfs_btree_balance_dirty(fs_info);
3413         }
3414
3415         if (!err) {
3416                 ret = relocate_file_extent_cluster(rc->data_inode,
3417                                                    &rc->cluster);
3418                 if (ret < 0)
3419                         err = ret;
3420         }
3421
3422         rc->create_reloc_tree = 0;
3423         set_reloc_control(rc);
3424
3425         btrfs_backref_release_cache(&rc->backref_cache);
3426         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3427
3428         /*
3429          * Even in the case when the relocation is cancelled, we should all go
3430          * through prepare_to_merge() and merge_reloc_roots().
3431          *
3432          * For error (including cancelled balance), prepare_to_merge() will
3433          * mark all reloc trees orphan, then queue them for cleanup in
3434          * merge_reloc_roots()
3435          */
3436         err = prepare_to_merge(rc, err);
3437
3438         merge_reloc_roots(rc);
3439
3440         rc->merge_reloc_tree = 0;
3441         unset_reloc_control(rc);
3442         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3443
3444         /* get rid of pinned extents */
3445         trans = btrfs_join_transaction(rc->extent_root);
3446         if (IS_ERR(trans)) {
3447                 err = PTR_ERR(trans);
3448                 goto out_free;
3449         }
3450         ret = btrfs_commit_transaction(trans);
3451         if (ret && !err)
3452                 err = ret;
3453 out_free:
3454         ret = clean_dirty_subvols(rc);
3455         if (ret < 0 && !err)
3456                 err = ret;
3457         btrfs_free_block_rsv(fs_info, rc->block_rsv);
3458         btrfs_free_path(path);
3459         return err;
3460 }
3461
3462 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3463                                  struct btrfs_root *root, u64 objectid)
3464 {
3465         struct btrfs_path *path;
3466         struct btrfs_inode_item *item;
3467         struct extent_buffer *leaf;
3468         int ret;
3469
3470         path = btrfs_alloc_path();
3471         if (!path)
3472                 return -ENOMEM;
3473
3474         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3475         if (ret)
3476                 goto out;
3477
3478         leaf = path->nodes[0];
3479         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3480         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3481         btrfs_set_inode_generation(leaf, item, 1);
3482         btrfs_set_inode_size(leaf, item, 0);
3483         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3484         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3485                                           BTRFS_INODE_PREALLOC);
3486         btrfs_mark_buffer_dirty(leaf);
3487 out:
3488         btrfs_free_path(path);
3489         return ret;
3490 }
3491
3492 /*
3493  * helper to create inode for data relocation.
3494  * the inode is in data relocation tree and its link count is 0
3495  */
3496 static noinline_for_stack
3497 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3498                                  struct btrfs_block_group *group)
3499 {
3500         struct inode *inode = NULL;
3501         struct btrfs_trans_handle *trans;
3502         struct btrfs_root *root;
3503         u64 objectid;
3504         int err = 0;
3505
3506         root = btrfs_grab_root(fs_info->data_reloc_root);
3507         trans = btrfs_start_transaction(root, 6);
3508         if (IS_ERR(trans)) {
3509                 btrfs_put_root(root);
3510                 return ERR_CAST(trans);
3511         }
3512
3513         err = btrfs_find_free_objectid(root, &objectid);
3514         if (err)
3515                 goto out;
3516
3517         err = __insert_orphan_inode(trans, root, objectid);
3518         BUG_ON(err);
3519
3520         inode = btrfs_iget(fs_info->sb, objectid, root);
3521         BUG_ON(IS_ERR(inode));
3522         BTRFS_I(inode)->index_cnt = group->start;
3523
3524         err = btrfs_orphan_add(trans, BTRFS_I(inode));
3525 out:
3526         btrfs_put_root(root);
3527         btrfs_end_transaction(trans);
3528         btrfs_btree_balance_dirty(fs_info);
3529         if (err) {
3530                 if (inode)
3531                         iput(inode);
3532                 inode = ERR_PTR(err);
3533         }
3534         return inode;
3535 }
3536
3537 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3538 {
3539         struct reloc_control *rc;
3540
3541         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3542         if (!rc)
3543                 return NULL;
3544
3545         INIT_LIST_HEAD(&rc->reloc_roots);
3546         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3547         btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3548         mapping_tree_init(&rc->reloc_root_tree);
3549         extent_io_tree_init(fs_info, &rc->processed_blocks,
3550                             IO_TREE_RELOC_BLOCKS, NULL);
3551         return rc;
3552 }
3553
3554 static void free_reloc_control(struct reloc_control *rc)
3555 {
3556         struct mapping_node *node, *tmp;
3557
3558         free_reloc_roots(&rc->reloc_roots);
3559         rbtree_postorder_for_each_entry_safe(node, tmp,
3560                         &rc->reloc_root_tree.rb_root, rb_node)
3561                 kfree(node);
3562
3563         kfree(rc);
3564 }
3565
3566 /*
3567  * Print the block group being relocated
3568  */
3569 static void describe_relocation(struct btrfs_fs_info *fs_info,
3570                                 struct btrfs_block_group *block_group)
3571 {
3572         char buf[128] = {'\0'};
3573
3574         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3575
3576         btrfs_info(fs_info,
3577                    "relocating block group %llu flags %s",
3578                    block_group->start, buf);
3579 }
3580
3581 static const char *stage_to_string(int stage)
3582 {
3583         if (stage == MOVE_DATA_EXTENTS)
3584                 return "move data extents";
3585         if (stage == UPDATE_DATA_PTRS)
3586                 return "update data pointers";
3587         return "unknown";
3588 }
3589
3590 /*
3591  * function to relocate all extents in a block group.
3592  */
3593 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3594 {
3595         struct btrfs_block_group *bg;
3596         struct btrfs_root *extent_root = fs_info->extent_root;
3597         struct reloc_control *rc;
3598         struct inode *inode;
3599         struct btrfs_path *path;
3600         int ret;
3601         int rw = 0;
3602         int err = 0;
3603
3604         bg = btrfs_lookup_block_group(fs_info, group_start);
3605         if (!bg)
3606                 return -ENOENT;
3607
3608         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3609                 btrfs_put_block_group(bg);
3610                 return -ETXTBSY;
3611         }
3612
3613         rc = alloc_reloc_control(fs_info);
3614         if (!rc) {
3615                 btrfs_put_block_group(bg);
3616                 return -ENOMEM;
3617         }
3618
3619         rc->extent_root = extent_root;
3620         rc->block_group = bg;
3621
3622         ret = btrfs_inc_block_group_ro(rc->block_group, true);
3623         if (ret) {
3624                 err = ret;
3625                 goto out;
3626         }
3627         rw = 1;
3628
3629         path = btrfs_alloc_path();
3630         if (!path) {
3631                 err = -ENOMEM;
3632                 goto out;
3633         }
3634
3635         inode = lookup_free_space_inode(rc->block_group, path);
3636         btrfs_free_path(path);
3637
3638         if (!IS_ERR(inode))
3639                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
3640         else
3641                 ret = PTR_ERR(inode);
3642
3643         if (ret && ret != -ENOENT) {
3644                 err = ret;
3645                 goto out;
3646         }
3647
3648         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3649         if (IS_ERR(rc->data_inode)) {
3650                 err = PTR_ERR(rc->data_inode);
3651                 rc->data_inode = NULL;
3652                 goto out;
3653         }
3654
3655         describe_relocation(fs_info, rc->block_group);
3656
3657         btrfs_wait_block_group_reservations(rc->block_group);
3658         btrfs_wait_nocow_writers(rc->block_group);
3659         btrfs_wait_ordered_roots(fs_info, U64_MAX,
3660                                  rc->block_group->start,
3661                                  rc->block_group->length);
3662
3663         while (1) {
3664                 int finishes_stage;
3665
3666                 mutex_lock(&fs_info->cleaner_mutex);
3667                 ret = relocate_block_group(rc);
3668                 mutex_unlock(&fs_info->cleaner_mutex);
3669                 if (ret < 0)
3670                         err = ret;
3671
3672                 finishes_stage = rc->stage;
3673                 /*
3674                  * We may have gotten ENOSPC after we already dirtied some
3675                  * extents.  If writeout happens while we're relocating a
3676                  * different block group we could end up hitting the
3677                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
3678                  * btrfs_reloc_cow_block.  Make sure we write everything out
3679                  * properly so we don't trip over this problem, and then break
3680                  * out of the loop if we hit an error.
3681                  */
3682                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
3683                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
3684                                                        (u64)-1);
3685                         if (ret)
3686                                 err = ret;
3687                         invalidate_mapping_pages(rc->data_inode->i_mapping,
3688                                                  0, -1);
3689                         rc->stage = UPDATE_DATA_PTRS;
3690                 }
3691
3692                 if (err < 0)
3693                         goto out;
3694
3695                 if (rc->extents_found == 0)
3696                         break;
3697
3698                 btrfs_info(fs_info, "found %llu extents, stage: %s",
3699                            rc->extents_found, stage_to_string(finishes_stage));
3700         }
3701
3702         WARN_ON(rc->block_group->pinned > 0);
3703         WARN_ON(rc->block_group->reserved > 0);
3704         WARN_ON(rc->block_group->used > 0);
3705 out:
3706         if (err && rw)
3707                 btrfs_dec_block_group_ro(rc->block_group);
3708         iput(rc->data_inode);
3709         btrfs_put_block_group(rc->block_group);
3710         free_reloc_control(rc);
3711         return err;
3712 }
3713
3714 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3715 {
3716         struct btrfs_fs_info *fs_info = root->fs_info;
3717         struct btrfs_trans_handle *trans;
3718         int ret, err;
3719
3720         trans = btrfs_start_transaction(fs_info->tree_root, 0);
3721         if (IS_ERR(trans))
3722                 return PTR_ERR(trans);
3723
3724         memset(&root->root_item.drop_progress, 0,
3725                 sizeof(root->root_item.drop_progress));
3726         root->root_item.drop_level = 0;
3727         btrfs_set_root_refs(&root->root_item, 0);
3728         ret = btrfs_update_root(trans, fs_info->tree_root,
3729                                 &root->root_key, &root->root_item);
3730
3731         err = btrfs_end_transaction(trans);
3732         if (err)
3733                 return err;
3734         return ret;
3735 }
3736
3737 /*
3738  * recover relocation interrupted by system crash.
3739  *
3740  * this function resumes merging reloc trees with corresponding fs trees.
3741  * this is important for keeping the sharing of tree blocks
3742  */
3743 int btrfs_recover_relocation(struct btrfs_root *root)
3744 {
3745         struct btrfs_fs_info *fs_info = root->fs_info;
3746         LIST_HEAD(reloc_roots);
3747         struct btrfs_key key;
3748         struct btrfs_root *fs_root;
3749         struct btrfs_root *reloc_root;
3750         struct btrfs_path *path;
3751         struct extent_buffer *leaf;
3752         struct reloc_control *rc = NULL;
3753         struct btrfs_trans_handle *trans;
3754         int ret;
3755         int err = 0;
3756
3757         path = btrfs_alloc_path();
3758         if (!path)
3759                 return -ENOMEM;
3760         path->reada = READA_BACK;
3761
3762         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
3763         key.type = BTRFS_ROOT_ITEM_KEY;
3764         key.offset = (u64)-1;
3765
3766         while (1) {
3767                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
3768                                         path, 0, 0);
3769                 if (ret < 0) {
3770                         err = ret;
3771                         goto out;
3772                 }
3773                 if (ret > 0) {
3774                         if (path->slots[0] == 0)
3775                                 break;
3776                         path->slots[0]--;
3777                 }
3778                 leaf = path->nodes[0];
3779                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3780                 btrfs_release_path(path);
3781
3782                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
3783                     key.type != BTRFS_ROOT_ITEM_KEY)
3784                         break;
3785
3786                 reloc_root = btrfs_read_tree_root(root, &key);
3787                 if (IS_ERR(reloc_root)) {
3788                         err = PTR_ERR(reloc_root);
3789                         goto out;
3790                 }
3791
3792                 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
3793                 list_add(&reloc_root->root_list, &reloc_roots);
3794
3795                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
3796                         fs_root = btrfs_get_fs_root(fs_info,
3797                                         reloc_root->root_key.offset, false);
3798                         if (IS_ERR(fs_root)) {
3799                                 ret = PTR_ERR(fs_root);
3800                                 if (ret != -ENOENT) {
3801                                         err = ret;
3802                                         goto out;
3803                                 }
3804                                 ret = mark_garbage_root(reloc_root);
3805                                 if (ret < 0) {
3806                                         err = ret;
3807                                         goto out;
3808                                 }
3809                         } else {
3810                                 btrfs_put_root(fs_root);
3811                         }
3812                 }
3813
3814                 if (key.offset == 0)
3815                         break;
3816
3817                 key.offset--;
3818         }
3819         btrfs_release_path(path);
3820
3821         if (list_empty(&reloc_roots))
3822                 goto out;
3823
3824         rc = alloc_reloc_control(fs_info);
3825         if (!rc) {
3826                 err = -ENOMEM;
3827                 goto out;
3828         }
3829
3830         rc->extent_root = fs_info->extent_root;
3831
3832         set_reloc_control(rc);
3833
3834         trans = btrfs_join_transaction(rc->extent_root);
3835         if (IS_ERR(trans)) {
3836                 err = PTR_ERR(trans);
3837                 goto out_unset;
3838         }
3839
3840         rc->merge_reloc_tree = 1;
3841
3842         while (!list_empty(&reloc_roots)) {
3843                 reloc_root = list_entry(reloc_roots.next,
3844                                         struct btrfs_root, root_list);
3845                 list_del(&reloc_root->root_list);
3846
3847                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
3848                         list_add_tail(&reloc_root->root_list,
3849                                       &rc->reloc_roots);
3850                         continue;
3851                 }
3852
3853                 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
3854                                             false);
3855                 if (IS_ERR(fs_root)) {
3856                         err = PTR_ERR(fs_root);
3857                         list_add_tail(&reloc_root->root_list, &reloc_roots);
3858                         btrfs_end_transaction(trans);
3859                         goto out_unset;
3860                 }
3861
3862                 err = __add_reloc_root(reloc_root);
3863                 BUG_ON(err < 0); /* -ENOMEM or logic error */
3864                 fs_root->reloc_root = btrfs_grab_root(reloc_root);
3865                 btrfs_put_root(fs_root);
3866         }
3867
3868         err = btrfs_commit_transaction(trans);
3869         if (err)
3870                 goto out_unset;
3871
3872         merge_reloc_roots(rc);
3873
3874         unset_reloc_control(rc);
3875
3876         trans = btrfs_join_transaction(rc->extent_root);
3877         if (IS_ERR(trans)) {
3878                 err = PTR_ERR(trans);
3879                 goto out_clean;
3880         }
3881         err = btrfs_commit_transaction(trans);
3882 out_clean:
3883         ret = clean_dirty_subvols(rc);
3884         if (ret < 0 && !err)
3885                 err = ret;
3886 out_unset:
3887         unset_reloc_control(rc);
3888         free_reloc_control(rc);
3889 out:
3890         free_reloc_roots(&reloc_roots);
3891
3892         btrfs_free_path(path);
3893
3894         if (err == 0) {
3895                 /* cleanup orphan inode in data relocation tree */
3896                 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
3897                 ASSERT(fs_root);
3898                 err = btrfs_orphan_cleanup(fs_root);
3899                 btrfs_put_root(fs_root);
3900         }
3901         return err;
3902 }
3903
3904 /*
3905  * helper to add ordered checksum for data relocation.
3906  *
3907  * cloning checksum properly handles the nodatasum extents.
3908  * it also saves CPU time to re-calculate the checksum.
3909  */
3910 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
3911 {
3912         struct btrfs_fs_info *fs_info = inode->root->fs_info;
3913         struct btrfs_ordered_sum *sums;
3914         struct btrfs_ordered_extent *ordered;
3915         int ret;
3916         u64 disk_bytenr;
3917         u64 new_bytenr;
3918         LIST_HEAD(list);
3919
3920         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
3921         BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
3922
3923         disk_bytenr = file_pos + inode->index_cnt;
3924         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
3925                                        disk_bytenr + len - 1, &list, 0);
3926         if (ret)
3927                 goto out;
3928
3929         while (!list_empty(&list)) {
3930                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
3931                 list_del_init(&sums->list);
3932
3933                 /*
3934                  * We need to offset the new_bytenr based on where the csum is.
3935                  * We need to do this because we will read in entire prealloc
3936                  * extents but we may have written to say the middle of the
3937                  * prealloc extent, so we need to make sure the csum goes with
3938                  * the right disk offset.
3939                  *
3940                  * We can do this because the data reloc inode refers strictly
3941                  * to the on disk bytes, so we don't have to worry about
3942                  * disk_len vs real len like with real inodes since it's all
3943                  * disk length.
3944                  */
3945                 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
3946                 sums->bytenr = new_bytenr;
3947
3948                 btrfs_add_ordered_sum(ordered, sums);
3949         }
3950 out:
3951         btrfs_put_ordered_extent(ordered);
3952         return ret;
3953 }
3954
3955 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
3956                           struct btrfs_root *root, struct extent_buffer *buf,
3957                           struct extent_buffer *cow)
3958 {
3959         struct btrfs_fs_info *fs_info = root->fs_info;
3960         struct reloc_control *rc;
3961         struct btrfs_backref_node *node;
3962         int first_cow = 0;
3963         int level;
3964         int ret = 0;
3965
3966         rc = fs_info->reloc_ctl;
3967         if (!rc)
3968                 return 0;
3969
3970         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
3971                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
3972
3973         level = btrfs_header_level(buf);
3974         if (btrfs_header_generation(buf) <=
3975             btrfs_root_last_snapshot(&root->root_item))
3976                 first_cow = 1;
3977
3978         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
3979             rc->create_reloc_tree) {
3980                 WARN_ON(!first_cow && level == 0);
3981
3982                 node = rc->backref_cache.path[level];
3983                 BUG_ON(node->bytenr != buf->start &&
3984                        node->new_bytenr != buf->start);
3985
3986                 btrfs_backref_drop_node_buffer(node);
3987                 atomic_inc(&cow->refs);
3988                 node->eb = cow;
3989                 node->new_bytenr = cow->start;
3990
3991                 if (!node->pending) {
3992                         list_move_tail(&node->list,
3993                                        &rc->backref_cache.pending[level]);
3994                         node->pending = 1;
3995                 }
3996
3997                 if (first_cow)
3998                         mark_block_processed(rc, node);
3999
4000                 if (first_cow && level > 0)
4001                         rc->nodes_relocated += buf->len;
4002         }
4003
4004         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
4005                 ret = replace_file_extents(trans, rc, root, cow);
4006         return ret;
4007 }
4008
4009 /*
4010  * called before creating snapshot. it calculates metadata reservation
4011  * required for relocating tree blocks in the snapshot
4012  */
4013 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
4014                               u64 *bytes_to_reserve)
4015 {
4016         struct btrfs_root *root = pending->root;
4017         struct reloc_control *rc = root->fs_info->reloc_ctl;
4018
4019         if (!rc || !have_reloc_root(root))
4020                 return;
4021
4022         if (!rc->merge_reloc_tree)
4023                 return;
4024
4025         root = root->reloc_root;
4026         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
4027         /*
4028          * relocation is in the stage of merging trees. the space
4029          * used by merging a reloc tree is twice the size of
4030          * relocated tree nodes in the worst case. half for cowing
4031          * the reloc tree, half for cowing the fs tree. the space
4032          * used by cowing the reloc tree will be freed after the
4033          * tree is dropped. if we create snapshot, cowing the fs
4034          * tree may use more space than it frees. so we need
4035          * reserve extra space.
4036          */
4037         *bytes_to_reserve += rc->nodes_relocated;
4038 }
4039
4040 /*
4041  * called after snapshot is created. migrate block reservation
4042  * and create reloc root for the newly created snapshot
4043  *
4044  * This is similar to btrfs_init_reloc_root(), we come out of here with two
4045  * references held on the reloc_root, one for root->reloc_root and one for
4046  * rc->reloc_roots.
4047  */
4048 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4049                                struct btrfs_pending_snapshot *pending)
4050 {
4051         struct btrfs_root *root = pending->root;
4052         struct btrfs_root *reloc_root;
4053         struct btrfs_root *new_root;
4054         struct reloc_control *rc = root->fs_info->reloc_ctl;
4055         int ret;
4056
4057         if (!rc || !have_reloc_root(root))
4058                 return 0;
4059
4060         rc = root->fs_info->reloc_ctl;
4061         rc->merging_rsv_size += rc->nodes_relocated;
4062
4063         if (rc->merge_reloc_tree) {
4064                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4065                                               rc->block_rsv,
4066                                               rc->nodes_relocated, true);
4067                 if (ret)
4068                         return ret;
4069         }
4070
4071         new_root = pending->snap;
4072         reloc_root = create_reloc_root(trans, root->reloc_root,
4073                                        new_root->root_key.objectid);
4074         if (IS_ERR(reloc_root))
4075                 return PTR_ERR(reloc_root);
4076
4077         ret = __add_reloc_root(reloc_root);
4078         BUG_ON(ret < 0);
4079         new_root->reloc_root = btrfs_grab_root(reloc_root);
4080
4081         if (rc->create_reloc_tree)
4082                 ret = clone_backref_node(trans, rc, root, reloc_root);
4083         return ret;
4084 }