Merge tag 'amdtee-fixes-for-5.10' of git://git.linaro.org:/people/jens.wiklander...
[linux-2.6-microblaze.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                 if (!node)
673                         return;
674                 BUG_ON((struct btrfs_root *)node->data != root);
675         }
676
677         /*
678          * We only put the reloc root here if it's on the list.  There's a lot
679          * of places where the pattern is to splice the rc->reloc_roots, process
680          * the reloc roots, and then add the reloc root back onto
681          * rc->reloc_roots.  If we call __del_reloc_root while it's off of the
682          * list we don't want the reference being dropped, because the guy
683          * messing with the list is in charge of the reference.
684          */
685         spin_lock(&fs_info->trans_lock);
686         if (!list_empty(&root->root_list)) {
687                 put_ref = true;
688                 list_del_init(&root->root_list);
689         }
690         spin_unlock(&fs_info->trans_lock);
691         if (put_ref)
692                 btrfs_put_root(root);
693         kfree(node);
694 }
695
696 /*
697  * helper to update the 'address of tree root -> reloc tree'
698  * mapping
699  */
700 static int __update_reloc_root(struct btrfs_root *root)
701 {
702         struct btrfs_fs_info *fs_info = root->fs_info;
703         struct rb_node *rb_node;
704         struct mapping_node *node = NULL;
705         struct reloc_control *rc = fs_info->reloc_ctl;
706
707         spin_lock(&rc->reloc_root_tree.lock);
708         rb_node = rb_simple_search(&rc->reloc_root_tree.rb_root,
709                                    root->commit_root->start);
710         if (rb_node) {
711                 node = rb_entry(rb_node, struct mapping_node, rb_node);
712                 rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root);
713         }
714         spin_unlock(&rc->reloc_root_tree.lock);
715
716         if (!node)
717                 return 0;
718         BUG_ON((struct btrfs_root *)node->data != root);
719
720         spin_lock(&rc->reloc_root_tree.lock);
721         node->bytenr = root->node->start;
722         rb_node = rb_simple_insert(&rc->reloc_root_tree.rb_root,
723                                    node->bytenr, &node->rb_node);
724         spin_unlock(&rc->reloc_root_tree.lock);
725         if (rb_node)
726                 btrfs_backref_panic(fs_info, node->bytenr, -EEXIST);
727         return 0;
728 }
729
730 static struct btrfs_root *create_reloc_root(struct btrfs_trans_handle *trans,
731                                         struct btrfs_root *root, u64 objectid)
732 {
733         struct btrfs_fs_info *fs_info = root->fs_info;
734         struct btrfs_root *reloc_root;
735         struct extent_buffer *eb;
736         struct btrfs_root_item *root_item;
737         struct btrfs_key root_key;
738         int ret;
739
740         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
741         BUG_ON(!root_item);
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                 BUG_ON(ret);
754                 /*
755                  * Set the last_snapshot field to the generation of the commit
756                  * root - like this ctree.c:btrfs_block_can_be_shared() behaves
757                  * correctly (returns true) when the relocation root is created
758                  * either inside the critical section of a transaction commit
759                  * (through transaction.c:qgroup_account_snapshot()) and when
760                  * it's created before the transaction commit is started.
761                  */
762                 commit_root_gen = btrfs_header_generation(root->commit_root);
763                 btrfs_set_root_last_snapshot(&root->root_item, commit_root_gen);
764         } else {
765                 /*
766                  * called by btrfs_reloc_post_snapshot_hook.
767                  * the source tree is a reloc tree, all tree blocks
768                  * modified after it was created have RELOC flag
769                  * set in their headers. so it's OK to not update
770                  * the 'last_snapshot'.
771                  */
772                 ret = btrfs_copy_root(trans, root, root->node, &eb,
773                                       BTRFS_TREE_RELOC_OBJECTID);
774                 BUG_ON(ret);
775         }
776
777         memcpy(root_item, &root->root_item, sizeof(*root_item));
778         btrfs_set_root_bytenr(root_item, eb->start);
779         btrfs_set_root_level(root_item, btrfs_header_level(eb));
780         btrfs_set_root_generation(root_item, trans->transid);
781
782         if (root->root_key.objectid == objectid) {
783                 btrfs_set_root_refs(root_item, 0);
784                 memset(&root_item->drop_progress, 0,
785                        sizeof(struct btrfs_disk_key));
786                 root_item->drop_level = 0;
787         }
788
789         btrfs_tree_unlock(eb);
790         free_extent_buffer(eb);
791
792         ret = btrfs_insert_root(trans, fs_info->tree_root,
793                                 &root_key, root_item);
794         BUG_ON(ret);
795         kfree(root_item);
796
797         reloc_root = btrfs_read_tree_root(fs_info->tree_root, &root_key);
798         BUG_ON(IS_ERR(reloc_root));
799         set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
800         reloc_root->last_trans = trans->transid;
801         return reloc_root;
802 }
803
804 /*
805  * create reloc tree for a given fs tree. reloc tree is just a
806  * snapshot of the fs tree with special root objectid.
807  *
808  * The reloc_root comes out of here with two references, one for
809  * root->reloc_root, and another for being on the rc->reloc_roots list.
810  */
811 int btrfs_init_reloc_root(struct btrfs_trans_handle *trans,
812                           struct btrfs_root *root)
813 {
814         struct btrfs_fs_info *fs_info = root->fs_info;
815         struct btrfs_root *reloc_root;
816         struct reloc_control *rc = fs_info->reloc_ctl;
817         struct btrfs_block_rsv *rsv;
818         int clear_rsv = 0;
819         int ret;
820
821         if (!rc)
822                 return 0;
823
824         /*
825          * The subvolume has reloc tree but the swap is finished, no need to
826          * create/update the dead reloc tree
827          */
828         if (reloc_root_is_dead(root))
829                 return 0;
830
831         /*
832          * This is subtle but important.  We do not do
833          * record_root_in_transaction for reloc roots, instead we record their
834          * corresponding fs root, and then here we update the last trans for the
835          * reloc root.  This means that we have to do this for the entire life
836          * of the reloc root, regardless of which stage of the relocation we are
837          * in.
838          */
839         if (root->reloc_root) {
840                 reloc_root = root->reloc_root;
841                 reloc_root->last_trans = trans->transid;
842                 return 0;
843         }
844
845         /*
846          * We are merging reloc roots, we do not need new reloc trees.  Also
847          * reloc trees never need their own reloc tree.
848          */
849         if (!rc->create_reloc_tree ||
850             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
851                 return 0;
852
853         if (!trans->reloc_reserved) {
854                 rsv = trans->block_rsv;
855                 trans->block_rsv = rc->block_rsv;
856                 clear_rsv = 1;
857         }
858         reloc_root = create_reloc_root(trans, root, root->root_key.objectid);
859         if (clear_rsv)
860                 trans->block_rsv = rsv;
861
862         ret = __add_reloc_root(reloc_root);
863         BUG_ON(ret < 0);
864         root->reloc_root = btrfs_grab_root(reloc_root);
865         return 0;
866 }
867
868 /*
869  * update root item of reloc tree
870  */
871 int btrfs_update_reloc_root(struct btrfs_trans_handle *trans,
872                             struct btrfs_root *root)
873 {
874         struct btrfs_fs_info *fs_info = root->fs_info;
875         struct btrfs_root *reloc_root;
876         struct btrfs_root_item *root_item;
877         int ret;
878
879         if (!have_reloc_root(root))
880                 goto out;
881
882         reloc_root = root->reloc_root;
883         root_item = &reloc_root->root_item;
884
885         /*
886          * We are probably ok here, but __del_reloc_root() will drop its ref of
887          * the root.  We have the ref for root->reloc_root, but just in case
888          * hold it while we update the reloc root.
889          */
890         btrfs_grab_root(reloc_root);
891
892         /* root->reloc_root will stay until current relocation finished */
893         if (fs_info->reloc_ctl->merge_reloc_tree &&
894             btrfs_root_refs(root_item) == 0) {
895                 set_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
896                 /*
897                  * Mark the tree as dead before we change reloc_root so
898                  * have_reloc_root will not touch it from now on.
899                  */
900                 smp_wmb();
901                 __del_reloc_root(reloc_root);
902         }
903
904         if (reloc_root->commit_root != reloc_root->node) {
905                 __update_reloc_root(reloc_root);
906                 btrfs_set_root_node(root_item, reloc_root->node);
907                 free_extent_buffer(reloc_root->commit_root);
908                 reloc_root->commit_root = btrfs_root_node(reloc_root);
909         }
910
911         ret = btrfs_update_root(trans, fs_info->tree_root,
912                                 &reloc_root->root_key, root_item);
913         BUG_ON(ret);
914         btrfs_put_root(reloc_root);
915 out:
916         return 0;
917 }
918
919 /*
920  * helper to find first cached inode with inode number >= objectid
921  * in a subvolume
922  */
923 static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid)
924 {
925         struct rb_node *node;
926         struct rb_node *prev;
927         struct btrfs_inode *entry;
928         struct inode *inode;
929
930         spin_lock(&root->inode_lock);
931 again:
932         node = root->inode_tree.rb_node;
933         prev = NULL;
934         while (node) {
935                 prev = node;
936                 entry = rb_entry(node, struct btrfs_inode, rb_node);
937
938                 if (objectid < btrfs_ino(entry))
939                         node = node->rb_left;
940                 else if (objectid > btrfs_ino(entry))
941                         node = node->rb_right;
942                 else
943                         break;
944         }
945         if (!node) {
946                 while (prev) {
947                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
948                         if (objectid <= btrfs_ino(entry)) {
949                                 node = prev;
950                                 break;
951                         }
952                         prev = rb_next(prev);
953                 }
954         }
955         while (node) {
956                 entry = rb_entry(node, struct btrfs_inode, rb_node);
957                 inode = igrab(&entry->vfs_inode);
958                 if (inode) {
959                         spin_unlock(&root->inode_lock);
960                         return inode;
961                 }
962
963                 objectid = btrfs_ino(entry) + 1;
964                 if (cond_resched_lock(&root->inode_lock))
965                         goto again;
966
967                 node = rb_next(node);
968         }
969         spin_unlock(&root->inode_lock);
970         return NULL;
971 }
972
973 /*
974  * get new location of data
975  */
976 static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr,
977                             u64 bytenr, u64 num_bytes)
978 {
979         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
980         struct btrfs_path *path;
981         struct btrfs_file_extent_item *fi;
982         struct extent_buffer *leaf;
983         int ret;
984
985         path = btrfs_alloc_path();
986         if (!path)
987                 return -ENOMEM;
988
989         bytenr -= BTRFS_I(reloc_inode)->index_cnt;
990         ret = btrfs_lookup_file_extent(NULL, root, path,
991                         btrfs_ino(BTRFS_I(reloc_inode)), bytenr, 0);
992         if (ret < 0)
993                 goto out;
994         if (ret > 0) {
995                 ret = -ENOENT;
996                 goto out;
997         }
998
999         leaf = path->nodes[0];
1000         fi = btrfs_item_ptr(leaf, path->slots[0],
1001                             struct btrfs_file_extent_item);
1002
1003         BUG_ON(btrfs_file_extent_offset(leaf, fi) ||
1004                btrfs_file_extent_compression(leaf, fi) ||
1005                btrfs_file_extent_encryption(leaf, fi) ||
1006                btrfs_file_extent_other_encoding(leaf, fi));
1007
1008         if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1009                 ret = -EINVAL;
1010                 goto out;
1011         }
1012
1013         *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1014         ret = 0;
1015 out:
1016         btrfs_free_path(path);
1017         return ret;
1018 }
1019
1020 /*
1021  * update file extent items in the tree leaf to point to
1022  * the new locations.
1023  */
1024 static noinline_for_stack
1025 int replace_file_extents(struct btrfs_trans_handle *trans,
1026                          struct reloc_control *rc,
1027                          struct btrfs_root *root,
1028                          struct extent_buffer *leaf)
1029 {
1030         struct btrfs_fs_info *fs_info = root->fs_info;
1031         struct btrfs_key key;
1032         struct btrfs_file_extent_item *fi;
1033         struct inode *inode = NULL;
1034         u64 parent;
1035         u64 bytenr;
1036         u64 new_bytenr = 0;
1037         u64 num_bytes;
1038         u64 end;
1039         u32 nritems;
1040         u32 i;
1041         int ret = 0;
1042         int first = 1;
1043         int dirty = 0;
1044
1045         if (rc->stage != UPDATE_DATA_PTRS)
1046                 return 0;
1047
1048         /* reloc trees always use full backref */
1049         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1050                 parent = leaf->start;
1051         else
1052                 parent = 0;
1053
1054         nritems = btrfs_header_nritems(leaf);
1055         for (i = 0; i < nritems; i++) {
1056                 struct btrfs_ref ref = { 0 };
1057
1058                 cond_resched();
1059                 btrfs_item_key_to_cpu(leaf, &key, i);
1060                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1061                         continue;
1062                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
1063                 if (btrfs_file_extent_type(leaf, fi) ==
1064                     BTRFS_FILE_EXTENT_INLINE)
1065                         continue;
1066                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1067                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1068                 if (bytenr == 0)
1069                         continue;
1070                 if (!in_range(bytenr, rc->block_group->start,
1071                               rc->block_group->length))
1072                         continue;
1073
1074                 /*
1075                  * if we are modifying block in fs tree, wait for readpage
1076                  * to complete and drop the extent cache
1077                  */
1078                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1079                         if (first) {
1080                                 inode = find_next_inode(root, key.objectid);
1081                                 first = 0;
1082                         } else if (inode && btrfs_ino(BTRFS_I(inode)) < key.objectid) {
1083                                 btrfs_add_delayed_iput(inode);
1084                                 inode = find_next_inode(root, key.objectid);
1085                         }
1086                         if (inode && btrfs_ino(BTRFS_I(inode)) == key.objectid) {
1087                                 end = key.offset +
1088                                       btrfs_file_extent_num_bytes(leaf, fi);
1089                                 WARN_ON(!IS_ALIGNED(key.offset,
1090                                                     fs_info->sectorsize));
1091                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1092                                 end--;
1093                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
1094                                                       key.offset, end);
1095                                 if (!ret)
1096                                         continue;
1097
1098                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1099                                                 key.offset,     end, 1);
1100                                 unlock_extent(&BTRFS_I(inode)->io_tree,
1101                                               key.offset, end);
1102                         }
1103                 }
1104
1105                 ret = get_new_location(rc->data_inode, &new_bytenr,
1106                                        bytenr, num_bytes);
1107                 if (ret) {
1108                         /*
1109                          * Don't have to abort since we've not changed anything
1110                          * in the file extent yet.
1111                          */
1112                         break;
1113                 }
1114
1115                 btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr);
1116                 dirty = 1;
1117
1118                 key.offset -= btrfs_file_extent_offset(leaf, fi);
1119                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1120                                        num_bytes, parent);
1121                 ref.real_root = root->root_key.objectid;
1122                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1123                                     key.objectid, key.offset);
1124                 ret = btrfs_inc_extent_ref(trans, &ref);
1125                 if (ret) {
1126                         btrfs_abort_transaction(trans, ret);
1127                         break;
1128                 }
1129
1130                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1131                                        num_bytes, parent);
1132                 ref.real_root = root->root_key.objectid;
1133                 btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
1134                                     key.objectid, key.offset);
1135                 ret = btrfs_free_extent(trans, &ref);
1136                 if (ret) {
1137                         btrfs_abort_transaction(trans, ret);
1138                         break;
1139                 }
1140         }
1141         if (dirty)
1142                 btrfs_mark_buffer_dirty(leaf);
1143         if (inode)
1144                 btrfs_add_delayed_iput(inode);
1145         return ret;
1146 }
1147
1148 static noinline_for_stack
1149 int memcmp_node_keys(struct extent_buffer *eb, int slot,
1150                      struct btrfs_path *path, int level)
1151 {
1152         struct btrfs_disk_key key1;
1153         struct btrfs_disk_key key2;
1154         btrfs_node_key(eb, &key1, slot);
1155         btrfs_node_key(path->nodes[level], &key2, path->slots[level]);
1156         return memcmp(&key1, &key2, sizeof(key1));
1157 }
1158
1159 /*
1160  * try to replace tree blocks in fs tree with the new blocks
1161  * in reloc tree. tree blocks haven't been modified since the
1162  * reloc tree was create can be replaced.
1163  *
1164  * if a block was replaced, level of the block + 1 is returned.
1165  * if no block got replaced, 0 is returned. if there are other
1166  * errors, a negative error number is returned.
1167  */
1168 static noinline_for_stack
1169 int replace_path(struct btrfs_trans_handle *trans, struct reloc_control *rc,
1170                  struct btrfs_root *dest, struct btrfs_root *src,
1171                  struct btrfs_path *path, struct btrfs_key *next_key,
1172                  int lowest_level, int max_level)
1173 {
1174         struct btrfs_fs_info *fs_info = dest->fs_info;
1175         struct extent_buffer *eb;
1176         struct extent_buffer *parent;
1177         struct btrfs_ref ref = { 0 };
1178         struct btrfs_key key;
1179         u64 old_bytenr;
1180         u64 new_bytenr;
1181         u64 old_ptr_gen;
1182         u64 new_ptr_gen;
1183         u64 last_snapshot;
1184         u32 blocksize;
1185         int cow = 0;
1186         int level;
1187         int ret;
1188         int slot;
1189
1190         BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1191         BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID);
1192
1193         last_snapshot = btrfs_root_last_snapshot(&src->root_item);
1194 again:
1195         slot = path->slots[lowest_level];
1196         btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot);
1197
1198         eb = btrfs_lock_root_node(dest);
1199         btrfs_set_lock_blocking_write(eb);
1200         level = btrfs_header_level(eb);
1201
1202         if (level < lowest_level) {
1203                 btrfs_tree_unlock(eb);
1204                 free_extent_buffer(eb);
1205                 return 0;
1206         }
1207
1208         if (cow) {
1209                 ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb,
1210                                       BTRFS_NESTING_COW);
1211                 BUG_ON(ret);
1212         }
1213         btrfs_set_lock_blocking_write(eb);
1214
1215         if (next_key) {
1216                 next_key->objectid = (u64)-1;
1217                 next_key->type = (u8)-1;
1218                 next_key->offset = (u64)-1;
1219         }
1220
1221         parent = eb;
1222         while (1) {
1223                 struct btrfs_key first_key;
1224
1225                 level = btrfs_header_level(parent);
1226                 BUG_ON(level < lowest_level);
1227
1228                 ret = btrfs_bin_search(parent, &key, &slot);
1229                 if (ret < 0)
1230                         break;
1231                 if (ret && slot > 0)
1232                         slot--;
1233
1234                 if (next_key && slot + 1 < btrfs_header_nritems(parent))
1235                         btrfs_node_key_to_cpu(parent, next_key, slot + 1);
1236
1237                 old_bytenr = btrfs_node_blockptr(parent, slot);
1238                 blocksize = fs_info->nodesize;
1239                 old_ptr_gen = btrfs_node_ptr_generation(parent, slot);
1240                 btrfs_node_key_to_cpu(parent, &first_key, slot);
1241
1242                 if (level <= max_level) {
1243                         eb = path->nodes[level];
1244                         new_bytenr = btrfs_node_blockptr(eb,
1245                                                         path->slots[level]);
1246                         new_ptr_gen = btrfs_node_ptr_generation(eb,
1247                                                         path->slots[level]);
1248                 } else {
1249                         new_bytenr = 0;
1250                         new_ptr_gen = 0;
1251                 }
1252
1253                 if (WARN_ON(new_bytenr > 0 && new_bytenr == old_bytenr)) {
1254                         ret = level;
1255                         break;
1256                 }
1257
1258                 if (new_bytenr == 0 || old_ptr_gen > last_snapshot ||
1259                     memcmp_node_keys(parent, slot, path, level)) {
1260                         if (level <= lowest_level) {
1261                                 ret = 0;
1262                                 break;
1263                         }
1264
1265                         eb = read_tree_block(fs_info, old_bytenr, old_ptr_gen,
1266                                              level - 1, &first_key);
1267                         if (IS_ERR(eb)) {
1268                                 ret = PTR_ERR(eb);
1269                                 break;
1270                         } else if (!extent_buffer_uptodate(eb)) {
1271                                 ret = -EIO;
1272                                 free_extent_buffer(eb);
1273                                 break;
1274                         }
1275                         btrfs_tree_lock(eb);
1276                         if (cow) {
1277                                 ret = btrfs_cow_block(trans, dest, eb, parent,
1278                                                       slot, &eb,
1279                                                       BTRFS_NESTING_COW);
1280                                 BUG_ON(ret);
1281                         }
1282                         btrfs_set_lock_blocking_write(eb);
1283
1284                         btrfs_tree_unlock(parent);
1285                         free_extent_buffer(parent);
1286
1287                         parent = eb;
1288                         continue;
1289                 }
1290
1291                 if (!cow) {
1292                         btrfs_tree_unlock(parent);
1293                         free_extent_buffer(parent);
1294                         cow = 1;
1295                         goto again;
1296                 }
1297
1298                 btrfs_node_key_to_cpu(path->nodes[level], &key,
1299                                       path->slots[level]);
1300                 btrfs_release_path(path);
1301
1302                 path->lowest_level = level;
1303                 ret = btrfs_search_slot(trans, src, &key, path, 0, 1);
1304                 path->lowest_level = 0;
1305                 BUG_ON(ret);
1306
1307                 /*
1308                  * Info qgroup to trace both subtrees.
1309                  *
1310                  * We must trace both trees.
1311                  * 1) Tree reloc subtree
1312                  *    If not traced, we will leak data numbers
1313                  * 2) Fs subtree
1314                  *    If not traced, we will double count old data
1315                  *
1316                  * We don't scan the subtree right now, but only record
1317                  * the swapped tree blocks.
1318                  * The real subtree rescan is delayed until we have new
1319                  * CoW on the subtree root node before transaction commit.
1320                  */
1321                 ret = btrfs_qgroup_add_swapped_blocks(trans, dest,
1322                                 rc->block_group, parent, slot,
1323                                 path->nodes[level], path->slots[level],
1324                                 last_snapshot);
1325                 if (ret < 0)
1326                         break;
1327                 /*
1328                  * swap blocks in fs tree and reloc tree.
1329                  */
1330                 btrfs_set_node_blockptr(parent, slot, new_bytenr);
1331                 btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen);
1332                 btrfs_mark_buffer_dirty(parent);
1333
1334                 btrfs_set_node_blockptr(path->nodes[level],
1335                                         path->slots[level], old_bytenr);
1336                 btrfs_set_node_ptr_generation(path->nodes[level],
1337                                               path->slots[level], old_ptr_gen);
1338                 btrfs_mark_buffer_dirty(path->nodes[level]);
1339
1340                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, old_bytenr,
1341                                        blocksize, path->nodes[level]->start);
1342                 ref.skip_qgroup = true;
1343                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1344                 ret = btrfs_inc_extent_ref(trans, &ref);
1345                 BUG_ON(ret);
1346                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new_bytenr,
1347                                        blocksize, 0);
1348                 ref.skip_qgroup = true;
1349                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1350                 ret = btrfs_inc_extent_ref(trans, &ref);
1351                 BUG_ON(ret);
1352
1353                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, new_bytenr,
1354                                        blocksize, path->nodes[level]->start);
1355                 btrfs_init_tree_ref(&ref, level - 1, src->root_key.objectid);
1356                 ref.skip_qgroup = true;
1357                 ret = btrfs_free_extent(trans, &ref);
1358                 BUG_ON(ret);
1359
1360                 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, old_bytenr,
1361                                        blocksize, 0);
1362                 btrfs_init_tree_ref(&ref, level - 1, dest->root_key.objectid);
1363                 ref.skip_qgroup = true;
1364                 ret = btrfs_free_extent(trans, &ref);
1365                 BUG_ON(ret);
1366
1367                 btrfs_unlock_up_safe(path, 0);
1368
1369                 ret = level;
1370                 break;
1371         }
1372         btrfs_tree_unlock(parent);
1373         free_extent_buffer(parent);
1374         return ret;
1375 }
1376
1377 /*
1378  * helper to find next relocated block in reloc tree
1379  */
1380 static noinline_for_stack
1381 int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1382                        int *level)
1383 {
1384         struct extent_buffer *eb;
1385         int i;
1386         u64 last_snapshot;
1387         u32 nritems;
1388
1389         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1390
1391         for (i = 0; i < *level; i++) {
1392                 free_extent_buffer(path->nodes[i]);
1393                 path->nodes[i] = NULL;
1394         }
1395
1396         for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
1397                 eb = path->nodes[i];
1398                 nritems = btrfs_header_nritems(eb);
1399                 while (path->slots[i] + 1 < nritems) {
1400                         path->slots[i]++;
1401                         if (btrfs_node_ptr_generation(eb, path->slots[i]) <=
1402                             last_snapshot)
1403                                 continue;
1404
1405                         *level = i;
1406                         return 0;
1407                 }
1408                 free_extent_buffer(path->nodes[i]);
1409                 path->nodes[i] = NULL;
1410         }
1411         return 1;
1412 }
1413
1414 /*
1415  * walk down reloc tree to find relocated block of lowest level
1416  */
1417 static noinline_for_stack
1418 int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path,
1419                          int *level)
1420 {
1421         struct btrfs_fs_info *fs_info = root->fs_info;
1422         struct extent_buffer *eb = NULL;
1423         int i;
1424         u64 bytenr;
1425         u64 ptr_gen = 0;
1426         u64 last_snapshot;
1427         u32 nritems;
1428
1429         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1430
1431         for (i = *level; i > 0; i--) {
1432                 struct btrfs_key first_key;
1433
1434                 eb = path->nodes[i];
1435                 nritems = btrfs_header_nritems(eb);
1436                 while (path->slots[i] < nritems) {
1437                         ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]);
1438                         if (ptr_gen > last_snapshot)
1439                                 break;
1440                         path->slots[i]++;
1441                 }
1442                 if (path->slots[i] >= nritems) {
1443                         if (i == *level)
1444                                 break;
1445                         *level = i + 1;
1446                         return 0;
1447                 }
1448                 if (i == 1) {
1449                         *level = i;
1450                         return 0;
1451                 }
1452
1453                 bytenr = btrfs_node_blockptr(eb, path->slots[i]);
1454                 btrfs_node_key_to_cpu(eb, &first_key, path->slots[i]);
1455                 eb = read_tree_block(fs_info, bytenr, ptr_gen, i - 1,
1456                                      &first_key);
1457                 if (IS_ERR(eb)) {
1458                         return PTR_ERR(eb);
1459                 } else if (!extent_buffer_uptodate(eb)) {
1460                         free_extent_buffer(eb);
1461                         return -EIO;
1462                 }
1463                 BUG_ON(btrfs_header_level(eb) != i - 1);
1464                 path->nodes[i - 1] = eb;
1465                 path->slots[i - 1] = 0;
1466         }
1467         return 1;
1468 }
1469
1470 /*
1471  * invalidate extent cache for file extents whose key in range of
1472  * [min_key, max_key)
1473  */
1474 static int invalidate_extent_cache(struct btrfs_root *root,
1475                                    struct btrfs_key *min_key,
1476                                    struct btrfs_key *max_key)
1477 {
1478         struct btrfs_fs_info *fs_info = root->fs_info;
1479         struct inode *inode = NULL;
1480         u64 objectid;
1481         u64 start, end;
1482         u64 ino;
1483
1484         objectid = min_key->objectid;
1485         while (1) {
1486                 cond_resched();
1487                 iput(inode);
1488
1489                 if (objectid > max_key->objectid)
1490                         break;
1491
1492                 inode = find_next_inode(root, objectid);
1493                 if (!inode)
1494                         break;
1495                 ino = btrfs_ino(BTRFS_I(inode));
1496
1497                 if (ino > max_key->objectid) {
1498                         iput(inode);
1499                         break;
1500                 }
1501
1502                 objectid = ino + 1;
1503                 if (!S_ISREG(inode->i_mode))
1504                         continue;
1505
1506                 if (unlikely(min_key->objectid == ino)) {
1507                         if (min_key->type > BTRFS_EXTENT_DATA_KEY)
1508                                 continue;
1509                         if (min_key->type < BTRFS_EXTENT_DATA_KEY)
1510                                 start = 0;
1511                         else {
1512                                 start = min_key->offset;
1513                                 WARN_ON(!IS_ALIGNED(start, fs_info->sectorsize));
1514                         }
1515                 } else {
1516                         start = 0;
1517                 }
1518
1519                 if (unlikely(max_key->objectid == ino)) {
1520                         if (max_key->type < BTRFS_EXTENT_DATA_KEY)
1521                                 continue;
1522                         if (max_key->type > BTRFS_EXTENT_DATA_KEY) {
1523                                 end = (u64)-1;
1524                         } else {
1525                                 if (max_key->offset == 0)
1526                                         continue;
1527                                 end = max_key->offset;
1528                                 WARN_ON(!IS_ALIGNED(end, fs_info->sectorsize));
1529                                 end--;
1530                         }
1531                 } else {
1532                         end = (u64)-1;
1533                 }
1534
1535                 /* the lock_extent waits for readpage to complete */
1536                 lock_extent(&BTRFS_I(inode)->io_tree, start, end);
1537                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 1);
1538                 unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1539         }
1540         return 0;
1541 }
1542
1543 static int find_next_key(struct btrfs_path *path, int level,
1544                          struct btrfs_key *key)
1545
1546 {
1547         while (level < BTRFS_MAX_LEVEL) {
1548                 if (!path->nodes[level])
1549                         break;
1550                 if (path->slots[level] + 1 <
1551                     btrfs_header_nritems(path->nodes[level])) {
1552                         btrfs_node_key_to_cpu(path->nodes[level], key,
1553                                               path->slots[level] + 1);
1554                         return 0;
1555                 }
1556                 level++;
1557         }
1558         return 1;
1559 }
1560
1561 /*
1562  * Insert current subvolume into reloc_control::dirty_subvol_roots
1563  */
1564 static void insert_dirty_subvol(struct btrfs_trans_handle *trans,
1565                                 struct reloc_control *rc,
1566                                 struct btrfs_root *root)
1567 {
1568         struct btrfs_root *reloc_root = root->reloc_root;
1569         struct btrfs_root_item *reloc_root_item;
1570
1571         /* @root must be a subvolume tree root with a valid reloc tree */
1572         ASSERT(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
1573         ASSERT(reloc_root);
1574
1575         reloc_root_item = &reloc_root->root_item;
1576         memset(&reloc_root_item->drop_progress, 0,
1577                 sizeof(reloc_root_item->drop_progress));
1578         reloc_root_item->drop_level = 0;
1579         btrfs_set_root_refs(reloc_root_item, 0);
1580         btrfs_update_reloc_root(trans, root);
1581
1582         if (list_empty(&root->reloc_dirty_list)) {
1583                 btrfs_grab_root(root);
1584                 list_add_tail(&root->reloc_dirty_list, &rc->dirty_subvol_roots);
1585         }
1586 }
1587
1588 static int clean_dirty_subvols(struct reloc_control *rc)
1589 {
1590         struct btrfs_root *root;
1591         struct btrfs_root *next;
1592         int ret = 0;
1593         int ret2;
1594
1595         list_for_each_entry_safe(root, next, &rc->dirty_subvol_roots,
1596                                  reloc_dirty_list) {
1597                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1598                         /* Merged subvolume, cleanup its reloc root */
1599                         struct btrfs_root *reloc_root = root->reloc_root;
1600
1601                         list_del_init(&root->reloc_dirty_list);
1602                         root->reloc_root = NULL;
1603                         /*
1604                          * Need barrier to ensure clear_bit() only happens after
1605                          * root->reloc_root = NULL. Pairs with have_reloc_root.
1606                          */
1607                         smp_wmb();
1608                         clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE, &root->state);
1609                         if (reloc_root) {
1610                                 /*
1611                                  * btrfs_drop_snapshot drops our ref we hold for
1612                                  * ->reloc_root.  If it fails however we must
1613                                  * drop the ref ourselves.
1614                                  */
1615                                 ret2 = btrfs_drop_snapshot(reloc_root, 0, 1);
1616                                 if (ret2 < 0) {
1617                                         btrfs_put_root(reloc_root);
1618                                         if (!ret)
1619                                                 ret = ret2;
1620                                 }
1621                         }
1622                         btrfs_put_root(root);
1623                 } else {
1624                         /* Orphan reloc tree, just clean it up */
1625                         ret2 = btrfs_drop_snapshot(root, 0, 1);
1626                         if (ret2 < 0) {
1627                                 btrfs_put_root(root);
1628                                 if (!ret)
1629                                         ret = ret2;
1630                         }
1631                 }
1632         }
1633         return ret;
1634 }
1635
1636 /*
1637  * merge the relocated tree blocks in reloc tree with corresponding
1638  * fs tree.
1639  */
1640 static noinline_for_stack int merge_reloc_root(struct reloc_control *rc,
1641                                                struct btrfs_root *root)
1642 {
1643         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1644         struct btrfs_key key;
1645         struct btrfs_key next_key;
1646         struct btrfs_trans_handle *trans = NULL;
1647         struct btrfs_root *reloc_root;
1648         struct btrfs_root_item *root_item;
1649         struct btrfs_path *path;
1650         struct extent_buffer *leaf;
1651         int level;
1652         int max_level;
1653         int replaced = 0;
1654         int ret;
1655         int err = 0;
1656         u32 min_reserved;
1657
1658         path = btrfs_alloc_path();
1659         if (!path)
1660                 return -ENOMEM;
1661         path->reada = READA_FORWARD;
1662
1663         reloc_root = root->reloc_root;
1664         root_item = &reloc_root->root_item;
1665
1666         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1667                 level = btrfs_root_level(root_item);
1668                 atomic_inc(&reloc_root->node->refs);
1669                 path->nodes[level] = reloc_root->node;
1670                 path->slots[level] = 0;
1671         } else {
1672                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1673
1674                 level = root_item->drop_level;
1675                 BUG_ON(level == 0);
1676                 path->lowest_level = level;
1677                 ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0);
1678                 path->lowest_level = 0;
1679                 if (ret < 0) {
1680                         btrfs_free_path(path);
1681                         return ret;
1682                 }
1683
1684                 btrfs_node_key_to_cpu(path->nodes[level], &next_key,
1685                                       path->slots[level]);
1686                 WARN_ON(memcmp(&key, &next_key, sizeof(key)));
1687
1688                 btrfs_unlock_up_safe(path, 0);
1689         }
1690
1691         /*
1692          * In merge_reloc_root(), we modify the upper level pointer to swap the
1693          * tree blocks between reloc tree and subvolume tree.  Thus for tree
1694          * block COW, we COW at most from level 1 to root level for each tree.
1695          *
1696          * Thus the needed metadata size is at most root_level * nodesize,
1697          * and * 2 since we have two trees to COW.
1698          */
1699         min_reserved = fs_info->nodesize * btrfs_root_level(root_item) * 2;
1700         memset(&next_key, 0, sizeof(next_key));
1701
1702         while (1) {
1703                 ret = btrfs_block_rsv_refill(root, rc->block_rsv, min_reserved,
1704                                              BTRFS_RESERVE_FLUSH_LIMIT);
1705                 if (ret) {
1706                         err = ret;
1707                         goto out;
1708                 }
1709                 trans = btrfs_start_transaction(root, 0);
1710                 if (IS_ERR(trans)) {
1711                         err = PTR_ERR(trans);
1712                         trans = NULL;
1713                         goto out;
1714                 }
1715
1716                 /*
1717                  * At this point we no longer have a reloc_control, so we can't
1718                  * depend on btrfs_init_reloc_root to update our last_trans.
1719                  *
1720                  * But that's ok, we started the trans handle on our
1721                  * corresponding fs_root, which means it's been added to the
1722                  * dirty list.  At commit time we'll still call
1723                  * btrfs_update_reloc_root() and update our root item
1724                  * appropriately.
1725                  */
1726                 reloc_root->last_trans = trans->transid;
1727                 trans->block_rsv = rc->block_rsv;
1728
1729                 replaced = 0;
1730                 max_level = level;
1731
1732                 ret = walk_down_reloc_tree(reloc_root, path, &level);
1733                 if (ret < 0) {
1734                         err = ret;
1735                         goto out;
1736                 }
1737                 if (ret > 0)
1738                         break;
1739
1740                 if (!find_next_key(path, level, &key) &&
1741                     btrfs_comp_cpu_keys(&next_key, &key) >= 0) {
1742                         ret = 0;
1743                 } else {
1744                         ret = replace_path(trans, rc, root, reloc_root, path,
1745                                            &next_key, level, max_level);
1746                 }
1747                 if (ret < 0) {
1748                         err = ret;
1749                         goto out;
1750                 }
1751
1752                 if (ret > 0) {
1753                         level = ret;
1754                         btrfs_node_key_to_cpu(path->nodes[level], &key,
1755                                               path->slots[level]);
1756                         replaced = 1;
1757                 }
1758
1759                 ret = walk_up_reloc_tree(reloc_root, path, &level);
1760                 if (ret > 0)
1761                         break;
1762
1763                 BUG_ON(level == 0);
1764                 /*
1765                  * save the merging progress in the drop_progress.
1766                  * this is OK since root refs == 1 in this case.
1767                  */
1768                 btrfs_node_key(path->nodes[level], &root_item->drop_progress,
1769                                path->slots[level]);
1770                 root_item->drop_level = level;
1771
1772                 btrfs_end_transaction_throttle(trans);
1773                 trans = NULL;
1774
1775                 btrfs_btree_balance_dirty(fs_info);
1776
1777                 if (replaced && rc->stage == UPDATE_DATA_PTRS)
1778                         invalidate_extent_cache(root, &key, &next_key);
1779         }
1780
1781         /*
1782          * handle the case only one block in the fs tree need to be
1783          * relocated and the block is tree root.
1784          */
1785         leaf = btrfs_lock_root_node(root);
1786         ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf,
1787                               BTRFS_NESTING_COW);
1788         btrfs_tree_unlock(leaf);
1789         free_extent_buffer(leaf);
1790         if (ret < 0)
1791                 err = ret;
1792 out:
1793         btrfs_free_path(path);
1794
1795         if (err == 0)
1796                 insert_dirty_subvol(trans, rc, root);
1797
1798         if (trans)
1799                 btrfs_end_transaction_throttle(trans);
1800
1801         btrfs_btree_balance_dirty(fs_info);
1802
1803         if (replaced && rc->stage == UPDATE_DATA_PTRS)
1804                 invalidate_extent_cache(root, &key, &next_key);
1805
1806         return err;
1807 }
1808
1809 static noinline_for_stack
1810 int prepare_to_merge(struct reloc_control *rc, int err)
1811 {
1812         struct btrfs_root *root = rc->extent_root;
1813         struct btrfs_fs_info *fs_info = root->fs_info;
1814         struct btrfs_root *reloc_root;
1815         struct btrfs_trans_handle *trans;
1816         LIST_HEAD(reloc_roots);
1817         u64 num_bytes = 0;
1818         int ret;
1819
1820         mutex_lock(&fs_info->reloc_mutex);
1821         rc->merging_rsv_size += fs_info->nodesize * (BTRFS_MAX_LEVEL - 1) * 2;
1822         rc->merging_rsv_size += rc->nodes_relocated * 2;
1823         mutex_unlock(&fs_info->reloc_mutex);
1824
1825 again:
1826         if (!err) {
1827                 num_bytes = rc->merging_rsv_size;
1828                 ret = btrfs_block_rsv_add(root, rc->block_rsv, num_bytes,
1829                                           BTRFS_RESERVE_FLUSH_ALL);
1830                 if (ret)
1831                         err = ret;
1832         }
1833
1834         trans = btrfs_join_transaction(rc->extent_root);
1835         if (IS_ERR(trans)) {
1836                 if (!err)
1837                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1838                                                 num_bytes, NULL);
1839                 return PTR_ERR(trans);
1840         }
1841
1842         if (!err) {
1843                 if (num_bytes != rc->merging_rsv_size) {
1844                         btrfs_end_transaction(trans);
1845                         btrfs_block_rsv_release(fs_info, rc->block_rsv,
1846                                                 num_bytes, NULL);
1847                         goto again;
1848                 }
1849         }
1850
1851         rc->merge_reloc_tree = 1;
1852
1853         while (!list_empty(&rc->reloc_roots)) {
1854                 reloc_root = list_entry(rc->reloc_roots.next,
1855                                         struct btrfs_root, root_list);
1856                 list_del_init(&reloc_root->root_list);
1857
1858                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1859                                 false);
1860                 BUG_ON(IS_ERR(root));
1861                 BUG_ON(root->reloc_root != reloc_root);
1862
1863                 /*
1864                  * set reference count to 1, so btrfs_recover_relocation
1865                  * knows it should resumes merging
1866                  */
1867                 if (!err)
1868                         btrfs_set_root_refs(&reloc_root->root_item, 1);
1869                 btrfs_update_reloc_root(trans, root);
1870
1871                 list_add(&reloc_root->root_list, &reloc_roots);
1872                 btrfs_put_root(root);
1873         }
1874
1875         list_splice(&reloc_roots, &rc->reloc_roots);
1876
1877         if (!err)
1878                 btrfs_commit_transaction(trans);
1879         else
1880                 btrfs_end_transaction(trans);
1881         return err;
1882 }
1883
1884 static noinline_for_stack
1885 void free_reloc_roots(struct list_head *list)
1886 {
1887         struct btrfs_root *reloc_root, *tmp;
1888
1889         list_for_each_entry_safe(reloc_root, tmp, list, root_list)
1890                 __del_reloc_root(reloc_root);
1891 }
1892
1893 static noinline_for_stack
1894 void merge_reloc_roots(struct reloc_control *rc)
1895 {
1896         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
1897         struct btrfs_root *root;
1898         struct btrfs_root *reloc_root;
1899         LIST_HEAD(reloc_roots);
1900         int found = 0;
1901         int ret = 0;
1902 again:
1903         root = rc->extent_root;
1904
1905         /*
1906          * this serializes us with btrfs_record_root_in_transaction,
1907          * we have to make sure nobody is in the middle of
1908          * adding their roots to the list while we are
1909          * doing this splice
1910          */
1911         mutex_lock(&fs_info->reloc_mutex);
1912         list_splice_init(&rc->reloc_roots, &reloc_roots);
1913         mutex_unlock(&fs_info->reloc_mutex);
1914
1915         while (!list_empty(&reloc_roots)) {
1916                 found = 1;
1917                 reloc_root = list_entry(reloc_roots.next,
1918                                         struct btrfs_root, root_list);
1919
1920                 root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
1921                                          false);
1922                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
1923                         BUG_ON(IS_ERR(root));
1924                         BUG_ON(root->reloc_root != reloc_root);
1925                         ret = merge_reloc_root(rc, root);
1926                         btrfs_put_root(root);
1927                         if (ret) {
1928                                 if (list_empty(&reloc_root->root_list))
1929                                         list_add_tail(&reloc_root->root_list,
1930                                                       &reloc_roots);
1931                                 goto out;
1932                         }
1933                 } else {
1934                         if (!IS_ERR(root)) {
1935                                 if (root->reloc_root == reloc_root) {
1936                                         root->reloc_root = NULL;
1937                                         btrfs_put_root(reloc_root);
1938                                 }
1939                                 clear_bit(BTRFS_ROOT_DEAD_RELOC_TREE,
1940                                           &root->state);
1941                                 btrfs_put_root(root);
1942                         }
1943
1944                         list_del_init(&reloc_root->root_list);
1945                         /* Don't forget to queue this reloc root for cleanup */
1946                         list_add_tail(&reloc_root->reloc_dirty_list,
1947                                       &rc->dirty_subvol_roots);
1948                 }
1949         }
1950
1951         if (found) {
1952                 found = 0;
1953                 goto again;
1954         }
1955 out:
1956         if (ret) {
1957                 btrfs_handle_fs_error(fs_info, ret, NULL);
1958                 free_reloc_roots(&reloc_roots);
1959
1960                 /* new reloc root may be added */
1961                 mutex_lock(&fs_info->reloc_mutex);
1962                 list_splice_init(&rc->reloc_roots, &reloc_roots);
1963                 mutex_unlock(&fs_info->reloc_mutex);
1964                 free_reloc_roots(&reloc_roots);
1965         }
1966
1967         /*
1968          * We used to have
1969          *
1970          * BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root));
1971          *
1972          * here, but it's wrong.  If we fail to start the transaction in
1973          * prepare_to_merge() we will have only 0 ref reloc roots, none of which
1974          * have actually been removed from the reloc_root_tree rb tree.  This is
1975          * fine because we're bailing here, and we hold a reference on the root
1976          * for the list that holds it, so these roots will be cleaned up when we
1977          * do the reloc_dirty_list afterwards.  Meanwhile the root->reloc_root
1978          * will be cleaned up on unmount.
1979          *
1980          * The remaining nodes will be cleaned up by free_reloc_control.
1981          */
1982 }
1983
1984 static void free_block_list(struct rb_root *blocks)
1985 {
1986         struct tree_block *block;
1987         struct rb_node *rb_node;
1988         while ((rb_node = rb_first(blocks))) {
1989                 block = rb_entry(rb_node, struct tree_block, rb_node);
1990                 rb_erase(rb_node, blocks);
1991                 kfree(block);
1992         }
1993 }
1994
1995 static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans,
1996                                       struct btrfs_root *reloc_root)
1997 {
1998         struct btrfs_fs_info *fs_info = reloc_root->fs_info;
1999         struct btrfs_root *root;
2000         int ret;
2001
2002         if (reloc_root->last_trans == trans->transid)
2003                 return 0;
2004
2005         root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset, false);
2006         BUG_ON(IS_ERR(root));
2007         BUG_ON(root->reloc_root != reloc_root);
2008         ret = btrfs_record_root_in_trans(trans, root);
2009         btrfs_put_root(root);
2010
2011         return ret;
2012 }
2013
2014 static noinline_for_stack
2015 struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans,
2016                                      struct reloc_control *rc,
2017                                      struct btrfs_backref_node *node,
2018                                      struct btrfs_backref_edge *edges[])
2019 {
2020         struct btrfs_backref_node *next;
2021         struct btrfs_root *root;
2022         int index = 0;
2023
2024         next = node;
2025         while (1) {
2026                 cond_resched();
2027                 next = walk_up_backref(next, edges, &index);
2028                 root = next->root;
2029                 BUG_ON(!root);
2030                 BUG_ON(!test_bit(BTRFS_ROOT_SHAREABLE, &root->state));
2031
2032                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
2033                         record_reloc_root_in_trans(trans, root);
2034                         break;
2035                 }
2036
2037                 btrfs_record_root_in_trans(trans, root);
2038                 root = root->reloc_root;
2039
2040                 if (next->new_bytenr != root->node->start) {
2041                         BUG_ON(next->new_bytenr);
2042                         BUG_ON(!list_empty(&next->list));
2043                         next->new_bytenr = root->node->start;
2044                         btrfs_put_root(next->root);
2045                         next->root = btrfs_grab_root(root);
2046                         ASSERT(next->root);
2047                         list_add_tail(&next->list,
2048                                       &rc->backref_cache.changed);
2049                         mark_block_processed(rc, next);
2050                         break;
2051                 }
2052
2053                 WARN_ON(1);
2054                 root = NULL;
2055                 next = walk_down_backref(edges, &index);
2056                 if (!next || next->level <= node->level)
2057                         break;
2058         }
2059         if (!root)
2060                 return NULL;
2061
2062         next = node;
2063         /* setup backref node path for btrfs_reloc_cow_block */
2064         while (1) {
2065                 rc->backref_cache.path[next->level] = next;
2066                 if (--index < 0)
2067                         break;
2068                 next = edges[index]->node[UPPER];
2069         }
2070         return root;
2071 }
2072
2073 /*
2074  * Select a tree root for relocation.
2075  *
2076  * Return NULL if the block is not shareable. We should use do_relocation() in
2077  * this case.
2078  *
2079  * Return a tree root pointer if the block is shareable.
2080  * Return -ENOENT if the block is root of reloc tree.
2081  */
2082 static noinline_for_stack
2083 struct btrfs_root *select_one_root(struct btrfs_backref_node *node)
2084 {
2085         struct btrfs_backref_node *next;
2086         struct btrfs_root *root;
2087         struct btrfs_root *fs_root = NULL;
2088         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2089         int index = 0;
2090
2091         next = node;
2092         while (1) {
2093                 cond_resched();
2094                 next = walk_up_backref(next, edges, &index);
2095                 root = next->root;
2096                 BUG_ON(!root);
2097
2098                 /* No other choice for non-shareable tree */
2099                 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
2100                         return root;
2101
2102                 if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID)
2103                         fs_root = root;
2104
2105                 if (next != node)
2106                         return NULL;
2107
2108                 next = walk_down_backref(edges, &index);
2109                 if (!next || next->level <= node->level)
2110                         break;
2111         }
2112
2113         if (!fs_root)
2114                 return ERR_PTR(-ENOENT);
2115         return fs_root;
2116 }
2117
2118 static noinline_for_stack
2119 u64 calcu_metadata_size(struct reloc_control *rc,
2120                         struct btrfs_backref_node *node, int reserve)
2121 {
2122         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2123         struct btrfs_backref_node *next = node;
2124         struct btrfs_backref_edge *edge;
2125         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2126         u64 num_bytes = 0;
2127         int index = 0;
2128
2129         BUG_ON(reserve && node->processed);
2130
2131         while (next) {
2132                 cond_resched();
2133                 while (1) {
2134                         if (next->processed && (reserve || next != node))
2135                                 break;
2136
2137                         num_bytes += fs_info->nodesize;
2138
2139                         if (list_empty(&next->upper))
2140                                 break;
2141
2142                         edge = list_entry(next->upper.next,
2143                                         struct btrfs_backref_edge, list[LOWER]);
2144                         edges[index++] = edge;
2145                         next = edge->node[UPPER];
2146                 }
2147                 next = walk_down_backref(edges, &index);
2148         }
2149         return num_bytes;
2150 }
2151
2152 static int reserve_metadata_space(struct btrfs_trans_handle *trans,
2153                                   struct reloc_control *rc,
2154                                   struct btrfs_backref_node *node)
2155 {
2156         struct btrfs_root *root = rc->extent_root;
2157         struct btrfs_fs_info *fs_info = root->fs_info;
2158         u64 num_bytes;
2159         int ret;
2160         u64 tmp;
2161
2162         num_bytes = calcu_metadata_size(rc, node, 1) * 2;
2163
2164         trans->block_rsv = rc->block_rsv;
2165         rc->reserved_bytes += num_bytes;
2166
2167         /*
2168          * We are under a transaction here so we can only do limited flushing.
2169          * If we get an enospc just kick back -EAGAIN so we know to drop the
2170          * transaction and try to refill when we can flush all the things.
2171          */
2172         ret = btrfs_block_rsv_refill(root, rc->block_rsv, num_bytes,
2173                                 BTRFS_RESERVE_FLUSH_LIMIT);
2174         if (ret) {
2175                 tmp = fs_info->nodesize * RELOCATION_RESERVED_NODES;
2176                 while (tmp <= rc->reserved_bytes)
2177                         tmp <<= 1;
2178                 /*
2179                  * only one thread can access block_rsv at this point,
2180                  * so we don't need hold lock to protect block_rsv.
2181                  * we expand more reservation size here to allow enough
2182                  * space for relocation and we will return earlier in
2183                  * enospc case.
2184                  */
2185                 rc->block_rsv->size = tmp + fs_info->nodesize *
2186                                       RELOCATION_RESERVED_NODES;
2187                 return -EAGAIN;
2188         }
2189
2190         return 0;
2191 }
2192
2193 /*
2194  * relocate a block tree, and then update pointers in upper level
2195  * blocks that reference the block to point to the new location.
2196  *
2197  * if called by link_to_upper, the block has already been relocated.
2198  * in that case this function just updates pointers.
2199  */
2200 static int do_relocation(struct btrfs_trans_handle *trans,
2201                          struct reloc_control *rc,
2202                          struct btrfs_backref_node *node,
2203                          struct btrfs_key *key,
2204                          struct btrfs_path *path, int lowest)
2205 {
2206         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2207         struct btrfs_backref_node *upper;
2208         struct btrfs_backref_edge *edge;
2209         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2210         struct btrfs_root *root;
2211         struct extent_buffer *eb;
2212         u32 blocksize;
2213         u64 bytenr;
2214         u64 generation;
2215         int slot;
2216         int ret;
2217         int err = 0;
2218
2219         BUG_ON(lowest && node->eb);
2220
2221         path->lowest_level = node->level + 1;
2222         rc->backref_cache.path[node->level] = node;
2223         list_for_each_entry(edge, &node->upper, list[LOWER]) {
2224                 struct btrfs_key first_key;
2225                 struct btrfs_ref ref = { 0 };
2226
2227                 cond_resched();
2228
2229                 upper = edge->node[UPPER];
2230                 root = select_reloc_root(trans, rc, upper, edges);
2231                 BUG_ON(!root);
2232
2233                 if (upper->eb && !upper->locked) {
2234                         if (!lowest) {
2235                                 ret = btrfs_bin_search(upper->eb, key, &slot);
2236                                 if (ret < 0) {
2237                                         err = ret;
2238                                         goto next;
2239                                 }
2240                                 BUG_ON(ret);
2241                                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2242                                 if (node->eb->start == bytenr)
2243                                         goto next;
2244                         }
2245                         btrfs_backref_drop_node_buffer(upper);
2246                 }
2247
2248                 if (!upper->eb) {
2249                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2250                         if (ret) {
2251                                 if (ret < 0)
2252                                         err = ret;
2253                                 else
2254                                         err = -ENOENT;
2255
2256                                 btrfs_release_path(path);
2257                                 break;
2258                         }
2259
2260                         if (!upper->eb) {
2261                                 upper->eb = path->nodes[upper->level];
2262                                 path->nodes[upper->level] = NULL;
2263                         } else {
2264                                 BUG_ON(upper->eb != path->nodes[upper->level]);
2265                         }
2266
2267                         upper->locked = 1;
2268                         path->locks[upper->level] = 0;
2269
2270                         slot = path->slots[upper->level];
2271                         btrfs_release_path(path);
2272                 } else {
2273                         ret = btrfs_bin_search(upper->eb, key, &slot);
2274                         if (ret < 0) {
2275                                 err = ret;
2276                                 goto next;
2277                         }
2278                         BUG_ON(ret);
2279                 }
2280
2281                 bytenr = btrfs_node_blockptr(upper->eb, slot);
2282                 if (lowest) {
2283                         if (bytenr != node->bytenr) {
2284                                 btrfs_err(root->fs_info,
2285                 "lowest leaf/node mismatch: bytenr %llu node->bytenr %llu slot %d upper %llu",
2286                                           bytenr, node->bytenr, slot,
2287                                           upper->eb->start);
2288                                 err = -EIO;
2289                                 goto next;
2290                         }
2291                 } else {
2292                         if (node->eb->start == bytenr)
2293                                 goto next;
2294                 }
2295
2296                 blocksize = root->fs_info->nodesize;
2297                 generation = btrfs_node_ptr_generation(upper->eb, slot);
2298                 btrfs_node_key_to_cpu(upper->eb, &first_key, slot);
2299                 eb = read_tree_block(fs_info, bytenr, generation,
2300                                      upper->level - 1, &first_key);
2301                 if (IS_ERR(eb)) {
2302                         err = PTR_ERR(eb);
2303                         goto next;
2304                 } else if (!extent_buffer_uptodate(eb)) {
2305                         free_extent_buffer(eb);
2306                         err = -EIO;
2307                         goto next;
2308                 }
2309                 btrfs_tree_lock(eb);
2310                 btrfs_set_lock_blocking_write(eb);
2311
2312                 if (!node->eb) {
2313                         ret = btrfs_cow_block(trans, root, eb, upper->eb,
2314                                               slot, &eb, BTRFS_NESTING_COW);
2315                         btrfs_tree_unlock(eb);
2316                         free_extent_buffer(eb);
2317                         if (ret < 0) {
2318                                 err = ret;
2319                                 goto next;
2320                         }
2321                         BUG_ON(node->eb != eb);
2322                 } else {
2323                         btrfs_set_node_blockptr(upper->eb, slot,
2324                                                 node->eb->start);
2325                         btrfs_set_node_ptr_generation(upper->eb, slot,
2326                                                       trans->transid);
2327                         btrfs_mark_buffer_dirty(upper->eb);
2328
2329                         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2330                                                node->eb->start, blocksize,
2331                                                upper->eb->start);
2332                         ref.real_root = root->root_key.objectid;
2333                         btrfs_init_tree_ref(&ref, node->level,
2334                                             btrfs_header_owner(upper->eb));
2335                         ret = btrfs_inc_extent_ref(trans, &ref);
2336                         BUG_ON(ret);
2337
2338                         ret = btrfs_drop_subtree(trans, root, eb, upper->eb);
2339                         BUG_ON(ret);
2340                 }
2341 next:
2342                 if (!upper->pending)
2343                         btrfs_backref_drop_node_buffer(upper);
2344                 else
2345                         btrfs_backref_unlock_node_buffer(upper);
2346                 if (err)
2347                         break;
2348         }
2349
2350         if (!err && node->pending) {
2351                 btrfs_backref_drop_node_buffer(node);
2352                 list_move_tail(&node->list, &rc->backref_cache.changed);
2353                 node->pending = 0;
2354         }
2355
2356         path->lowest_level = 0;
2357         BUG_ON(err == -ENOSPC);
2358         return err;
2359 }
2360
2361 static int link_to_upper(struct btrfs_trans_handle *trans,
2362                          struct reloc_control *rc,
2363                          struct btrfs_backref_node *node,
2364                          struct btrfs_path *path)
2365 {
2366         struct btrfs_key key;
2367
2368         btrfs_node_key_to_cpu(node->eb, &key, 0);
2369         return do_relocation(trans, rc, node, &key, path, 0);
2370 }
2371
2372 static int finish_pending_nodes(struct btrfs_trans_handle *trans,
2373                                 struct reloc_control *rc,
2374                                 struct btrfs_path *path, int err)
2375 {
2376         LIST_HEAD(list);
2377         struct btrfs_backref_cache *cache = &rc->backref_cache;
2378         struct btrfs_backref_node *node;
2379         int level;
2380         int ret;
2381
2382         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2383                 while (!list_empty(&cache->pending[level])) {
2384                         node = list_entry(cache->pending[level].next,
2385                                           struct btrfs_backref_node, list);
2386                         list_move_tail(&node->list, &list);
2387                         BUG_ON(!node->pending);
2388
2389                         if (!err) {
2390                                 ret = link_to_upper(trans, rc, node, path);
2391                                 if (ret < 0)
2392                                         err = ret;
2393                         }
2394                 }
2395                 list_splice_init(&list, &cache->pending[level]);
2396         }
2397         return err;
2398 }
2399
2400 /*
2401  * mark a block and all blocks directly/indirectly reference the block
2402  * as processed.
2403  */
2404 static void update_processed_blocks(struct reloc_control *rc,
2405                                     struct btrfs_backref_node *node)
2406 {
2407         struct btrfs_backref_node *next = node;
2408         struct btrfs_backref_edge *edge;
2409         struct btrfs_backref_edge *edges[BTRFS_MAX_LEVEL - 1];
2410         int index = 0;
2411
2412         while (next) {
2413                 cond_resched();
2414                 while (1) {
2415                         if (next->processed)
2416                                 break;
2417
2418                         mark_block_processed(rc, next);
2419
2420                         if (list_empty(&next->upper))
2421                                 break;
2422
2423                         edge = list_entry(next->upper.next,
2424                                         struct btrfs_backref_edge, list[LOWER]);
2425                         edges[index++] = edge;
2426                         next = edge->node[UPPER];
2427                 }
2428                 next = walk_down_backref(edges, &index);
2429         }
2430 }
2431
2432 static int tree_block_processed(u64 bytenr, struct reloc_control *rc)
2433 {
2434         u32 blocksize = rc->extent_root->fs_info->nodesize;
2435
2436         if (test_range_bit(&rc->processed_blocks, bytenr,
2437                            bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL))
2438                 return 1;
2439         return 0;
2440 }
2441
2442 static int get_tree_block_key(struct btrfs_fs_info *fs_info,
2443                               struct tree_block *block)
2444 {
2445         struct extent_buffer *eb;
2446
2447         eb = read_tree_block(fs_info, block->bytenr, block->key.offset,
2448                              block->level, NULL);
2449         if (IS_ERR(eb)) {
2450                 return PTR_ERR(eb);
2451         } else if (!extent_buffer_uptodate(eb)) {
2452                 free_extent_buffer(eb);
2453                 return -EIO;
2454         }
2455         if (block->level == 0)
2456                 btrfs_item_key_to_cpu(eb, &block->key, 0);
2457         else
2458                 btrfs_node_key_to_cpu(eb, &block->key, 0);
2459         free_extent_buffer(eb);
2460         block->key_ready = 1;
2461         return 0;
2462 }
2463
2464 /*
2465  * helper function to relocate a tree block
2466  */
2467 static int relocate_tree_block(struct btrfs_trans_handle *trans,
2468                                 struct reloc_control *rc,
2469                                 struct btrfs_backref_node *node,
2470                                 struct btrfs_key *key,
2471                                 struct btrfs_path *path)
2472 {
2473         struct btrfs_root *root;
2474         int ret = 0;
2475
2476         if (!node)
2477                 return 0;
2478
2479         /*
2480          * If we fail here we want to drop our backref_node because we are going
2481          * to start over and regenerate the tree for it.
2482          */
2483         ret = reserve_metadata_space(trans, rc, node);
2484         if (ret)
2485                 goto out;
2486
2487         BUG_ON(node->processed);
2488         root = select_one_root(node);
2489         if (root == ERR_PTR(-ENOENT)) {
2490                 update_processed_blocks(rc, node);
2491                 goto out;
2492         }
2493
2494         if (root) {
2495                 if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state)) {
2496                         BUG_ON(node->new_bytenr);
2497                         BUG_ON(!list_empty(&node->list));
2498                         btrfs_record_root_in_trans(trans, root);
2499                         root = root->reloc_root;
2500                         node->new_bytenr = root->node->start;
2501                         btrfs_put_root(node->root);
2502                         node->root = btrfs_grab_root(root);
2503                         ASSERT(node->root);
2504                         list_add_tail(&node->list, &rc->backref_cache.changed);
2505                 } else {
2506                         path->lowest_level = node->level;
2507                         ret = btrfs_search_slot(trans, root, key, path, 0, 1);
2508                         btrfs_release_path(path);
2509                         if (ret > 0)
2510                                 ret = 0;
2511                 }
2512                 if (!ret)
2513                         update_processed_blocks(rc, node);
2514         } else {
2515                 ret = do_relocation(trans, rc, node, key, path, 1);
2516         }
2517 out:
2518         if (ret || node->level == 0 || node->cowonly)
2519                 btrfs_backref_cleanup_node(&rc->backref_cache, node);
2520         return ret;
2521 }
2522
2523 /*
2524  * relocate a list of blocks
2525  */
2526 static noinline_for_stack
2527 int relocate_tree_blocks(struct btrfs_trans_handle *trans,
2528                          struct reloc_control *rc, struct rb_root *blocks)
2529 {
2530         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2531         struct btrfs_backref_node *node;
2532         struct btrfs_path *path;
2533         struct tree_block *block;
2534         struct tree_block *next;
2535         int ret;
2536         int err = 0;
2537
2538         path = btrfs_alloc_path();
2539         if (!path) {
2540                 err = -ENOMEM;
2541                 goto out_free_blocks;
2542         }
2543
2544         /* Kick in readahead for tree blocks with missing keys */
2545         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2546                 if (!block->key_ready)
2547                         readahead_tree_block(fs_info, block->bytenr);
2548         }
2549
2550         /* Get first keys */
2551         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2552                 if (!block->key_ready) {
2553                         err = get_tree_block_key(fs_info, block);
2554                         if (err)
2555                                 goto out_free_path;
2556                 }
2557         }
2558
2559         /* Do tree relocation */
2560         rbtree_postorder_for_each_entry_safe(block, next, blocks, rb_node) {
2561                 node = build_backref_tree(rc, &block->key,
2562                                           block->level, block->bytenr);
2563                 if (IS_ERR(node)) {
2564                         err = PTR_ERR(node);
2565                         goto out;
2566                 }
2567
2568                 ret = relocate_tree_block(trans, rc, node, &block->key,
2569                                           path);
2570                 if (ret < 0) {
2571                         err = ret;
2572                         break;
2573                 }
2574         }
2575 out:
2576         err = finish_pending_nodes(trans, rc, path, err);
2577
2578 out_free_path:
2579         btrfs_free_path(path);
2580 out_free_blocks:
2581         free_block_list(blocks);
2582         return err;
2583 }
2584
2585 static noinline_for_stack int prealloc_file_extent_cluster(
2586                                 struct btrfs_inode *inode,
2587                                 struct file_extent_cluster *cluster)
2588 {
2589         u64 alloc_hint = 0;
2590         u64 start;
2591         u64 end;
2592         u64 offset = inode->index_cnt;
2593         u64 num_bytes;
2594         int nr;
2595         int ret = 0;
2596         u64 prealloc_start = cluster->start - offset;
2597         u64 prealloc_end = cluster->end - offset;
2598         u64 cur_offset = prealloc_start;
2599
2600         BUG_ON(cluster->start != cluster->boundary[0]);
2601         ret = btrfs_alloc_data_chunk_ondemand(inode,
2602                                               prealloc_end + 1 - prealloc_start);
2603         if (ret)
2604                 return ret;
2605
2606         inode_lock(&inode->vfs_inode);
2607         for (nr = 0; nr < cluster->nr; nr++) {
2608                 start = cluster->boundary[nr] - offset;
2609                 if (nr + 1 < cluster->nr)
2610                         end = cluster->boundary[nr + 1] - 1 - offset;
2611                 else
2612                         end = cluster->end - offset;
2613
2614                 lock_extent(&inode->io_tree, start, end);
2615                 num_bytes = end + 1 - start;
2616                 ret = btrfs_prealloc_file_range(&inode->vfs_inode, 0, start,
2617                                                 num_bytes, num_bytes,
2618                                                 end + 1, &alloc_hint);
2619                 cur_offset = end + 1;
2620                 unlock_extent(&inode->io_tree, start, end);
2621                 if (ret)
2622                         break;
2623         }
2624         inode_unlock(&inode->vfs_inode);
2625
2626         if (cur_offset < prealloc_end)
2627                 btrfs_free_reserved_data_space_noquota(inode->root->fs_info,
2628                                                prealloc_end + 1 - cur_offset);
2629         return ret;
2630 }
2631
2632 static noinline_for_stack
2633 int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
2634                          u64 block_start)
2635 {
2636         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2637         struct extent_map *em;
2638         int ret = 0;
2639
2640         em = alloc_extent_map();
2641         if (!em)
2642                 return -ENOMEM;
2643
2644         em->start = start;
2645         em->len = end + 1 - start;
2646         em->block_len = em->len;
2647         em->block_start = block_start;
2648         set_bit(EXTENT_FLAG_PINNED, &em->flags);
2649
2650         lock_extent(&BTRFS_I(inode)->io_tree, start, end);
2651         while (1) {
2652                 write_lock(&em_tree->lock);
2653                 ret = add_extent_mapping(em_tree, em, 0);
2654                 write_unlock(&em_tree->lock);
2655                 if (ret != -EEXIST) {
2656                         free_extent_map(em);
2657                         break;
2658                 }
2659                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2660         }
2661         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
2662         return ret;
2663 }
2664
2665 /*
2666  * Allow error injection to test balance cancellation
2667  */
2668 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
2669 {
2670         return atomic_read(&fs_info->balance_cancel_req) ||
2671                 fatal_signal_pending(current);
2672 }
2673 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
2674
2675 static int relocate_file_extent_cluster(struct inode *inode,
2676                                         struct file_extent_cluster *cluster)
2677 {
2678         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2679         u64 page_start;
2680         u64 page_end;
2681         u64 offset = BTRFS_I(inode)->index_cnt;
2682         unsigned long index;
2683         unsigned long last_index;
2684         struct page *page;
2685         struct file_ra_state *ra;
2686         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
2687         int nr = 0;
2688         int ret = 0;
2689
2690         if (!cluster->nr)
2691                 return 0;
2692
2693         ra = kzalloc(sizeof(*ra), GFP_NOFS);
2694         if (!ra)
2695                 return -ENOMEM;
2696
2697         ret = prealloc_file_extent_cluster(BTRFS_I(inode), cluster);
2698         if (ret)
2699                 goto out;
2700
2701         file_ra_state_init(ra, inode->i_mapping);
2702
2703         ret = setup_extent_mapping(inode, cluster->start - offset,
2704                                    cluster->end - offset, cluster->start);
2705         if (ret)
2706                 goto out;
2707
2708         index = (cluster->start - offset) >> PAGE_SHIFT;
2709         last_index = (cluster->end - offset) >> PAGE_SHIFT;
2710         while (index <= last_index) {
2711                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
2712                                 PAGE_SIZE);
2713                 if (ret)
2714                         goto out;
2715
2716                 page = find_lock_page(inode->i_mapping, index);
2717                 if (!page) {
2718                         page_cache_sync_readahead(inode->i_mapping,
2719                                                   ra, NULL, index,
2720                                                   last_index + 1 - index);
2721                         page = find_or_create_page(inode->i_mapping, index,
2722                                                    mask);
2723                         if (!page) {
2724                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2725                                                         PAGE_SIZE, true);
2726                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2727                                                         PAGE_SIZE);
2728                                 ret = -ENOMEM;
2729                                 goto out;
2730                         }
2731                 }
2732
2733                 if (PageReadahead(page)) {
2734                         page_cache_async_readahead(inode->i_mapping,
2735                                                    ra, NULL, page, index,
2736                                                    last_index + 1 - index);
2737                 }
2738
2739                 if (!PageUptodate(page)) {
2740                         btrfs_readpage(NULL, page);
2741                         lock_page(page);
2742                         if (!PageUptodate(page)) {
2743                                 unlock_page(page);
2744                                 put_page(page);
2745                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2746                                                         PAGE_SIZE, true);
2747                                 btrfs_delalloc_release_extents(BTRFS_I(inode),
2748                                                                PAGE_SIZE);
2749                                 ret = -EIO;
2750                                 goto out;
2751                         }
2752                 }
2753
2754                 page_start = page_offset(page);
2755                 page_end = page_start + PAGE_SIZE - 1;
2756
2757                 lock_extent(&BTRFS_I(inode)->io_tree, page_start, page_end);
2758
2759                 set_page_extent_mapped(page);
2760
2761                 if (nr < cluster->nr &&
2762                     page_start + offset == cluster->boundary[nr]) {
2763                         set_extent_bits(&BTRFS_I(inode)->io_tree,
2764                                         page_start, page_end,
2765                                         EXTENT_BOUNDARY);
2766                         nr++;
2767                 }
2768
2769                 ret = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start,
2770                                                 page_end, 0, NULL);
2771                 if (ret) {
2772                         unlock_page(page);
2773                         put_page(page);
2774                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
2775                                                          PAGE_SIZE, true);
2776                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2777                                                        PAGE_SIZE);
2778
2779                         clear_extent_bits(&BTRFS_I(inode)->io_tree,
2780                                           page_start, page_end,
2781                                           EXTENT_LOCKED | EXTENT_BOUNDARY);
2782                         goto out;
2783
2784                 }
2785                 set_page_dirty(page);
2786
2787                 unlock_extent(&BTRFS_I(inode)->io_tree,
2788                               page_start, page_end);
2789                 unlock_page(page);
2790                 put_page(page);
2791
2792                 index++;
2793                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2794                 balance_dirty_pages_ratelimited(inode->i_mapping);
2795                 btrfs_throttle(fs_info);
2796                 if (btrfs_should_cancel_balance(fs_info)) {
2797                         ret = -ECANCELED;
2798                         goto out;
2799                 }
2800         }
2801         WARN_ON(nr != cluster->nr);
2802 out:
2803         kfree(ra);
2804         return ret;
2805 }
2806
2807 static noinline_for_stack
2808 int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key,
2809                          struct file_extent_cluster *cluster)
2810 {
2811         int ret;
2812
2813         if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) {
2814                 ret = relocate_file_extent_cluster(inode, cluster);
2815                 if (ret)
2816                         return ret;
2817                 cluster->nr = 0;
2818         }
2819
2820         if (!cluster->nr)
2821                 cluster->start = extent_key->objectid;
2822         else
2823                 BUG_ON(cluster->nr >= MAX_EXTENTS);
2824         cluster->end = extent_key->objectid + extent_key->offset - 1;
2825         cluster->boundary[cluster->nr] = extent_key->objectid;
2826         cluster->nr++;
2827
2828         if (cluster->nr >= MAX_EXTENTS) {
2829                 ret = relocate_file_extent_cluster(inode, cluster);
2830                 if (ret)
2831                         return ret;
2832                 cluster->nr = 0;
2833         }
2834         return 0;
2835 }
2836
2837 /*
2838  * helper to add a tree block to the list.
2839  * the major work is getting the generation and level of the block
2840  */
2841 static int add_tree_block(struct reloc_control *rc,
2842                           struct btrfs_key *extent_key,
2843                           struct btrfs_path *path,
2844                           struct rb_root *blocks)
2845 {
2846         struct extent_buffer *eb;
2847         struct btrfs_extent_item *ei;
2848         struct btrfs_tree_block_info *bi;
2849         struct tree_block *block;
2850         struct rb_node *rb_node;
2851         u32 item_size;
2852         int level = -1;
2853         u64 generation;
2854
2855         eb =  path->nodes[0];
2856         item_size = btrfs_item_size_nr(eb, path->slots[0]);
2857
2858         if (extent_key->type == BTRFS_METADATA_ITEM_KEY ||
2859             item_size >= sizeof(*ei) + sizeof(*bi)) {
2860                 ei = btrfs_item_ptr(eb, path->slots[0],
2861                                 struct btrfs_extent_item);
2862                 if (extent_key->type == BTRFS_EXTENT_ITEM_KEY) {
2863                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2864                         level = btrfs_tree_block_level(eb, bi);
2865                 } else {
2866                         level = (int)extent_key->offset;
2867                 }
2868                 generation = btrfs_extent_generation(eb, ei);
2869         } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
2870                 btrfs_print_v0_err(eb->fs_info);
2871                 btrfs_handle_fs_error(eb->fs_info, -EINVAL, NULL);
2872                 return -EINVAL;
2873         } else {
2874                 BUG();
2875         }
2876
2877         btrfs_release_path(path);
2878
2879         BUG_ON(level == -1);
2880
2881         block = kmalloc(sizeof(*block), GFP_NOFS);
2882         if (!block)
2883                 return -ENOMEM;
2884
2885         block->bytenr = extent_key->objectid;
2886         block->key.objectid = rc->extent_root->fs_info->nodesize;
2887         block->key.offset = generation;
2888         block->level = level;
2889         block->key_ready = 0;
2890
2891         rb_node = rb_simple_insert(blocks, block->bytenr, &block->rb_node);
2892         if (rb_node)
2893                 btrfs_backref_panic(rc->extent_root->fs_info, block->bytenr,
2894                                     -EEXIST);
2895
2896         return 0;
2897 }
2898
2899 /*
2900  * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY
2901  */
2902 static int __add_tree_block(struct reloc_control *rc,
2903                             u64 bytenr, u32 blocksize,
2904                             struct rb_root *blocks)
2905 {
2906         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
2907         struct btrfs_path *path;
2908         struct btrfs_key key;
2909         int ret;
2910         bool skinny = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2911
2912         if (tree_block_processed(bytenr, rc))
2913                 return 0;
2914
2915         if (rb_simple_search(blocks, bytenr))
2916                 return 0;
2917
2918         path = btrfs_alloc_path();
2919         if (!path)
2920                 return -ENOMEM;
2921 again:
2922         key.objectid = bytenr;
2923         if (skinny) {
2924                 key.type = BTRFS_METADATA_ITEM_KEY;
2925                 key.offset = (u64)-1;
2926         } else {
2927                 key.type = BTRFS_EXTENT_ITEM_KEY;
2928                 key.offset = blocksize;
2929         }
2930
2931         path->search_commit_root = 1;
2932         path->skip_locking = 1;
2933         ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0);
2934         if (ret < 0)
2935                 goto out;
2936
2937         if (ret > 0 && skinny) {
2938                 if (path->slots[0]) {
2939                         path->slots[0]--;
2940                         btrfs_item_key_to_cpu(path->nodes[0], &key,
2941                                               path->slots[0]);
2942                         if (key.objectid == bytenr &&
2943                             (key.type == BTRFS_METADATA_ITEM_KEY ||
2944                              (key.type == BTRFS_EXTENT_ITEM_KEY &&
2945                               key.offset == blocksize)))
2946                                 ret = 0;
2947                 }
2948
2949                 if (ret) {
2950                         skinny = false;
2951                         btrfs_release_path(path);
2952                         goto again;
2953                 }
2954         }
2955         if (ret) {
2956                 ASSERT(ret == 1);
2957                 btrfs_print_leaf(path->nodes[0]);
2958                 btrfs_err(fs_info,
2959              "tree block extent item (%llu) is not found in extent tree",
2960                      bytenr);
2961                 WARN_ON(1);
2962                 ret = -EINVAL;
2963                 goto out;
2964         }
2965
2966         ret = add_tree_block(rc, &key, path, blocks);
2967 out:
2968         btrfs_free_path(path);
2969         return ret;
2970 }
2971
2972 static int delete_block_group_cache(struct btrfs_fs_info *fs_info,
2973                                     struct btrfs_block_group *block_group,
2974                                     struct inode *inode,
2975                                     u64 ino)
2976 {
2977         struct btrfs_root *root = fs_info->tree_root;
2978         struct btrfs_trans_handle *trans;
2979         int ret = 0;
2980
2981         if (inode)
2982                 goto truncate;
2983
2984         inode = btrfs_iget(fs_info->sb, ino, root);
2985         if (IS_ERR(inode))
2986                 return -ENOENT;
2987
2988 truncate:
2989         ret = btrfs_check_trunc_cache_free_space(fs_info,
2990                                                  &fs_info->global_block_rsv);
2991         if (ret)
2992                 goto out;
2993
2994         trans = btrfs_join_transaction(root);
2995         if (IS_ERR(trans)) {
2996                 ret = PTR_ERR(trans);
2997                 goto out;
2998         }
2999
3000         ret = btrfs_truncate_free_space_cache(trans, block_group, inode);
3001
3002         btrfs_end_transaction(trans);
3003         btrfs_btree_balance_dirty(fs_info);
3004 out:
3005         iput(inode);
3006         return ret;
3007 }
3008
3009 /*
3010  * Locate the free space cache EXTENT_DATA in root tree leaf and delete the
3011  * cache inode, to avoid free space cache data extent blocking data relocation.
3012  */
3013 static int delete_v1_space_cache(struct extent_buffer *leaf,
3014                                  struct btrfs_block_group *block_group,
3015                                  u64 data_bytenr)
3016 {
3017         u64 space_cache_ino;
3018         struct btrfs_file_extent_item *ei;
3019         struct btrfs_key key;
3020         bool found = false;
3021         int i;
3022         int ret;
3023
3024         if (btrfs_header_owner(leaf) != BTRFS_ROOT_TREE_OBJECTID)
3025                 return 0;
3026
3027         for (i = 0; i < btrfs_header_nritems(leaf); i++) {
3028                 btrfs_item_key_to_cpu(leaf, &key, i);
3029                 if (key.type != BTRFS_EXTENT_DATA_KEY)
3030                         continue;
3031                 ei = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3032                 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_REG &&
3033                     btrfs_file_extent_disk_bytenr(leaf, ei) == data_bytenr) {
3034                         found = true;
3035                         space_cache_ino = key.objectid;
3036                         break;
3037                 }
3038         }
3039         if (!found)
3040                 return -ENOENT;
3041         ret = delete_block_group_cache(leaf->fs_info, block_group, NULL,
3042                                         space_cache_ino);
3043         return ret;
3044 }
3045
3046 /*
3047  * helper to find all tree blocks that reference a given data extent
3048  */
3049 static noinline_for_stack
3050 int add_data_references(struct reloc_control *rc,
3051                         struct btrfs_key *extent_key,
3052                         struct btrfs_path *path,
3053                         struct rb_root *blocks)
3054 {
3055         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3056         struct ulist *leaves = NULL;
3057         struct ulist_iterator leaf_uiter;
3058         struct ulist_node *ref_node = NULL;
3059         const u32 blocksize = fs_info->nodesize;
3060         int ret = 0;
3061
3062         btrfs_release_path(path);
3063         ret = btrfs_find_all_leafs(NULL, fs_info, extent_key->objectid,
3064                                    0, &leaves, NULL, true);
3065         if (ret < 0)
3066                 return ret;
3067
3068         ULIST_ITER_INIT(&leaf_uiter);
3069         while ((ref_node = ulist_next(leaves, &leaf_uiter))) {
3070                 struct extent_buffer *eb;
3071
3072                 eb = read_tree_block(fs_info, ref_node->val, 0, 0, NULL);
3073                 if (IS_ERR(eb)) {
3074                         ret = PTR_ERR(eb);
3075                         break;
3076                 }
3077                 ret = delete_v1_space_cache(eb, rc->block_group,
3078                                             extent_key->objectid);
3079                 free_extent_buffer(eb);
3080                 if (ret < 0)
3081                         break;
3082                 ret = __add_tree_block(rc, ref_node->val, blocksize, blocks);
3083                 if (ret < 0)
3084                         break;
3085         }
3086         if (ret < 0)
3087                 free_block_list(blocks);
3088         ulist_free(leaves);
3089         return ret;
3090 }
3091
3092 /*
3093  * helper to find next unprocessed extent
3094  */
3095 static noinline_for_stack
3096 int find_next_extent(struct reloc_control *rc, struct btrfs_path *path,
3097                      struct btrfs_key *extent_key)
3098 {
3099         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3100         struct btrfs_key key;
3101         struct extent_buffer *leaf;
3102         u64 start, end, last;
3103         int ret;
3104
3105         last = rc->block_group->start + rc->block_group->length;
3106         while (1) {
3107                 cond_resched();
3108                 if (rc->search_start >= last) {
3109                         ret = 1;
3110                         break;
3111                 }
3112
3113                 key.objectid = rc->search_start;
3114                 key.type = BTRFS_EXTENT_ITEM_KEY;
3115                 key.offset = 0;
3116
3117                 path->search_commit_root = 1;
3118                 path->skip_locking = 1;
3119                 ret = btrfs_search_slot(NULL, rc->extent_root, &key, path,
3120                                         0, 0);
3121                 if (ret < 0)
3122                         break;
3123 next:
3124                 leaf = path->nodes[0];
3125                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3126                         ret = btrfs_next_leaf(rc->extent_root, path);
3127                         if (ret != 0)
3128                                 break;
3129                         leaf = path->nodes[0];
3130                 }
3131
3132                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3133                 if (key.objectid >= last) {
3134                         ret = 1;
3135                         break;
3136                 }
3137
3138                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3139                     key.type != BTRFS_METADATA_ITEM_KEY) {
3140                         path->slots[0]++;
3141                         goto next;
3142                 }
3143
3144                 if (key.type == BTRFS_EXTENT_ITEM_KEY &&
3145                     key.objectid + key.offset <= rc->search_start) {
3146                         path->slots[0]++;
3147                         goto next;
3148                 }
3149
3150                 if (key.type == BTRFS_METADATA_ITEM_KEY &&
3151                     key.objectid + fs_info->nodesize <=
3152                     rc->search_start) {
3153                         path->slots[0]++;
3154                         goto next;
3155                 }
3156
3157                 ret = find_first_extent_bit(&rc->processed_blocks,
3158                                             key.objectid, &start, &end,
3159                                             EXTENT_DIRTY, NULL);
3160
3161                 if (ret == 0 && start <= key.objectid) {
3162                         btrfs_release_path(path);
3163                         rc->search_start = end + 1;
3164                 } else {
3165                         if (key.type == BTRFS_EXTENT_ITEM_KEY)
3166                                 rc->search_start = key.objectid + key.offset;
3167                         else
3168                                 rc->search_start = key.objectid +
3169                                         fs_info->nodesize;
3170                         memcpy(extent_key, &key, sizeof(key));
3171                         return 0;
3172                 }
3173         }
3174         btrfs_release_path(path);
3175         return ret;
3176 }
3177
3178 static void set_reloc_control(struct reloc_control *rc)
3179 {
3180         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3181
3182         mutex_lock(&fs_info->reloc_mutex);
3183         fs_info->reloc_ctl = rc;
3184         mutex_unlock(&fs_info->reloc_mutex);
3185 }
3186
3187 static void unset_reloc_control(struct reloc_control *rc)
3188 {
3189         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3190
3191         mutex_lock(&fs_info->reloc_mutex);
3192         fs_info->reloc_ctl = NULL;
3193         mutex_unlock(&fs_info->reloc_mutex);
3194 }
3195
3196 static int check_extent_flags(u64 flags)
3197 {
3198         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3199             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3200                 return 1;
3201         if (!(flags & BTRFS_EXTENT_FLAG_DATA) &&
3202             !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
3203                 return 1;
3204         if ((flags & BTRFS_EXTENT_FLAG_DATA) &&
3205             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
3206                 return 1;
3207         return 0;
3208 }
3209
3210 static noinline_for_stack
3211 int prepare_to_relocate(struct reloc_control *rc)
3212 {
3213         struct btrfs_trans_handle *trans;
3214         int ret;
3215
3216         rc->block_rsv = btrfs_alloc_block_rsv(rc->extent_root->fs_info,
3217                                               BTRFS_BLOCK_RSV_TEMP);
3218         if (!rc->block_rsv)
3219                 return -ENOMEM;
3220
3221         memset(&rc->cluster, 0, sizeof(rc->cluster));
3222         rc->search_start = rc->block_group->start;
3223         rc->extents_found = 0;
3224         rc->nodes_relocated = 0;
3225         rc->merging_rsv_size = 0;
3226         rc->reserved_bytes = 0;
3227         rc->block_rsv->size = rc->extent_root->fs_info->nodesize *
3228                               RELOCATION_RESERVED_NODES;
3229         ret = btrfs_block_rsv_refill(rc->extent_root,
3230                                      rc->block_rsv, rc->block_rsv->size,
3231                                      BTRFS_RESERVE_FLUSH_ALL);
3232         if (ret)
3233                 return ret;
3234
3235         rc->create_reloc_tree = 1;
3236         set_reloc_control(rc);
3237
3238         trans = btrfs_join_transaction(rc->extent_root);
3239         if (IS_ERR(trans)) {
3240                 unset_reloc_control(rc);
3241                 /*
3242                  * extent tree is not a ref_cow tree and has no reloc_root to
3243                  * cleanup.  And callers are responsible to free the above
3244                  * block rsv.
3245                  */
3246                 return PTR_ERR(trans);
3247         }
3248         btrfs_commit_transaction(trans);
3249         return 0;
3250 }
3251
3252 static noinline_for_stack int relocate_block_group(struct reloc_control *rc)
3253 {
3254         struct btrfs_fs_info *fs_info = rc->extent_root->fs_info;
3255         struct rb_root blocks = RB_ROOT;
3256         struct btrfs_key key;
3257         struct btrfs_trans_handle *trans = NULL;
3258         struct btrfs_path *path;
3259         struct btrfs_extent_item *ei;
3260         u64 flags;
3261         u32 item_size;
3262         int ret;
3263         int err = 0;
3264         int progress = 0;
3265
3266         path = btrfs_alloc_path();
3267         if (!path)
3268                 return -ENOMEM;
3269         path->reada = READA_FORWARD;
3270
3271         ret = prepare_to_relocate(rc);
3272         if (ret) {
3273                 err = ret;
3274                 goto out_free;
3275         }
3276
3277         while (1) {
3278                 rc->reserved_bytes = 0;
3279                 ret = btrfs_block_rsv_refill(rc->extent_root,
3280                                         rc->block_rsv, rc->block_rsv->size,
3281                                         BTRFS_RESERVE_FLUSH_ALL);
3282                 if (ret) {
3283                         err = ret;
3284                         break;
3285                 }
3286                 progress++;
3287                 trans = btrfs_start_transaction(rc->extent_root, 0);
3288                 if (IS_ERR(trans)) {
3289                         err = PTR_ERR(trans);
3290                         trans = NULL;
3291                         break;
3292                 }
3293 restart:
3294                 if (update_backref_cache(trans, &rc->backref_cache)) {
3295                         btrfs_end_transaction(trans);
3296                         trans = NULL;
3297                         continue;
3298                 }
3299
3300                 ret = find_next_extent(rc, path, &key);
3301                 if (ret < 0)
3302                         err = ret;
3303                 if (ret != 0)
3304                         break;
3305
3306                 rc->extents_found++;
3307
3308                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
3309                                     struct btrfs_extent_item);
3310                 item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
3311                 if (item_size >= sizeof(*ei)) {
3312                         flags = btrfs_extent_flags(path->nodes[0], ei);
3313                         ret = check_extent_flags(flags);
3314                         BUG_ON(ret);
3315                 } else if (unlikely(item_size == sizeof(struct btrfs_extent_item_v0))) {
3316                         err = -EINVAL;
3317                         btrfs_print_v0_err(trans->fs_info);
3318                         btrfs_abort_transaction(trans, err);
3319                         break;
3320                 } else {
3321                         BUG();
3322                 }
3323
3324                 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
3325                         ret = add_tree_block(rc, &key, path, &blocks);
3326                 } else if (rc->stage == UPDATE_DATA_PTRS &&
3327                            (flags & BTRFS_EXTENT_FLAG_DATA)) {
3328                         ret = add_data_references(rc, &key, path, &blocks);
3329                 } else {
3330                         btrfs_release_path(path);
3331                         ret = 0;
3332                 }
3333                 if (ret < 0) {
3334                         err = ret;
3335                         break;
3336                 }
3337
3338                 if (!RB_EMPTY_ROOT(&blocks)) {
3339                         ret = relocate_tree_blocks(trans, rc, &blocks);
3340                         if (ret < 0) {
3341                                 if (ret != -EAGAIN) {
3342                                         err = ret;
3343                                         break;
3344                                 }
3345                                 rc->extents_found--;
3346                                 rc->search_start = key.objectid;
3347                         }
3348                 }
3349
3350                 btrfs_end_transaction_throttle(trans);
3351                 btrfs_btree_balance_dirty(fs_info);
3352                 trans = NULL;
3353
3354                 if (rc->stage == MOVE_DATA_EXTENTS &&
3355                     (flags & BTRFS_EXTENT_FLAG_DATA)) {
3356                         rc->found_file_extent = 1;
3357                         ret = relocate_data_extent(rc->data_inode,
3358                                                    &key, &rc->cluster);
3359                         if (ret < 0) {
3360                                 err = ret;
3361                                 break;
3362                         }
3363                 }
3364                 if (btrfs_should_cancel_balance(fs_info)) {
3365                         err = -ECANCELED;
3366                         break;
3367                 }
3368         }
3369         if (trans && progress && err == -ENOSPC) {
3370                 ret = btrfs_force_chunk_alloc(trans, rc->block_group->flags);
3371                 if (ret == 1) {
3372                         err = 0;
3373                         progress = 0;
3374                         goto restart;
3375                 }
3376         }
3377
3378         btrfs_release_path(path);
3379         clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY);
3380
3381         if (trans) {
3382                 btrfs_end_transaction_throttle(trans);
3383                 btrfs_btree_balance_dirty(fs_info);
3384         }
3385
3386         if (!err) {
3387                 ret = relocate_file_extent_cluster(rc->data_inode,
3388                                                    &rc->cluster);
3389                 if (ret < 0)
3390                         err = ret;
3391         }
3392
3393         rc->create_reloc_tree = 0;
3394         set_reloc_control(rc);
3395
3396         btrfs_backref_release_cache(&rc->backref_cache);
3397         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3398
3399         /*
3400          * Even in the case when the relocation is cancelled, we should all go
3401          * through prepare_to_merge() and merge_reloc_roots().
3402          *
3403          * For error (including cancelled balance), prepare_to_merge() will
3404          * mark all reloc trees orphan, then queue them for cleanup in
3405          * merge_reloc_roots()
3406          */
3407         err = prepare_to_merge(rc, err);
3408
3409         merge_reloc_roots(rc);
3410
3411         rc->merge_reloc_tree = 0;
3412         unset_reloc_control(rc);
3413         btrfs_block_rsv_release(fs_info, rc->block_rsv, (u64)-1, NULL);
3414
3415         /* get rid of pinned extents */
3416         trans = btrfs_join_transaction(rc->extent_root);
3417         if (IS_ERR(trans)) {
3418                 err = PTR_ERR(trans);
3419                 goto out_free;
3420         }
3421         btrfs_commit_transaction(trans);
3422 out_free:
3423         ret = clean_dirty_subvols(rc);
3424         if (ret < 0 && !err)
3425                 err = ret;
3426         btrfs_free_block_rsv(fs_info, rc->block_rsv);
3427         btrfs_free_path(path);
3428         return err;
3429 }
3430
3431 static int __insert_orphan_inode(struct btrfs_trans_handle *trans,
3432                                  struct btrfs_root *root, u64 objectid)
3433 {
3434         struct btrfs_path *path;
3435         struct btrfs_inode_item *item;
3436         struct extent_buffer *leaf;
3437         int ret;
3438
3439         path = btrfs_alloc_path();
3440         if (!path)
3441                 return -ENOMEM;
3442
3443         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
3444         if (ret)
3445                 goto out;
3446
3447         leaf = path->nodes[0];
3448         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item);
3449         memzero_extent_buffer(leaf, (unsigned long)item, sizeof(*item));
3450         btrfs_set_inode_generation(leaf, item, 1);
3451         btrfs_set_inode_size(leaf, item, 0);
3452         btrfs_set_inode_mode(leaf, item, S_IFREG | 0600);
3453         btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS |
3454                                           BTRFS_INODE_PREALLOC);
3455         btrfs_mark_buffer_dirty(leaf);
3456 out:
3457         btrfs_free_path(path);
3458         return ret;
3459 }
3460
3461 /*
3462  * helper to create inode for data relocation.
3463  * the inode is in data relocation tree and its link count is 0
3464  */
3465 static noinline_for_stack
3466 struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info,
3467                                  struct btrfs_block_group *group)
3468 {
3469         struct inode *inode = NULL;
3470         struct btrfs_trans_handle *trans;
3471         struct btrfs_root *root;
3472         u64 objectid;
3473         int err = 0;
3474
3475         root = btrfs_grab_root(fs_info->data_reloc_root);
3476         trans = btrfs_start_transaction(root, 6);
3477         if (IS_ERR(trans)) {
3478                 btrfs_put_root(root);
3479                 return ERR_CAST(trans);
3480         }
3481
3482         err = btrfs_find_free_objectid(root, &objectid);
3483         if (err)
3484                 goto out;
3485
3486         err = __insert_orphan_inode(trans, root, objectid);
3487         BUG_ON(err);
3488
3489         inode = btrfs_iget(fs_info->sb, objectid, root);
3490         BUG_ON(IS_ERR(inode));
3491         BTRFS_I(inode)->index_cnt = group->start;
3492
3493         err = btrfs_orphan_add(trans, BTRFS_I(inode));
3494 out:
3495         btrfs_put_root(root);
3496         btrfs_end_transaction(trans);
3497         btrfs_btree_balance_dirty(fs_info);
3498         if (err) {
3499                 if (inode)
3500                         iput(inode);
3501                 inode = ERR_PTR(err);
3502         }
3503         return inode;
3504 }
3505
3506 static struct reloc_control *alloc_reloc_control(struct btrfs_fs_info *fs_info)
3507 {
3508         struct reloc_control *rc;
3509
3510         rc = kzalloc(sizeof(*rc), GFP_NOFS);
3511         if (!rc)
3512                 return NULL;
3513
3514         INIT_LIST_HEAD(&rc->reloc_roots);
3515         INIT_LIST_HEAD(&rc->dirty_subvol_roots);
3516         btrfs_backref_init_cache(fs_info, &rc->backref_cache, 1);
3517         mapping_tree_init(&rc->reloc_root_tree);
3518         extent_io_tree_init(fs_info, &rc->processed_blocks,
3519                             IO_TREE_RELOC_BLOCKS, NULL);
3520         return rc;
3521 }
3522
3523 static void free_reloc_control(struct reloc_control *rc)
3524 {
3525         struct mapping_node *node, *tmp;
3526
3527         free_reloc_roots(&rc->reloc_roots);
3528         rbtree_postorder_for_each_entry_safe(node, tmp,
3529                         &rc->reloc_root_tree.rb_root, rb_node)
3530                 kfree(node);
3531
3532         kfree(rc);
3533 }
3534
3535 /*
3536  * Print the block group being relocated
3537  */
3538 static void describe_relocation(struct btrfs_fs_info *fs_info,
3539                                 struct btrfs_block_group *block_group)
3540 {
3541         char buf[128] = {'\0'};
3542
3543         btrfs_describe_block_groups(block_group->flags, buf, sizeof(buf));
3544
3545         btrfs_info(fs_info,
3546                    "relocating block group %llu flags %s",
3547                    block_group->start, buf);
3548 }
3549
3550 static const char *stage_to_string(int stage)
3551 {
3552         if (stage == MOVE_DATA_EXTENTS)
3553                 return "move data extents";
3554         if (stage == UPDATE_DATA_PTRS)
3555                 return "update data pointers";
3556         return "unknown";
3557 }
3558
3559 /*
3560  * function to relocate all extents in a block group.
3561  */
3562 int btrfs_relocate_block_group(struct btrfs_fs_info *fs_info, u64 group_start)
3563 {
3564         struct btrfs_block_group *bg;
3565         struct btrfs_root *extent_root = fs_info->extent_root;
3566         struct reloc_control *rc;
3567         struct inode *inode;
3568         struct btrfs_path *path;
3569         int ret;
3570         int rw = 0;
3571         int err = 0;
3572
3573         bg = btrfs_lookup_block_group(fs_info, group_start);
3574         if (!bg)
3575                 return -ENOENT;
3576
3577         if (btrfs_pinned_by_swapfile(fs_info, bg)) {
3578                 btrfs_put_block_group(bg);
3579                 return -ETXTBSY;
3580         }
3581
3582         rc = alloc_reloc_control(fs_info);
3583         if (!rc) {
3584                 btrfs_put_block_group(bg);
3585                 return -ENOMEM;
3586         }
3587
3588         rc->extent_root = extent_root;
3589         rc->block_group = bg;
3590
3591         ret = btrfs_inc_block_group_ro(rc->block_group, true);
3592         if (ret) {
3593                 err = ret;
3594                 goto out;
3595         }
3596         rw = 1;
3597
3598         path = btrfs_alloc_path();
3599         if (!path) {
3600                 err = -ENOMEM;
3601                 goto out;
3602         }
3603
3604         inode = lookup_free_space_inode(rc->block_group, path);
3605         btrfs_free_path(path);
3606
3607         if (!IS_ERR(inode))
3608                 ret = delete_block_group_cache(fs_info, rc->block_group, inode, 0);
3609         else
3610                 ret = PTR_ERR(inode);
3611
3612         if (ret && ret != -ENOENT) {
3613                 err = ret;
3614                 goto out;
3615         }
3616
3617         rc->data_inode = create_reloc_inode(fs_info, rc->block_group);
3618         if (IS_ERR(rc->data_inode)) {
3619                 err = PTR_ERR(rc->data_inode);
3620                 rc->data_inode = NULL;
3621                 goto out;
3622         }
3623
3624         describe_relocation(fs_info, rc->block_group);
3625
3626         btrfs_wait_block_group_reservations(rc->block_group);
3627         btrfs_wait_nocow_writers(rc->block_group);
3628         btrfs_wait_ordered_roots(fs_info, U64_MAX,
3629                                  rc->block_group->start,
3630                                  rc->block_group->length);
3631
3632         while (1) {
3633                 int finishes_stage;
3634
3635                 mutex_lock(&fs_info->cleaner_mutex);
3636                 ret = relocate_block_group(rc);
3637                 mutex_unlock(&fs_info->cleaner_mutex);
3638                 if (ret < 0)
3639                         err = ret;
3640
3641                 finishes_stage = rc->stage;
3642                 /*
3643                  * We may have gotten ENOSPC after we already dirtied some
3644                  * extents.  If writeout happens while we're relocating a
3645                  * different block group we could end up hitting the
3646                  * BUG_ON(rc->stage == UPDATE_DATA_PTRS) in
3647                  * btrfs_reloc_cow_block.  Make sure we write everything out
3648                  * properly so we don't trip over this problem, and then break
3649                  * out of the loop if we hit an error.
3650                  */
3651                 if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) {
3652                         ret = btrfs_wait_ordered_range(rc->data_inode, 0,
3653                                                        (u64)-1);
3654                         if (ret)
3655                                 err = ret;
3656                         invalidate_mapping_pages(rc->data_inode->i_mapping,
3657                                                  0, -1);
3658                         rc->stage = UPDATE_DATA_PTRS;
3659                 }
3660
3661                 if (err < 0)
3662                         goto out;
3663
3664                 if (rc->extents_found == 0)
3665                         break;
3666
3667                 btrfs_info(fs_info, "found %llu extents, stage: %s",
3668                            rc->extents_found, stage_to_string(finishes_stage));
3669         }
3670
3671         WARN_ON(rc->block_group->pinned > 0);
3672         WARN_ON(rc->block_group->reserved > 0);
3673         WARN_ON(rc->block_group->used > 0);
3674 out:
3675         if (err && rw)
3676                 btrfs_dec_block_group_ro(rc->block_group);
3677         iput(rc->data_inode);
3678         btrfs_put_block_group(rc->block_group);
3679         free_reloc_control(rc);
3680         return err;
3681 }
3682
3683 static noinline_for_stack int mark_garbage_root(struct btrfs_root *root)
3684 {
3685         struct btrfs_fs_info *fs_info = root->fs_info;
3686         struct btrfs_trans_handle *trans;
3687         int ret, err;
3688
3689         trans = btrfs_start_transaction(fs_info->tree_root, 0);
3690         if (IS_ERR(trans))
3691                 return PTR_ERR(trans);
3692
3693         memset(&root->root_item.drop_progress, 0,
3694                 sizeof(root->root_item.drop_progress));
3695         root->root_item.drop_level = 0;
3696         btrfs_set_root_refs(&root->root_item, 0);
3697         ret = btrfs_update_root(trans, fs_info->tree_root,
3698                                 &root->root_key, &root->root_item);
3699
3700         err = btrfs_end_transaction(trans);
3701         if (err)
3702                 return err;
3703         return ret;
3704 }
3705
3706 /*
3707  * recover relocation interrupted by system crash.
3708  *
3709  * this function resumes merging reloc trees with corresponding fs trees.
3710  * this is important for keeping the sharing of tree blocks
3711  */
3712 int btrfs_recover_relocation(struct btrfs_root *root)
3713 {
3714         struct btrfs_fs_info *fs_info = root->fs_info;
3715         LIST_HEAD(reloc_roots);
3716         struct btrfs_key key;
3717         struct btrfs_root *fs_root;
3718         struct btrfs_root *reloc_root;
3719         struct btrfs_path *path;
3720         struct extent_buffer *leaf;
3721         struct reloc_control *rc = NULL;
3722         struct btrfs_trans_handle *trans;
3723         int ret;
3724         int err = 0;
3725
3726         path = btrfs_alloc_path();
3727         if (!path)
3728                 return -ENOMEM;
3729         path->reada = READA_BACK;
3730
3731         key.objectid = BTRFS_TREE_RELOC_OBJECTID;
3732         key.type = BTRFS_ROOT_ITEM_KEY;
3733         key.offset = (u64)-1;
3734
3735         while (1) {
3736                 ret = btrfs_search_slot(NULL, fs_info->tree_root, &key,
3737                                         path, 0, 0);
3738                 if (ret < 0) {
3739                         err = ret;
3740                         goto out;
3741                 }
3742                 if (ret > 0) {
3743                         if (path->slots[0] == 0)
3744                                 break;
3745                         path->slots[0]--;
3746                 }
3747                 leaf = path->nodes[0];
3748                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3749                 btrfs_release_path(path);
3750
3751                 if (key.objectid != BTRFS_TREE_RELOC_OBJECTID ||
3752                     key.type != BTRFS_ROOT_ITEM_KEY)
3753                         break;
3754
3755                 reloc_root = btrfs_read_tree_root(root, &key);
3756                 if (IS_ERR(reloc_root)) {
3757                         err = PTR_ERR(reloc_root);
3758                         goto out;
3759                 }
3760
3761                 set_bit(BTRFS_ROOT_SHAREABLE, &reloc_root->state);
3762                 list_add(&reloc_root->root_list, &reloc_roots);
3763
3764                 if (btrfs_root_refs(&reloc_root->root_item) > 0) {
3765                         fs_root = btrfs_get_fs_root(fs_info,
3766                                         reloc_root->root_key.offset, false);
3767                         if (IS_ERR(fs_root)) {
3768                                 ret = PTR_ERR(fs_root);
3769                                 if (ret != -ENOENT) {
3770                                         err = ret;
3771                                         goto out;
3772                                 }
3773                                 ret = mark_garbage_root(reloc_root);
3774                                 if (ret < 0) {
3775                                         err = ret;
3776                                         goto out;
3777                                 }
3778                         } else {
3779                                 btrfs_put_root(fs_root);
3780                         }
3781                 }
3782
3783                 if (key.offset == 0)
3784                         break;
3785
3786                 key.offset--;
3787         }
3788         btrfs_release_path(path);
3789
3790         if (list_empty(&reloc_roots))
3791                 goto out;
3792
3793         rc = alloc_reloc_control(fs_info);
3794         if (!rc) {
3795                 err = -ENOMEM;
3796                 goto out;
3797         }
3798
3799         rc->extent_root = fs_info->extent_root;
3800
3801         set_reloc_control(rc);
3802
3803         trans = btrfs_join_transaction(rc->extent_root);
3804         if (IS_ERR(trans)) {
3805                 err = PTR_ERR(trans);
3806                 goto out_unset;
3807         }
3808
3809         rc->merge_reloc_tree = 1;
3810
3811         while (!list_empty(&reloc_roots)) {
3812                 reloc_root = list_entry(reloc_roots.next,
3813                                         struct btrfs_root, root_list);
3814                 list_del(&reloc_root->root_list);
3815
3816                 if (btrfs_root_refs(&reloc_root->root_item) == 0) {
3817                         list_add_tail(&reloc_root->root_list,
3818                                       &rc->reloc_roots);
3819                         continue;
3820                 }
3821
3822                 fs_root = btrfs_get_fs_root(fs_info, reloc_root->root_key.offset,
3823                                             false);
3824                 if (IS_ERR(fs_root)) {
3825                         err = PTR_ERR(fs_root);
3826                         list_add_tail(&reloc_root->root_list, &reloc_roots);
3827                         btrfs_end_transaction(trans);
3828                         goto out_unset;
3829                 }
3830
3831                 err = __add_reloc_root(reloc_root);
3832                 BUG_ON(err < 0); /* -ENOMEM or logic error */
3833                 fs_root->reloc_root = btrfs_grab_root(reloc_root);
3834                 btrfs_put_root(fs_root);
3835         }
3836
3837         err = btrfs_commit_transaction(trans);
3838         if (err)
3839                 goto out_unset;
3840
3841         merge_reloc_roots(rc);
3842
3843         unset_reloc_control(rc);
3844
3845         trans = btrfs_join_transaction(rc->extent_root);
3846         if (IS_ERR(trans)) {
3847                 err = PTR_ERR(trans);
3848                 goto out_clean;
3849         }
3850         err = btrfs_commit_transaction(trans);
3851 out_clean:
3852         ret = clean_dirty_subvols(rc);
3853         if (ret < 0 && !err)
3854                 err = ret;
3855 out_unset:
3856         unset_reloc_control(rc);
3857         free_reloc_control(rc);
3858 out:
3859         free_reloc_roots(&reloc_roots);
3860
3861         btrfs_free_path(path);
3862
3863         if (err == 0) {
3864                 /* cleanup orphan inode in data relocation tree */
3865                 fs_root = btrfs_grab_root(fs_info->data_reloc_root);
3866                 ASSERT(fs_root);
3867                 err = btrfs_orphan_cleanup(fs_root);
3868                 btrfs_put_root(fs_root);
3869         }
3870         return err;
3871 }
3872
3873 /*
3874  * helper to add ordered checksum for data relocation.
3875  *
3876  * cloning checksum properly handles the nodatasum extents.
3877  * it also saves CPU time to re-calculate the checksum.
3878  */
3879 int btrfs_reloc_clone_csums(struct btrfs_inode *inode, u64 file_pos, u64 len)
3880 {
3881         struct btrfs_fs_info *fs_info = inode->root->fs_info;
3882         struct btrfs_ordered_sum *sums;
3883         struct btrfs_ordered_extent *ordered;
3884         int ret;
3885         u64 disk_bytenr;
3886         u64 new_bytenr;
3887         LIST_HEAD(list);
3888
3889         ordered = btrfs_lookup_ordered_extent(inode, file_pos);
3890         BUG_ON(ordered->file_offset != file_pos || ordered->num_bytes != len);
3891
3892         disk_bytenr = file_pos + inode->index_cnt;
3893         ret = btrfs_lookup_csums_range(fs_info->csum_root, disk_bytenr,
3894                                        disk_bytenr + len - 1, &list, 0);
3895         if (ret)
3896                 goto out;
3897
3898         while (!list_empty(&list)) {
3899                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
3900                 list_del_init(&sums->list);
3901
3902                 /*
3903                  * We need to offset the new_bytenr based on where the csum is.
3904                  * We need to do this because we will read in entire prealloc
3905                  * extents but we may have written to say the middle of the
3906                  * prealloc extent, so we need to make sure the csum goes with
3907                  * the right disk offset.
3908                  *
3909                  * We can do this because the data reloc inode refers strictly
3910                  * to the on disk bytes, so we don't have to worry about
3911                  * disk_len vs real len like with real inodes since it's all
3912                  * disk length.
3913                  */
3914                 new_bytenr = ordered->disk_bytenr + sums->bytenr - disk_bytenr;
3915                 sums->bytenr = new_bytenr;
3916
3917                 btrfs_add_ordered_sum(ordered, sums);
3918         }
3919 out:
3920         btrfs_put_ordered_extent(ordered);
3921         return ret;
3922 }
3923
3924 int btrfs_reloc_cow_block(struct btrfs_trans_handle *trans,
3925                           struct btrfs_root *root, struct extent_buffer *buf,
3926                           struct extent_buffer *cow)
3927 {
3928         struct btrfs_fs_info *fs_info = root->fs_info;
3929         struct reloc_control *rc;
3930         struct btrfs_backref_node *node;
3931         int first_cow = 0;
3932         int level;
3933         int ret = 0;
3934
3935         rc = fs_info->reloc_ctl;
3936         if (!rc)
3937                 return 0;
3938
3939         BUG_ON(rc->stage == UPDATE_DATA_PTRS &&
3940                root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID);
3941
3942         level = btrfs_header_level(buf);
3943         if (btrfs_header_generation(buf) <=
3944             btrfs_root_last_snapshot(&root->root_item))
3945                 first_cow = 1;
3946
3947         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID &&
3948             rc->create_reloc_tree) {
3949                 WARN_ON(!first_cow && level == 0);
3950
3951                 node = rc->backref_cache.path[level];
3952                 BUG_ON(node->bytenr != buf->start &&
3953                        node->new_bytenr != buf->start);
3954
3955                 btrfs_backref_drop_node_buffer(node);
3956                 atomic_inc(&cow->refs);
3957                 node->eb = cow;
3958                 node->new_bytenr = cow->start;
3959
3960                 if (!node->pending) {
3961                         list_move_tail(&node->list,
3962                                        &rc->backref_cache.pending[level]);
3963                         node->pending = 1;
3964                 }
3965
3966                 if (first_cow)
3967                         mark_block_processed(rc, node);
3968
3969                 if (first_cow && level > 0)
3970                         rc->nodes_relocated += buf->len;
3971         }
3972
3973         if (level == 0 && first_cow && rc->stage == UPDATE_DATA_PTRS)
3974                 ret = replace_file_extents(trans, rc, root, cow);
3975         return ret;
3976 }
3977
3978 /*
3979  * called before creating snapshot. it calculates metadata reservation
3980  * required for relocating tree blocks in the snapshot
3981  */
3982 void btrfs_reloc_pre_snapshot(struct btrfs_pending_snapshot *pending,
3983                               u64 *bytes_to_reserve)
3984 {
3985         struct btrfs_root *root = pending->root;
3986         struct reloc_control *rc = root->fs_info->reloc_ctl;
3987
3988         if (!rc || !have_reloc_root(root))
3989                 return;
3990
3991         if (!rc->merge_reloc_tree)
3992                 return;
3993
3994         root = root->reloc_root;
3995         BUG_ON(btrfs_root_refs(&root->root_item) == 0);
3996         /*
3997          * relocation is in the stage of merging trees. the space
3998          * used by merging a reloc tree is twice the size of
3999          * relocated tree nodes in the worst case. half for cowing
4000          * the reloc tree, half for cowing the fs tree. the space
4001          * used by cowing the reloc tree will be freed after the
4002          * tree is dropped. if we create snapshot, cowing the fs
4003          * tree may use more space than it frees. so we need
4004          * reserve extra space.
4005          */
4006         *bytes_to_reserve += rc->nodes_relocated;
4007 }
4008
4009 /*
4010  * called after snapshot is created. migrate block reservation
4011  * and create reloc root for the newly created snapshot
4012  *
4013  * This is similar to btrfs_init_reloc_root(), we come out of here with two
4014  * references held on the reloc_root, one for root->reloc_root and one for
4015  * rc->reloc_roots.
4016  */
4017 int btrfs_reloc_post_snapshot(struct btrfs_trans_handle *trans,
4018                                struct btrfs_pending_snapshot *pending)
4019 {
4020         struct btrfs_root *root = pending->root;
4021         struct btrfs_root *reloc_root;
4022         struct btrfs_root *new_root;
4023         struct reloc_control *rc = root->fs_info->reloc_ctl;
4024         int ret;
4025
4026         if (!rc || !have_reloc_root(root))
4027                 return 0;
4028
4029         rc = root->fs_info->reloc_ctl;
4030         rc->merging_rsv_size += rc->nodes_relocated;
4031
4032         if (rc->merge_reloc_tree) {
4033                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
4034                                               rc->block_rsv,
4035                                               rc->nodes_relocated, true);
4036                 if (ret)
4037                         return ret;
4038         }
4039
4040         new_root = pending->snap;
4041         reloc_root = create_reloc_root(trans, root->reloc_root,
4042                                        new_root->root_key.objectid);
4043         if (IS_ERR(reloc_root))
4044                 return PTR_ERR(reloc_root);
4045
4046         ret = __add_reloc_root(reloc_root);
4047         BUG_ON(ret < 0);
4048         new_root->reloc_root = btrfs_grab_root(reloc_root);
4049
4050         if (rc->create_reloc_tree)
4051                 ret = clone_backref_node(trans, rc, root, reloc_root);
4052         return ret;
4053 }